1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright 2019 Joyent, Inc.
25 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
26 */
27
28 /*
29 * DTrace - Dynamic Tracing for Solaris
30 *
31 * This is the implementation of the Solaris Dynamic Tracing framework
32 * (DTrace). The user-visible interface to DTrace is described at length in
33 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
34 * library, the in-kernel DTrace framework, and the DTrace providers are
35 * described in the block comments in the <sys/dtrace.h> header file. The
36 * internal architecture of DTrace is described in the block comments in the
37 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
38 * implementation very much assume mastery of all of these sources; if one has
39 * an unanswered question about the implementation, one should consult them
40 * first.
41 *
42 * The functions here are ordered roughly as follows:
43 *
44 * - Probe context functions
45 * - Probe hashing functions
46 * - Non-probe context utility functions
47 * - Matching functions
48 * - Provider-to-Framework API functions
49 * - Probe management functions
50 * - DIF object functions
51 * - Format functions
52 * - Predicate functions
53 * - ECB functions
54 * - Buffer functions
55 * - Enabling functions
56 * - DOF functions
57 * - Anonymous enabling functions
58 * - Consumer state functions
59 * - Helper functions
60 * - Hook functions
61 * - Driver cookbook functions
62 *
63 * Each group of functions begins with a block comment labelled the "DTrace
64 * [Group] Functions", allowing one to find each block by searching forward
65 * on capital-f functions.
66 */
67 #include <sys/errno.h>
68 #include <sys/stat.h>
69 #include <sys/modctl.h>
70 #include <sys/conf.h>
71 #include <sys/systm.h>
72 #include <sys/ddi.h>
73 #include <sys/sunddi.h>
74 #include <sys/cpuvar.h>
75 #include <sys/kmem.h>
76 #include <sys/strsubr.h>
77 #include <sys/sysmacros.h>
78 #include <sys/dtrace_impl.h>
79 #include <sys/atomic.h>
80 #include <sys/cmn_err.h>
81 #include <sys/mutex_impl.h>
82 #include <sys/rwlock_impl.h>
83 #include <sys/ctf_api.h>
84 #include <sys/panic.h>
85 #include <sys/priv_impl.h>
86 #include <sys/policy.h>
87 #include <sys/cred_impl.h>
88 #include <sys/procfs_isa.h>
89 #include <sys/taskq.h>
90 #include <sys/mkdev.h>
91 #include <sys/kdi.h>
92 #include <sys/zone.h>
93 #include <sys/socket.h>
94 #include <netinet/in.h>
95 #include "strtolctype.h"
96
97 /*
98 * DTrace Tunable Variables
99 *
100 * The following variables may be tuned by adding a line to /etc/system that
101 * includes both the name of the DTrace module ("dtrace") and the name of the
102 * variable. For example:
103 *
104 * set dtrace:dtrace_destructive_disallow = 1
105 *
106 * In general, the only variables that one should be tuning this way are those
107 * that affect system-wide DTrace behavior, and for which the default behavior
108 * is undesirable. Most of these variables are tunable on a per-consumer
109 * basis using DTrace options, and need not be tuned on a system-wide basis.
110 * When tuning these variables, avoid pathological values; while some attempt
111 * is made to verify the integrity of these variables, they are not considered
112 * part of the supported interface to DTrace, and they are therefore not
113 * checked comprehensively. Further, these variables should not be tuned
114 * dynamically via "mdb -kw" or other means; they should only be tuned via
115 * /etc/system.
116 */
117 int dtrace_destructive_disallow = 0;
118 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
119 size_t dtrace_difo_maxsize = (256 * 1024);
120 dtrace_optval_t dtrace_dof_maxsize = (8 * 1024 * 1024);
121 size_t dtrace_statvar_maxsize = (16 * 1024);
122 size_t dtrace_actions_max = (16 * 1024);
123 size_t dtrace_retain_max = 1024;
124 dtrace_optval_t dtrace_helper_actions_max = 1024;
125 dtrace_optval_t dtrace_helper_providers_max = 32;
126 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
127 size_t dtrace_strsize_default = 256;
128 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
129 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
130 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
131 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
132 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
133 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
134 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
135 dtrace_optval_t dtrace_nspec_default = 1;
136 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
137 dtrace_optval_t dtrace_stackframes_default = 20;
138 dtrace_optval_t dtrace_ustackframes_default = 20;
139 dtrace_optval_t dtrace_jstackframes_default = 50;
140 dtrace_optval_t dtrace_jstackstrsize_default = 512;
141 int dtrace_msgdsize_max = 128;
142 hrtime_t dtrace_chill_max = MSEC2NSEC(500); /* 500 ms */
143 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
144 int dtrace_devdepth_max = 32;
145 int dtrace_err_verbose;
146 hrtime_t dtrace_deadman_interval = NANOSEC;
147 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
148 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
149 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
150
151 /*
152 * DTrace External Variables
153 *
154 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
155 * available to DTrace consumers via the backtick (`) syntax. One of these,
156 * dtrace_zero, is made deliberately so: it is provided as a source of
157 * well-known, zero-filled memory. While this variable is not documented,
158 * it is used by some translators as an implementation detail.
159 */
160 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
161
162 /*
163 * DTrace Internal Variables
164 */
165 static dev_info_t *dtrace_devi; /* device info */
166 static vmem_t *dtrace_arena; /* probe ID arena */
167 static vmem_t *dtrace_minor; /* minor number arena */
168 static taskq_t *dtrace_taskq; /* task queue */
169 static dtrace_probe_t **dtrace_probes; /* array of all probes */
170 static int dtrace_nprobes; /* number of probes */
171 static dtrace_provider_t *dtrace_provider; /* provider list */
172 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
173 static int dtrace_opens; /* number of opens */
174 static int dtrace_helpers; /* number of helpers */
175 static int dtrace_getf; /* number of unpriv getf()s */
176 static void *dtrace_softstate; /* softstate pointer */
177 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
178 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
179 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
180 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
181 static int dtrace_toxranges; /* number of toxic ranges */
182 static int dtrace_toxranges_max; /* size of toxic range array */
183 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
184 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
185 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
186 static kthread_t *dtrace_panicked; /* panicking thread */
187 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
188 static dtrace_genid_t dtrace_probegen; /* current probe generation */
189 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
190 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
191 static dtrace_genid_t dtrace_retained_gen; /* current retained enab gen */
192 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
193 static int dtrace_dynvar_failclean; /* dynvars failed to clean */
194
195 /*
196 * DTrace Locking
197 * DTrace is protected by three (relatively coarse-grained) locks:
198 *
199 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
200 * including enabling state, probes, ECBs, consumer state, helper state,
201 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
202 * probe context is lock-free -- synchronization is handled via the
203 * dtrace_sync() cross call mechanism.
204 *
205 * (2) dtrace_provider_lock is required when manipulating provider state, or
206 * when provider state must be held constant.
207 *
208 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
209 * when meta provider state must be held constant.
210 *
211 * The lock ordering between these three locks is dtrace_meta_lock before
212 * dtrace_provider_lock before dtrace_lock. (In particular, there are
213 * several places where dtrace_provider_lock is held by the framework as it
214 * calls into the providers -- which then call back into the framework,
215 * grabbing dtrace_lock.)
216 *
217 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
218 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
219 * role as a coarse-grained lock; it is acquired before both of these locks.
220 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
221 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
222 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
223 * acquired _between_ dtrace_provider_lock and dtrace_lock.
224 */
225 static kmutex_t dtrace_lock; /* probe state lock */
226 static kmutex_t dtrace_provider_lock; /* provider state lock */
227 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
228
229 /*
230 * DTrace Provider Variables
231 *
232 * These are the variables relating to DTrace as a provider (that is, the
233 * provider of the BEGIN, END, and ERROR probes).
234 */
235 static dtrace_pattr_t dtrace_provider_attr = {
236 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
237 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
238 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
239 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
240 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
241 };
242
243 static void
244 dtrace_nullop_provide(void *arg __unused,
245 const dtrace_probedesc_t *spec __unused)
246 {
247 }
248
249 static void
250 dtrace_nullop_module(void *arg __unused, struct modctl *mp __unused)
251 {
252 }
253
254 static void
255 dtrace_nullop(void *arg __unused, dtrace_id_t id __unused, void *parg __unused)
256 {
257 }
258
259 static int
260 dtrace_enable_nullop(void *arg __unused, dtrace_id_t id __unused,
261 void *parg __unused)
262 {
263 return (0);
264 }
265
266 static dtrace_pops_t dtrace_provider_ops = {
267 .dtps_provide = dtrace_nullop_provide,
268 .dtps_provide_module = dtrace_nullop_module,
269 .dtps_enable = dtrace_enable_nullop,
270 .dtps_disable = dtrace_nullop,
271 .dtps_suspend = dtrace_nullop,
272 .dtps_resume = dtrace_nullop,
273 .dtps_getargdesc = NULL,
274 .dtps_getargval = NULL,
275 .dtps_mode = NULL,
276 .dtps_destroy = dtrace_nullop
277 };
278
279 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
280 static dtrace_id_t dtrace_probeid_end; /* special END probe */
281 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
282
283 /*
284 * DTrace Helper Tracing Variables
285 *
286 * These variables should be set dynamically to enable helper tracing. The
287 * only variables that should be set are dtrace_helptrace_enable (which should
288 * be set to a non-zero value to allocate helper tracing buffers on the next
289 * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
290 * non-zero value to deallocate helper tracing buffers on the next close of
291 * /dev/dtrace). When (and only when) helper tracing is disabled, the
292 * buffer size may also be set via dtrace_helptrace_bufsize.
293 */
294 int dtrace_helptrace_enable = 0;
295 int dtrace_helptrace_disable = 0;
296 int dtrace_helptrace_bufsize = 16 * 1024 * 1024;
297 uint32_t dtrace_helptrace_nlocals;
298 static dtrace_helptrace_t *dtrace_helptrace_buffer;
299 static uint32_t dtrace_helptrace_next = 0;
300 static int dtrace_helptrace_wrapped = 0;
301
302 /*
303 * DTrace Error Hashing
304 *
305 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
306 * table. This is very useful for checking coverage of tests that are
307 * expected to induce DIF or DOF processing errors, and may be useful for
308 * debugging problems in the DIF code generator or in DOF generation . The
309 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
310 */
311 #ifdef DEBUG
312 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
313 static const char *dtrace_errlast;
314 static kthread_t *dtrace_errthread;
315 static kmutex_t dtrace_errlock;
316 #endif
317
318 /*
319 * DTrace Macros and Constants
320 *
321 * These are various macros that are useful in various spots in the
322 * implementation, along with a few random constants that have no meaning
323 * outside of the implementation. There is no real structure to this cpp
324 * mishmash -- but is there ever?
325 */
326 #define DTRACE_HASHSTR(hash, probe) \
327 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
328
329 #define DTRACE_HASHNEXT(hash, probe) \
330 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
331
332 #define DTRACE_HASHPREV(hash, probe) \
333 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
334
335 #define DTRACE_HASHEQ(hash, lhs, rhs) \
336 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
337 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
338
339 #define DTRACE_AGGHASHSIZE_SLEW 17
340
341 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
342
343 /*
344 * The key for a thread-local variable consists of the lower 61 bits of the
345 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
346 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
347 * equal to a variable identifier. This is necessary (but not sufficient) to
348 * assure that global associative arrays never collide with thread-local
349 * variables. To guarantee that they cannot collide, we must also define the
350 * order for keying dynamic variables. That order is:
351 *
352 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
353 *
354 * Because the variable-key and the tls-key are in orthogonal spaces, there is
355 * no way for a global variable key signature to match a thread-local key
356 * signature.
357 */
358 #define DTRACE_TLS_THRKEY(where) { \
359 uint_t intr = 0; \
360 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
361 for (; actv; actv >>= 1) \
362 intr++; \
363 ASSERT(intr < (1 << 3)); \
364 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
365 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
366 }
367
368 #define DT_BSWAP_8(x) ((x) & 0xff)
369 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
370 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
371 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
372
373 #define DT_MASK_LO 0x00000000FFFFFFFFULL
374
375 #define DTRACE_STORE(type, tomax, offset, what) \
376 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
377
378 #ifndef __x86
379 #define DTRACE_ALIGNCHECK(addr, size, flags) \
380 if (addr & (size - 1)) { \
381 *flags |= CPU_DTRACE_BADALIGN; \
382 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
383 return (0); \
384 }
385 #else
386 #define DTRACE_ALIGNCHECK(addr, size, flags)
387 #endif
388
389 /*
390 * Test whether a range of memory starting at testaddr of size testsz falls
391 * within the range of memory described by addr, sz. We take care to avoid
392 * problems with overflow and underflow of the unsigned quantities, and
393 * disallow all negative sizes. Ranges of size 0 are allowed.
394 */
395 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
396 ((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
397 (testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
398 (testaddr) + (testsz) >= (testaddr))
399
400 #define DTRACE_RANGE_REMAIN(remp, addr, baseaddr, basesz) \
401 do { \
402 if ((remp) != NULL) { \
403 *(remp) = (uintptr_t)(baseaddr) + (basesz) - (addr); \
404 } \
405 _NOTE(CONSTCOND) } while (0)
406
407
408 /*
409 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
410 * alloc_sz on the righthand side of the comparison in order to avoid overflow
411 * or underflow in the comparison with it. This is simpler than the INRANGE
412 * check above, because we know that the dtms_scratch_ptr is valid in the
413 * range. Allocations of size zero are allowed.
414 */
415 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
416 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
417 (mstate)->dtms_scratch_ptr >= (alloc_sz))
418
419 #define DTRACE_LOADFUNC(bits) \
420 /*CSTYLED*/ \
421 uint##bits##_t \
422 dtrace_load##bits(uintptr_t addr) \
423 { \
424 size_t size = bits / NBBY; \
425 /*CSTYLED*/ \
426 uint##bits##_t rval; \
427 int i; \
428 volatile uint16_t *flags = (volatile uint16_t *) \
429 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; \
430 \
431 DTRACE_ALIGNCHECK(addr, size, flags); \
432 \
433 for (i = 0; i < dtrace_toxranges; i++) { \
434 if (addr >= dtrace_toxrange[i].dtt_limit) \
435 continue; \
436 \
437 if (addr + size <= dtrace_toxrange[i].dtt_base) \
438 continue; \
439 \
440 /* \
441 * This address falls within a toxic region; return 0. \
442 */ \
443 *flags |= CPU_DTRACE_BADADDR; \
444 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
445 return (0); \
446 } \
447 \
448 *flags |= CPU_DTRACE_NOFAULT; \
449 /*CSTYLED*/ \
450 rval = *((volatile uint##bits##_t *)addr); \
451 *flags &= ~CPU_DTRACE_NOFAULT; \
452 \
453 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
454 }
455
456 #ifdef _LP64
457 #define dtrace_loadptr dtrace_load64
458 #else
459 #define dtrace_loadptr dtrace_load32
460 #endif
461
462 #define DTRACE_DYNHASH_FREE 0
463 #define DTRACE_DYNHASH_SINK 1
464 #define DTRACE_DYNHASH_VALID 2
465
466 #define DTRACE_MATCH_FAIL -1
467 #define DTRACE_MATCH_NEXT 0
468 #define DTRACE_MATCH_DONE 1
469 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
470 #define DTRACE_STATE_ALIGN 64
471
472 #define DTRACE_FLAGS2FLT(flags) \
473 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
474 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
475 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
476 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
477 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
478 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
479 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
480 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
481 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
482 DTRACEFLT_UNKNOWN)
483
484 #define DTRACEACT_ISSTRING(act) \
485 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
486 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
487
488 static size_t dtrace_strlen(const char *, size_t);
489 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
490 static void dtrace_enabling_provide(dtrace_provider_t *);
491 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
492 static void dtrace_enabling_matchall(void);
493 static void dtrace_enabling_reap(void);
494 static dtrace_state_t *dtrace_anon_grab(void);
495 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
496 dtrace_state_t *, uint64_t, uint64_t);
497 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
498 static void dtrace_buffer_drop(dtrace_buffer_t *);
499 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
500 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
501 dtrace_state_t *, dtrace_mstate_t *);
502 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
503 dtrace_optval_t);
504 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
505 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
506 static int dtrace_priv_proc(dtrace_state_t *, dtrace_mstate_t *);
507 static void dtrace_getf_barrier(void);
508 static int dtrace_canload_remains(uint64_t, size_t, size_t *,
509 dtrace_mstate_t *, dtrace_vstate_t *);
510 static int dtrace_canstore_remains(uint64_t, size_t, size_t *,
511 dtrace_mstate_t *, dtrace_vstate_t *);
512
513 /*
514 * DTrace Probe Context Functions
515 *
516 * These functions are called from probe context. Because probe context is
517 * any context in which C may be called, arbitrarily locks may be held,
518 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
519 * As a result, functions called from probe context may only call other DTrace
520 * support functions -- they may not interact at all with the system at large.
521 * (Note that the ASSERT macro is made probe-context safe by redefining it in
522 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
523 * loads are to be performed from probe context, they _must_ be in terms of
524 * the safe dtrace_load*() variants.
525 *
526 * Some functions in this block are not actually called from probe context;
527 * for these functions, there will be a comment above the function reading
528 * "Note: not called from probe context."
529 */
530 void
531 dtrace_panic(const char *format, ...)
532 {
533 va_list alist;
534
535 va_start(alist, format);
536 dtrace_vpanic(format, alist);
537 va_end(alist);
538 }
539
540 int
541 dtrace_assfail(const char *a, const char *f, int l)
542 {
543 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
544
545 /*
546 * We just need something here that even the most clever compiler
547 * cannot optimize away.
548 */
549 return (a[(uintptr_t)f]);
550 }
551
552 /*
553 * Atomically increment a specified error counter from probe context.
554 */
555 static void
556 dtrace_error(uint32_t *counter)
557 {
558 /*
559 * Most counters stored to in probe context are per-CPU counters.
560 * However, there are some error conditions that are sufficiently
561 * arcane that they don't merit per-CPU storage. If these counters
562 * are incremented concurrently on different CPUs, scalability will be
563 * adversely affected -- but we don't expect them to be white-hot in a
564 * correctly constructed enabling...
565 */
566 uint32_t oval, nval;
567
568 do {
569 oval = *counter;
570
571 if ((nval = oval + 1) == 0) {
572 /*
573 * If the counter would wrap, set it to 1 -- assuring
574 * that the counter is never zero when we have seen
575 * errors. (The counter must be 32-bits because we
576 * aren't guaranteed a 64-bit compare&swap operation.)
577 * To save this code both the infamy of being fingered
578 * by a priggish news story and the indignity of being
579 * the target of a neo-puritan witch trial, we're
580 * carefully avoiding any colorful description of the
581 * likelihood of this condition -- but suffice it to
582 * say that it is only slightly more likely than the
583 * overflow of predicate cache IDs, as discussed in
584 * dtrace_predicate_create().
585 */
586 nval = 1;
587 }
588 } while (dtrace_cas32(counter, oval, nval) != oval);
589 }
590
591 /*
592 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
593 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
594 */
595 /* BEGIN CSTYLED */
596 DTRACE_LOADFUNC(8)
597 DTRACE_LOADFUNC(16)
598 DTRACE_LOADFUNC(32)
599 DTRACE_LOADFUNC(64)
600 /* END CSTYLED */
601
602 static int
603 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
604 {
605 if (dest < mstate->dtms_scratch_base)
606 return (0);
607
608 if (dest + size < dest)
609 return (0);
610
611 if (dest + size > mstate->dtms_scratch_ptr)
612 return (0);
613
614 return (1);
615 }
616
617 static int
618 dtrace_canstore_statvar(uint64_t addr, size_t sz, size_t *remain,
619 dtrace_statvar_t **svars, int nsvars)
620 {
621 int i;
622 size_t maxglobalsize, maxlocalsize;
623
624 if (nsvars == 0)
625 return (0);
626
627 maxglobalsize = dtrace_statvar_maxsize + sizeof (uint64_t);
628 maxlocalsize = maxglobalsize * NCPU;
629
630 for (i = 0; i < nsvars; i++) {
631 dtrace_statvar_t *svar = svars[i];
632 uint8_t scope;
633 size_t size;
634
635 if (svar == NULL || (size = svar->dtsv_size) == 0)
636 continue;
637
638 scope = svar->dtsv_var.dtdv_scope;
639
640 /*
641 * We verify that our size is valid in the spirit of providing
642 * defense in depth: we want to prevent attackers from using
643 * DTrace to escalate an orthogonal kernel heap corruption bug
644 * into the ability to store to arbitrary locations in memory.
645 */
646 VERIFY((scope == DIFV_SCOPE_GLOBAL && size <= maxglobalsize) ||
647 (scope == DIFV_SCOPE_LOCAL && size <= maxlocalsize));
648
649 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data,
650 svar->dtsv_size)) {
651 DTRACE_RANGE_REMAIN(remain, addr, svar->dtsv_data,
652 svar->dtsv_size);
653 return (1);
654 }
655 }
656
657 return (0);
658 }
659
660 /*
661 * Check to see if the address is within a memory region to which a store may
662 * be issued. This includes the DTrace scratch areas, and any DTrace variable
663 * region. The caller of dtrace_canstore() is responsible for performing any
664 * alignment checks that are needed before stores are actually executed.
665 */
666 static int
667 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
668 dtrace_vstate_t *vstate)
669 {
670 return (dtrace_canstore_remains(addr, sz, NULL, mstate, vstate));
671 }
672
673 /*
674 * Implementation of dtrace_canstore which communicates the upper bound of the
675 * allowed memory region.
676 */
677 static int
678 dtrace_canstore_remains(uint64_t addr, size_t sz, size_t *remain,
679 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
680 {
681 /*
682 * First, check to see if the address is in scratch space...
683 */
684 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
685 mstate->dtms_scratch_size)) {
686 DTRACE_RANGE_REMAIN(remain, addr, mstate->dtms_scratch_base,
687 mstate->dtms_scratch_size);
688 return (1);
689 }
690
691 /*
692 * Now check to see if it's a dynamic variable. This check will pick
693 * up both thread-local variables and any global dynamically-allocated
694 * variables.
695 */
696 if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
697 vstate->dtvs_dynvars.dtds_size)) {
698 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
699 uintptr_t base = (uintptr_t)dstate->dtds_base +
700 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
701 uintptr_t chunkoffs;
702 dtrace_dynvar_t *dvar;
703
704 /*
705 * Before we assume that we can store here, we need to make
706 * sure that it isn't in our metadata -- storing to our
707 * dynamic variable metadata would corrupt our state. For
708 * the range to not include any dynamic variable metadata,
709 * it must:
710 *
711 * (1) Start above the hash table that is at the base of
712 * the dynamic variable space
713 *
714 * (2) Have a starting chunk offset that is beyond the
715 * dtrace_dynvar_t that is at the base of every chunk
716 *
717 * (3) Not span a chunk boundary
718 *
719 * (4) Not be in the tuple space of a dynamic variable
720 *
721 */
722 if (addr < base)
723 return (0);
724
725 chunkoffs = (addr - base) % dstate->dtds_chunksize;
726
727 if (chunkoffs < sizeof (dtrace_dynvar_t))
728 return (0);
729
730 if (chunkoffs + sz > dstate->dtds_chunksize)
731 return (0);
732
733 dvar = (dtrace_dynvar_t *)((uintptr_t)addr - chunkoffs);
734
735 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE)
736 return (0);
737
738 if (chunkoffs < sizeof (dtrace_dynvar_t) +
739 ((dvar->dtdv_tuple.dtt_nkeys - 1) * sizeof (dtrace_key_t)))
740 return (0);
741
742 DTRACE_RANGE_REMAIN(remain, addr, dvar, dstate->dtds_chunksize);
743 return (1);
744 }
745
746 /*
747 * Finally, check the static local and global variables. These checks
748 * take the longest, so we perform them last.
749 */
750 if (dtrace_canstore_statvar(addr, sz, remain,
751 vstate->dtvs_locals, vstate->dtvs_nlocals))
752 return (1);
753
754 if (dtrace_canstore_statvar(addr, sz, remain,
755 vstate->dtvs_globals, vstate->dtvs_nglobals))
756 return (1);
757
758 return (0);
759 }
760
761
762 /*
763 * Convenience routine to check to see if the address is within a memory
764 * region in which a load may be issued given the user's privilege level;
765 * if not, it sets the appropriate error flags and loads 'addr' into the
766 * illegal value slot.
767 *
768 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
769 * appropriate memory access protection.
770 */
771 static int
772 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
773 dtrace_vstate_t *vstate)
774 {
775 return (dtrace_canload_remains(addr, sz, NULL, mstate, vstate));
776 }
777
778 /*
779 * Implementation of dtrace_canload which communicates the upper bound of the
780 * allowed memory region.
781 */
782 static int
783 dtrace_canload_remains(uint64_t addr, size_t sz, size_t *remain,
784 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
785 {
786 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
787 file_t *fp;
788
789 /*
790 * If we hold the privilege to read from kernel memory, then
791 * everything is readable.
792 */
793 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
794 DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
795 return (1);
796 }
797
798 /*
799 * You can obviously read that which you can store.
800 */
801 if (dtrace_canstore_remains(addr, sz, remain, mstate, vstate))
802 return (1);
803
804 /*
805 * We're allowed to read from our own string table.
806 */
807 if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
808 mstate->dtms_difo->dtdo_strlen)) {
809 DTRACE_RANGE_REMAIN(remain, addr,
810 mstate->dtms_difo->dtdo_strtab,
811 mstate->dtms_difo->dtdo_strlen);
812 return (1);
813 }
814
815 if (vstate->dtvs_state != NULL &&
816 dtrace_priv_proc(vstate->dtvs_state, mstate)) {
817 proc_t *p;
818
819 /*
820 * When we have privileges to the current process, there are
821 * several context-related kernel structures that are safe to
822 * read, even absent the privilege to read from kernel memory.
823 * These reads are safe because these structures contain only
824 * state that (1) we're permitted to read, (2) is harmless or
825 * (3) contains pointers to additional kernel state that we're
826 * not permitted to read (and as such, do not present an
827 * opportunity for privilege escalation). Finally (and
828 * critically), because of the nature of their relation with
829 * the current thread context, the memory associated with these
830 * structures cannot change over the duration of probe context,
831 * and it is therefore impossible for this memory to be
832 * deallocated and reallocated as something else while it's
833 * being operated upon.
834 */
835 if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t))) {
836 DTRACE_RANGE_REMAIN(remain, addr, curthread,
837 sizeof (kthread_t));
838 return (1);
839 }
840
841 if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
842 sz, curthread->t_procp, sizeof (proc_t))) {
843 DTRACE_RANGE_REMAIN(remain, addr, curthread->t_procp,
844 sizeof (proc_t));
845 return (1);
846 }
847
848 if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
849 curthread->t_cred, sizeof (cred_t))) {
850 DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cred,
851 sizeof (cred_t));
852 return (1);
853 }
854
855 if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
856 &(p->p_pidp->pid_id), sizeof (pid_t))) {
857 DTRACE_RANGE_REMAIN(remain, addr, &(p->p_pidp->pid_id),
858 sizeof (pid_t));
859 return (1);
860 }
861
862 if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
863 curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
864 DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cpu,
865 offsetof(cpu_t, cpu_pause_thread));
866 return (1);
867 }
868 }
869
870 if ((fp = mstate->dtms_getf) != NULL) {
871 uintptr_t psz = sizeof (void *);
872 vnode_t *vp;
873 vnodeops_t *op;
874
875 /*
876 * When getf() returns a file_t, the enabling is implicitly
877 * granted the (transient) right to read the returned file_t
878 * as well as the v_path and v_op->vnop_name of the underlying
879 * vnode. These accesses are allowed after a successful
880 * getf() because the members that they refer to cannot change
881 * once set -- and the barrier logic in the kernel's closef()
882 * path assures that the file_t and its referenced vode_t
883 * cannot themselves be stale (that is, it impossible for
884 * either dtms_getf itself or its f_vnode member to reference
885 * freed memory).
886 */
887 if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t))) {
888 DTRACE_RANGE_REMAIN(remain, addr, fp, sizeof (file_t));
889 return (1);
890 }
891
892 if ((vp = fp->f_vnode) != NULL) {
893 size_t slen;
894
895 if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz)) {
896 DTRACE_RANGE_REMAIN(remain, addr, &vp->v_path,
897 psz);
898 return (1);
899 }
900
901 slen = strlen(vp->v_path) + 1;
902 if (DTRACE_INRANGE(addr, sz, vp->v_path, slen)) {
903 DTRACE_RANGE_REMAIN(remain, addr, vp->v_path,
904 slen);
905 return (1);
906 }
907
908 if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz)) {
909 DTRACE_RANGE_REMAIN(remain, addr, &vp->v_op,
910 psz);
911 return (1);
912 }
913
914 if ((op = vp->v_op) != NULL &&
915 DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
916 DTRACE_RANGE_REMAIN(remain, addr,
917 &op->vnop_name, psz);
918 return (1);
919 }
920
921 if (op != NULL && op->vnop_name != NULL &&
922 DTRACE_INRANGE(addr, sz, op->vnop_name,
923 (slen = strlen(op->vnop_name) + 1))) {
924 DTRACE_RANGE_REMAIN(remain, addr,
925 op->vnop_name, slen);
926 return (1);
927 }
928 }
929 }
930
931 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
932 *illval = addr;
933 return (0);
934 }
935
936 /*
937 * Convenience routine to check to see if a given string is within a memory
938 * region in which a load may be issued given the user's privilege level;
939 * this exists so that we don't need to issue unnecessary dtrace_strlen()
940 * calls in the event that the user has all privileges.
941 */
942 static int
943 dtrace_strcanload(uint64_t addr, size_t sz, size_t *remain,
944 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
945 {
946 size_t rsize;
947
948 /*
949 * If we hold the privilege to read from kernel memory, then
950 * everything is readable.
951 */
952 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
953 DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
954 return (1);
955 }
956
957 /*
958 * Even if the caller is uninterested in querying the remaining valid
959 * range, it is required to ensure that the access is allowed.
960 */
961 if (remain == NULL) {
962 remain = &rsize;
963 }
964 if (dtrace_canload_remains(addr, 0, remain, mstate, vstate)) {
965 size_t strsz;
966 /*
967 * Perform the strlen after determining the length of the
968 * memory region which is accessible. This prevents timing
969 * information from being used to find NULs in memory which is
970 * not accessible to the caller.
971 */
972 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr,
973 MIN(sz, *remain));
974 if (strsz <= *remain) {
975 return (1);
976 }
977 }
978
979 return (0);
980 }
981
982 /*
983 * Convenience routine to check to see if a given variable is within a memory
984 * region in which a load may be issued given the user's privilege level.
985 */
986 static int
987 dtrace_vcanload(void *src, dtrace_diftype_t *type, size_t *remain,
988 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
989 {
990 size_t sz;
991 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
992
993 /*
994 * Calculate the max size before performing any checks since even
995 * DTRACE_ACCESS_KERNEL-credentialed callers expect that this function
996 * return the max length via 'remain'.
997 */
998 if (type->dtdt_kind == DIF_TYPE_STRING) {
999 dtrace_state_t *state = vstate->dtvs_state;
1000
1001 if (state != NULL) {
1002 sz = state->dts_options[DTRACEOPT_STRSIZE];
1003 } else {
1004 /*
1005 * In helper context, we have a NULL state; fall back
1006 * to using the system-wide default for the string size
1007 * in this case.
1008 */
1009 sz = dtrace_strsize_default;
1010 }
1011 } else {
1012 sz = type->dtdt_size;
1013 }
1014
1015 /*
1016 * If we hold the privilege to read from kernel memory, then
1017 * everything is readable.
1018 */
1019 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
1020 DTRACE_RANGE_REMAIN(remain, (uintptr_t)src, src, sz);
1021 return (1);
1022 }
1023
1024 if (type->dtdt_kind == DIF_TYPE_STRING) {
1025 return (dtrace_strcanload((uintptr_t)src, sz, remain, mstate,
1026 vstate));
1027 }
1028 return (dtrace_canload_remains((uintptr_t)src, sz, remain, mstate,
1029 vstate));
1030 }
1031
1032 /*
1033 * Convert a string to a signed integer using safe loads.
1034 *
1035 * NOTE: This function uses various macros from strtolctype.h to manipulate
1036 * digit values, etc -- these have all been checked to ensure they make
1037 * no additional function calls.
1038 */
1039 static int64_t
1040 dtrace_strtoll(char *input, int base, size_t limit)
1041 {
1042 uintptr_t pos = (uintptr_t)input;
1043 int64_t val = 0;
1044 int x;
1045 boolean_t neg = B_FALSE;
1046 char c, cc, ccc;
1047 uintptr_t end = pos + limit;
1048
1049 /*
1050 * Consume any whitespace preceding digits.
1051 */
1052 while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
1053 pos++;
1054
1055 /*
1056 * Handle an explicit sign if one is present.
1057 */
1058 if (c == '-' || c == '+') {
1059 if (c == '-')
1060 neg = B_TRUE;
1061 c = dtrace_load8(++pos);
1062 }
1063
1064 /*
1065 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
1066 * if present.
1067 */
1068 if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
1069 cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
1070 pos += 2;
1071 c = ccc;
1072 }
1073
1074 /*
1075 * Read in contiguous digits until the first non-digit character.
1076 */
1077 for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
1078 c = dtrace_load8(++pos))
1079 val = val * base + x;
1080
1081 return (neg ? -val : val);
1082 }
1083
1084 /*
1085 * Compare two strings using safe loads.
1086 */
1087 static int
1088 dtrace_strncmp(char *s1, char *s2, size_t limit)
1089 {
1090 uint8_t c1, c2;
1091 volatile uint16_t *flags;
1092
1093 if (s1 == s2 || limit == 0)
1094 return (0);
1095
1096 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1097
1098 do {
1099 if (s1 == NULL) {
1100 c1 = '\0';
1101 } else {
1102 c1 = dtrace_load8((uintptr_t)s1++);
1103 }
1104
1105 if (s2 == NULL) {
1106 c2 = '\0';
1107 } else {
1108 c2 = dtrace_load8((uintptr_t)s2++);
1109 }
1110
1111 if (c1 != c2)
1112 return (c1 - c2);
1113 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
1114
1115 return (0);
1116 }
1117
1118 /*
1119 * Compute strlen(s) for a string using safe memory accesses. The additional
1120 * len parameter is used to specify a maximum length to ensure completion.
1121 */
1122 static size_t
1123 dtrace_strlen(const char *s, size_t lim)
1124 {
1125 uint_t len;
1126
1127 for (len = 0; len != lim; len++) {
1128 if (dtrace_load8((uintptr_t)s++) == '\0')
1129 break;
1130 }
1131
1132 return (len);
1133 }
1134
1135 /*
1136 * Check if an address falls within a toxic region.
1137 */
1138 static int
1139 dtrace_istoxic(uintptr_t kaddr, size_t size)
1140 {
1141 uintptr_t taddr, tsize;
1142 int i;
1143
1144 for (i = 0; i < dtrace_toxranges; i++) {
1145 taddr = dtrace_toxrange[i].dtt_base;
1146 tsize = dtrace_toxrange[i].dtt_limit - taddr;
1147
1148 if (kaddr - taddr < tsize) {
1149 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1150 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
1151 return (1);
1152 }
1153
1154 if (taddr - kaddr < size) {
1155 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1156 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
1157 return (1);
1158 }
1159 }
1160
1161 return (0);
1162 }
1163
1164 /*
1165 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
1166 * memory specified by the DIF program. The dst is assumed to be safe memory
1167 * that we can store to directly because it is managed by DTrace. As with
1168 * standard bcopy, overlapping copies are handled properly.
1169 */
1170 static void
1171 dtrace_bcopy(const void *src, void *dst, size_t len)
1172 {
1173 if (len != 0) {
1174 uint8_t *s1 = dst;
1175 const uint8_t *s2 = src;
1176
1177 if (s1 <= s2) {
1178 do {
1179 *s1++ = dtrace_load8((uintptr_t)s2++);
1180 } while (--len != 0);
1181 } else {
1182 s2 += len;
1183 s1 += len;
1184
1185 do {
1186 *--s1 = dtrace_load8((uintptr_t)--s2);
1187 } while (--len != 0);
1188 }
1189 }
1190 }
1191
1192 /*
1193 * Copy src to dst using safe memory accesses, up to either the specified
1194 * length, or the point that a nul byte is encountered. The src is assumed to
1195 * be unsafe memory specified by the DIF program. The dst is assumed to be
1196 * safe memory that we can store to directly because it is managed by DTrace.
1197 * Unlike dtrace_bcopy(), overlapping regions are not handled.
1198 */
1199 static void
1200 dtrace_strcpy(const void *src, void *dst, size_t len)
1201 {
1202 if (len != 0) {
1203 uint8_t *s1 = dst, c;
1204 const uint8_t *s2 = src;
1205
1206 do {
1207 *s1++ = c = dtrace_load8((uintptr_t)s2++);
1208 } while (--len != 0 && c != '\0');
1209 }
1210 }
1211
1212 /*
1213 * Copy src to dst, deriving the size and type from the specified (BYREF)
1214 * variable type. The src is assumed to be unsafe memory specified by the DIF
1215 * program. The dst is assumed to be DTrace variable memory that is of the
1216 * specified type; we assume that we can store to directly.
1217 */
1218 static void
1219 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type, size_t limit)
1220 {
1221 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1222
1223 if (type->dtdt_kind == DIF_TYPE_STRING) {
1224 dtrace_strcpy(src, dst, MIN(type->dtdt_size, limit));
1225 } else {
1226 dtrace_bcopy(src, dst, MIN(type->dtdt_size, limit));
1227 }
1228 }
1229
1230 /*
1231 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1232 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1233 * safe memory that we can access directly because it is managed by DTrace.
1234 */
1235 static int
1236 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1237 {
1238 volatile uint16_t *flags;
1239
1240 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1241
1242 if (s1 == s2)
1243 return (0);
1244
1245 if (s1 == NULL || s2 == NULL)
1246 return (1);
1247
1248 if (s1 != s2 && len != 0) {
1249 const uint8_t *ps1 = s1;
1250 const uint8_t *ps2 = s2;
1251
1252 do {
1253 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1254 return (1);
1255 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1256 }
1257 return (0);
1258 }
1259
1260 /*
1261 * Zero the specified region using a simple byte-by-byte loop. Note that this
1262 * is for safe DTrace-managed memory only.
1263 */
1264 static void
1265 dtrace_bzero(void *dst, size_t len)
1266 {
1267 uchar_t *cp;
1268
1269 for (cp = dst; len != 0; len--)
1270 *cp++ = 0;
1271 }
1272
1273 static void
1274 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1275 {
1276 uint64_t result[2];
1277
1278 result[0] = addend1[0] + addend2[0];
1279 result[1] = addend1[1] + addend2[1] +
1280 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1281
1282 sum[0] = result[0];
1283 sum[1] = result[1];
1284 }
1285
1286 /*
1287 * Shift the 128-bit value in a by b. If b is positive, shift left.
1288 * If b is negative, shift right.
1289 */
1290 static void
1291 dtrace_shift_128(uint64_t *a, int b)
1292 {
1293 uint64_t mask;
1294
1295 if (b == 0)
1296 return;
1297
1298 if (b < 0) {
1299 b = -b;
1300 if (b >= 64) {
1301 a[0] = a[1] >> (b - 64);
1302 a[1] = 0;
1303 } else {
1304 a[0] >>= b;
1305 mask = 1LL << (64 - b);
1306 mask -= 1;
1307 a[0] |= ((a[1] & mask) << (64 - b));
1308 a[1] >>= b;
1309 }
1310 } else {
1311 if (b >= 64) {
1312 a[1] = a[0] << (b - 64);
1313 a[0] = 0;
1314 } else {
1315 a[1] <<= b;
1316 mask = a[0] >> (64 - b);
1317 a[1] |= mask;
1318 a[0] <<= b;
1319 }
1320 }
1321 }
1322
1323 /*
1324 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1325 * use native multiplication on those, and then re-combine into the
1326 * resulting 128-bit value.
1327 *
1328 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1329 * hi1 * hi2 << 64 +
1330 * hi1 * lo2 << 32 +
1331 * hi2 * lo1 << 32 +
1332 * lo1 * lo2
1333 */
1334 static void
1335 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1336 {
1337 uint64_t hi1, hi2, lo1, lo2;
1338 uint64_t tmp[2];
1339
1340 hi1 = factor1 >> 32;
1341 hi2 = factor2 >> 32;
1342
1343 lo1 = factor1 & DT_MASK_LO;
1344 lo2 = factor2 & DT_MASK_LO;
1345
1346 product[0] = lo1 * lo2;
1347 product[1] = hi1 * hi2;
1348
1349 tmp[0] = hi1 * lo2;
1350 tmp[1] = 0;
1351 dtrace_shift_128(tmp, 32);
1352 dtrace_add_128(product, tmp, product);
1353
1354 tmp[0] = hi2 * lo1;
1355 tmp[1] = 0;
1356 dtrace_shift_128(tmp, 32);
1357 dtrace_add_128(product, tmp, product);
1358 }
1359
1360 /*
1361 * This privilege check should be used by actions and subroutines to
1362 * verify that the user credentials of the process that enabled the
1363 * invoking ECB match the target credentials
1364 */
1365 static int
1366 dtrace_priv_proc_common_user(dtrace_state_t *state)
1367 {
1368 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1369
1370 /*
1371 * We should always have a non-NULL state cred here, since if cred
1372 * is null (anonymous tracing), we fast-path bypass this routine.
1373 */
1374 ASSERT(s_cr != NULL);
1375
1376 if ((cr = CRED()) != NULL &&
1377 s_cr->cr_uid == cr->cr_uid &&
1378 s_cr->cr_uid == cr->cr_ruid &&
1379 s_cr->cr_uid == cr->cr_suid &&
1380 s_cr->cr_gid == cr->cr_gid &&
1381 s_cr->cr_gid == cr->cr_rgid &&
1382 s_cr->cr_gid == cr->cr_sgid)
1383 return (1);
1384
1385 return (0);
1386 }
1387
1388 /*
1389 * This privilege check should be used by actions and subroutines to
1390 * verify that the zone of the process that enabled the invoking ECB
1391 * matches the target credentials
1392 */
1393 static int
1394 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1395 {
1396 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1397
1398 /*
1399 * We should always have a non-NULL state cred here, since if cred
1400 * is null (anonymous tracing), we fast-path bypass this routine.
1401 */
1402 ASSERT(s_cr != NULL);
1403
1404 if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1405 return (1);
1406
1407 return (0);
1408 }
1409
1410 /*
1411 * This privilege check should be used by actions and subroutines to
1412 * verify that the process has not setuid or changed credentials.
1413 */
1414 static int
1415 dtrace_priv_proc_common_nocd()
1416 {
1417 proc_t *proc;
1418
1419 if ((proc = ttoproc(curthread)) != NULL &&
1420 !(proc->p_flag & SNOCD))
1421 return (1);
1422
1423 return (0);
1424 }
1425
1426 static int
1427 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1428 {
1429 int action = state->dts_cred.dcr_action;
1430
1431 if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1432 goto bad;
1433
1434 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1435 dtrace_priv_proc_common_zone(state) == 0)
1436 goto bad;
1437
1438 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1439 dtrace_priv_proc_common_user(state) == 0)
1440 goto bad;
1441
1442 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1443 dtrace_priv_proc_common_nocd() == 0)
1444 goto bad;
1445
1446 return (1);
1447
1448 bad:
1449 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1450
1451 return (0);
1452 }
1453
1454 static int
1455 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1456 {
1457 if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1458 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1459 return (1);
1460
1461 if (dtrace_priv_proc_common_zone(state) &&
1462 dtrace_priv_proc_common_user(state) &&
1463 dtrace_priv_proc_common_nocd())
1464 return (1);
1465 }
1466
1467 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1468
1469 return (0);
1470 }
1471
1472 static int
1473 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1474 {
1475 if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1476 (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1477 return (1);
1478
1479 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1480
1481 return (0);
1482 }
1483
1484 static int
1485 dtrace_priv_kernel(dtrace_state_t *state)
1486 {
1487 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1488 return (1);
1489
1490 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1491
1492 return (0);
1493 }
1494
1495 static int
1496 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1497 {
1498 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1499 return (1);
1500
1501 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1502
1503 return (0);
1504 }
1505
1506 /*
1507 * Determine if the dte_cond of the specified ECB allows for processing of
1508 * the current probe to continue. Note that this routine may allow continued
1509 * processing, but with access(es) stripped from the mstate's dtms_access
1510 * field.
1511 */
1512 static int
1513 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1514 dtrace_ecb_t *ecb)
1515 {
1516 dtrace_probe_t *probe = ecb->dte_probe;
1517 dtrace_provider_t *prov = probe->dtpr_provider;
1518 dtrace_pops_t *pops = &prov->dtpv_pops;
1519 int mode = DTRACE_MODE_NOPRIV_DROP;
1520
1521 ASSERT(ecb->dte_cond);
1522
1523 if (pops->dtps_mode != NULL) {
1524 mode = pops->dtps_mode(prov->dtpv_arg,
1525 probe->dtpr_id, probe->dtpr_arg);
1526
1527 ASSERT(mode & (DTRACE_MODE_USER | DTRACE_MODE_KERNEL));
1528 ASSERT(mode & (DTRACE_MODE_NOPRIV_RESTRICT |
1529 DTRACE_MODE_NOPRIV_DROP));
1530 }
1531
1532 /*
1533 * If the dte_cond bits indicate that this consumer is only allowed to
1534 * see user-mode firings of this probe, check that the probe was fired
1535 * while in a user context. If that's not the case, use the policy
1536 * specified by the provider to determine if we drop the probe or
1537 * merely restrict operation.
1538 */
1539 if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1540 ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1541
1542 if (!(mode & DTRACE_MODE_USER)) {
1543 if (mode & DTRACE_MODE_NOPRIV_DROP)
1544 return (0);
1545
1546 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1547 }
1548 }
1549
1550 /*
1551 * This is more subtle than it looks. We have to be absolutely certain
1552 * that CRED() isn't going to change out from under us so it's only
1553 * legit to examine that structure if we're in constrained situations.
1554 * Currently, the only times we'll this check is if a non-super-user
1555 * has enabled the profile or syscall providers -- providers that
1556 * allow visibility of all processes. For the profile case, the check
1557 * above will ensure that we're examining a user context.
1558 */
1559 if (ecb->dte_cond & DTRACE_COND_OWNER) {
1560 cred_t *cr;
1561 cred_t *s_cr = state->dts_cred.dcr_cred;
1562 proc_t *proc;
1563
1564 ASSERT(s_cr != NULL);
1565
1566 if ((cr = CRED()) == NULL ||
1567 s_cr->cr_uid != cr->cr_uid ||
1568 s_cr->cr_uid != cr->cr_ruid ||
1569 s_cr->cr_uid != cr->cr_suid ||
1570 s_cr->cr_gid != cr->cr_gid ||
1571 s_cr->cr_gid != cr->cr_rgid ||
1572 s_cr->cr_gid != cr->cr_sgid ||
1573 (proc = ttoproc(curthread)) == NULL ||
1574 (proc->p_flag & SNOCD)) {
1575 if (mode & DTRACE_MODE_NOPRIV_DROP)
1576 return (0);
1577
1578 mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1579 }
1580 }
1581
1582 /*
1583 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1584 * in our zone, check to see if our mode policy is to restrict rather
1585 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1586 * and DTRACE_ACCESS_ARGS
1587 */
1588 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1589 cred_t *cr;
1590 cred_t *s_cr = state->dts_cred.dcr_cred;
1591
1592 ASSERT(s_cr != NULL);
1593
1594 if ((cr = CRED()) == NULL ||
1595 s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1596 if (mode & DTRACE_MODE_NOPRIV_DROP)
1597 return (0);
1598
1599 mstate->dtms_access &=
1600 ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1601 }
1602 }
1603
1604 /*
1605 * By merits of being in this code path at all, we have limited
1606 * privileges. If the provider has indicated that limited privileges
1607 * are to denote restricted operation, strip off the ability to access
1608 * arguments.
1609 */
1610 if (mode & DTRACE_MODE_LIMITEDPRIV_RESTRICT)
1611 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1612
1613 return (1);
1614 }
1615
1616 /*
1617 * Note: not called from probe context. This function is called
1618 * asynchronously (and at a regular interval) from outside of probe context to
1619 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1620 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1621 */
1622 void
1623 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1624 {
1625 dtrace_dynvar_t *dirty;
1626 dtrace_dstate_percpu_t *dcpu;
1627 dtrace_dynvar_t **rinsep;
1628 int i, j, work = 0;
1629
1630 for (i = 0; i < NCPU; i++) {
1631 dcpu = &dstate->dtds_percpu[i];
1632 rinsep = &dcpu->dtdsc_rinsing;
1633
1634 /*
1635 * If the dirty list is NULL, there is no dirty work to do.
1636 */
1637 if (dcpu->dtdsc_dirty == NULL)
1638 continue;
1639
1640 if (dcpu->dtdsc_rinsing != NULL) {
1641 /*
1642 * If the rinsing list is non-NULL, then it is because
1643 * this CPU was selected to accept another CPU's
1644 * dirty list -- and since that time, dirty buffers
1645 * have accumulated. This is a highly unlikely
1646 * condition, but we choose to ignore the dirty
1647 * buffers -- they'll be picked up a future cleanse.
1648 */
1649 continue;
1650 }
1651
1652 if (dcpu->dtdsc_clean != NULL) {
1653 /*
1654 * If the clean list is non-NULL, then we're in a
1655 * situation where a CPU has done deallocations (we
1656 * have a non-NULL dirty list) but no allocations (we
1657 * also have a non-NULL clean list). We can't simply
1658 * move the dirty list into the clean list on this
1659 * CPU, yet we also don't want to allow this condition
1660 * to persist, lest a short clean list prevent a
1661 * massive dirty list from being cleaned (which in
1662 * turn could lead to otherwise avoidable dynamic
1663 * drops). To deal with this, we look for some CPU
1664 * with a NULL clean list, NULL dirty list, and NULL
1665 * rinsing list -- and then we borrow this CPU to
1666 * rinse our dirty list.
1667 */
1668 for (j = 0; j < NCPU; j++) {
1669 dtrace_dstate_percpu_t *rinser;
1670
1671 rinser = &dstate->dtds_percpu[j];
1672
1673 if (rinser->dtdsc_rinsing != NULL)
1674 continue;
1675
1676 if (rinser->dtdsc_dirty != NULL)
1677 continue;
1678
1679 if (rinser->dtdsc_clean != NULL)
1680 continue;
1681
1682 rinsep = &rinser->dtdsc_rinsing;
1683 break;
1684 }
1685
1686 if (j == NCPU) {
1687 /*
1688 * We were unable to find another CPU that
1689 * could accept this dirty list -- we are
1690 * therefore unable to clean it now.
1691 */
1692 dtrace_dynvar_failclean++;
1693 continue;
1694 }
1695 }
1696
1697 work = 1;
1698
1699 /*
1700 * Atomically move the dirty list aside.
1701 */
1702 do {
1703 dirty = dcpu->dtdsc_dirty;
1704
1705 /*
1706 * Before we zap the dirty list, set the rinsing list.
1707 * (This allows for a potential assertion in
1708 * dtrace_dynvar(): if a free dynamic variable appears
1709 * on a hash chain, either the dirty list or the
1710 * rinsing list for some CPU must be non-NULL.)
1711 */
1712 *rinsep = dirty;
1713 dtrace_membar_producer();
1714 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1715 dirty, NULL) != dirty);
1716 }
1717
1718 if (!work) {
1719 /*
1720 * We have no work to do; we can simply return.
1721 */
1722 return;
1723 }
1724
1725 dtrace_sync();
1726
1727 for (i = 0; i < NCPU; i++) {
1728 dcpu = &dstate->dtds_percpu[i];
1729
1730 if (dcpu->dtdsc_rinsing == NULL)
1731 continue;
1732
1733 /*
1734 * We are now guaranteed that no hash chain contains a pointer
1735 * into this dirty list; we can make it clean.
1736 */
1737 ASSERT(dcpu->dtdsc_clean == NULL);
1738 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1739 dcpu->dtdsc_rinsing = NULL;
1740 }
1741
1742 /*
1743 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1744 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1745 * This prevents a race whereby a CPU incorrectly decides that
1746 * the state should be something other than DTRACE_DSTATE_CLEAN
1747 * after dtrace_dynvar_clean() has completed.
1748 */
1749 dtrace_sync();
1750
1751 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1752 }
1753
1754 /*
1755 * Depending on the value of the op parameter, this function looks-up,
1756 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1757 * allocation is requested, this function will return a pointer to a
1758 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1759 * variable can be allocated. If NULL is returned, the appropriate counter
1760 * will be incremented.
1761 */
1762 dtrace_dynvar_t *
1763 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1764 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1765 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1766 {
1767 uint64_t hashval = DTRACE_DYNHASH_VALID;
1768 dtrace_dynhash_t *hash = dstate->dtds_hash;
1769 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1770 processorid_t me = CPU->cpu_id, cpu = me;
1771 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1772 size_t bucket, ksize;
1773 size_t chunksize = dstate->dtds_chunksize;
1774 uintptr_t kdata, lock, nstate;
1775 uint_t i;
1776
1777 ASSERT(nkeys != 0);
1778
1779 /*
1780 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1781 * algorithm. For the by-value portions, we perform the algorithm in
1782 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1783 * bit, and seems to have only a minute effect on distribution. For
1784 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1785 * over each referenced byte. It's painful to do this, but it's much
1786 * better than pathological hash distribution. The efficacy of the
1787 * hashing algorithm (and a comparison with other algorithms) may be
1788 * found by running the ::dtrace_dynstat MDB dcmd.
1789 */
1790 for (i = 0; i < nkeys; i++) {
1791 if (key[i].dttk_size == 0) {
1792 uint64_t val = key[i].dttk_value;
1793
1794 hashval += (val >> 48) & 0xffff;
1795 hashval += (hashval << 10);
1796 hashval ^= (hashval >> 6);
1797
1798 hashval += (val >> 32) & 0xffff;
1799 hashval += (hashval << 10);
1800 hashval ^= (hashval >> 6);
1801
1802 hashval += (val >> 16) & 0xffff;
1803 hashval += (hashval << 10);
1804 hashval ^= (hashval >> 6);
1805
1806 hashval += val & 0xffff;
1807 hashval += (hashval << 10);
1808 hashval ^= (hashval >> 6);
1809 } else {
1810 /*
1811 * This is incredibly painful, but it beats the hell
1812 * out of the alternative.
1813 */
1814 uint64_t j, size = key[i].dttk_size;
1815 uintptr_t base = (uintptr_t)key[i].dttk_value;
1816
1817 if (!dtrace_canload(base, size, mstate, vstate))
1818 break;
1819
1820 for (j = 0; j < size; j++) {
1821 hashval += dtrace_load8(base + j);
1822 hashval += (hashval << 10);
1823 hashval ^= (hashval >> 6);
1824 }
1825 }
1826 }
1827
1828 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1829 return (NULL);
1830
1831 hashval += (hashval << 3);
1832 hashval ^= (hashval >> 11);
1833 hashval += (hashval << 15);
1834
1835 /*
1836 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1837 * comes out to be one of our two sentinel hash values. If this
1838 * actually happens, we set the hashval to be a value known to be a
1839 * non-sentinel value.
1840 */
1841 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1842 hashval = DTRACE_DYNHASH_VALID;
1843
1844 /*
1845 * Yes, it's painful to do a divide here. If the cycle count becomes
1846 * important here, tricks can be pulled to reduce it. (However, it's
1847 * critical that hash collisions be kept to an absolute minimum;
1848 * they're much more painful than a divide.) It's better to have a
1849 * solution that generates few collisions and still keeps things
1850 * relatively simple.
1851 */
1852 bucket = hashval % dstate->dtds_hashsize;
1853
1854 if (op == DTRACE_DYNVAR_DEALLOC) {
1855 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1856
1857 for (;;) {
1858 while ((lock = *lockp) & 1)
1859 continue;
1860
1861 if (dtrace_casptr((void *)lockp,
1862 (void *)lock, (void *)(lock + 1)) == (void *)lock)
1863 break;
1864 }
1865
1866 dtrace_membar_producer();
1867 }
1868
1869 top:
1870 prev = NULL;
1871 lock = hash[bucket].dtdh_lock;
1872
1873 dtrace_membar_consumer();
1874
1875 start = hash[bucket].dtdh_chain;
1876 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1877 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1878 op != DTRACE_DYNVAR_DEALLOC));
1879
1880 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1881 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1882 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1883
1884 if (dvar->dtdv_hashval != hashval) {
1885 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1886 /*
1887 * We've reached the sink, and therefore the
1888 * end of the hash chain; we can kick out of
1889 * the loop knowing that we have seen a valid
1890 * snapshot of state.
1891 */
1892 ASSERT(dvar->dtdv_next == NULL);
1893 ASSERT(dvar == &dtrace_dynhash_sink);
1894 break;
1895 }
1896
1897 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1898 /*
1899 * We've gone off the rails: somewhere along
1900 * the line, one of the members of this hash
1901 * chain was deleted. Note that we could also
1902 * detect this by simply letting this loop run
1903 * to completion, as we would eventually hit
1904 * the end of the dirty list. However, we
1905 * want to avoid running the length of the
1906 * dirty list unnecessarily (it might be quite
1907 * long), so we catch this as early as
1908 * possible by detecting the hash marker. In
1909 * this case, we simply set dvar to NULL and
1910 * break; the conditional after the loop will
1911 * send us back to top.
1912 */
1913 dvar = NULL;
1914 break;
1915 }
1916
1917 goto next;
1918 }
1919
1920 if (dtuple->dtt_nkeys != nkeys)
1921 goto next;
1922
1923 for (i = 0; i < nkeys; i++, dkey++) {
1924 if (dkey->dttk_size != key[i].dttk_size)
1925 goto next; /* size or type mismatch */
1926
1927 if (dkey->dttk_size != 0) {
1928 if (dtrace_bcmp(
1929 (void *)(uintptr_t)key[i].dttk_value,
1930 (void *)(uintptr_t)dkey->dttk_value,
1931 dkey->dttk_size))
1932 goto next;
1933 } else {
1934 if (dkey->dttk_value != key[i].dttk_value)
1935 goto next;
1936 }
1937 }
1938
1939 if (op != DTRACE_DYNVAR_DEALLOC)
1940 return (dvar);
1941
1942 ASSERT(dvar->dtdv_next == NULL ||
1943 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1944
1945 if (prev != NULL) {
1946 ASSERT(hash[bucket].dtdh_chain != dvar);
1947 ASSERT(start != dvar);
1948 ASSERT(prev->dtdv_next == dvar);
1949 prev->dtdv_next = dvar->dtdv_next;
1950 } else {
1951 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1952 start, dvar->dtdv_next) != start) {
1953 /*
1954 * We have failed to atomically swing the
1955 * hash table head pointer, presumably because
1956 * of a conflicting allocation on another CPU.
1957 * We need to reread the hash chain and try
1958 * again.
1959 */
1960 goto top;
1961 }
1962 }
1963
1964 dtrace_membar_producer();
1965
1966 /*
1967 * Now set the hash value to indicate that it's free.
1968 */
1969 ASSERT(hash[bucket].dtdh_chain != dvar);
1970 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1971
1972 dtrace_membar_producer();
1973
1974 /*
1975 * Set the next pointer to point at the dirty list, and
1976 * atomically swing the dirty pointer to the newly freed dvar.
1977 */
1978 do {
1979 next = dcpu->dtdsc_dirty;
1980 dvar->dtdv_next = next;
1981 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1982
1983 /*
1984 * Finally, unlock this hash bucket.
1985 */
1986 ASSERT(hash[bucket].dtdh_lock == lock);
1987 ASSERT(lock & 1);
1988 hash[bucket].dtdh_lock++;
1989
1990 return (NULL);
1991 next:
1992 prev = dvar;
1993 continue;
1994 }
1995
1996 if (dvar == NULL) {
1997 /*
1998 * If dvar is NULL, it is because we went off the rails:
1999 * one of the elements that we traversed in the hash chain
2000 * was deleted while we were traversing it. In this case,
2001 * we assert that we aren't doing a dealloc (deallocs lock
2002 * the hash bucket to prevent themselves from racing with
2003 * one another), and retry the hash chain traversal.
2004 */
2005 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
2006 goto top;
2007 }
2008
2009 if (op != DTRACE_DYNVAR_ALLOC) {
2010 /*
2011 * If we are not to allocate a new variable, we want to
2012 * return NULL now. Before we return, check that the value
2013 * of the lock word hasn't changed. If it has, we may have
2014 * seen an inconsistent snapshot.
2015 */
2016 if (op == DTRACE_DYNVAR_NOALLOC) {
2017 if (hash[bucket].dtdh_lock != lock)
2018 goto top;
2019 } else {
2020 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
2021 ASSERT(hash[bucket].dtdh_lock == lock);
2022 ASSERT(lock & 1);
2023 hash[bucket].dtdh_lock++;
2024 }
2025
2026 return (NULL);
2027 }
2028
2029 /*
2030 * We need to allocate a new dynamic variable. The size we need is the
2031 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
2032 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
2033 * the size of any referred-to data (dsize). We then round the final
2034 * size up to the chunksize for allocation.
2035 */
2036 for (ksize = 0, i = 0; i < nkeys; i++)
2037 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
2038
2039 /*
2040 * This should be pretty much impossible, but could happen if, say,
2041 * strange DIF specified the tuple. Ideally, this should be an
2042 * assertion and not an error condition -- but that requires that the
2043 * chunksize calculation in dtrace_difo_chunksize() be absolutely
2044 * bullet-proof. (That is, it must not be able to be fooled by
2045 * malicious DIF.) Given the lack of backwards branches in DIF,
2046 * solving this would presumably not amount to solving the Halting
2047 * Problem -- but it still seems awfully hard.
2048 */
2049 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
2050 ksize + dsize > chunksize) {
2051 dcpu->dtdsc_drops++;
2052 return (NULL);
2053 }
2054
2055 nstate = DTRACE_DSTATE_EMPTY;
2056
2057 do {
2058 retry:
2059 free = dcpu->dtdsc_free;
2060
2061 if (free == NULL) {
2062 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
2063 void *rval;
2064
2065 if (clean == NULL) {
2066 /*
2067 * We're out of dynamic variable space on
2068 * this CPU. Unless we have tried all CPUs,
2069 * we'll try to allocate from a different
2070 * CPU.
2071 */
2072 switch (dstate->dtds_state) {
2073 case DTRACE_DSTATE_CLEAN: {
2074 void *sp = &dstate->dtds_state;
2075
2076 if (++cpu >= NCPU)
2077 cpu = 0;
2078
2079 if (dcpu->dtdsc_dirty != NULL &&
2080 nstate == DTRACE_DSTATE_EMPTY)
2081 nstate = DTRACE_DSTATE_DIRTY;
2082
2083 if (dcpu->dtdsc_rinsing != NULL)
2084 nstate = DTRACE_DSTATE_RINSING;
2085
2086 dcpu = &dstate->dtds_percpu[cpu];
2087
2088 if (cpu != me)
2089 goto retry;
2090
2091 (void) dtrace_cas32(sp,
2092 DTRACE_DSTATE_CLEAN, nstate);
2093
2094 /*
2095 * To increment the correct bean
2096 * counter, take another lap.
2097 */
2098 goto retry;
2099 }
2100
2101 case DTRACE_DSTATE_DIRTY:
2102 dcpu->dtdsc_dirty_drops++;
2103 break;
2104
2105 case DTRACE_DSTATE_RINSING:
2106 dcpu->dtdsc_rinsing_drops++;
2107 break;
2108
2109 case DTRACE_DSTATE_EMPTY:
2110 dcpu->dtdsc_drops++;
2111 break;
2112 }
2113
2114 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
2115 return (NULL);
2116 }
2117
2118 /*
2119 * The clean list appears to be non-empty. We want to
2120 * move the clean list to the free list; we start by
2121 * moving the clean pointer aside.
2122 */
2123 if (dtrace_casptr(&dcpu->dtdsc_clean,
2124 clean, NULL) != clean) {
2125 /*
2126 * We are in one of two situations:
2127 *
2128 * (a) The clean list was switched to the
2129 * free list by another CPU.
2130 *
2131 * (b) The clean list was added to by the
2132 * cleansing cyclic.
2133 *
2134 * In either of these situations, we can
2135 * just reattempt the free list allocation.
2136 */
2137 goto retry;
2138 }
2139
2140 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
2141
2142 /*
2143 * Now we'll move the clean list to our free list.
2144 * It's impossible for this to fail: the only way
2145 * the free list can be updated is through this
2146 * code path, and only one CPU can own the clean list.
2147 * Thus, it would only be possible for this to fail if
2148 * this code were racing with dtrace_dynvar_clean().
2149 * (That is, if dtrace_dynvar_clean() updated the clean
2150 * list, and we ended up racing to update the free
2151 * list.) This race is prevented by the dtrace_sync()
2152 * in dtrace_dynvar_clean() -- which flushes the
2153 * owners of the clean lists out before resetting
2154 * the clean lists.
2155 */
2156 dcpu = &dstate->dtds_percpu[me];
2157 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
2158 ASSERT(rval == NULL);
2159 goto retry;
2160 }
2161
2162 dvar = free;
2163 new_free = dvar->dtdv_next;
2164 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2165
2166 /*
2167 * We have now allocated a new chunk. We copy the tuple keys into the
2168 * tuple array and copy any referenced key data into the data space
2169 * following the tuple array. As we do this, we relocate dttk_value
2170 * in the final tuple to point to the key data address in the chunk.
2171 */
2172 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2173 dvar->dtdv_data = (void *)(kdata + ksize);
2174 dvar->dtdv_tuple.dtt_nkeys = nkeys;
2175
2176 for (i = 0; i < nkeys; i++) {
2177 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2178 size_t kesize = key[i].dttk_size;
2179
2180 if (kesize != 0) {
2181 dtrace_bcopy(
2182 (const void *)(uintptr_t)key[i].dttk_value,
2183 (void *)kdata, kesize);
2184 dkey->dttk_value = kdata;
2185 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2186 } else {
2187 dkey->dttk_value = key[i].dttk_value;
2188 }
2189
2190 dkey->dttk_size = kesize;
2191 }
2192
2193 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2194 dvar->dtdv_hashval = hashval;
2195 dvar->dtdv_next = start;
2196
2197 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2198 return (dvar);
2199
2200 /*
2201 * The cas has failed. Either another CPU is adding an element to
2202 * this hash chain, or another CPU is deleting an element from this
2203 * hash chain. The simplest way to deal with both of these cases
2204 * (though not necessarily the most efficient) is to free our
2205 * allocated block and re-attempt it all. Note that the free is
2206 * to the dirty list and _not_ to the free list. This is to prevent
2207 * races with allocators, above.
2208 */
2209 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2210
2211 dtrace_membar_producer();
2212
2213 do {
2214 free = dcpu->dtdsc_dirty;
2215 dvar->dtdv_next = free;
2216 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2217
2218 goto top;
2219 }
2220
2221 /*ARGSUSED*/
2222 static void
2223 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2224 {
2225 if ((int64_t)nval < (int64_t)*oval)
2226 *oval = nval;
2227 }
2228
2229 /*ARGSUSED*/
2230 static void
2231 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2232 {
2233 if ((int64_t)nval > (int64_t)*oval)
2234 *oval = nval;
2235 }
2236
2237 static void
2238 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2239 {
2240 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2241 int64_t val = (int64_t)nval;
2242
2243 if (val < 0) {
2244 for (i = 0; i < zero; i++) {
2245 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2246 quanta[i] += incr;
2247 return;
2248 }
2249 }
2250 } else {
2251 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2252 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2253 quanta[i - 1] += incr;
2254 return;
2255 }
2256 }
2257
2258 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2259 return;
2260 }
2261
2262 ASSERT(0);
2263 }
2264
2265 static void
2266 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2267 {
2268 uint64_t arg = *lquanta++;
2269 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2270 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2271 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2272 int32_t val = (int32_t)nval, level;
2273
2274 ASSERT(step != 0);
2275 ASSERT(levels != 0);
2276
2277 if (val < base) {
2278 /*
2279 * This is an underflow.
2280 */
2281 lquanta[0] += incr;
2282 return;
2283 }
2284
2285 level = (val - base) / step;
2286
2287 if (level < levels) {
2288 lquanta[level + 1] += incr;
2289 return;
2290 }
2291
2292 /*
2293 * This is an overflow.
2294 */
2295 lquanta[levels + 1] += incr;
2296 }
2297
2298 static int
2299 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2300 uint16_t high, uint16_t nsteps, int64_t value)
2301 {
2302 int64_t this = 1, last, next;
2303 int base = 1, order;
2304
2305 ASSERT(factor <= nsteps);
2306 ASSERT(nsteps % factor == 0);
2307
2308 for (order = 0; order < low; order++)
2309 this *= factor;
2310
2311 /*
2312 * If our value is less than our factor taken to the power of the
2313 * low order of magnitude, it goes into the zeroth bucket.
2314 */
2315 if (value < (last = this))
2316 return (0);
2317
2318 for (this *= factor; order <= high; order++) {
2319 int nbuckets = this > nsteps ? nsteps : this;
2320
2321 if ((next = this * factor) < this) {
2322 /*
2323 * We should not generally get log/linear quantizations
2324 * with a high magnitude that allows 64-bits to
2325 * overflow, but we nonetheless protect against this
2326 * by explicitly checking for overflow, and clamping
2327 * our value accordingly.
2328 */
2329 value = this - 1;
2330 }
2331
2332 if (value < this) {
2333 /*
2334 * If our value lies within this order of magnitude,
2335 * determine its position by taking the offset within
2336 * the order of magnitude, dividing by the bucket
2337 * width, and adding to our (accumulated) base.
2338 */
2339 return (base + (value - last) / (this / nbuckets));
2340 }
2341
2342 base += nbuckets - (nbuckets / factor);
2343 last = this;
2344 this = next;
2345 }
2346
2347 /*
2348 * Our value is greater than or equal to our factor taken to the
2349 * power of one plus the high magnitude -- return the top bucket.
2350 */
2351 return (base);
2352 }
2353
2354 static void
2355 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2356 {
2357 uint64_t arg = *llquanta++;
2358 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2359 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2360 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2361 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2362
2363 llquanta[dtrace_aggregate_llquantize_bucket(factor,
2364 low, high, nsteps, nval)] += incr;
2365 }
2366
2367 /*ARGSUSED*/
2368 static void
2369 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2370 {
2371 data[0]++;
2372 data[1] += nval;
2373 }
2374
2375 /*ARGSUSED*/
2376 static void
2377 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2378 {
2379 int64_t snval = (int64_t)nval;
2380 uint64_t tmp[2];
2381
2382 data[0]++;
2383 data[1] += nval;
2384
2385 /*
2386 * What we want to say here is:
2387 *
2388 * data[2] += nval * nval;
2389 *
2390 * But given that nval is 64-bit, we could easily overflow, so
2391 * we do this as 128-bit arithmetic.
2392 */
2393 if (snval < 0)
2394 snval = -snval;
2395
2396 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2397 dtrace_add_128(data + 2, tmp, data + 2);
2398 }
2399
2400 /*ARGSUSED*/
2401 static void
2402 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2403 {
2404 *oval = *oval + 1;
2405 }
2406
2407 /*ARGSUSED*/
2408 static void
2409 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2410 {
2411 *oval += nval;
2412 }
2413
2414 /*
2415 * Aggregate given the tuple in the principal data buffer, and the aggregating
2416 * action denoted by the specified dtrace_aggregation_t. The aggregation
2417 * buffer is specified as the buf parameter. This routine does not return
2418 * failure; if there is no space in the aggregation buffer, the data will be
2419 * dropped, and a corresponding counter incremented.
2420 */
2421 static void
2422 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2423 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2424 {
2425 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2426 uint32_t i, ndx, size, fsize;
2427 uint32_t align = sizeof (uint64_t) - 1;
2428 dtrace_aggbuffer_t *agb;
2429 dtrace_aggkey_t *key;
2430 uint32_t hashval = 0, limit, isstr;
2431 caddr_t tomax, data, kdata;
2432 dtrace_actkind_t action;
2433 dtrace_action_t *act;
2434 uintptr_t offs;
2435
2436 if (buf == NULL)
2437 return;
2438
2439 if (!agg->dtag_hasarg) {
2440 /*
2441 * Currently, only quantize() and lquantize() take additional
2442 * arguments, and they have the same semantics: an increment
2443 * value that defaults to 1 when not present. If additional
2444 * aggregating actions take arguments, the setting of the
2445 * default argument value will presumably have to become more
2446 * sophisticated...
2447 */
2448 arg = 1;
2449 }
2450
2451 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2452 size = rec->dtrd_offset - agg->dtag_base;
2453 fsize = size + rec->dtrd_size;
2454
2455 ASSERT(dbuf->dtb_tomax != NULL);
2456 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2457
2458 if ((tomax = buf->dtb_tomax) == NULL) {
2459 dtrace_buffer_drop(buf);
2460 return;
2461 }
2462
2463 /*
2464 * The metastructure is always at the bottom of the buffer.
2465 */
2466 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2467 sizeof (dtrace_aggbuffer_t));
2468
2469 if (buf->dtb_offset == 0) {
2470 /*
2471 * We just kludge up approximately 1/8th of the size to be
2472 * buckets. If this guess ends up being routinely
2473 * off-the-mark, we may need to dynamically readjust this
2474 * based on past performance.
2475 */
2476 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2477
2478 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2479 (uintptr_t)tomax || hashsize == 0) {
2480 /*
2481 * We've been given a ludicrously small buffer;
2482 * increment our drop count and leave.
2483 */
2484 dtrace_buffer_drop(buf);
2485 return;
2486 }
2487
2488 /*
2489 * And now, a pathetic attempt to try to get a an odd (or
2490 * perchance, a prime) hash size for better hash distribution.
2491 */
2492 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2493 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2494
2495 agb->dtagb_hashsize = hashsize;
2496 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2497 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2498 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2499
2500 for (i = 0; i < agb->dtagb_hashsize; i++)
2501 agb->dtagb_hash[i] = NULL;
2502 }
2503
2504 ASSERT(agg->dtag_first != NULL);
2505 ASSERT(agg->dtag_first->dta_intuple);
2506
2507 /*
2508 * Calculate the hash value based on the key. Note that we _don't_
2509 * include the aggid in the hashing (but we will store it as part of
2510 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2511 * algorithm: a simple, quick algorithm that has no known funnels, and
2512 * gets good distribution in practice. The efficacy of the hashing
2513 * algorithm (and a comparison with other algorithms) may be found by
2514 * running the ::dtrace_aggstat MDB dcmd.
2515 */
2516 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2517 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2518 limit = i + act->dta_rec.dtrd_size;
2519 ASSERT(limit <= size);
2520 isstr = DTRACEACT_ISSTRING(act);
2521
2522 for (; i < limit; i++) {
2523 hashval += data[i];
2524 hashval += (hashval << 10);
2525 hashval ^= (hashval >> 6);
2526
2527 if (isstr && data[i] == '\0')
2528 break;
2529 }
2530 }
2531
2532 hashval += (hashval << 3);
2533 hashval ^= (hashval >> 11);
2534 hashval += (hashval << 15);
2535
2536 /*
2537 * Yes, the divide here is expensive -- but it's generally the least
2538 * of the performance issues given the amount of data that we iterate
2539 * over to compute hash values, compare data, etc.
2540 */
2541 ndx = hashval % agb->dtagb_hashsize;
2542
2543 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2544 ASSERT((caddr_t)key >= tomax);
2545 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2546
2547 if (hashval != key->dtak_hashval || key->dtak_size != size)
2548 continue;
2549
2550 kdata = key->dtak_data;
2551 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2552
2553 for (act = agg->dtag_first; act->dta_intuple;
2554 act = act->dta_next) {
2555 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2556 limit = i + act->dta_rec.dtrd_size;
2557 ASSERT(limit <= size);
2558 isstr = DTRACEACT_ISSTRING(act);
2559
2560 for (; i < limit; i++) {
2561 if (kdata[i] != data[i])
2562 goto next;
2563
2564 if (isstr && data[i] == '\0')
2565 break;
2566 }
2567 }
2568
2569 if (action != key->dtak_action) {
2570 /*
2571 * We are aggregating on the same value in the same
2572 * aggregation with two different aggregating actions.
2573 * (This should have been picked up in the compiler,
2574 * so we may be dealing with errant or devious DIF.)
2575 * This is an error condition; we indicate as much,
2576 * and return.
2577 */
2578 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2579 return;
2580 }
2581
2582 /*
2583 * This is a hit: we need to apply the aggregator to
2584 * the value at this key.
2585 */
2586 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2587 return;
2588 next:
2589 continue;
2590 }
2591
2592 /*
2593 * We didn't find it. We need to allocate some zero-filled space,
2594 * link it into the hash table appropriately, and apply the aggregator
2595 * to the (zero-filled) value.
2596 */
2597 offs = buf->dtb_offset;
2598 while (offs & (align - 1))
2599 offs += sizeof (uint32_t);
2600
2601 /*
2602 * If we don't have enough room to both allocate a new key _and_
2603 * its associated data, increment the drop count and return.
2604 */
2605 if ((uintptr_t)tomax + offs + fsize >
2606 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2607 dtrace_buffer_drop(buf);
2608 return;
2609 }
2610
2611 /*CONSTCOND*/
2612 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2613 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2614 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2615
2616 key->dtak_data = kdata = tomax + offs;
2617 buf->dtb_offset = offs + fsize;
2618
2619 /*
2620 * Now copy the data across.
2621 */
2622 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2623
2624 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2625 kdata[i] = data[i];
2626
2627 /*
2628 * Because strings are not zeroed out by default, we need to iterate
2629 * looking for actions that store strings, and we need to explicitly
2630 * pad these strings out with zeroes.
2631 */
2632 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2633 int nul;
2634
2635 if (!DTRACEACT_ISSTRING(act))
2636 continue;
2637
2638 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2639 limit = i + act->dta_rec.dtrd_size;
2640 ASSERT(limit <= size);
2641
2642 for (nul = 0; i < limit; i++) {
2643 if (nul) {
2644 kdata[i] = '\0';
2645 continue;
2646 }
2647
2648 if (data[i] != '\0')
2649 continue;
2650
2651 nul = 1;
2652 }
2653 }
2654
2655 for (i = size; i < fsize; i++)
2656 kdata[i] = 0;
2657
2658 key->dtak_hashval = hashval;
2659 key->dtak_size = size;
2660 key->dtak_action = action;
2661 key->dtak_next = agb->dtagb_hash[ndx];
2662 agb->dtagb_hash[ndx] = key;
2663
2664 /*
2665 * Finally, apply the aggregator.
2666 */
2667 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2668 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2669 }
2670
2671 /*
2672 * Given consumer state, this routine finds a speculation in the INACTIVE
2673 * state and transitions it into the ACTIVE state. If there is no speculation
2674 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2675 * incremented -- it is up to the caller to take appropriate action.
2676 */
2677 static int
2678 dtrace_speculation(dtrace_state_t *state)
2679 {
2680 int i = 0;
2681 dtrace_speculation_state_t current;
2682 uint32_t *stat = &state->dts_speculations_unavail, count;
2683
2684 while (i < state->dts_nspeculations) {
2685 dtrace_speculation_t *spec = &state->dts_speculations[i];
2686
2687 current = spec->dtsp_state;
2688
2689 if (current != DTRACESPEC_INACTIVE) {
2690 if (current == DTRACESPEC_COMMITTINGMANY ||
2691 current == DTRACESPEC_COMMITTING ||
2692 current == DTRACESPEC_DISCARDING)
2693 stat = &state->dts_speculations_busy;
2694 i++;
2695 continue;
2696 }
2697
2698 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2699 current, DTRACESPEC_ACTIVE) == current)
2700 return (i + 1);
2701 }
2702
2703 /*
2704 * We couldn't find a speculation. If we found as much as a single
2705 * busy speculation buffer, we'll attribute this failure as "busy"
2706 * instead of "unavail".
2707 */
2708 do {
2709 count = *stat;
2710 } while (dtrace_cas32(stat, count, count + 1) != count);
2711
2712 return (0);
2713 }
2714
2715 /*
2716 * This routine commits an active speculation. If the specified speculation
2717 * is not in a valid state to perform a commit(), this routine will silently do
2718 * nothing. The state of the specified speculation is transitioned according
2719 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2720 */
2721 static void
2722 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2723 dtrace_specid_t which)
2724 {
2725 dtrace_speculation_t *spec;
2726 dtrace_buffer_t *src, *dest;
2727 uintptr_t daddr, saddr, dlimit, slimit;
2728 dtrace_speculation_state_t current, new;
2729 intptr_t offs;
2730 uint64_t timestamp;
2731
2732 if (which == 0)
2733 return;
2734
2735 if (which > state->dts_nspeculations) {
2736 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2737 return;
2738 }
2739
2740 spec = &state->dts_speculations[which - 1];
2741 src = &spec->dtsp_buffer[cpu];
2742 dest = &state->dts_buffer[cpu];
2743
2744 do {
2745 current = spec->dtsp_state;
2746
2747 if (current == DTRACESPEC_COMMITTINGMANY)
2748 break;
2749
2750 switch (current) {
2751 case DTRACESPEC_INACTIVE:
2752 case DTRACESPEC_DISCARDING:
2753 return;
2754
2755 case DTRACESPEC_COMMITTING:
2756 /*
2757 * This is only possible if we are (a) commit()'ing
2758 * without having done a prior speculate() on this CPU
2759 * and (b) racing with another commit() on a different
2760 * CPU. There's nothing to do -- we just assert that
2761 * our offset is 0.
2762 */
2763 ASSERT(src->dtb_offset == 0);
2764 return;
2765
2766 case DTRACESPEC_ACTIVE:
2767 new = DTRACESPEC_COMMITTING;
2768 break;
2769
2770 case DTRACESPEC_ACTIVEONE:
2771 /*
2772 * This speculation is active on one CPU. If our
2773 * buffer offset is non-zero, we know that the one CPU
2774 * must be us. Otherwise, we are committing on a
2775 * different CPU from the speculate(), and we must
2776 * rely on being asynchronously cleaned.
2777 */
2778 if (src->dtb_offset != 0) {
2779 new = DTRACESPEC_COMMITTING;
2780 break;
2781 }
2782 /*FALLTHROUGH*/
2783
2784 case DTRACESPEC_ACTIVEMANY:
2785 new = DTRACESPEC_COMMITTINGMANY;
2786 break;
2787
2788 default:
2789 ASSERT(0);
2790 }
2791 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2792 current, new) != current);
2793
2794 /*
2795 * We have set the state to indicate that we are committing this
2796 * speculation. Now reserve the necessary space in the destination
2797 * buffer.
2798 */
2799 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2800 sizeof (uint64_t), state, NULL)) < 0) {
2801 dtrace_buffer_drop(dest);
2802 goto out;
2803 }
2804
2805 /*
2806 * We have sufficient space to copy the speculative buffer into the
2807 * primary buffer. First, modify the speculative buffer, filling
2808 * in the timestamp of all entries with the current time. The data
2809 * must have the commit() time rather than the time it was traced,
2810 * so that all entries in the primary buffer are in timestamp order.
2811 */
2812 timestamp = dtrace_gethrtime();
2813 saddr = (uintptr_t)src->dtb_tomax;
2814 slimit = saddr + src->dtb_offset;
2815 while (saddr < slimit) {
2816 size_t size;
2817 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2818
2819 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2820 saddr += sizeof (dtrace_epid_t);
2821 continue;
2822 }
2823 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2824 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2825
2826 ASSERT3U(saddr + size, <=, slimit);
2827 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2828 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2829
2830 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2831
2832 saddr += size;
2833 }
2834
2835 /*
2836 * Copy the buffer across. (Note that this is a
2837 * highly subobtimal bcopy(); in the unlikely event that this becomes
2838 * a serious performance issue, a high-performance DTrace-specific
2839 * bcopy() should obviously be invented.)
2840 */
2841 daddr = (uintptr_t)dest->dtb_tomax + offs;
2842 dlimit = daddr + src->dtb_offset;
2843 saddr = (uintptr_t)src->dtb_tomax;
2844
2845 /*
2846 * First, the aligned portion.
2847 */
2848 while (dlimit - daddr >= sizeof (uint64_t)) {
2849 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2850
2851 daddr += sizeof (uint64_t);
2852 saddr += sizeof (uint64_t);
2853 }
2854
2855 /*
2856 * Now any left-over bit...
2857 */
2858 while (dlimit - daddr)
2859 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2860
2861 /*
2862 * Finally, commit the reserved space in the destination buffer.
2863 */
2864 dest->dtb_offset = offs + src->dtb_offset;
2865
2866 out:
2867 /*
2868 * If we're lucky enough to be the only active CPU on this speculation
2869 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2870 */
2871 if (current == DTRACESPEC_ACTIVE ||
2872 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2873 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2874 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2875
2876 ASSERT(rval == DTRACESPEC_COMMITTING);
2877 }
2878
2879 src->dtb_offset = 0;
2880 src->dtb_xamot_drops += src->dtb_drops;
2881 src->dtb_drops = 0;
2882 }
2883
2884 /*
2885 * This routine discards an active speculation. If the specified speculation
2886 * is not in a valid state to perform a discard(), this routine will silently
2887 * do nothing. The state of the specified speculation is transitioned
2888 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2889 */
2890 static void
2891 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2892 dtrace_specid_t which)
2893 {
2894 dtrace_speculation_t *spec;
2895 dtrace_speculation_state_t current, new;
2896 dtrace_buffer_t *buf;
2897
2898 if (which == 0)
2899 return;
2900
2901 if (which > state->dts_nspeculations) {
2902 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2903 return;
2904 }
2905
2906 spec = &state->dts_speculations[which - 1];
2907 buf = &spec->dtsp_buffer[cpu];
2908
2909 do {
2910 current = spec->dtsp_state;
2911
2912 switch (current) {
2913 case DTRACESPEC_INACTIVE:
2914 case DTRACESPEC_COMMITTINGMANY:
2915 case DTRACESPEC_COMMITTING:
2916 case DTRACESPEC_DISCARDING:
2917 return;
2918
2919 case DTRACESPEC_ACTIVE:
2920 case DTRACESPEC_ACTIVEMANY:
2921 new = DTRACESPEC_DISCARDING;
2922 break;
2923
2924 case DTRACESPEC_ACTIVEONE:
2925 if (buf->dtb_offset != 0) {
2926 new = DTRACESPEC_INACTIVE;
2927 } else {
2928 new = DTRACESPEC_DISCARDING;
2929 }
2930 break;
2931
2932 default:
2933 ASSERT(0);
2934 }
2935 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2936 current, new) != current);
2937
2938 buf->dtb_offset = 0;
2939 buf->dtb_drops = 0;
2940 }
2941
2942 /*
2943 * Note: not called from probe context. This function is called
2944 * asynchronously from cross call context to clean any speculations that are
2945 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2946 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2947 * speculation.
2948 */
2949 static void
2950 dtrace_speculation_clean_here(dtrace_state_t *state)
2951 {
2952 dtrace_icookie_t cookie;
2953 processorid_t cpu = CPU->cpu_id;
2954 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2955 dtrace_specid_t i;
2956
2957 cookie = dtrace_interrupt_disable();
2958
2959 if (dest->dtb_tomax == NULL) {
2960 dtrace_interrupt_enable(cookie);
2961 return;
2962 }
2963
2964 for (i = 0; i < state->dts_nspeculations; i++) {
2965 dtrace_speculation_t *spec = &state->dts_speculations[i];
2966 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2967
2968 if (src->dtb_tomax == NULL)
2969 continue;
2970
2971 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2972 src->dtb_offset = 0;
2973 continue;
2974 }
2975
2976 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2977 continue;
2978
2979 if (src->dtb_offset == 0)
2980 continue;
2981
2982 dtrace_speculation_commit(state, cpu, i + 1);
2983 }
2984
2985 dtrace_interrupt_enable(cookie);
2986 }
2987
2988 /*
2989 * Note: not called from probe context. This function is called
2990 * asynchronously (and at a regular interval) to clean any speculations that
2991 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2992 * is work to be done, it cross calls all CPUs to perform that work;
2993 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2994 * INACTIVE state until they have been cleaned by all CPUs.
2995 */
2996 static void
2997 dtrace_speculation_clean(dtrace_state_t *state)
2998 {
2999 int work = 0, rv;
3000 dtrace_specid_t i;
3001
3002 for (i = 0; i < state->dts_nspeculations; i++) {
3003 dtrace_speculation_t *spec = &state->dts_speculations[i];
3004
3005 ASSERT(!spec->dtsp_cleaning);
3006
3007 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
3008 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
3009 continue;
3010
3011 work++;
3012 spec->dtsp_cleaning = 1;
3013 }
3014
3015 if (!work)
3016 return;
3017
3018 dtrace_xcall(DTRACE_CPUALL,
3019 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
3020
3021 /*
3022 * We now know that all CPUs have committed or discarded their
3023 * speculation buffers, as appropriate. We can now set the state
3024 * to inactive.
3025 */
3026 for (i = 0; i < state->dts_nspeculations; i++) {
3027 dtrace_speculation_t *spec = &state->dts_speculations[i];
3028 dtrace_speculation_state_t current, new;
3029
3030 if (!spec->dtsp_cleaning)
3031 continue;
3032
3033 current = spec->dtsp_state;
3034 ASSERT(current == DTRACESPEC_DISCARDING ||
3035 current == DTRACESPEC_COMMITTINGMANY);
3036
3037 new = DTRACESPEC_INACTIVE;
3038
3039 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
3040 ASSERT(rv == current);
3041 spec->dtsp_cleaning = 0;
3042 }
3043 }
3044
3045 /*
3046 * Called as part of a speculate() to get the speculative buffer associated
3047 * with a given speculation. Returns NULL if the specified speculation is not
3048 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
3049 * the active CPU is not the specified CPU -- the speculation will be
3050 * atomically transitioned into the ACTIVEMANY state.
3051 */
3052 static dtrace_buffer_t *
3053 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
3054 dtrace_specid_t which)
3055 {
3056 dtrace_speculation_t *spec;
3057 dtrace_speculation_state_t current, new;
3058 dtrace_buffer_t *buf;
3059
3060 if (which == 0)
3061 return (NULL);
3062
3063 if (which > state->dts_nspeculations) {
3064 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3065 return (NULL);
3066 }
3067
3068 spec = &state->dts_speculations[which - 1];
3069 buf = &spec->dtsp_buffer[cpuid];
3070
3071 do {
3072 current = spec->dtsp_state;
3073
3074 switch (current) {
3075 case DTRACESPEC_INACTIVE:
3076 case DTRACESPEC_COMMITTINGMANY:
3077 case DTRACESPEC_DISCARDING:
3078 return (NULL);
3079
3080 case DTRACESPEC_COMMITTING:
3081 ASSERT(buf->dtb_offset == 0);
3082 return (NULL);
3083
3084 case DTRACESPEC_ACTIVEONE:
3085 /*
3086 * This speculation is currently active on one CPU.
3087 * Check the offset in the buffer; if it's non-zero,
3088 * that CPU must be us (and we leave the state alone).
3089 * If it's zero, assume that we're starting on a new
3090 * CPU -- and change the state to indicate that the
3091 * speculation is active on more than one CPU.
3092 */
3093 if (buf->dtb_offset != 0)
3094 return (buf);
3095
3096 new = DTRACESPEC_ACTIVEMANY;
3097 break;
3098
3099 case DTRACESPEC_ACTIVEMANY:
3100 return (buf);
3101
3102 case DTRACESPEC_ACTIVE:
3103 new = DTRACESPEC_ACTIVEONE;
3104 break;
3105
3106 default:
3107 ASSERT(0);
3108 }
3109 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3110 current, new) != current);
3111
3112 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
3113 return (buf);
3114 }
3115
3116 /*
3117 * Return a string. In the event that the user lacks the privilege to access
3118 * arbitrary kernel memory, we copy the string out to scratch memory so that we
3119 * don't fail access checking.
3120 *
3121 * dtrace_dif_variable() uses this routine as a helper for various
3122 * builtin values such as 'execname' and 'probefunc.'
3123 */
3124 uintptr_t
3125 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
3126 dtrace_mstate_t *mstate)
3127 {
3128 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3129 uintptr_t ret;
3130 size_t strsz;
3131
3132 /*
3133 * The easy case: this probe is allowed to read all of memory, so
3134 * we can just return this as a vanilla pointer.
3135 */
3136 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
3137 return (addr);
3138
3139 /*
3140 * This is the tougher case: we copy the string in question from
3141 * kernel memory into scratch memory and return it that way: this
3142 * ensures that we won't trip up when access checking tests the
3143 * BYREF return value.
3144 */
3145 strsz = dtrace_strlen((char *)addr, size) + 1;
3146
3147 if (mstate->dtms_scratch_ptr + strsz >
3148 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3149 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3150 return (0);
3151 }
3152
3153 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3154 strsz);
3155 ret = mstate->dtms_scratch_ptr;
3156 mstate->dtms_scratch_ptr += strsz;
3157 return (ret);
3158 }
3159
3160 /*
3161 * This function implements the DIF emulator's variable lookups. The emulator
3162 * passes a reserved variable identifier and optional built-in array index.
3163 */
3164 static uint64_t
3165 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3166 uint64_t ndx)
3167 {
3168 /*
3169 * If we're accessing one of the uncached arguments, we'll turn this
3170 * into a reference in the args array.
3171 */
3172 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3173 ndx = v - DIF_VAR_ARG0;
3174 v = DIF_VAR_ARGS;
3175 }
3176
3177 switch (v) {
3178 case DIF_VAR_ARGS:
3179 if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
3180 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
3181 CPU_DTRACE_KPRIV;
3182 return (0);
3183 }
3184
3185 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3186 if (ndx >= sizeof (mstate->dtms_arg) /
3187 sizeof (mstate->dtms_arg[0])) {
3188 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3189 dtrace_provider_t *pv;
3190 uint64_t val;
3191
3192 pv = mstate->dtms_probe->dtpr_provider;
3193 if (pv->dtpv_pops.dtps_getargval != NULL)
3194 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3195 mstate->dtms_probe->dtpr_id,
3196 mstate->dtms_probe->dtpr_arg, ndx, aframes);
3197 else
3198 val = dtrace_getarg(ndx, aframes);
3199
3200 /*
3201 * This is regrettably required to keep the compiler
3202 * from tail-optimizing the call to dtrace_getarg().
3203 * The condition always evaluates to true, but the
3204 * compiler has no way of figuring that out a priori.
3205 * (None of this would be necessary if the compiler
3206 * could be relied upon to _always_ tail-optimize
3207 * the call to dtrace_getarg() -- but it can't.)
3208 */
3209 if (mstate->dtms_probe != NULL)
3210 return (val);
3211
3212 ASSERT(0);
3213 }
3214
3215 return (mstate->dtms_arg[ndx]);
3216
3217 case DIF_VAR_UREGS: {
3218 klwp_t *lwp;
3219
3220 if (!dtrace_priv_proc(state, mstate))
3221 return (0);
3222
3223 if ((lwp = curthread->t_lwp) == NULL) {
3224 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3225 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = 0;
3226 return (0);
3227 }
3228
3229 return (dtrace_getreg(lwp->lwp_regs, ndx));
3230 }
3231
3232 case DIF_VAR_VMREGS: {
3233 uint64_t rval;
3234
3235 if (!dtrace_priv_kernel(state))
3236 return (0);
3237
3238 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3239
3240 rval = dtrace_getvmreg(ndx,
3241 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
3242
3243 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3244
3245 return (rval);
3246 }
3247
3248 case DIF_VAR_CURTHREAD:
3249 if (!dtrace_priv_proc(state, mstate))
3250 return (0);
3251 return ((uint64_t)(uintptr_t)curthread);
3252
3253 case DIF_VAR_TIMESTAMP:
3254 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3255 mstate->dtms_timestamp = dtrace_gethrtime();
3256 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3257 }
3258 return (mstate->dtms_timestamp);
3259
3260 case DIF_VAR_VTIMESTAMP:
3261 ASSERT(dtrace_vtime_references != 0);
3262 return (curthread->t_dtrace_vtime);
3263
3264 case DIF_VAR_WALLTIMESTAMP:
3265 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3266 mstate->dtms_walltimestamp = dtrace_gethrestime();
3267 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3268 }
3269 return (mstate->dtms_walltimestamp);
3270
3271 case DIF_VAR_IPL:
3272 if (!dtrace_priv_kernel(state))
3273 return (0);
3274 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3275 mstate->dtms_ipl = dtrace_getipl();
3276 mstate->dtms_present |= DTRACE_MSTATE_IPL;
3277 }
3278 return (mstate->dtms_ipl);
3279
3280 case DIF_VAR_EPID:
3281 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3282 return (mstate->dtms_epid);
3283
3284 case DIF_VAR_ID:
3285 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3286 return (mstate->dtms_probe->dtpr_id);
3287
3288 case DIF_VAR_STACKDEPTH:
3289 if (!dtrace_priv_kernel(state))
3290 return (0);
3291 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3292 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3293
3294 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3295 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3296 }
3297 return (mstate->dtms_stackdepth);
3298
3299 case DIF_VAR_USTACKDEPTH:
3300 if (!dtrace_priv_proc(state, mstate))
3301 return (0);
3302 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3303 /*
3304 * See comment in DIF_VAR_PID.
3305 */
3306 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3307 CPU_ON_INTR(CPU)) {
3308 mstate->dtms_ustackdepth = 0;
3309 } else {
3310 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3311 mstate->dtms_ustackdepth =
3312 dtrace_getustackdepth();
3313 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3314 }
3315 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3316 }
3317 return (mstate->dtms_ustackdepth);
3318
3319 case DIF_VAR_CALLER:
3320 if (!dtrace_priv_kernel(state))
3321 return (0);
3322 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3323 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3324
3325 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3326 /*
3327 * If this is an unanchored probe, we are
3328 * required to go through the slow path:
3329 * dtrace_caller() only guarantees correct
3330 * results for anchored probes.
3331 */
3332 pc_t caller[2];
3333
3334 dtrace_getpcstack(caller, 2, aframes,
3335 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3336 mstate->dtms_caller = caller[1];
3337 } else if ((mstate->dtms_caller =
3338 dtrace_caller(aframes)) == -1) {
3339 /*
3340 * We have failed to do this the quick way;
3341 * we must resort to the slower approach of
3342 * calling dtrace_getpcstack().
3343 */
3344 pc_t caller;
3345
3346 dtrace_getpcstack(&caller, 1, aframes, NULL);
3347 mstate->dtms_caller = caller;
3348 }
3349
3350 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3351 }
3352 return (mstate->dtms_caller);
3353
3354 case DIF_VAR_UCALLER:
3355 if (!dtrace_priv_proc(state, mstate))
3356 return (0);
3357
3358 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3359 uint64_t ustack[3];
3360
3361 /*
3362 * dtrace_getupcstack() fills in the first uint64_t
3363 * with the current PID. The second uint64_t will
3364 * be the program counter at user-level. The third
3365 * uint64_t will contain the caller, which is what
3366 * we're after.
3367 */
3368 ustack[2] = 0;
3369 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3370 dtrace_getupcstack(ustack, 3);
3371 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3372 mstate->dtms_ucaller = ustack[2];
3373 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3374 }
3375
3376 return (mstate->dtms_ucaller);
3377
3378 case DIF_VAR_PROBEPROV:
3379 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3380 return (dtrace_dif_varstr(
3381 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3382 state, mstate));
3383
3384 case DIF_VAR_PROBEMOD:
3385 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3386 return (dtrace_dif_varstr(
3387 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3388 state, mstate));
3389
3390 case DIF_VAR_PROBEFUNC:
3391 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3392 return (dtrace_dif_varstr(
3393 (uintptr_t)mstate->dtms_probe->dtpr_func,
3394 state, mstate));
3395
3396 case DIF_VAR_PROBENAME:
3397 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3398 return (dtrace_dif_varstr(
3399 (uintptr_t)mstate->dtms_probe->dtpr_name,
3400 state, mstate));
3401
3402 case DIF_VAR_PID:
3403 if (!dtrace_priv_proc(state, mstate))
3404 return (0);
3405
3406 /*
3407 * Note that we are assuming that an unanchored probe is
3408 * always due to a high-level interrupt. (And we're assuming
3409 * that there is only a single high level interrupt.)
3410 */
3411 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3412 return (pid0.pid_id);
3413
3414 /*
3415 * It is always safe to dereference one's own t_procp pointer:
3416 * it always points to a valid, allocated proc structure.
3417 * Further, it is always safe to dereference the p_pidp member
3418 * of one's own proc structure. (These are truisms becuase
3419 * threads and processes don't clean up their own state --
3420 * they leave that task to whomever reaps them.)
3421 */
3422 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3423
3424 case DIF_VAR_PPID:
3425 if (!dtrace_priv_proc(state, mstate))
3426 return (0);
3427
3428 /*
3429 * See comment in DIF_VAR_PID.
3430 */
3431 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3432 return (pid0.pid_id);
3433
3434 /*
3435 * It is always safe to dereference one's own t_procp pointer:
3436 * it always points to a valid, allocated proc structure.
3437 * (This is true because threads don't clean up their own
3438 * state -- they leave that task to whomever reaps them.)
3439 */
3440 return ((uint64_t)curthread->t_procp->p_ppid);
3441
3442 case DIF_VAR_TID:
3443 /*
3444 * See comment in DIF_VAR_PID.
3445 */
3446 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3447 return (0);
3448
3449 return ((uint64_t)curthread->t_tid);
3450
3451 case DIF_VAR_EXECNAME:
3452 if (!dtrace_priv_proc(state, mstate))
3453 return (0);
3454
3455 /*
3456 * See comment in DIF_VAR_PID.
3457 */
3458 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3459 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3460
3461 /*
3462 * It is always safe to dereference one's own t_procp pointer:
3463 * it always points to a valid, allocated proc structure.
3464 * (This is true because threads don't clean up their own
3465 * state -- they leave that task to whomever reaps them.)
3466 */
3467 return (dtrace_dif_varstr(
3468 (uintptr_t)curthread->t_procp->p_user.u_comm,
3469 state, mstate));
3470
3471 case DIF_VAR_ZONENAME:
3472 if (!dtrace_priv_proc(state, mstate))
3473 return (0);
3474
3475 /*
3476 * See comment in DIF_VAR_PID.
3477 */
3478 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3479 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3480
3481 /*
3482 * It is always safe to dereference one's own t_procp pointer:
3483 * it always points to a valid, allocated proc structure.
3484 * (This is true because threads don't clean up their own
3485 * state -- they leave that task to whomever reaps them.)
3486 */
3487 return (dtrace_dif_varstr(
3488 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3489 state, mstate));
3490
3491 case DIF_VAR_UID:
3492 if (!dtrace_priv_proc(state, mstate))
3493 return (0);
3494
3495 /*
3496 * See comment in DIF_VAR_PID.
3497 */
3498 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3499 return ((uint64_t)p0.p_cred->cr_uid);
3500
3501 /*
3502 * It is always safe to dereference one's own t_procp pointer:
3503 * it always points to a valid, allocated proc structure.
3504 * (This is true because threads don't clean up their own
3505 * state -- they leave that task to whomever reaps them.)
3506 *
3507 * Additionally, it is safe to dereference one's own process
3508 * credential, since this is never NULL after process birth.
3509 */
3510 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3511
3512 case DIF_VAR_GID:
3513 if (!dtrace_priv_proc(state, mstate))
3514 return (0);
3515
3516 /*
3517 * See comment in DIF_VAR_PID.
3518 */
3519 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3520 return ((uint64_t)p0.p_cred->cr_gid);
3521
3522 /*
3523 * It is always safe to dereference one's own t_procp pointer:
3524 * it always points to a valid, allocated proc structure.
3525 * (This is true because threads don't clean up their own
3526 * state -- they leave that task to whomever reaps them.)
3527 *
3528 * Additionally, it is safe to dereference one's own process
3529 * credential, since this is never NULL after process birth.
3530 */
3531 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3532
3533 case DIF_VAR_ERRNO: {
3534 klwp_t *lwp;
3535 if (!dtrace_priv_proc(state, mstate))
3536 return (0);
3537
3538 /*
3539 * See comment in DIF_VAR_PID.
3540 */
3541 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3542 return (0);
3543
3544 /*
3545 * It is always safe to dereference one's own t_lwp pointer in
3546 * the event that this pointer is non-NULL. (This is true
3547 * because threads and lwps don't clean up their own state --
3548 * they leave that task to whomever reaps them.)
3549 */
3550 if ((lwp = curthread->t_lwp) == NULL)
3551 return (0);
3552
3553 return ((uint64_t)lwp->lwp_errno);
3554 }
3555
3556 case DIF_VAR_THREADNAME:
3557 /*
3558 * See comment in DIF_VAR_PID.
3559 */
3560 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3561 return (0);
3562
3563 if (curthread->t_name == NULL)
3564 return (0);
3565
3566 /*
3567 * Once set, ->t_name itself is never changed: any updates are
3568 * made to the same buffer that we are pointing out. So we are
3569 * safe to dereference it here.
3570 */
3571 return (dtrace_dif_varstr((uintptr_t)curthread->t_name,
3572 state, mstate));
3573
3574 default:
3575 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3576 return (0);
3577 }
3578 }
3579
3580 static void
3581 dtrace_dif_variable_write(dtrace_mstate_t *mstate, dtrace_state_t *state,
3582 uint64_t v, uint64_t ndx, uint64_t data)
3583 {
3584 switch (v) {
3585 case DIF_VAR_UREGS: {
3586 klwp_t *lwp;
3587
3588 if (dtrace_destructive_disallow ||
3589 !dtrace_priv_proc_control(state, mstate)) {
3590 return;
3591 }
3592
3593 if ((lwp = curthread->t_lwp) == NULL) {
3594 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3595 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = 0;
3596 return;
3597 }
3598
3599 dtrace_setreg(lwp->lwp_regs, ndx, data);
3600 return;
3601 }
3602
3603 default:
3604 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3605 return;
3606 }
3607 }
3608
3609 typedef enum dtrace_json_state {
3610 DTRACE_JSON_REST = 1,
3611 DTRACE_JSON_OBJECT,
3612 DTRACE_JSON_STRING,
3613 DTRACE_JSON_STRING_ESCAPE,
3614 DTRACE_JSON_STRING_ESCAPE_UNICODE,
3615 DTRACE_JSON_COLON,
3616 DTRACE_JSON_COMMA,
3617 DTRACE_JSON_VALUE,
3618 DTRACE_JSON_IDENTIFIER,
3619 DTRACE_JSON_NUMBER,
3620 DTRACE_JSON_NUMBER_FRAC,
3621 DTRACE_JSON_NUMBER_EXP,
3622 DTRACE_JSON_COLLECT_OBJECT
3623 } dtrace_json_state_t;
3624
3625 /*
3626 * This function possesses just enough knowledge about JSON to extract a single
3627 * value from a JSON string and store it in the scratch buffer. It is able
3628 * to extract nested object values, and members of arrays by index.
3629 *
3630 * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3631 * be looked up as we descend into the object tree. e.g.
3632 *
3633 * foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3634 * with nelems = 5.
3635 *
3636 * The run time of this function must be bounded above by strsize to limit the
3637 * amount of work done in probe context. As such, it is implemented as a
3638 * simple state machine, reading one character at a time using safe loads
3639 * until we find the requested element, hit a parsing error or run off the
3640 * end of the object or string.
3641 *
3642 * As there is no way for a subroutine to return an error without interrupting
3643 * clause execution, we simply return NULL in the event of a missing key or any
3644 * other error condition. Each NULL return in this function is commented with
3645 * the error condition it represents -- parsing or otherwise.
3646 *
3647 * The set of states for the state machine closely matches the JSON
3648 * specification (http://json.org/). Briefly:
3649 *
3650 * DTRACE_JSON_REST:
3651 * Skip whitespace until we find either a top-level Object, moving
3652 * to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3653 *
3654 * DTRACE_JSON_OBJECT:
3655 * Locate the next key String in an Object. Sets a flag to denote
3656 * the next String as a key string and moves to DTRACE_JSON_STRING.
3657 *
3658 * DTRACE_JSON_COLON:
3659 * Skip whitespace until we find the colon that separates key Strings
3660 * from their values. Once found, move to DTRACE_JSON_VALUE.
3661 *
3662 * DTRACE_JSON_VALUE:
3663 * Detects the type of the next value (String, Number, Identifier, Object
3664 * or Array) and routes to the states that process that type. Here we also
3665 * deal with the element selector list if we are requested to traverse down
3666 * into the object tree.
3667 *
3668 * DTRACE_JSON_COMMA:
3669 * Skip whitespace until we find the comma that separates key-value pairs
3670 * in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3671 * (similarly DTRACE_JSON_VALUE). All following literal value processing
3672 * states return to this state at the end of their value, unless otherwise
3673 * noted.
3674 *
3675 * DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3676 * Processes a Number literal from the JSON, including any exponent
3677 * component that may be present. Numbers are returned as strings, which
3678 * may be passed to strtoll() if an integer is required.
3679 *
3680 * DTRACE_JSON_IDENTIFIER:
3681 * Processes a "true", "false" or "null" literal in the JSON.
3682 *
3683 * DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3684 * DTRACE_JSON_STRING_ESCAPE_UNICODE:
3685 * Processes a String literal from the JSON, whether the String denotes
3686 * a key, a value or part of a larger Object. Handles all escape sequences
3687 * present in the specification, including four-digit unicode characters,
3688 * but merely includes the escape sequence without converting it to the
3689 * actual escaped character. If the String is flagged as a key, we
3690 * move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3691 *
3692 * DTRACE_JSON_COLLECT_OBJECT:
3693 * This state collects an entire Object (or Array), correctly handling
3694 * embedded strings. If the full element selector list matches this nested
3695 * object, we return the Object in full as a string. If not, we use this
3696 * state to skip to the next value at this level and continue processing.
3697 *
3698 * NOTE: This function uses various macros from strtolctype.h to manipulate
3699 * digit values, etc -- these have all been checked to ensure they make
3700 * no additional function calls.
3701 */
3702 static char *
3703 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3704 char *dest)
3705 {
3706 dtrace_json_state_t state = DTRACE_JSON_REST;
3707 int64_t array_elem = INT64_MIN;
3708 int64_t array_pos = 0;
3709 uint8_t escape_unicount = 0;
3710 boolean_t string_is_key = B_FALSE;
3711 boolean_t collect_object = B_FALSE;
3712 boolean_t found_key = B_FALSE;
3713 boolean_t in_array = B_FALSE;
3714 uint32_t braces = 0, brackets = 0;
3715 char *elem = elemlist;
3716 char *dd = dest;
3717 uintptr_t cur;
3718
3719 for (cur = json; cur < json + size; cur++) {
3720 char cc = dtrace_load8(cur);
3721 if (cc == '\0')
3722 return (NULL);
3723
3724 switch (state) {
3725 case DTRACE_JSON_REST:
3726 if (isspace(cc))
3727 break;
3728
3729 if (cc == '{') {
3730 state = DTRACE_JSON_OBJECT;
3731 break;
3732 }
3733
3734 if (cc == '[') {
3735 in_array = B_TRUE;
3736 array_pos = 0;
3737 array_elem = dtrace_strtoll(elem, 10, size);
3738 found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3739 state = DTRACE_JSON_VALUE;
3740 break;
3741 }
3742
3743 /*
3744 * ERROR: expected to find a top-level object or array.
3745 */
3746 return (NULL);
3747 case DTRACE_JSON_OBJECT:
3748 if (isspace(cc))
3749 break;
3750
3751 if (cc == '"') {
3752 state = DTRACE_JSON_STRING;
3753 string_is_key = B_TRUE;
3754 break;
3755 }
3756
3757 /*
3758 * ERROR: either the object did not start with a key
3759 * string, or we've run off the end of the object
3760 * without finding the requested key.
3761 */
3762 return (NULL);
3763 case DTRACE_JSON_STRING:
3764 if (cc == '\\') {
3765 *dd++ = '\\';
3766 state = DTRACE_JSON_STRING_ESCAPE;
3767 break;
3768 }
3769
3770 if (cc == '"') {
3771 if (collect_object) {
3772 /*
3773 * We don't reset the dest here, as
3774 * the string is part of a larger
3775 * object being collected.
3776 */
3777 *dd++ = cc;
3778 collect_object = B_FALSE;
3779 state = DTRACE_JSON_COLLECT_OBJECT;
3780 break;
3781 }
3782 *dd = '\0';
3783 dd = dest; /* reset string buffer */
3784 if (string_is_key) {
3785 if (dtrace_strncmp(dest, elem,
3786 size) == 0)
3787 found_key = B_TRUE;
3788 } else if (found_key) {
3789 if (nelems > 1) {
3790 /*
3791 * We expected an object, not
3792 * this string.
3793 */
3794 return (NULL);
3795 }
3796 return (dest);
3797 }
3798 state = string_is_key ? DTRACE_JSON_COLON :
3799 DTRACE_JSON_COMMA;
3800 string_is_key = B_FALSE;
3801 break;
3802 }
3803
3804 *dd++ = cc;
3805 break;
3806 case DTRACE_JSON_STRING_ESCAPE:
3807 *dd++ = cc;
3808 if (cc == 'u') {
3809 escape_unicount = 0;
3810 state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3811 } else {
3812 state = DTRACE_JSON_STRING;
3813 }
3814 break;
3815 case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3816 if (!isxdigit(cc)) {
3817 /*
3818 * ERROR: invalid unicode escape, expected
3819 * four valid hexidecimal digits.
3820 */
3821 return (NULL);
3822 }
3823
3824 *dd++ = cc;
3825 if (++escape_unicount == 4)
3826 state = DTRACE_JSON_STRING;
3827 break;
3828 case DTRACE_JSON_COLON:
3829 if (isspace(cc))
3830 break;
3831
3832 if (cc == ':') {
3833 state = DTRACE_JSON_VALUE;
3834 break;
3835 }
3836
3837 /*
3838 * ERROR: expected a colon.
3839 */
3840 return (NULL);
3841 case DTRACE_JSON_COMMA:
3842 if (isspace(cc))
3843 break;
3844
3845 if (cc == ',') {
3846 if (in_array) {
3847 state = DTRACE_JSON_VALUE;
3848 if (++array_pos == array_elem)
3849 found_key = B_TRUE;
3850 } else {
3851 state = DTRACE_JSON_OBJECT;
3852 }
3853 break;
3854 }
3855
3856 /*
3857 * ERROR: either we hit an unexpected character, or
3858 * we reached the end of the object or array without
3859 * finding the requested key.
3860 */
3861 return (NULL);
3862 case DTRACE_JSON_IDENTIFIER:
3863 if (islower(cc)) {
3864 *dd++ = cc;
3865 break;
3866 }
3867
3868 *dd = '\0';
3869 dd = dest; /* reset string buffer */
3870
3871 if (dtrace_strncmp(dest, "true", 5) == 0 ||
3872 dtrace_strncmp(dest, "false", 6) == 0 ||
3873 dtrace_strncmp(dest, "null", 5) == 0) {
3874 if (found_key) {
3875 if (nelems > 1) {
3876 /*
3877 * ERROR: We expected an object,
3878 * not this identifier.
3879 */
3880 return (NULL);
3881 }
3882 return (dest);
3883 } else {
3884 cur--;
3885 state = DTRACE_JSON_COMMA;
3886 break;
3887 }
3888 }
3889
3890 /*
3891 * ERROR: we did not recognise the identifier as one
3892 * of those in the JSON specification.
3893 */
3894 return (NULL);
3895 case DTRACE_JSON_NUMBER:
3896 if (cc == '.') {
3897 *dd++ = cc;
3898 state = DTRACE_JSON_NUMBER_FRAC;
3899 break;
3900 }
3901
3902 if (cc == 'x' || cc == 'X') {
3903 /*
3904 * ERROR: specification explicitly excludes
3905 * hexidecimal or octal numbers.
3906 */
3907 return (NULL);
3908 }
3909
3910 /* FALLTHRU */
3911 case DTRACE_JSON_NUMBER_FRAC:
3912 if (cc == 'e' || cc == 'E') {
3913 *dd++ = cc;
3914 state = DTRACE_JSON_NUMBER_EXP;
3915 break;
3916 }
3917
3918 if (cc == '+' || cc == '-') {
3919 /*
3920 * ERROR: expect sign as part of exponent only.
3921 */
3922 return (NULL);
3923 }
3924 /* FALLTHRU */
3925 case DTRACE_JSON_NUMBER_EXP:
3926 if (isdigit(cc) || cc == '+' || cc == '-') {
3927 *dd++ = cc;
3928 break;
3929 }
3930
3931 *dd = '\0';
3932 dd = dest; /* reset string buffer */
3933 if (found_key) {
3934 if (nelems > 1) {
3935 /*
3936 * ERROR: We expected an object, not
3937 * this number.
3938 */
3939 return (NULL);
3940 }
3941 return (dest);
3942 }
3943
3944 cur--;
3945 state = DTRACE_JSON_COMMA;
3946 break;
3947 case DTRACE_JSON_VALUE:
3948 if (isspace(cc))
3949 break;
3950
3951 if (cc == '{' || cc == '[') {
3952 if (nelems > 1 && found_key) {
3953 in_array = cc == '[' ? B_TRUE : B_FALSE;
3954 /*
3955 * If our element selector directs us
3956 * to descend into this nested object,
3957 * then move to the next selector
3958 * element in the list and restart the
3959 * state machine.
3960 */
3961 while (*elem != '\0')
3962 elem++;
3963 elem++; /* skip the inter-element NUL */
3964 nelems--;
3965 dd = dest;
3966 if (in_array) {
3967 state = DTRACE_JSON_VALUE;
3968 array_pos = 0;
3969 array_elem = dtrace_strtoll(
3970 elem, 10, size);
3971 found_key = array_elem == 0 ?
3972 B_TRUE : B_FALSE;
3973 } else {
3974 found_key = B_FALSE;
3975 state = DTRACE_JSON_OBJECT;
3976 }
3977 break;
3978 }
3979
3980 /*
3981 * Otherwise, we wish to either skip this
3982 * nested object or return it in full.
3983 */
3984 if (cc == '[')
3985 brackets = 1;
3986 else
3987 braces = 1;
3988 *dd++ = cc;
3989 state = DTRACE_JSON_COLLECT_OBJECT;
3990 break;
3991 }
3992
3993 if (cc == '"') {
3994 state = DTRACE_JSON_STRING;
3995 break;
3996 }
3997
3998 if (islower(cc)) {
3999 /*
4000 * Here we deal with true, false and null.
4001 */
4002 *dd++ = cc;
4003 state = DTRACE_JSON_IDENTIFIER;
4004 break;
4005 }
4006
4007 if (cc == '-' || isdigit(cc)) {
4008 *dd++ = cc;
4009 state = DTRACE_JSON_NUMBER;
4010 break;
4011 }
4012
4013 /*
4014 * ERROR: unexpected character at start of value.
4015 */
4016 return (NULL);
4017 case DTRACE_JSON_COLLECT_OBJECT:
4018 if (cc == '\0')
4019 /*
4020 * ERROR: unexpected end of input.
4021 */
4022 return (NULL);
4023
4024 *dd++ = cc;
4025 if (cc == '"') {
4026 collect_object = B_TRUE;
4027 state = DTRACE_JSON_STRING;
4028 break;
4029 }
4030
4031 if (cc == ']') {
4032 if (brackets-- == 0) {
4033 /*
4034 * ERROR: unbalanced brackets.
4035 */
4036 return (NULL);
4037 }
4038 } else if (cc == '}') {
4039 if (braces-- == 0) {
4040 /*
4041 * ERROR: unbalanced braces.
4042 */
4043 return (NULL);
4044 }
4045 } else if (cc == '{') {
4046 braces++;
4047 } else if (cc == '[') {
4048 brackets++;
4049 }
4050
4051 if (brackets == 0 && braces == 0) {
4052 if (found_key) {
4053 *dd = '\0';
4054 return (dest);
4055 }
4056 dd = dest; /* reset string buffer */
4057 state = DTRACE_JSON_COMMA;
4058 }
4059 break;
4060 }
4061 }
4062 return (NULL);
4063 }
4064
4065 /*
4066 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
4067 * Notice that we don't bother validating the proper number of arguments or
4068 * their types in the tuple stack. This isn't needed because all argument
4069 * interpretation is safe because of our load safety -- the worst that can
4070 * happen is that a bogus program can obtain bogus results.
4071 */
4072 static void
4073 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
4074 dtrace_key_t *tupregs, int nargs,
4075 dtrace_mstate_t *mstate, dtrace_state_t *state)
4076 {
4077 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4078 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
4079 dtrace_vstate_t *vstate = &state->dts_vstate;
4080
4081 union {
4082 mutex_impl_t mi;
4083 uint64_t mx;
4084 } m;
4085
4086 union {
4087 krwlock_t ri;
4088 uintptr_t rw;
4089 } r;
4090
4091 switch (subr) {
4092 case DIF_SUBR_RAND:
4093 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
4094 break;
4095
4096 case DIF_SUBR_MUTEX_OWNED:
4097 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4098 mstate, vstate)) {
4099 regs[rd] = 0;
4100 break;
4101 }
4102
4103 m.mx = dtrace_load64(tupregs[0].dttk_value);
4104 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
4105 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
4106 else
4107 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
4108 break;
4109
4110 case DIF_SUBR_MUTEX_OWNER:
4111 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4112 mstate, vstate)) {
4113 regs[rd] = 0;
4114 break;
4115 }
4116
4117 m.mx = dtrace_load64(tupregs[0].dttk_value);
4118 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
4119 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
4120 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
4121 else
4122 regs[rd] = 0;
4123 break;
4124
4125 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4126 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4127 mstate, vstate)) {
4128 regs[rd] = 0;
4129 break;
4130 }
4131
4132 m.mx = dtrace_load64(tupregs[0].dttk_value);
4133 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
4134 break;
4135
4136 case DIF_SUBR_MUTEX_TYPE_SPIN:
4137 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4138 mstate, vstate)) {
4139 regs[rd] = 0;
4140 break;
4141 }
4142
4143 m.mx = dtrace_load64(tupregs[0].dttk_value);
4144 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
4145 break;
4146
4147 case DIF_SUBR_RW_READ_HELD: {
4148 uintptr_t tmp;
4149
4150 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4151 mstate, vstate)) {
4152 regs[rd] = 0;
4153 break;
4154 }
4155
4156 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4157 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
4158 break;
4159 }
4160
4161 case DIF_SUBR_RW_WRITE_HELD:
4162 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4163 mstate, vstate)) {
4164 regs[rd] = 0;
4165 break;
4166 }
4167
4168 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4169 regs[rd] = _RW_WRITE_HELD(&r.ri);
4170 break;
4171
4172 case DIF_SUBR_RW_ISWRITER:
4173 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4174 mstate, vstate)) {
4175 regs[rd] = 0;
4176 break;
4177 }
4178
4179 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4180 regs[rd] = _RW_ISWRITER(&r.ri);
4181 break;
4182
4183 case DIF_SUBR_BCOPY: {
4184 /*
4185 * We need to be sure that the destination is in the scratch
4186 * region -- no other region is allowed.
4187 */
4188 uintptr_t src = tupregs[0].dttk_value;
4189 uintptr_t dest = tupregs[1].dttk_value;
4190 size_t size = tupregs[2].dttk_value;
4191
4192 if (!dtrace_inscratch(dest, size, mstate)) {
4193 *flags |= CPU_DTRACE_BADADDR;
4194 *illval = regs[rd];
4195 break;
4196 }
4197
4198 if (!dtrace_canload(src, size, mstate, vstate)) {
4199 regs[rd] = 0;
4200 break;
4201 }
4202
4203 dtrace_bcopy((void *)src, (void *)dest, size);
4204 break;
4205 }
4206
4207 case DIF_SUBR_ALLOCA:
4208 case DIF_SUBR_COPYIN: {
4209 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4210 uint64_t size =
4211 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
4212 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
4213
4214 /*
4215 * This action doesn't require any credential checks since
4216 * probes will not activate in user contexts to which the
4217 * enabling user does not have permissions.
4218 */
4219
4220 /*
4221 * Rounding up the user allocation size could have overflowed
4222 * a large, bogus allocation (like -1ULL) to 0.
4223 */
4224 if (scratch_size < size ||
4225 !DTRACE_INSCRATCH(mstate, scratch_size)) {
4226 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4227 regs[rd] = 0;
4228 break;
4229 }
4230
4231 if (subr == DIF_SUBR_COPYIN) {
4232 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4233 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4234 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4235 }
4236
4237 mstate->dtms_scratch_ptr += scratch_size;
4238 regs[rd] = dest;
4239 break;
4240 }
4241
4242 case DIF_SUBR_COPYINTO: {
4243 uint64_t size = tupregs[1].dttk_value;
4244 uintptr_t dest = tupregs[2].dttk_value;
4245
4246 /*
4247 * This action doesn't require any credential checks since
4248 * probes will not activate in user contexts to which the
4249 * enabling user does not have permissions.
4250 */
4251 if (!dtrace_inscratch(dest, size, mstate)) {
4252 *flags |= CPU_DTRACE_BADADDR;
4253 *illval = regs[rd];
4254 break;
4255 }
4256
4257 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4258 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4259 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4260 break;
4261 }
4262
4263 case DIF_SUBR_COPYINSTR: {
4264 uintptr_t dest = mstate->dtms_scratch_ptr;
4265 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4266
4267 if (nargs > 1 && tupregs[1].dttk_value < size)
4268 size = tupregs[1].dttk_value + 1;
4269
4270 /*
4271 * This action doesn't require any credential checks since
4272 * probes will not activate in user contexts to which the
4273 * enabling user does not have permissions.
4274 */
4275 if (!DTRACE_INSCRATCH(mstate, size)) {
4276 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4277 regs[rd] = 0;
4278 break;
4279 }
4280
4281 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4282 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4283 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4284
4285 ((char *)dest)[size - 1] = '\0';
4286 mstate->dtms_scratch_ptr += size;
4287 regs[rd] = dest;
4288 break;
4289 }
4290
4291 case DIF_SUBR_MSGSIZE:
4292 case DIF_SUBR_MSGDSIZE: {
4293 uintptr_t baddr = tupregs[0].dttk_value, daddr;
4294 uintptr_t wptr, rptr;
4295 size_t count = 0;
4296 int cont = 0;
4297
4298 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
4299
4300 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4301 vstate)) {
4302 regs[rd] = 0;
4303 break;
4304 }
4305
4306 wptr = dtrace_loadptr(baddr +
4307 offsetof(mblk_t, b_wptr));
4308
4309 rptr = dtrace_loadptr(baddr +
4310 offsetof(mblk_t, b_rptr));
4311
4312 if (wptr < rptr) {
4313 *flags |= CPU_DTRACE_BADADDR;
4314 *illval = tupregs[0].dttk_value;
4315 break;
4316 }
4317
4318 daddr = dtrace_loadptr(baddr +
4319 offsetof(mblk_t, b_datap));
4320
4321 baddr = dtrace_loadptr(baddr +
4322 offsetof(mblk_t, b_cont));
4323
4324 /*
4325 * We want to prevent against denial-of-service here,
4326 * so we're only going to search the list for
4327 * dtrace_msgdsize_max mblks.
4328 */
4329 if (cont++ > dtrace_msgdsize_max) {
4330 *flags |= CPU_DTRACE_ILLOP;
4331 break;
4332 }
4333
4334 if (subr == DIF_SUBR_MSGDSIZE) {
4335 if (dtrace_load8(daddr +
4336 offsetof(dblk_t, db_type)) != M_DATA)
4337 continue;
4338 }
4339
4340 count += wptr - rptr;
4341 }
4342
4343 if (!(*flags & CPU_DTRACE_FAULT))
4344 regs[rd] = count;
4345
4346 break;
4347 }
4348
4349 case DIF_SUBR_PROGENYOF: {
4350 pid_t pid = tupregs[0].dttk_value;
4351 proc_t *p;
4352 int rval = 0;
4353
4354 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4355
4356 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4357 if (p->p_pidp->pid_id == pid) {
4358 rval = 1;
4359 break;
4360 }
4361 }
4362
4363 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4364
4365 regs[rd] = rval;
4366 break;
4367 }
4368
4369 case DIF_SUBR_SPECULATION:
4370 regs[rd] = dtrace_speculation(state);
4371 break;
4372
4373 case DIF_SUBR_COPYOUT: {
4374 uintptr_t kaddr = tupregs[0].dttk_value;
4375 uintptr_t uaddr = tupregs[1].dttk_value;
4376 uint64_t size = tupregs[2].dttk_value;
4377
4378 if (!dtrace_destructive_disallow &&
4379 dtrace_priv_proc_control(state, mstate) &&
4380 !dtrace_istoxic(kaddr, size) &&
4381 dtrace_canload(kaddr, size, mstate, vstate)) {
4382 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4383 dtrace_copyout(kaddr, uaddr, size, flags);
4384 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4385 }
4386 break;
4387 }
4388
4389 case DIF_SUBR_COPYOUTSTR: {
4390 uintptr_t kaddr = tupregs[0].dttk_value;
4391 uintptr_t uaddr = tupregs[1].dttk_value;
4392 uint64_t size = tupregs[2].dttk_value;
4393 size_t lim;
4394
4395 if (!dtrace_destructive_disallow &&
4396 dtrace_priv_proc_control(state, mstate) &&
4397 !dtrace_istoxic(kaddr, size) &&
4398 dtrace_strcanload(kaddr, size, &lim, mstate, vstate)) {
4399 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4400 dtrace_copyoutstr(kaddr, uaddr, lim, flags);
4401 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4402 }
4403 break;
4404 }
4405
4406 case DIF_SUBR_STRLEN: {
4407 size_t size = state->dts_options[DTRACEOPT_STRSIZE];
4408 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4409 size_t lim;
4410
4411 if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4412 regs[rd] = 0;
4413 break;
4414 }
4415 regs[rd] = dtrace_strlen((char *)addr, lim);
4416
4417 break;
4418 }
4419
4420 case DIF_SUBR_STRCHR:
4421 case DIF_SUBR_STRRCHR: {
4422 /*
4423 * We're going to iterate over the string looking for the
4424 * specified character. We will iterate until we have reached
4425 * the string length or we have found the character. If this
4426 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4427 * of the specified character instead of the first.
4428 */
4429 uintptr_t addr = tupregs[0].dttk_value;
4430 uintptr_t addr_limit;
4431 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4432 size_t lim;
4433 char c, target = (char)tupregs[1].dttk_value;
4434
4435 if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4436 regs[rd] = 0;
4437 break;
4438 }
4439 addr_limit = addr + lim;
4440
4441 for (regs[rd] = 0; addr < addr_limit; addr++) {
4442 if ((c = dtrace_load8(addr)) == target) {
4443 regs[rd] = addr;
4444
4445 if (subr == DIF_SUBR_STRCHR)
4446 break;
4447 }
4448 if (c == '\0')
4449 break;
4450 }
4451
4452 break;
4453 }
4454
4455 case DIF_SUBR_STRSTR:
4456 case DIF_SUBR_INDEX:
4457 case DIF_SUBR_RINDEX: {
4458 /*
4459 * We're going to iterate over the string looking for the
4460 * specified string. We will iterate until we have reached
4461 * the string length or we have found the string. (Yes, this
4462 * is done in the most naive way possible -- but considering
4463 * that the string we're searching for is likely to be
4464 * relatively short, the complexity of Rabin-Karp or similar
4465 * hardly seems merited.)
4466 */
4467 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4468 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4469 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4470 size_t len = dtrace_strlen(addr, size);
4471 size_t sublen = dtrace_strlen(substr, size);
4472 char *limit = addr + len, *orig = addr;
4473 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4474 int inc = 1;
4475
4476 regs[rd] = notfound;
4477
4478 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4479 regs[rd] = 0;
4480 break;
4481 }
4482
4483 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4484 vstate)) {
4485 regs[rd] = 0;
4486 break;
4487 }
4488
4489 /*
4490 * strstr() and index()/rindex() have similar semantics if
4491 * both strings are the empty string: strstr() returns a
4492 * pointer to the (empty) string, and index() and rindex()
4493 * both return index 0 (regardless of any position argument).
4494 */
4495 if (sublen == 0 && len == 0) {
4496 if (subr == DIF_SUBR_STRSTR)
4497 regs[rd] = (uintptr_t)addr;
4498 else
4499 regs[rd] = 0;
4500 break;
4501 }
4502
4503 if (subr != DIF_SUBR_STRSTR) {
4504 if (subr == DIF_SUBR_RINDEX) {
4505 limit = orig - 1;
4506 addr += len;
4507 inc = -1;
4508 }
4509
4510 /*
4511 * Both index() and rindex() take an optional position
4512 * argument that denotes the starting position.
4513 */
4514 if (nargs == 3) {
4515 int64_t pos = (int64_t)tupregs[2].dttk_value;
4516
4517 /*
4518 * If the position argument to index() is
4519 * negative, Perl implicitly clamps it at
4520 * zero. This semantic is a little surprising
4521 * given the special meaning of negative
4522 * positions to similar Perl functions like
4523 * substr(), but it appears to reflect a
4524 * notion that index() can start from a
4525 * negative index and increment its way up to
4526 * the string. Given this notion, Perl's
4527 * rindex() is at least self-consistent in
4528 * that it implicitly clamps positions greater
4529 * than the string length to be the string
4530 * length. Where Perl completely loses
4531 * coherence, however, is when the specified
4532 * substring is the empty string (""). In
4533 * this case, even if the position is
4534 * negative, rindex() returns 0 -- and even if
4535 * the position is greater than the length,
4536 * index() returns the string length. These
4537 * semantics violate the notion that index()
4538 * should never return a value less than the
4539 * specified position and that rindex() should
4540 * never return a value greater than the
4541 * specified position. (One assumes that
4542 * these semantics are artifacts of Perl's
4543 * implementation and not the results of
4544 * deliberate design -- it beggars belief that
4545 * even Larry Wall could desire such oddness.)
4546 * While in the abstract one would wish for
4547 * consistent position semantics across
4548 * substr(), index() and rindex() -- or at the
4549 * very least self-consistent position
4550 * semantics for index() and rindex() -- we
4551 * instead opt to keep with the extant Perl
4552 * semantics, in all their broken glory. (Do
4553 * we have more desire to maintain Perl's
4554 * semantics than Perl does? Probably.)
4555 */
4556 if (subr == DIF_SUBR_RINDEX) {
4557 if (pos < 0) {
4558 if (sublen == 0)
4559 regs[rd] = 0;
4560 break;
4561 }
4562
4563 if (pos > len)
4564 pos = len;
4565 } else {
4566 if (pos < 0)
4567 pos = 0;
4568
4569 if (pos >= len) {
4570 if (sublen == 0)
4571 regs[rd] = len;
4572 break;
4573 }
4574 }
4575
4576 addr = orig + pos;
4577 }
4578 }
4579
4580 for (regs[rd] = notfound; addr != limit; addr += inc) {
4581 if (dtrace_strncmp(addr, substr, sublen) == 0) {
4582 if (subr != DIF_SUBR_STRSTR) {
4583 /*
4584 * As D index() and rindex() are
4585 * modeled on Perl (and not on awk),
4586 * we return a zero-based (and not a
4587 * one-based) index. (For you Perl
4588 * weenies: no, we're not going to add
4589 * $[ -- and shouldn't you be at a con
4590 * or something?)
4591 */
4592 regs[rd] = (uintptr_t)(addr - orig);
4593 break;
4594 }
4595
4596 ASSERT(subr == DIF_SUBR_STRSTR);
4597 regs[rd] = (uintptr_t)addr;
4598 break;
4599 }
4600 }
4601
4602 break;
4603 }
4604
4605 case DIF_SUBR_STRTOK: {
4606 uintptr_t addr = tupregs[0].dttk_value;
4607 uintptr_t tokaddr = tupregs[1].dttk_value;
4608 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4609 uintptr_t limit, toklimit;
4610 size_t clim;
4611 uint8_t c, tokmap[32]; /* 256 / 8 */
4612 char *dest = (char *)mstate->dtms_scratch_ptr;
4613 int i;
4614
4615 /*
4616 * Check both the token buffer and (later) the input buffer,
4617 * since both could be non-scratch addresses.
4618 */
4619 if (!dtrace_strcanload(tokaddr, size, &clim, mstate, vstate)) {
4620 regs[rd] = 0;
4621 break;
4622 }
4623 toklimit = tokaddr + clim;
4624
4625 if (!DTRACE_INSCRATCH(mstate, size)) {
4626 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4627 regs[rd] = 0;
4628 break;
4629 }
4630
4631 if (addr == 0) {
4632 /*
4633 * If the address specified is NULL, we use our saved
4634 * strtok pointer from the mstate. Note that this
4635 * means that the saved strtok pointer is _only_
4636 * valid within multiple enablings of the same probe --
4637 * it behaves like an implicit clause-local variable.
4638 */
4639 addr = mstate->dtms_strtok;
4640 limit = mstate->dtms_strtok_limit;
4641 } else {
4642 /*
4643 * If the user-specified address is non-NULL we must
4644 * access check it. This is the only time we have
4645 * a chance to do so, since this address may reside
4646 * in the string table of this clause-- future calls
4647 * (when we fetch addr from mstate->dtms_strtok)
4648 * would fail this access check.
4649 */
4650 if (!dtrace_strcanload(addr, size, &clim, mstate,
4651 vstate)) {
4652 regs[rd] = 0;
4653 break;
4654 }
4655 limit = addr + clim;
4656 }
4657
4658 /*
4659 * First, zero the token map, and then process the token
4660 * string -- setting a bit in the map for every character
4661 * found in the token string.
4662 */
4663 for (i = 0; i < sizeof (tokmap); i++)
4664 tokmap[i] = 0;
4665
4666 for (; tokaddr < toklimit; tokaddr++) {
4667 if ((c = dtrace_load8(tokaddr)) == '\0')
4668 break;
4669
4670 ASSERT((c >> 3) < sizeof (tokmap));
4671 tokmap[c >> 3] |= (1 << (c & 0x7));
4672 }
4673
4674 for (; addr < limit; addr++) {
4675 /*
4676 * We're looking for a character that is _not_
4677 * contained in the token string.
4678 */
4679 if ((c = dtrace_load8(addr)) == '\0')
4680 break;
4681
4682 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4683 break;
4684 }
4685
4686 if (c == '\0') {
4687 /*
4688 * We reached the end of the string without finding
4689 * any character that was not in the token string.
4690 * We return NULL in this case, and we set the saved
4691 * address to NULL as well.
4692 */
4693 regs[rd] = 0;
4694 mstate->dtms_strtok = 0;
4695 mstate->dtms_strtok_limit = 0;
4696 break;
4697 }
4698
4699 /*
4700 * From here on, we're copying into the destination string.
4701 */
4702 for (i = 0; addr < limit && i < size - 1; addr++) {
4703 if ((c = dtrace_load8(addr)) == '\0')
4704 break;
4705
4706 if (tokmap[c >> 3] & (1 << (c & 0x7)))
4707 break;
4708
4709 ASSERT(i < size);
4710 dest[i++] = c;
4711 }
4712
4713 ASSERT(i < size);
4714 dest[i] = '\0';
4715 regs[rd] = (uintptr_t)dest;
4716 mstate->dtms_scratch_ptr += size;
4717 mstate->dtms_strtok = addr;
4718 mstate->dtms_strtok_limit = limit;
4719 break;
4720 }
4721
4722 case DIF_SUBR_SUBSTR: {
4723 uintptr_t s = tupregs[0].dttk_value;
4724 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4725 char *d = (char *)mstate->dtms_scratch_ptr;
4726 int64_t index = (int64_t)tupregs[1].dttk_value;
4727 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4728 size_t len = dtrace_strlen((char *)s, size);
4729 int64_t i;
4730
4731 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4732 regs[rd] = 0;
4733 break;
4734 }
4735
4736 if (!DTRACE_INSCRATCH(mstate, size)) {
4737 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4738 regs[rd] = 0;
4739 break;
4740 }
4741
4742 if (nargs <= 2)
4743 remaining = (int64_t)size;
4744
4745 if (index < 0) {
4746 index += len;
4747
4748 if (index < 0 && index + remaining > 0) {
4749 remaining += index;
4750 index = 0;
4751 }
4752 }
4753
4754 if (index >= len || index < 0) {
4755 remaining = 0;
4756 } else if (remaining < 0) {
4757 remaining += len - index;
4758 } else if (index + remaining > size) {
4759 remaining = size - index;
4760 }
4761
4762 for (i = 0; i < remaining; i++) {
4763 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4764 break;
4765 }
4766
4767 d[i] = '\0';
4768
4769 mstate->dtms_scratch_ptr += size;
4770 regs[rd] = (uintptr_t)d;
4771 break;
4772 }
4773
4774 case DIF_SUBR_JSON: {
4775 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4776 uintptr_t json = tupregs[0].dttk_value;
4777 size_t jsonlen = dtrace_strlen((char *)json, size);
4778 uintptr_t elem = tupregs[1].dttk_value;
4779 size_t elemlen = dtrace_strlen((char *)elem, size);
4780
4781 char *dest = (char *)mstate->dtms_scratch_ptr;
4782 char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
4783 char *ee = elemlist;
4784 int nelems = 1;
4785 uintptr_t cur;
4786
4787 if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
4788 !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
4789 regs[rd] = 0;
4790 break;
4791 }
4792
4793 if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
4794 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4795 regs[rd] = 0;
4796 break;
4797 }
4798
4799 /*
4800 * Read the element selector and split it up into a packed list
4801 * of strings.
4802 */
4803 for (cur = elem; cur < elem + elemlen; cur++) {
4804 char cc = dtrace_load8(cur);
4805
4806 if (cur == elem && cc == '[') {
4807 /*
4808 * If the first element selector key is
4809 * actually an array index then ignore the
4810 * bracket.
4811 */
4812 continue;
4813 }
4814
4815 if (cc == ']')
4816 continue;
4817
4818 if (cc == '.' || cc == '[') {
4819 nelems++;
4820 cc = '\0';
4821 }
4822
4823 *ee++ = cc;
4824 }
4825 *ee++ = '\0';
4826
4827 if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
4828 nelems, dest)) != 0)
4829 mstate->dtms_scratch_ptr += jsonlen + 1;
4830 break;
4831 }
4832
4833 case DIF_SUBR_TOUPPER:
4834 case DIF_SUBR_TOLOWER: {
4835 uintptr_t s = tupregs[0].dttk_value;
4836 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4837 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4838 size_t len = dtrace_strlen((char *)s, size);
4839 char lower, upper, convert;
4840 int64_t i;
4841
4842 if (subr == DIF_SUBR_TOUPPER) {
4843 lower = 'a';
4844 upper = 'z';
4845 convert = 'A';
4846 } else {
4847 lower = 'A';
4848 upper = 'Z';
4849 convert = 'a';
4850 }
4851
4852 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4853 regs[rd] = 0;
4854 break;
4855 }
4856
4857 if (!DTRACE_INSCRATCH(mstate, size)) {
4858 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4859 regs[rd] = 0;
4860 break;
4861 }
4862
4863 for (i = 0; i < size - 1; i++) {
4864 if ((c = dtrace_load8(s + i)) == '\0')
4865 break;
4866
4867 if (c >= lower && c <= upper)
4868 c = convert + (c - lower);
4869
4870 dest[i] = c;
4871 }
4872
4873 ASSERT(i < size);
4874 dest[i] = '\0';
4875 regs[rd] = (uintptr_t)dest;
4876 mstate->dtms_scratch_ptr += size;
4877 break;
4878 }
4879
4880 case DIF_SUBR_GETMAJOR:
4881 #ifdef _LP64
4882 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4883 #else
4884 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4885 #endif
4886 break;
4887
4888 case DIF_SUBR_GETMINOR:
4889 #ifdef _LP64
4890 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4891 #else
4892 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4893 #endif
4894 break;
4895
4896 case DIF_SUBR_DDI_PATHNAME: {
4897 /*
4898 * This one is a galactic mess. We are going to roughly
4899 * emulate ddi_pathname(), but it's made more complicated
4900 * by the fact that we (a) want to include the minor name and
4901 * (b) must proceed iteratively instead of recursively.
4902 */
4903 uintptr_t dest = mstate->dtms_scratch_ptr;
4904 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4905 char *start = (char *)dest, *end = start + size - 1;
4906 uintptr_t daddr = tupregs[0].dttk_value;
4907 int64_t minor = (int64_t)tupregs[1].dttk_value;
4908 char *s;
4909 int i, len, depth = 0;
4910
4911 /*
4912 * Due to all the pointer jumping we do and context we must
4913 * rely upon, we just mandate that the user must have kernel
4914 * read privileges to use this routine.
4915 */
4916 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4917 *flags |= CPU_DTRACE_KPRIV;
4918 *illval = daddr;
4919 regs[rd] = 0;
4920 }
4921
4922 if (!DTRACE_INSCRATCH(mstate, size)) {
4923 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4924 regs[rd] = 0;
4925 break;
4926 }
4927
4928 *end = '\0';
4929
4930 /*
4931 * We want to have a name for the minor. In order to do this,
4932 * we need to walk the minor list from the devinfo. We want
4933 * to be sure that we don't infinitely walk a circular list,
4934 * so we check for circularity by sending a scout pointer
4935 * ahead two elements for every element that we iterate over;
4936 * if the list is circular, these will ultimately point to the
4937 * same element. You may recognize this little trick as the
4938 * answer to a stupid interview question -- one that always
4939 * seems to be asked by those who had to have it laboriously
4940 * explained to them, and who can't even concisely describe
4941 * the conditions under which one would be forced to resort to
4942 * this technique. Needless to say, those conditions are
4943 * found here -- and probably only here. Is this the only use
4944 * of this infamous trick in shipping, production code? If it
4945 * isn't, it probably should be...
4946 */
4947 if (minor != -1) {
4948 uintptr_t maddr = dtrace_loadptr(daddr +
4949 offsetof(struct dev_info, devi_minor));
4950
4951 uintptr_t next = offsetof(struct ddi_minor_data, next);
4952 uintptr_t name = offsetof(struct ddi_minor_data,
4953 d_minor) + offsetof(struct ddi_minor, name);
4954 uintptr_t dev = offsetof(struct ddi_minor_data,
4955 d_minor) + offsetof(struct ddi_minor, dev);
4956 uintptr_t scout;
4957
4958 if (maddr != 0)
4959 scout = dtrace_loadptr(maddr + next);
4960
4961 while (maddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
4962 uint64_t m;
4963 #ifdef _LP64
4964 m = dtrace_load64(maddr + dev) & MAXMIN64;
4965 #else
4966 m = dtrace_load32(maddr + dev) & MAXMIN;
4967 #endif
4968 if (m != minor) {
4969 maddr = dtrace_loadptr(maddr + next);
4970
4971 if (scout == 0)
4972 continue;
4973
4974 scout = dtrace_loadptr(scout + next);
4975
4976 if (scout == 0)
4977 continue;
4978
4979 scout = dtrace_loadptr(scout + next);
4980
4981 if (scout == 0)
4982 continue;
4983
4984 if (scout == maddr) {
4985 *flags |= CPU_DTRACE_ILLOP;
4986 break;
4987 }
4988
4989 continue;
4990 }
4991
4992 /*
4993 * We have the minor data. Now we need to
4994 * copy the minor's name into the end of the
4995 * pathname.
4996 */
4997 s = (char *)dtrace_loadptr(maddr + name);
4998 len = dtrace_strlen(s, size);
4999
5000 if (*flags & CPU_DTRACE_FAULT)
5001 break;
5002
5003 if (len != 0) {
5004 if ((end -= (len + 1)) < start)
5005 break;
5006
5007 *end = ':';
5008 }
5009
5010 for (i = 1; i <= len; i++)
5011 end[i] = dtrace_load8((uintptr_t)s++);
5012 break;
5013 }
5014 }
5015
5016 while (daddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
5017 ddi_node_state_t devi_state;
5018
5019 devi_state = dtrace_load32(daddr +
5020 offsetof(struct dev_info, devi_node_state));
5021
5022 if (*flags & CPU_DTRACE_FAULT)
5023 break;
5024
5025 if (devi_state >= DS_INITIALIZED) {
5026 s = (char *)dtrace_loadptr(daddr +
5027 offsetof(struct dev_info, devi_addr));
5028 len = dtrace_strlen(s, size);
5029
5030 if (*flags & CPU_DTRACE_FAULT)
5031 break;
5032
5033 if (len != 0) {
5034 if ((end -= (len + 1)) < start)
5035 break;
5036
5037 *end = '@';
5038 }
5039
5040 for (i = 1; i <= len; i++)
5041 end[i] = dtrace_load8((uintptr_t)s++);
5042 }
5043
5044 /*
5045 * Now for the node name...
5046 */
5047 s = (char *)dtrace_loadptr(daddr +
5048 offsetof(struct dev_info, devi_node_name));
5049
5050 daddr = dtrace_loadptr(daddr +
5051 offsetof(struct dev_info, devi_parent));
5052
5053 /*
5054 * If our parent is NULL (that is, if we're the root
5055 * node), we're going to use the special path
5056 * "devices".
5057 */
5058 if (daddr == 0)
5059 s = "devices";
5060
5061 len = dtrace_strlen(s, size);
5062 if (*flags & CPU_DTRACE_FAULT)
5063 break;
5064
5065 if ((end -= (len + 1)) < start)
5066 break;
5067
5068 for (i = 1; i <= len; i++)
5069 end[i] = dtrace_load8((uintptr_t)s++);
5070 *end = '/';
5071
5072 if (depth++ > dtrace_devdepth_max) {
5073 *flags |= CPU_DTRACE_ILLOP;
5074 break;
5075 }
5076 }
5077
5078 if (end < start)
5079 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5080
5081 if (daddr == 0) {
5082 regs[rd] = (uintptr_t)end;
5083 mstate->dtms_scratch_ptr += size;
5084 }
5085
5086 break;
5087 }
5088
5089 case DIF_SUBR_STRJOIN: {
5090 char *d = (char *)mstate->dtms_scratch_ptr;
5091 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5092 uintptr_t s1 = tupregs[0].dttk_value;
5093 uintptr_t s2 = tupregs[1].dttk_value;
5094 int i = 0, j = 0;
5095 size_t lim1, lim2;
5096 char c;
5097
5098 if (!dtrace_strcanload(s1, size, &lim1, mstate, vstate) ||
5099 !dtrace_strcanload(s2, size, &lim2, mstate, vstate)) {
5100 regs[rd] = 0;
5101 break;
5102 }
5103
5104 if (!DTRACE_INSCRATCH(mstate, size)) {
5105 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5106 regs[rd] = 0;
5107 break;
5108 }
5109
5110 for (;;) {
5111 if (i >= size) {
5112 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5113 regs[rd] = 0;
5114 break;
5115 }
5116 c = (i >= lim1) ? '\0' : dtrace_load8(s1++);
5117 if ((d[i++] = c) == '\0') {
5118 i--;
5119 break;
5120 }
5121 }
5122
5123 for (;;) {
5124 if (i >= size) {
5125 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5126 regs[rd] = 0;
5127 break;
5128 }
5129
5130 c = (j++ >= lim2) ? '\0' : dtrace_load8(s2++);
5131 if ((d[i++] = c) == '\0')
5132 break;
5133 }
5134
5135 if (i < size) {
5136 mstate->dtms_scratch_ptr += i;
5137 regs[rd] = (uintptr_t)d;
5138 }
5139
5140 break;
5141 }
5142
5143 case DIF_SUBR_STRTOLL: {
5144 uintptr_t s = tupregs[0].dttk_value;
5145 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5146 size_t lim;
5147 int base = 10;
5148
5149 if (nargs > 1) {
5150 if ((base = tupregs[1].dttk_value) <= 1 ||
5151 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5152 *flags |= CPU_DTRACE_ILLOP;
5153 break;
5154 }
5155 }
5156
5157 if (!dtrace_strcanload(s, size, &lim, mstate, vstate)) {
5158 regs[rd] = INT64_MIN;
5159 break;
5160 }
5161
5162 regs[rd] = dtrace_strtoll((char *)s, base, lim);
5163 break;
5164 }
5165
5166 case DIF_SUBR_LLTOSTR: {
5167 int64_t i = (int64_t)tupregs[0].dttk_value;
5168 uint64_t val, digit;
5169 uint64_t size = 65; /* enough room for 2^64 in binary */
5170 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
5171 int base = 10;
5172
5173 if (nargs > 1) {
5174 if ((base = tupregs[1].dttk_value) <= 1 ||
5175 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5176 *flags |= CPU_DTRACE_ILLOP;
5177 break;
5178 }
5179 }
5180
5181 val = (base == 10 && i < 0) ? i * -1 : i;
5182
5183 if (!DTRACE_INSCRATCH(mstate, size)) {
5184 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5185 regs[rd] = 0;
5186 break;
5187 }
5188
5189 for (*end-- = '\0'; val; val /= base) {
5190 if ((digit = val % base) <= '9' - '0') {
5191 *end-- = '0' + digit;
5192 } else {
5193 *end-- = 'a' + (digit - ('9' - '0') - 1);
5194 }
5195 }
5196
5197 if (i == 0 && base == 16)
5198 *end-- = '0';
5199
5200 if (base == 16)
5201 *end-- = 'x';
5202
5203 if (i == 0 || base == 8 || base == 16)
5204 *end-- = '0';
5205
5206 if (i < 0 && base == 10)
5207 *end-- = '-';
5208
5209 regs[rd] = (uintptr_t)end + 1;
5210 mstate->dtms_scratch_ptr += size;
5211 break;
5212 }
5213
5214 case DIF_SUBR_HTONS:
5215 case DIF_SUBR_NTOHS:
5216 #ifdef _BIG_ENDIAN
5217 regs[rd] = (uint16_t)tupregs[0].dttk_value;
5218 #else
5219 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
5220 #endif
5221 break;
5222
5223
5224 case DIF_SUBR_HTONL:
5225 case DIF_SUBR_NTOHL:
5226 #ifdef _BIG_ENDIAN
5227 regs[rd] = (uint32_t)tupregs[0].dttk_value;
5228 #else
5229 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5230 #endif
5231 break;
5232
5233
5234 case DIF_SUBR_HTONLL:
5235 case DIF_SUBR_NTOHLL:
5236 #ifdef _BIG_ENDIAN
5237 regs[rd] = (uint64_t)tupregs[0].dttk_value;
5238 #else
5239 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5240 #endif
5241 break;
5242
5243
5244 case DIF_SUBR_DIRNAME:
5245 case DIF_SUBR_BASENAME: {
5246 char *dest = (char *)mstate->dtms_scratch_ptr;
5247 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5248 uintptr_t src = tupregs[0].dttk_value;
5249 int i, j, len = dtrace_strlen((char *)src, size);
5250 int lastbase = -1, firstbase = -1, lastdir = -1;
5251 int start, end;
5252
5253 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5254 regs[rd] = 0;
5255 break;
5256 }
5257
5258 if (!DTRACE_INSCRATCH(mstate, size)) {
5259 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5260 regs[rd] = 0;
5261 break;
5262 }
5263
5264 /*
5265 * The basename and dirname for a zero-length string is
5266 * defined to be "."
5267 */
5268 if (len == 0) {
5269 len = 1;
5270 src = (uintptr_t)".";
5271 }
5272
5273 /*
5274 * Start from the back of the string, moving back toward the
5275 * front until we see a character that isn't a slash. That
5276 * character is the last character in the basename.
5277 */
5278 for (i = len - 1; i >= 0; i--) {
5279 if (dtrace_load8(src + i) != '/')
5280 break;
5281 }
5282
5283 if (i >= 0)
5284 lastbase = i;
5285
5286 /*
5287 * Starting from the last character in the basename, move
5288 * towards the front until we find a slash. The character
5289 * that we processed immediately before that is the first
5290 * character in the basename.
5291 */
5292 for (; i >= 0; i--) {
5293 if (dtrace_load8(src + i) == '/')
5294 break;
5295 }
5296
5297 if (i >= 0)
5298 firstbase = i + 1;
5299
5300 /*
5301 * Now keep going until we find a non-slash character. That
5302 * character is the last character in the dirname.
5303 */
5304 for (; i >= 0; i--) {
5305 if (dtrace_load8(src + i) != '/')
5306 break;
5307 }
5308
5309 if (i >= 0)
5310 lastdir = i;
5311
5312 ASSERT(!(lastbase == -1 && firstbase != -1));
5313 ASSERT(!(firstbase == -1 && lastdir != -1));
5314
5315 if (lastbase == -1) {
5316 /*
5317 * We didn't find a non-slash character. We know that
5318 * the length is non-zero, so the whole string must be
5319 * slashes. In either the dirname or the basename
5320 * case, we return '/'.
5321 */
5322 ASSERT(firstbase == -1);
5323 firstbase = lastbase = lastdir = 0;
5324 }
5325
5326 if (firstbase == -1) {
5327 /*
5328 * The entire string consists only of a basename
5329 * component. If we're looking for dirname, we need
5330 * to change our string to be just "."; if we're
5331 * looking for a basename, we'll just set the first
5332 * character of the basename to be 0.
5333 */
5334 if (subr == DIF_SUBR_DIRNAME) {
5335 ASSERT(lastdir == -1);
5336 src = (uintptr_t)".";
5337 lastdir = 0;
5338 } else {
5339 firstbase = 0;
5340 }
5341 }
5342
5343 if (subr == DIF_SUBR_DIRNAME) {
5344 if (lastdir == -1) {
5345 /*
5346 * We know that we have a slash in the name --
5347 * or lastdir would be set to 0, above. And
5348 * because lastdir is -1, we know that this
5349 * slash must be the first character. (That
5350 * is, the full string must be of the form
5351 * "/basename".) In this case, the last
5352 * character of the directory name is 0.
5353 */
5354 lastdir = 0;
5355 }
5356
5357 start = 0;
5358 end = lastdir;
5359 } else {
5360 ASSERT(subr == DIF_SUBR_BASENAME);
5361 ASSERT(firstbase != -1 && lastbase != -1);
5362 start = firstbase;
5363 end = lastbase;
5364 }
5365
5366 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5367 dest[j] = dtrace_load8(src + i);
5368
5369 dest[j] = '\0';
5370 regs[rd] = (uintptr_t)dest;
5371 mstate->dtms_scratch_ptr += size;
5372 break;
5373 }
5374
5375 case DIF_SUBR_GETF: {
5376 uintptr_t fd = tupregs[0].dttk_value;
5377 uf_info_t *finfo = &curthread->t_procp->p_user.u_finfo;
5378 file_t *fp;
5379
5380 if (!dtrace_priv_proc(state, mstate)) {
5381 regs[rd] = 0;
5382 break;
5383 }
5384
5385 /*
5386 * This is safe because fi_nfiles only increases, and the
5387 * fi_list array is not freed when the array size doubles.
5388 * (See the comment in flist_grow() for details on the
5389 * management of the u_finfo structure.)
5390 */
5391 fp = fd < finfo->fi_nfiles ? finfo->fi_list[fd].uf_file : NULL;
5392
5393 mstate->dtms_getf = fp;
5394 regs[rd] = (uintptr_t)fp;
5395 break;
5396 }
5397
5398 case DIF_SUBR_CLEANPATH: {
5399 char *dest = (char *)mstate->dtms_scratch_ptr, c;
5400 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5401 uintptr_t src = tupregs[0].dttk_value;
5402 size_t lim;
5403 int i = 0, j = 0;
5404 zone_t *z;
5405
5406 if (!dtrace_strcanload(src, size, &lim, mstate, vstate)) {
5407 regs[rd] = 0;
5408 break;
5409 }
5410
5411 if (!DTRACE_INSCRATCH(mstate, size)) {
5412 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5413 regs[rd] = 0;
5414 break;
5415 }
5416
5417 /*
5418 * Move forward, loading each character.
5419 */
5420 do {
5421 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5422 next:
5423 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
5424 break;
5425
5426 if (c != '/') {
5427 dest[j++] = c;
5428 continue;
5429 }
5430
5431 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5432
5433 if (c == '/') {
5434 /*
5435 * We have two slashes -- we can just advance
5436 * to the next character.
5437 */
5438 goto next;
5439 }
5440
5441 if (c != '.') {
5442 /*
5443 * This is not "." and it's not ".." -- we can
5444 * just store the "/" and this character and
5445 * drive on.
5446 */
5447 dest[j++] = '/';
5448 dest[j++] = c;
5449 continue;
5450 }
5451
5452 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5453
5454 if (c == '/') {
5455 /*
5456 * This is a "/./" component. We're not going
5457 * to store anything in the destination buffer;
5458 * we're just going to go to the next component.
5459 */
5460 goto next;
5461 }
5462
5463 if (c != '.') {
5464 /*
5465 * This is not ".." -- we can just store the
5466 * "/." and this character and continue
5467 * processing.
5468 */
5469 dest[j++] = '/';
5470 dest[j++] = '.';
5471 dest[j++] = c;
5472 continue;
5473 }
5474
5475 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5476
5477 if (c != '/' && c != '\0') {
5478 /*
5479 * This is not ".." -- it's "..[mumble]".
5480 * We'll store the "/.." and this character
5481 * and continue processing.
5482 */
5483 dest[j++] = '/';
5484 dest[j++] = '.';
5485 dest[j++] = '.';
5486 dest[j++] = c;
5487 continue;
5488 }
5489
5490 /*
5491 * This is "/../" or "/..\0". We need to back up
5492 * our destination pointer until we find a "/".
5493 */
5494 i--;
5495 while (j != 0 && dest[--j] != '/')
5496 continue;
5497
5498 if (c == '\0')
5499 dest[++j] = '/';
5500 } while (c != '\0');
5501
5502 dest[j] = '\0';
5503
5504 if (mstate->dtms_getf != NULL &&
5505 !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5506 (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5507 /*
5508 * If we've done a getf() as a part of this ECB and we
5509 * don't have kernel access (and we're not in the global
5510 * zone), check if the path we cleaned up begins with
5511 * the zone's root path, and trim it off if so. Note
5512 * that this is an output cleanliness issue, not a
5513 * security issue: knowing one's zone root path does
5514 * not enable privilege escalation.
5515 */
5516 if (strstr(dest, z->zone_rootpath) == dest)
5517 dest += strlen(z->zone_rootpath) - 1;
5518 }
5519
5520 regs[rd] = (uintptr_t)dest;
5521 mstate->dtms_scratch_ptr += size;
5522 break;
5523 }
5524
5525 case DIF_SUBR_INET_NTOA:
5526 case DIF_SUBR_INET_NTOA6:
5527 case DIF_SUBR_INET_NTOP: {
5528 size_t size;
5529 int af, argi, i;
5530 char *base, *end;
5531
5532 if (subr == DIF_SUBR_INET_NTOP) {
5533 af = (int)tupregs[0].dttk_value;
5534 argi = 1;
5535 } else {
5536 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5537 argi = 0;
5538 }
5539
5540 if (af == AF_INET) {
5541 ipaddr_t ip4;
5542 uint8_t *ptr8, val;
5543
5544 if (!dtrace_canload(tupregs[argi].dttk_value,
5545 sizeof (ipaddr_t), mstate, vstate)) {
5546 regs[rd] = 0;
5547 break;
5548 }
5549
5550 /*
5551 * Safely load the IPv4 address.
5552 */
5553 ip4 = dtrace_load32(tupregs[argi].dttk_value);
5554
5555 /*
5556 * Check an IPv4 string will fit in scratch.
5557 */
5558 size = INET_ADDRSTRLEN;
5559 if (!DTRACE_INSCRATCH(mstate, size)) {
5560 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5561 regs[rd] = 0;
5562 break;
5563 }
5564 base = (char *)mstate->dtms_scratch_ptr;
5565 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5566
5567 /*
5568 * Stringify as a dotted decimal quad.
5569 */
5570 *end-- = '\0';
5571 ptr8 = (uint8_t *)&ip4;
5572 for (i = 3; i >= 0; i--) {
5573 val = ptr8[i];
5574
5575 if (val == 0) {
5576 *end-- = '0';
5577 } else {
5578 for (; val; val /= 10) {
5579 *end-- = '0' + (val % 10);
5580 }
5581 }
5582
5583 if (i > 0)
5584 *end-- = '.';
5585 }
5586 ASSERT(end + 1 >= base);
5587
5588 } else if (af == AF_INET6) {
5589 struct in6_addr ip6;
5590 int firstzero, tryzero, numzero, v6end;
5591 uint16_t val;
5592 const char digits[] = "0123456789abcdef";
5593
5594 /*
5595 * Stringify using RFC 1884 convention 2 - 16 bit
5596 * hexadecimal values with a zero-run compression.
5597 * Lower case hexadecimal digits are used.
5598 * eg, fe80::214:4fff:fe0b:76c8.
5599 * The IPv4 embedded form is returned for inet_ntop,
5600 * just the IPv4 string is returned for inet_ntoa6.
5601 */
5602
5603 if (!dtrace_canload(tupregs[argi].dttk_value,
5604 sizeof (struct in6_addr), mstate, vstate)) {
5605 regs[rd] = 0;
5606 break;
5607 }
5608
5609 /*
5610 * Safely load the IPv6 address.
5611 */
5612 dtrace_bcopy(
5613 (void *)(uintptr_t)tupregs[argi].dttk_value,
5614 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5615
5616 /*
5617 * Check an IPv6 string will fit in scratch.
5618 */
5619 size = INET6_ADDRSTRLEN;
5620 if (!DTRACE_INSCRATCH(mstate, size)) {
5621 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5622 regs[rd] = 0;
5623 break;
5624 }
5625 base = (char *)mstate->dtms_scratch_ptr;
5626 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5627 *end-- = '\0';
5628
5629 /*
5630 * Find the longest run of 16 bit zero values
5631 * for the single allowed zero compression - "::".
5632 */
5633 firstzero = -1;
5634 tryzero = -1;
5635 numzero = 1;
5636 for (i = 0; i < sizeof (struct in6_addr); i++) {
5637 if (ip6._S6_un._S6_u8[i] == 0 &&
5638 tryzero == -1 && i % 2 == 0) {
5639 tryzero = i;
5640 continue;
5641 }
5642
5643 if (tryzero != -1 &&
5644 (ip6._S6_un._S6_u8[i] != 0 ||
5645 i == sizeof (struct in6_addr) - 1)) {
5646
5647 if (i - tryzero <= numzero) {
5648 tryzero = -1;
5649 continue;
5650 }
5651
5652 firstzero = tryzero;
5653 numzero = i - i % 2 - tryzero;
5654 tryzero = -1;
5655
5656 if (ip6._S6_un._S6_u8[i] == 0 &&
5657 i == sizeof (struct in6_addr) - 1)
5658 numzero += 2;
5659 }
5660 }
5661 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5662
5663 /*
5664 * Check for an IPv4 embedded address.
5665 */
5666 v6end = sizeof (struct in6_addr) - 2;
5667 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5668 IN6_IS_ADDR_V4COMPAT(&ip6)) {
5669 for (i = sizeof (struct in6_addr) - 1;
5670 i >= DTRACE_V4MAPPED_OFFSET; i--) {
5671 ASSERT(end >= base);
5672
5673 val = ip6._S6_un._S6_u8[i];
5674
5675 if (val == 0) {
5676 *end-- = '0';
5677 } else {
5678 for (; val; val /= 10) {
5679 *end-- = '0' + val % 10;
5680 }
5681 }
5682
5683 if (i > DTRACE_V4MAPPED_OFFSET)
5684 *end-- = '.';
5685 }
5686
5687 if (subr == DIF_SUBR_INET_NTOA6)
5688 goto inetout;
5689
5690 /*
5691 * Set v6end to skip the IPv4 address that
5692 * we have already stringified.
5693 */
5694 v6end = 10;
5695 }
5696
5697 /*
5698 * Build the IPv6 string by working through the
5699 * address in reverse.
5700 */
5701 for (i = v6end; i >= 0; i -= 2) {
5702 ASSERT(end >= base);
5703
5704 if (i == firstzero + numzero - 2) {
5705 *end-- = ':';
5706 *end-- = ':';
5707 i -= numzero - 2;
5708 continue;
5709 }
5710
5711 if (i < 14 && i != firstzero - 2)
5712 *end-- = ':';
5713
5714 val = (ip6._S6_un._S6_u8[i] << 8) +
5715 ip6._S6_un._S6_u8[i + 1];
5716
5717 if (val == 0) {
5718 *end-- = '0';
5719 } else {
5720 for (; val; val /= 16) {
5721 *end-- = digits[val % 16];
5722 }
5723 }
5724 }
5725 ASSERT(end + 1 >= base);
5726
5727 } else {
5728 /*
5729 * The user didn't use AH_INET or AH_INET6.
5730 */
5731 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5732 regs[rd] = 0;
5733 break;
5734 }
5735
5736 inetout: regs[rd] = (uintptr_t)end + 1;
5737 mstate->dtms_scratch_ptr += size;
5738 break;
5739 }
5740
5741 }
5742 }
5743
5744 /*
5745 * Emulate the execution of DTrace IR instructions specified by the given
5746 * DIF object. This function is deliberately void of assertions as all of
5747 * the necessary checks are handled by a call to dtrace_difo_validate().
5748 */
5749 static uint64_t
5750 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
5751 dtrace_vstate_t *vstate, dtrace_state_t *state)
5752 {
5753 const dif_instr_t *text = difo->dtdo_buf;
5754 const uint_t textlen = difo->dtdo_len;
5755 const char *strtab = difo->dtdo_strtab;
5756 const uint64_t *inttab = difo->dtdo_inttab;
5757
5758 uint64_t rval = 0;
5759 dtrace_statvar_t *svar;
5760 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5761 dtrace_difv_t *v;
5762 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5763 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
5764
5765 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5766 uint64_t regs[DIF_DIR_NREGS];
5767 uint64_t *tmp;
5768
5769 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5770 int64_t cc_r;
5771 uint_t pc = 0, id, opc;
5772 uint8_t ttop = 0;
5773 dif_instr_t instr;
5774 uint_t r1, r2, rd;
5775
5776 /*
5777 * We stash the current DIF object into the machine state: we need it
5778 * for subsequent access checking.
5779 */
5780 mstate->dtms_difo = difo;
5781
5782 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
5783
5784 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5785 opc = pc;
5786
5787 instr = text[pc++];
5788 r1 = DIF_INSTR_R1(instr);
5789 r2 = DIF_INSTR_R2(instr);
5790 rd = DIF_INSTR_RD(instr);
5791
5792 switch (DIF_INSTR_OP(instr)) {
5793 case DIF_OP_OR:
5794 regs[rd] = regs[r1] | regs[r2];
5795 break;
5796 case DIF_OP_XOR:
5797 regs[rd] = regs[r1] ^ regs[r2];
5798 break;
5799 case DIF_OP_AND:
5800 regs[rd] = regs[r1] & regs[r2];
5801 break;
5802 case DIF_OP_SLL:
5803 regs[rd] = regs[r1] << regs[r2];
5804 break;
5805 case DIF_OP_SRL:
5806 regs[rd] = regs[r1] >> regs[r2];
5807 break;
5808 case DIF_OP_SUB:
5809 regs[rd] = regs[r1] - regs[r2];
5810 break;
5811 case DIF_OP_ADD:
5812 regs[rd] = regs[r1] + regs[r2];
5813 break;
5814 case DIF_OP_MUL:
5815 regs[rd] = regs[r1] * regs[r2];
5816 break;
5817 case DIF_OP_SDIV:
5818 if (regs[r2] == 0) {
5819 regs[rd] = 0;
5820 *flags |= CPU_DTRACE_DIVZERO;
5821 } else {
5822 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5823 regs[rd] = (int64_t)regs[r1] /
5824 (int64_t)regs[r2];
5825 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5826 }
5827 break;
5828
5829 case DIF_OP_UDIV:
5830 if (regs[r2] == 0) {
5831 regs[rd] = 0;
5832 *flags |= CPU_DTRACE_DIVZERO;
5833 } else {
5834 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5835 regs[rd] = regs[r1] / regs[r2];
5836 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5837 }
5838 break;
5839
5840 case DIF_OP_SREM:
5841 if (regs[r2] == 0) {
5842 regs[rd] = 0;
5843 *flags |= CPU_DTRACE_DIVZERO;
5844 } else {
5845 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5846 regs[rd] = (int64_t)regs[r1] %
5847 (int64_t)regs[r2];
5848 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5849 }
5850 break;
5851
5852 case DIF_OP_UREM:
5853 if (regs[r2] == 0) {
5854 regs[rd] = 0;
5855 *flags |= CPU_DTRACE_DIVZERO;
5856 } else {
5857 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5858 regs[rd] = regs[r1] % regs[r2];
5859 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5860 }
5861 break;
5862
5863 case DIF_OP_NOT:
5864 regs[rd] = ~regs[r1];
5865 break;
5866 case DIF_OP_MOV:
5867 regs[rd] = regs[r1];
5868 break;
5869 case DIF_OP_CMP:
5870 cc_r = regs[r1] - regs[r2];
5871 cc_n = cc_r < 0;
5872 cc_z = cc_r == 0;
5873 cc_v = 0;
5874 cc_c = regs[r1] < regs[r2];
5875 break;
5876 case DIF_OP_TST:
5877 cc_n = cc_v = cc_c = 0;
5878 cc_z = regs[r1] == 0;
5879 break;
5880 case DIF_OP_BA:
5881 pc = DIF_INSTR_LABEL(instr);
5882 break;
5883 case DIF_OP_BE:
5884 if (cc_z)
5885 pc = DIF_INSTR_LABEL(instr);
5886 break;
5887 case DIF_OP_BNE:
5888 if (cc_z == 0)
5889 pc = DIF_INSTR_LABEL(instr);
5890 break;
5891 case DIF_OP_BG:
5892 if ((cc_z | (cc_n ^ cc_v)) == 0)
5893 pc = DIF_INSTR_LABEL(instr);
5894 break;
5895 case DIF_OP_BGU:
5896 if ((cc_c | cc_z) == 0)
5897 pc = DIF_INSTR_LABEL(instr);
5898 break;
5899 case DIF_OP_BGE:
5900 if ((cc_n ^ cc_v) == 0)
5901 pc = DIF_INSTR_LABEL(instr);
5902 break;
5903 case DIF_OP_BGEU:
5904 if (cc_c == 0)
5905 pc = DIF_INSTR_LABEL(instr);
5906 break;
5907 case DIF_OP_BL:
5908 if (cc_n ^ cc_v)
5909 pc = DIF_INSTR_LABEL(instr);
5910 break;
5911 case DIF_OP_BLU:
5912 if (cc_c)
5913 pc = DIF_INSTR_LABEL(instr);
5914 break;
5915 case DIF_OP_BLE:
5916 if (cc_z | (cc_n ^ cc_v))
5917 pc = DIF_INSTR_LABEL(instr);
5918 break;
5919 case DIF_OP_BLEU:
5920 if (cc_c | cc_z)
5921 pc = DIF_INSTR_LABEL(instr);
5922 break;
5923 case DIF_OP_RLDSB:
5924 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5925 break;
5926 /*FALLTHROUGH*/
5927 case DIF_OP_LDSB:
5928 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5929 break;
5930 case DIF_OP_RLDSH:
5931 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5932 break;
5933 /*FALLTHROUGH*/
5934 case DIF_OP_LDSH:
5935 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5936 break;
5937 case DIF_OP_RLDSW:
5938 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5939 break;
5940 /*FALLTHROUGH*/
5941 case DIF_OP_LDSW:
5942 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5943 break;
5944 case DIF_OP_RLDUB:
5945 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5946 break;
5947 /*FALLTHROUGH*/
5948 case DIF_OP_LDUB:
5949 regs[rd] = dtrace_load8(regs[r1]);
5950 break;
5951 case DIF_OP_RLDUH:
5952 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5953 break;
5954 /*FALLTHROUGH*/
5955 case DIF_OP_LDUH:
5956 regs[rd] = dtrace_load16(regs[r1]);
5957 break;
5958 case DIF_OP_RLDUW:
5959 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5960 break;
5961 /*FALLTHROUGH*/
5962 case DIF_OP_LDUW:
5963 regs[rd] = dtrace_load32(regs[r1]);
5964 break;
5965 case DIF_OP_RLDX:
5966 if (!dtrace_canload(regs[r1], 8, mstate, vstate))
5967 break;
5968 /*FALLTHROUGH*/
5969 case DIF_OP_LDX:
5970 regs[rd] = dtrace_load64(regs[r1]);
5971 break;
5972 case DIF_OP_ULDSB:
5973 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5974 regs[rd] = (int8_t)
5975 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5976 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5977 break;
5978 case DIF_OP_ULDSH:
5979 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5980 regs[rd] = (int16_t)
5981 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5982 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5983 break;
5984 case DIF_OP_ULDSW:
5985 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5986 regs[rd] = (int32_t)
5987 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5988 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5989 break;
5990 case DIF_OP_ULDUB:
5991 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5992 regs[rd] =
5993 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5994 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5995 break;
5996 case DIF_OP_ULDUH:
5997 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5998 regs[rd] =
5999 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6000 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6001 break;
6002 case DIF_OP_ULDUW:
6003 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6004 regs[rd] =
6005 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6006 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6007 break;
6008 case DIF_OP_ULDX:
6009 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6010 regs[rd] =
6011 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
6012 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6013 break;
6014 case DIF_OP_RET:
6015 rval = regs[rd];
6016 pc = textlen;
6017 break;
6018 case DIF_OP_NOP:
6019 break;
6020 case DIF_OP_SETX:
6021 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
6022 break;
6023 case DIF_OP_SETS:
6024 regs[rd] = (uint64_t)(uintptr_t)
6025 (strtab + DIF_INSTR_STRING(instr));
6026 break;
6027 case DIF_OP_SCMP: {
6028 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
6029 uintptr_t s1 = regs[r1];
6030 uintptr_t s2 = regs[r2];
6031 size_t lim1, lim2;
6032
6033 if (s1 != 0 &&
6034 !dtrace_strcanload(s1, sz, &lim1, mstate, vstate))
6035 break;
6036 if (s2 != 0 &&
6037 !dtrace_strcanload(s2, sz, &lim2, mstate, vstate))
6038 break;
6039
6040 cc_r = dtrace_strncmp((char *)s1, (char *)s2,
6041 MIN(lim1, lim2));
6042
6043 cc_n = cc_r < 0;
6044 cc_z = cc_r == 0;
6045 cc_v = cc_c = 0;
6046 break;
6047 }
6048 case DIF_OP_LDGA:
6049 regs[rd] = dtrace_dif_variable(mstate, state,
6050 r1, regs[r2]);
6051 break;
6052 case DIF_OP_LDGS:
6053 id = DIF_INSTR_VAR(instr);
6054
6055 if (id >= DIF_VAR_OTHER_UBASE) {
6056 uintptr_t a;
6057
6058 id -= DIF_VAR_OTHER_UBASE;
6059 svar = vstate->dtvs_globals[id];
6060 ASSERT(svar != NULL);
6061 v = &svar->dtsv_var;
6062
6063 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
6064 regs[rd] = svar->dtsv_data;
6065 break;
6066 }
6067
6068 a = (uintptr_t)svar->dtsv_data;
6069
6070 if (*(uint8_t *)a == UINT8_MAX) {
6071 /*
6072 * If the 0th byte is set to UINT8_MAX
6073 * then this is to be treated as a
6074 * reference to a NULL variable.
6075 */
6076 regs[rd] = 0;
6077 } else {
6078 regs[rd] = a + sizeof (uint64_t);
6079 }
6080
6081 break;
6082 }
6083
6084 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
6085 break;
6086
6087 case DIF_OP_STGA:
6088 dtrace_dif_variable_write(mstate, state, r1, regs[r2],
6089 regs[rd]);
6090 break;
6091
6092 case DIF_OP_STGS:
6093 id = DIF_INSTR_VAR(instr);
6094
6095 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6096 id -= DIF_VAR_OTHER_UBASE;
6097
6098 VERIFY(id < vstate->dtvs_nglobals);
6099 svar = vstate->dtvs_globals[id];
6100 ASSERT(svar != NULL);
6101 v = &svar->dtsv_var;
6102
6103 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6104 uintptr_t a = (uintptr_t)svar->dtsv_data;
6105 size_t lim;
6106
6107 ASSERT(a != (uintptr_t)NULL);
6108 ASSERT(svar->dtsv_size != 0);
6109
6110 if (regs[rd] == 0) {
6111 *(uint8_t *)a = UINT8_MAX;
6112 break;
6113 } else {
6114 *(uint8_t *)a = 0;
6115 a += sizeof (uint64_t);
6116 }
6117 if (!dtrace_vcanload(
6118 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6119 &lim, mstate, vstate))
6120 break;
6121
6122 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6123 (void *)a, &v->dtdv_type, lim);
6124 break;
6125 }
6126
6127 svar->dtsv_data = regs[rd];
6128 break;
6129
6130 case DIF_OP_LDTA:
6131 /*
6132 * There are no DTrace built-in thread-local arrays at
6133 * present. This opcode is saved for future work.
6134 */
6135 *flags |= CPU_DTRACE_ILLOP;
6136 regs[rd] = 0;
6137 break;
6138
6139 case DIF_OP_LDLS:
6140 id = DIF_INSTR_VAR(instr);
6141
6142 if (id < DIF_VAR_OTHER_UBASE) {
6143 /*
6144 * For now, this has no meaning.
6145 */
6146 regs[rd] = 0;
6147 break;
6148 }
6149
6150 id -= DIF_VAR_OTHER_UBASE;
6151
6152 ASSERT(id < vstate->dtvs_nlocals);
6153 ASSERT(vstate->dtvs_locals != NULL);
6154
6155 svar = vstate->dtvs_locals[id];
6156 ASSERT(svar != NULL);
6157 v = &svar->dtsv_var;
6158
6159 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6160 uintptr_t a = (uintptr_t)svar->dtsv_data;
6161 size_t sz = v->dtdv_type.dtdt_size;
6162
6163 sz += sizeof (uint64_t);
6164 ASSERT(svar->dtsv_size == NCPU * sz);
6165 a += CPU->cpu_id * sz;
6166
6167 if (*(uint8_t *)a == UINT8_MAX) {
6168 /*
6169 * If the 0th byte is set to UINT8_MAX
6170 * then this is to be treated as a
6171 * reference to a NULL variable.
6172 */
6173 regs[rd] = 0;
6174 } else {
6175 regs[rd] = a + sizeof (uint64_t);
6176 }
6177
6178 break;
6179 }
6180
6181 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6182 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6183 regs[rd] = tmp[CPU->cpu_id];
6184 break;
6185
6186 case DIF_OP_STLS:
6187 id = DIF_INSTR_VAR(instr);
6188
6189 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6190 id -= DIF_VAR_OTHER_UBASE;
6191 VERIFY(id < vstate->dtvs_nlocals);
6192
6193 ASSERT(vstate->dtvs_locals != NULL);
6194 svar = vstate->dtvs_locals[id];
6195 ASSERT(svar != NULL);
6196 v = &svar->dtsv_var;
6197
6198 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6199 uintptr_t a = (uintptr_t)svar->dtsv_data;
6200 size_t sz = v->dtdv_type.dtdt_size;
6201 size_t lim;
6202
6203 sz += sizeof (uint64_t);
6204 ASSERT(svar->dtsv_size == NCPU * sz);
6205 a += CPU->cpu_id * sz;
6206
6207 if (regs[rd] == 0) {
6208 *(uint8_t *)a = UINT8_MAX;
6209 break;
6210 } else {
6211 *(uint8_t *)a = 0;
6212 a += sizeof (uint64_t);
6213 }
6214
6215 if (!dtrace_vcanload(
6216 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6217 &lim, mstate, vstate))
6218 break;
6219
6220 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6221 (void *)a, &v->dtdv_type, lim);
6222 break;
6223 }
6224
6225 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6226 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6227 tmp[CPU->cpu_id] = regs[rd];
6228 break;
6229
6230 case DIF_OP_LDTS: {
6231 dtrace_dynvar_t *dvar;
6232 dtrace_key_t *key;
6233
6234 id = DIF_INSTR_VAR(instr);
6235 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6236 id -= DIF_VAR_OTHER_UBASE;
6237 v = &vstate->dtvs_tlocals[id];
6238
6239 key = &tupregs[DIF_DTR_NREGS];
6240 key[0].dttk_value = (uint64_t)id;
6241 key[0].dttk_size = 0;
6242 DTRACE_TLS_THRKEY(key[1].dttk_value);
6243 key[1].dttk_size = 0;
6244
6245 dvar = dtrace_dynvar(dstate, 2, key,
6246 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
6247 mstate, vstate);
6248
6249 if (dvar == NULL) {
6250 regs[rd] = 0;
6251 break;
6252 }
6253
6254 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6255 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6256 } else {
6257 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6258 }
6259
6260 break;
6261 }
6262
6263 case DIF_OP_STTS: {
6264 dtrace_dynvar_t *dvar;
6265 dtrace_key_t *key;
6266
6267 id = DIF_INSTR_VAR(instr);
6268 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6269 id -= DIF_VAR_OTHER_UBASE;
6270 VERIFY(id < vstate->dtvs_ntlocals);
6271
6272 key = &tupregs[DIF_DTR_NREGS];
6273 key[0].dttk_value = (uint64_t)id;
6274 key[0].dttk_size = 0;
6275 DTRACE_TLS_THRKEY(key[1].dttk_value);
6276 key[1].dttk_size = 0;
6277 v = &vstate->dtvs_tlocals[id];
6278
6279 dvar = dtrace_dynvar(dstate, 2, key,
6280 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6281 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6282 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6283 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6284
6285 /*
6286 * Given that we're storing to thread-local data,
6287 * we need to flush our predicate cache.
6288 */
6289 curthread->t_predcache = DTRACE_CACHEIDNONE;
6290
6291 if (dvar == NULL)
6292 break;
6293
6294 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6295 size_t lim;
6296
6297 if (!dtrace_vcanload(
6298 (void *)(uintptr_t)regs[rd],
6299 &v->dtdv_type, &lim, mstate, vstate))
6300 break;
6301
6302 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6303 dvar->dtdv_data, &v->dtdv_type, lim);
6304 } else {
6305 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6306 }
6307
6308 break;
6309 }
6310
6311 case DIF_OP_SRA:
6312 regs[rd] = (int64_t)regs[r1] >> regs[r2];
6313 break;
6314
6315 case DIF_OP_CALL:
6316 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6317 regs, tupregs, ttop, mstate, state);
6318 break;
6319
6320 case DIF_OP_PUSHTR:
6321 if (ttop == DIF_DTR_NREGS) {
6322 *flags |= CPU_DTRACE_TUPOFLOW;
6323 break;
6324 }
6325
6326 if (r1 == DIF_TYPE_STRING) {
6327 /*
6328 * If this is a string type and the size is 0,
6329 * we'll use the system-wide default string
6330 * size. Note that we are _not_ looking at
6331 * the value of the DTRACEOPT_STRSIZE option;
6332 * had this been set, we would expect to have
6333 * a non-zero size value in the "pushtr".
6334 */
6335 tupregs[ttop].dttk_size =
6336 dtrace_strlen((char *)(uintptr_t)regs[rd],
6337 regs[r2] ? regs[r2] :
6338 dtrace_strsize_default) + 1;
6339 } else {
6340 if (regs[r2] > LONG_MAX) {
6341 *flags |= CPU_DTRACE_ILLOP;
6342 break;
6343 }
6344
6345 tupregs[ttop].dttk_size = regs[r2];
6346 }
6347
6348 tupregs[ttop++].dttk_value = regs[rd];
6349 break;
6350
6351 case DIF_OP_PUSHTV:
6352 if (ttop == DIF_DTR_NREGS) {
6353 *flags |= CPU_DTRACE_TUPOFLOW;
6354 break;
6355 }
6356
6357 tupregs[ttop].dttk_value = regs[rd];
6358 tupregs[ttop++].dttk_size = 0;
6359 break;
6360
6361 case DIF_OP_POPTS:
6362 if (ttop != 0)
6363 ttop--;
6364 break;
6365
6366 case DIF_OP_FLUSHTS:
6367 ttop = 0;
6368 break;
6369
6370 case DIF_OP_LDGAA:
6371 case DIF_OP_LDTAA: {
6372 dtrace_dynvar_t *dvar;
6373 dtrace_key_t *key = tupregs;
6374 uint_t nkeys = ttop;
6375
6376 id = DIF_INSTR_VAR(instr);
6377 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6378 id -= DIF_VAR_OTHER_UBASE;
6379
6380 key[nkeys].dttk_value = (uint64_t)id;
6381 key[nkeys++].dttk_size = 0;
6382
6383 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6384 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6385 key[nkeys++].dttk_size = 0;
6386 VERIFY(id < vstate->dtvs_ntlocals);
6387 v = &vstate->dtvs_tlocals[id];
6388 } else {
6389 VERIFY(id < vstate->dtvs_nglobals);
6390 v = &vstate->dtvs_globals[id]->dtsv_var;
6391 }
6392
6393 dvar = dtrace_dynvar(dstate, nkeys, key,
6394 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6395 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6396 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6397
6398 if (dvar == NULL) {
6399 regs[rd] = 0;
6400 break;
6401 }
6402
6403 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6404 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6405 } else {
6406 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6407 }
6408
6409 break;
6410 }
6411
6412 case DIF_OP_STGAA:
6413 case DIF_OP_STTAA: {
6414 dtrace_dynvar_t *dvar;
6415 dtrace_key_t *key = tupregs;
6416 uint_t nkeys = ttop;
6417
6418 id = DIF_INSTR_VAR(instr);
6419 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6420 id -= DIF_VAR_OTHER_UBASE;
6421
6422 key[nkeys].dttk_value = (uint64_t)id;
6423 key[nkeys++].dttk_size = 0;
6424
6425 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6426 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6427 key[nkeys++].dttk_size = 0;
6428 VERIFY(id < vstate->dtvs_ntlocals);
6429 v = &vstate->dtvs_tlocals[id];
6430 } else {
6431 VERIFY(id < vstate->dtvs_nglobals);
6432 v = &vstate->dtvs_globals[id]->dtsv_var;
6433 }
6434
6435 dvar = dtrace_dynvar(dstate, nkeys, key,
6436 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6437 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6438 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6439 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6440
6441 if (dvar == NULL)
6442 break;
6443
6444 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6445 size_t lim;
6446
6447 if (!dtrace_vcanload(
6448 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6449 &lim, mstate, vstate))
6450 break;
6451
6452 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6453 dvar->dtdv_data, &v->dtdv_type, lim);
6454 } else {
6455 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6456 }
6457
6458 break;
6459 }
6460
6461 case DIF_OP_ALLOCS: {
6462 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6463 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6464
6465 /*
6466 * Rounding up the user allocation size could have
6467 * overflowed large, bogus allocations (like -1ULL) to
6468 * 0.
6469 */
6470 if (size < regs[r1] ||
6471 !DTRACE_INSCRATCH(mstate, size)) {
6472 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6473 regs[rd] = 0;
6474 break;
6475 }
6476
6477 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6478 mstate->dtms_scratch_ptr += size;
6479 regs[rd] = ptr;
6480 break;
6481 }
6482
6483 case DIF_OP_COPYS:
6484 if (!dtrace_canstore(regs[rd], regs[r2],
6485 mstate, vstate)) {
6486 *flags |= CPU_DTRACE_BADADDR;
6487 *illval = regs[rd];
6488 break;
6489 }
6490
6491 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6492 break;
6493
6494 dtrace_bcopy((void *)(uintptr_t)regs[r1],
6495 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6496 break;
6497
6498 case DIF_OP_STB:
6499 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6500 *flags |= CPU_DTRACE_BADADDR;
6501 *illval = regs[rd];
6502 break;
6503 }
6504 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6505 break;
6506
6507 case DIF_OP_STH:
6508 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6509 *flags |= CPU_DTRACE_BADADDR;
6510 *illval = regs[rd];
6511 break;
6512 }
6513 if (regs[rd] & 1) {
6514 *flags |= CPU_DTRACE_BADALIGN;
6515 *illval = regs[rd];
6516 break;
6517 }
6518 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6519 break;
6520
6521 case DIF_OP_STW:
6522 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6523 *flags |= CPU_DTRACE_BADADDR;
6524 *illval = regs[rd];
6525 break;
6526 }
6527 if (regs[rd] & 3) {
6528 *flags |= CPU_DTRACE_BADALIGN;
6529 *illval = regs[rd];
6530 break;
6531 }
6532 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6533 break;
6534
6535 case DIF_OP_STX:
6536 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6537 *flags |= CPU_DTRACE_BADADDR;
6538 *illval = regs[rd];
6539 break;
6540 }
6541 if (regs[rd] & 7) {
6542 *flags |= CPU_DTRACE_BADALIGN;
6543 *illval = regs[rd];
6544 break;
6545 }
6546 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6547 break;
6548 }
6549 }
6550
6551 if (!(*flags & CPU_DTRACE_FAULT))
6552 return (rval);
6553
6554 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6555 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6556
6557 return (0);
6558 }
6559
6560 static void
6561 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6562 {
6563 dtrace_probe_t *probe = ecb->dte_probe;
6564 dtrace_provider_t *prov = probe->dtpr_provider;
6565 char c[DTRACE_FULLNAMELEN + 80], *str;
6566 char *msg = "dtrace: breakpoint action at probe ";
6567 char *ecbmsg = " (ecb ";
6568 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6569 uintptr_t val = (uintptr_t)ecb;
6570 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6571
6572 if (dtrace_destructive_disallow)
6573 return;
6574
6575 /*
6576 * It's impossible to be taking action on the NULL probe.
6577 */
6578 ASSERT(probe != NULL);
6579
6580 /*
6581 * This is a poor man's (destitute man's?) sprintf(): we want to
6582 * print the provider name, module name, function name and name of
6583 * the probe, along with the hex address of the ECB with the breakpoint
6584 * action -- all of which we must place in the character buffer by
6585 * hand.
6586 */
6587 while (*msg != '\0')
6588 c[i++] = *msg++;
6589
6590 for (str = prov->dtpv_name; *str != '\0'; str++)
6591 c[i++] = *str;
6592 c[i++] = ':';
6593
6594 for (str = probe->dtpr_mod; *str != '\0'; str++)
6595 c[i++] = *str;
6596 c[i++] = ':';
6597
6598 for (str = probe->dtpr_func; *str != '\0'; str++)
6599 c[i++] = *str;
6600 c[i++] = ':';
6601
6602 for (str = probe->dtpr_name; *str != '\0'; str++)
6603 c[i++] = *str;
6604
6605 while (*ecbmsg != '\0')
6606 c[i++] = *ecbmsg++;
6607
6608 while (shift >= 0) {
6609 mask = (uintptr_t)0xf << shift;
6610
6611 if (val >= ((uintptr_t)1 << shift))
6612 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6613 shift -= 4;
6614 }
6615
6616 c[i++] = ')';
6617 c[i] = '\0';
6618
6619 debug_enter(c);
6620 }
6621
6622 static void
6623 dtrace_action_panic(dtrace_ecb_t *ecb)
6624 {
6625 dtrace_probe_t *probe = ecb->dte_probe;
6626
6627 /*
6628 * It's impossible to be taking action on the NULL probe.
6629 */
6630 ASSERT(probe != NULL);
6631
6632 if (dtrace_destructive_disallow)
6633 return;
6634
6635 if (dtrace_panicked != NULL)
6636 return;
6637
6638 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6639 return;
6640
6641 /*
6642 * We won the right to panic. (We want to be sure that only one
6643 * thread calls panic() from dtrace_probe(), and that panic() is
6644 * called exactly once.)
6645 */
6646 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6647 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6648 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6649 }
6650
6651 static void
6652 dtrace_action_raise(uint64_t sig)
6653 {
6654 if (dtrace_destructive_disallow)
6655 return;
6656
6657 if (sig >= NSIG) {
6658 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6659 return;
6660 }
6661
6662 /*
6663 * raise() has a queue depth of 1 -- we ignore all subsequent
6664 * invocations of the raise() action.
6665 */
6666 if (curthread->t_dtrace_sig == 0)
6667 curthread->t_dtrace_sig = (uint8_t)sig;
6668
6669 curthread->t_sig_check = 1;
6670 aston(curthread);
6671 }
6672
6673 static void
6674 dtrace_action_stop(void)
6675 {
6676 if (dtrace_destructive_disallow)
6677 return;
6678
6679 if (!curthread->t_dtrace_stop) {
6680 curthread->t_dtrace_stop = 1;
6681 curthread->t_sig_check = 1;
6682 aston(curthread);
6683 }
6684 }
6685
6686 static void
6687 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
6688 {
6689 hrtime_t now;
6690 volatile uint16_t *flags;
6691 cpu_t *cpu = CPU;
6692
6693 if (dtrace_destructive_disallow)
6694 return;
6695
6696 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
6697
6698 now = dtrace_gethrtime();
6699
6700 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
6701 /*
6702 * We need to advance the mark to the current time.
6703 */
6704 cpu->cpu_dtrace_chillmark = now;
6705 cpu->cpu_dtrace_chilled = 0;
6706 }
6707
6708 /*
6709 * Now check to see if the requested chill time would take us over
6710 * the maximum amount of time allowed in the chill interval. (Or
6711 * worse, if the calculation itself induces overflow.)
6712 */
6713 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
6714 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
6715 *flags |= CPU_DTRACE_ILLOP;
6716 return;
6717 }
6718
6719 while (dtrace_gethrtime() - now < val)
6720 continue;
6721
6722 /*
6723 * Normally, we assure that the value of the variable "timestamp" does
6724 * not change within an ECB. The presence of chill() represents an
6725 * exception to this rule, however.
6726 */
6727 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
6728 cpu->cpu_dtrace_chilled += val;
6729 }
6730
6731 static void
6732 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
6733 uint64_t *buf, uint64_t arg)
6734 {
6735 int nframes = DTRACE_USTACK_NFRAMES(arg);
6736 int strsize = DTRACE_USTACK_STRSIZE(arg);
6737 uint64_t *pcs = &buf[1], *fps;
6738 char *str = (char *)&pcs[nframes];
6739 int size, offs = 0, i, j;
6740 size_t rem;
6741 uintptr_t old = mstate->dtms_scratch_ptr, saved;
6742 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6743 char *sym;
6744
6745 /*
6746 * Should be taking a faster path if string space has not been
6747 * allocated.
6748 */
6749 ASSERT(strsize != 0);
6750
6751 /*
6752 * We will first allocate some temporary space for the frame pointers.
6753 */
6754 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6755 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
6756 (nframes * sizeof (uint64_t));
6757
6758 if (!DTRACE_INSCRATCH(mstate, size)) {
6759 /*
6760 * Not enough room for our frame pointers -- need to indicate
6761 * that we ran out of scratch space.
6762 */
6763 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6764 return;
6765 }
6766
6767 mstate->dtms_scratch_ptr += size;
6768 saved = mstate->dtms_scratch_ptr;
6769
6770 /*
6771 * Now get a stack with both program counters and frame pointers.
6772 */
6773 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6774 dtrace_getufpstack(buf, fps, nframes + 1);
6775 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6776
6777 /*
6778 * If that faulted, we're cooked.
6779 */
6780 if (*flags & CPU_DTRACE_FAULT)
6781 goto out;
6782
6783 /*
6784 * Now we want to walk up the stack, calling the USTACK helper. For
6785 * each iteration, we restore the scratch pointer.
6786 */
6787 for (i = 0; i < nframes; i++) {
6788 mstate->dtms_scratch_ptr = saved;
6789
6790 if (offs >= strsize)
6791 break;
6792
6793 sym = (char *)(uintptr_t)dtrace_helper(
6794 DTRACE_HELPER_ACTION_USTACK,
6795 mstate, state, pcs[i], fps[i]);
6796
6797 /*
6798 * If we faulted while running the helper, we're going to
6799 * clear the fault and null out the corresponding string.
6800 */
6801 if (*flags & CPU_DTRACE_FAULT) {
6802 *flags &= ~CPU_DTRACE_FAULT;
6803 str[offs++] = '\0';
6804 continue;
6805 }
6806
6807 if (sym == NULL) {
6808 str[offs++] = '\0';
6809 continue;
6810 }
6811
6812 if (!dtrace_strcanload((uintptr_t)sym, strsize, &rem, mstate,
6813 &(state->dts_vstate))) {
6814 str[offs++] = '\0';
6815 continue;
6816 }
6817
6818 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6819
6820 /*
6821 * Now copy in the string that the helper returned to us.
6822 */
6823 for (j = 0; offs + j < strsize && j < rem; j++) {
6824 if ((str[offs + j] = sym[j]) == '\0')
6825 break;
6826 }
6827
6828 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6829
6830 offs += j + 1;
6831 }
6832
6833 if (offs >= strsize) {
6834 /*
6835 * If we didn't have room for all of the strings, we don't
6836 * abort processing -- this needn't be a fatal error -- but we
6837 * still want to increment a counter (dts_stkstroverflows) to
6838 * allow this condition to be warned about. (If this is from
6839 * a jstack() action, it is easily tuned via jstackstrsize.)
6840 */
6841 dtrace_error(&state->dts_stkstroverflows);
6842 }
6843
6844 while (offs < strsize)
6845 str[offs++] = '\0';
6846
6847 out:
6848 mstate->dtms_scratch_ptr = old;
6849 }
6850
6851 static void
6852 dtrace_store_by_ref(dtrace_difo_t *dp, caddr_t tomax, size_t size,
6853 size_t *valoffsp, uint64_t *valp, uint64_t end, int intuple, int dtkind)
6854 {
6855 volatile uint16_t *flags;
6856 uint64_t val = *valp;
6857 size_t valoffs = *valoffsp;
6858
6859 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6860 ASSERT(dtkind == DIF_TF_BYREF || dtkind == DIF_TF_BYUREF);
6861
6862 /*
6863 * If this is a string, we're going to only load until we find the zero
6864 * byte -- after which we'll store zero bytes.
6865 */
6866 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
6867 char c = '\0' + 1;
6868 size_t s;
6869
6870 for (s = 0; s < size; s++) {
6871 if (c != '\0' && dtkind == DIF_TF_BYREF) {
6872 c = dtrace_load8(val++);
6873 } else if (c != '\0' && dtkind == DIF_TF_BYUREF) {
6874 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6875 c = dtrace_fuword8((void *)(uintptr_t)val++);
6876 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6877 if (*flags & CPU_DTRACE_FAULT)
6878 break;
6879 }
6880
6881 DTRACE_STORE(uint8_t, tomax, valoffs++, c);
6882
6883 if (c == '\0' && intuple)
6884 break;
6885 }
6886 } else {
6887 uint8_t c;
6888 while (valoffs < end) {
6889 if (dtkind == DIF_TF_BYREF) {
6890 c = dtrace_load8(val++);
6891 } else if (dtkind == DIF_TF_BYUREF) {
6892 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6893 c = dtrace_fuword8((void *)(uintptr_t)val++);
6894 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6895 if (*flags & CPU_DTRACE_FAULT)
6896 break;
6897 }
6898
6899 DTRACE_STORE(uint8_t, tomax,
6900 valoffs++, c);
6901 }
6902 }
6903
6904 *valp = val;
6905 *valoffsp = valoffs;
6906 }
6907
6908 /*
6909 * If you're looking for the epicenter of DTrace, you just found it. This
6910 * is the function called by the provider to fire a probe -- from which all
6911 * subsequent probe-context DTrace activity emanates.
6912 */
6913 void
6914 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6915 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6916 {
6917 processorid_t cpuid;
6918 dtrace_icookie_t cookie;
6919 dtrace_probe_t *probe;
6920 dtrace_mstate_t mstate;
6921 dtrace_ecb_t *ecb;
6922 dtrace_action_t *act;
6923 intptr_t offs;
6924 size_t size;
6925 int vtime, onintr;
6926 volatile uint16_t *flags;
6927 hrtime_t now, end;
6928
6929 /*
6930 * Kick out immediately if this CPU is still being born (in which case
6931 * curthread will be set to -1) or the current thread can't allow
6932 * probes in its current context.
6933 */
6934 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6935 return;
6936
6937 cookie = dtrace_interrupt_disable();
6938
6939 /*
6940 * Also refuse to process any probe firings that might happen on a
6941 * disabled CPU.
6942 */
6943 if (CPU->cpu_flags & CPU_DISABLED) {
6944 dtrace_interrupt_enable(cookie);
6945 return;
6946 }
6947
6948 probe = dtrace_probes[id - 1];
6949 cpuid = CPU->cpu_id;
6950 onintr = CPU_ON_INTR(CPU);
6951
6952 CPU->cpu_dtrace_probes++;
6953
6954 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6955 probe->dtpr_predcache == curthread->t_predcache) {
6956 /*
6957 * We have hit in the predicate cache; we know that
6958 * this predicate would evaluate to be false.
6959 */
6960 dtrace_interrupt_enable(cookie);
6961 return;
6962 }
6963
6964 if (panic_quiesce) {
6965 /*
6966 * We don't trace anything if we're panicking.
6967 */
6968 dtrace_interrupt_enable(cookie);
6969 return;
6970 }
6971
6972 now = mstate.dtms_timestamp = dtrace_gethrtime();
6973 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6974 vtime = dtrace_vtime_references != 0;
6975
6976 if (vtime && curthread->t_dtrace_start)
6977 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6978
6979 mstate.dtms_difo = NULL;
6980 mstate.dtms_probe = probe;
6981 mstate.dtms_strtok = 0;
6982 mstate.dtms_arg[0] = arg0;
6983 mstate.dtms_arg[1] = arg1;
6984 mstate.dtms_arg[2] = arg2;
6985 mstate.dtms_arg[3] = arg3;
6986 mstate.dtms_arg[4] = arg4;
6987
6988 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6989
6990 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6991 dtrace_predicate_t *pred = ecb->dte_predicate;
6992 dtrace_state_t *state = ecb->dte_state;
6993 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6994 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6995 dtrace_vstate_t *vstate = &state->dts_vstate;
6996 dtrace_provider_t *prov = probe->dtpr_provider;
6997 uint64_t tracememsize = 0;
6998 int committed = 0;
6999 caddr_t tomax;
7000
7001 /*
7002 * A little subtlety with the following (seemingly innocuous)
7003 * declaration of the automatic 'val': by looking at the
7004 * code, you might think that it could be declared in the
7005 * action processing loop, below. (That is, it's only used in
7006 * the action processing loop.) However, it must be declared
7007 * out of that scope because in the case of DIF expression
7008 * arguments to aggregating actions, one iteration of the
7009 * action loop will use the last iteration's value.
7010 */
7011 #ifdef lint
7012 uint64_t val = 0;
7013 #else
7014 uint64_t val;
7015 #endif
7016
7017 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
7018 mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
7019 mstate.dtms_getf = NULL;
7020
7021 *flags &= ~CPU_DTRACE_ERROR;
7022
7023 if (prov == dtrace_provider) {
7024 /*
7025 * If dtrace itself is the provider of this probe,
7026 * we're only going to continue processing the ECB if
7027 * arg0 (the dtrace_state_t) is equal to the ECB's
7028 * creating state. (This prevents disjoint consumers
7029 * from seeing one another's metaprobes.)
7030 */
7031 if (arg0 != (uint64_t)(uintptr_t)state)
7032 continue;
7033 }
7034
7035 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
7036 /*
7037 * We're not currently active. If our provider isn't
7038 * the dtrace pseudo provider, we're not interested.
7039 */
7040 if (prov != dtrace_provider)
7041 continue;
7042
7043 /*
7044 * Now we must further check if we are in the BEGIN
7045 * probe. If we are, we will only continue processing
7046 * if we're still in WARMUP -- if one BEGIN enabling
7047 * has invoked the exit() action, we don't want to
7048 * evaluate subsequent BEGIN enablings.
7049 */
7050 if (probe->dtpr_id == dtrace_probeid_begin &&
7051 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
7052 ASSERT(state->dts_activity ==
7053 DTRACE_ACTIVITY_DRAINING);
7054 continue;
7055 }
7056 }
7057
7058 if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
7059 continue;
7060
7061 if (now - state->dts_alive > dtrace_deadman_timeout) {
7062 /*
7063 * We seem to be dead. Unless we (a) have kernel
7064 * destructive permissions (b) have explicitly enabled
7065 * destructive actions and (c) destructive actions have
7066 * not been disabled, we're going to transition into
7067 * the KILLED state, from which no further processing
7068 * on this state will be performed.
7069 */
7070 if (!dtrace_priv_kernel_destructive(state) ||
7071 !state->dts_cred.dcr_destructive ||
7072 dtrace_destructive_disallow) {
7073 void *activity = &state->dts_activity;
7074 dtrace_activity_t current;
7075
7076 do {
7077 current = state->dts_activity;
7078 } while (dtrace_cas32(activity, current,
7079 DTRACE_ACTIVITY_KILLED) != current);
7080
7081 continue;
7082 }
7083 }
7084
7085 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
7086 ecb->dte_alignment, state, &mstate)) < 0)
7087 continue;
7088
7089 tomax = buf->dtb_tomax;
7090 ASSERT(tomax != NULL);
7091
7092 if (ecb->dte_size != 0) {
7093 dtrace_rechdr_t dtrh;
7094 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
7095 mstate.dtms_timestamp = dtrace_gethrtime();
7096 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7097 }
7098 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
7099 dtrh.dtrh_epid = ecb->dte_epid;
7100 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
7101 mstate.dtms_timestamp);
7102 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
7103 }
7104
7105 mstate.dtms_epid = ecb->dte_epid;
7106 mstate.dtms_present |= DTRACE_MSTATE_EPID;
7107
7108 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
7109 mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
7110
7111 if (pred != NULL) {
7112 dtrace_difo_t *dp = pred->dtp_difo;
7113 int rval;
7114
7115 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
7116
7117 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
7118 dtrace_cacheid_t cid = probe->dtpr_predcache;
7119
7120 if (cid != DTRACE_CACHEIDNONE && !onintr) {
7121 /*
7122 * Update the predicate cache...
7123 */
7124 ASSERT(cid == pred->dtp_cacheid);
7125 curthread->t_predcache = cid;
7126 }
7127
7128 continue;
7129 }
7130 }
7131
7132 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
7133 act != NULL; act = act->dta_next) {
7134 size_t valoffs;
7135 dtrace_difo_t *dp;
7136 dtrace_recdesc_t *rec = &act->dta_rec;
7137
7138 size = rec->dtrd_size;
7139 valoffs = offs + rec->dtrd_offset;
7140
7141 if (DTRACEACT_ISAGG(act->dta_kind)) {
7142 uint64_t v = 0xbad;
7143 dtrace_aggregation_t *agg;
7144
7145 agg = (dtrace_aggregation_t *)act;
7146
7147 if ((dp = act->dta_difo) != NULL)
7148 v = dtrace_dif_emulate(dp,
7149 &mstate, vstate, state);
7150
7151 if (*flags & CPU_DTRACE_ERROR)
7152 continue;
7153
7154 /*
7155 * Note that we always pass the expression
7156 * value from the previous iteration of the
7157 * action loop. This value will only be used
7158 * if there is an expression argument to the
7159 * aggregating action, denoted by the
7160 * dtag_hasarg field.
7161 */
7162 dtrace_aggregate(agg, buf,
7163 offs, aggbuf, v, val);
7164 continue;
7165 }
7166
7167 switch (act->dta_kind) {
7168 case DTRACEACT_STOP:
7169 if (dtrace_priv_proc_destructive(state,
7170 &mstate))
7171 dtrace_action_stop();
7172 continue;
7173
7174 case DTRACEACT_BREAKPOINT:
7175 if (dtrace_priv_kernel_destructive(state))
7176 dtrace_action_breakpoint(ecb);
7177 continue;
7178
7179 case DTRACEACT_PANIC:
7180 if (dtrace_priv_kernel_destructive(state))
7181 dtrace_action_panic(ecb);
7182 continue;
7183
7184 case DTRACEACT_STACK:
7185 if (!dtrace_priv_kernel(state))
7186 continue;
7187
7188 dtrace_getpcstack((pc_t *)(tomax + valoffs),
7189 size / sizeof (pc_t), probe->dtpr_aframes,
7190 DTRACE_ANCHORED(probe) ? NULL :
7191 (uint32_t *)arg0);
7192
7193 continue;
7194
7195 case DTRACEACT_JSTACK:
7196 case DTRACEACT_USTACK:
7197 if (!dtrace_priv_proc(state, &mstate))
7198 continue;
7199
7200 /*
7201 * See comment in DIF_VAR_PID.
7202 */
7203 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
7204 CPU_ON_INTR(CPU)) {
7205 int depth = DTRACE_USTACK_NFRAMES(
7206 rec->dtrd_arg) + 1;
7207
7208 dtrace_bzero((void *)(tomax + valoffs),
7209 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
7210 + depth * sizeof (uint64_t));
7211
7212 continue;
7213 }
7214
7215 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
7216 curproc->p_dtrace_helpers != NULL) {
7217 /*
7218 * This is the slow path -- we have
7219 * allocated string space, and we're
7220 * getting the stack of a process that
7221 * has helpers. Call into a separate
7222 * routine to perform this processing.
7223 */
7224 dtrace_action_ustack(&mstate, state,
7225 (uint64_t *)(tomax + valoffs),
7226 rec->dtrd_arg);
7227 continue;
7228 }
7229
7230 /*
7231 * Clear the string space, since there's no
7232 * helper to do it for us.
7233 */
7234 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
7235 int depth = DTRACE_USTACK_NFRAMES(
7236 rec->dtrd_arg);
7237 size_t strsize = DTRACE_USTACK_STRSIZE(
7238 rec->dtrd_arg);
7239 uint64_t *buf = (uint64_t *)(tomax +
7240 valoffs);
7241 void *strspace = &buf[depth + 1];
7242
7243 dtrace_bzero(strspace,
7244 MIN(depth, strsize));
7245 }
7246
7247 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7248 dtrace_getupcstack((uint64_t *)
7249 (tomax + valoffs),
7250 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
7251 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7252 continue;
7253
7254 default:
7255 break;
7256 }
7257
7258 dp = act->dta_difo;
7259 ASSERT(dp != NULL);
7260
7261 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
7262
7263 if (*flags & CPU_DTRACE_ERROR)
7264 continue;
7265
7266 switch (act->dta_kind) {
7267 case DTRACEACT_SPECULATE: {
7268 dtrace_rechdr_t *dtrh;
7269
7270 ASSERT(buf == &state->dts_buffer[cpuid]);
7271 buf = dtrace_speculation_buffer(state,
7272 cpuid, val);
7273
7274 if (buf == NULL) {
7275 *flags |= CPU_DTRACE_DROP;
7276 continue;
7277 }
7278
7279 offs = dtrace_buffer_reserve(buf,
7280 ecb->dte_needed, ecb->dte_alignment,
7281 state, NULL);
7282
7283 if (offs < 0) {
7284 *flags |= CPU_DTRACE_DROP;
7285 continue;
7286 }
7287
7288 tomax = buf->dtb_tomax;
7289 ASSERT(tomax != NULL);
7290
7291 if (ecb->dte_size == 0)
7292 continue;
7293
7294 ASSERT3U(ecb->dte_size, >=,
7295 sizeof (dtrace_rechdr_t));
7296 dtrh = ((void *)(tomax + offs));
7297 dtrh->dtrh_epid = ecb->dte_epid;
7298 /*
7299 * When the speculation is committed, all of
7300 * the records in the speculative buffer will
7301 * have their timestamps set to the commit
7302 * time. Until then, it is set to a sentinel
7303 * value, for debugability.
7304 */
7305 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
7306 continue;
7307 }
7308
7309 case DTRACEACT_CHILL:
7310 if (dtrace_priv_kernel_destructive(state))
7311 dtrace_action_chill(&mstate, val);
7312 continue;
7313
7314 case DTRACEACT_RAISE:
7315 if (dtrace_priv_proc_destructive(state,
7316 &mstate))
7317 dtrace_action_raise(val);
7318 continue;
7319
7320 case DTRACEACT_COMMIT:
7321 ASSERT(!committed);
7322
7323 /*
7324 * We need to commit our buffer state.
7325 */
7326 if (ecb->dte_size)
7327 buf->dtb_offset = offs + ecb->dte_size;
7328 buf = &state->dts_buffer[cpuid];
7329 dtrace_speculation_commit(state, cpuid, val);
7330 committed = 1;
7331 continue;
7332
7333 case DTRACEACT_DISCARD:
7334 dtrace_speculation_discard(state, cpuid, val);
7335 continue;
7336
7337 case DTRACEACT_DIFEXPR:
7338 case DTRACEACT_LIBACT:
7339 case DTRACEACT_PRINTF:
7340 case DTRACEACT_PRINTA:
7341 case DTRACEACT_SYSTEM:
7342 case DTRACEACT_FREOPEN:
7343 case DTRACEACT_TRACEMEM:
7344 break;
7345
7346 case DTRACEACT_TRACEMEM_DYNSIZE:
7347 tracememsize = val;
7348 break;
7349
7350 case DTRACEACT_SYM:
7351 case DTRACEACT_MOD:
7352 if (!dtrace_priv_kernel(state))
7353 continue;
7354 break;
7355
7356 case DTRACEACT_USYM:
7357 case DTRACEACT_UMOD:
7358 case DTRACEACT_UADDR: {
7359 struct pid *pid = curthread->t_procp->p_pidp;
7360
7361 if (!dtrace_priv_proc(state, &mstate))
7362 continue;
7363
7364 DTRACE_STORE(uint64_t, tomax,
7365 valoffs, (uint64_t)pid->pid_id);
7366 DTRACE_STORE(uint64_t, tomax,
7367 valoffs + sizeof (uint64_t), val);
7368
7369 continue;
7370 }
7371
7372 case DTRACEACT_EXIT: {
7373 /*
7374 * For the exit action, we are going to attempt
7375 * to atomically set our activity to be
7376 * draining. If this fails (either because
7377 * another CPU has beat us to the exit action,
7378 * or because our current activity is something
7379 * other than ACTIVE or WARMUP), we will
7380 * continue. This assures that the exit action
7381 * can be successfully recorded at most once
7382 * when we're in the ACTIVE state. If we're
7383 * encountering the exit() action while in
7384 * COOLDOWN, however, we want to honor the new
7385 * status code. (We know that we're the only
7386 * thread in COOLDOWN, so there is no race.)
7387 */
7388 void *activity = &state->dts_activity;
7389 dtrace_activity_t current = state->dts_activity;
7390
7391 if (current == DTRACE_ACTIVITY_COOLDOWN)
7392 break;
7393
7394 if (current != DTRACE_ACTIVITY_WARMUP)
7395 current = DTRACE_ACTIVITY_ACTIVE;
7396
7397 if (dtrace_cas32(activity, current,
7398 DTRACE_ACTIVITY_DRAINING) != current) {
7399 *flags |= CPU_DTRACE_DROP;
7400 continue;
7401 }
7402
7403 break;
7404 }
7405
7406 default:
7407 ASSERT(0);
7408 }
7409
7410 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ||
7411 dp->dtdo_rtype.dtdt_flags & DIF_TF_BYUREF) {
7412 uintptr_t end = valoffs + size;
7413
7414 if (tracememsize != 0 &&
7415 valoffs + tracememsize < end) {
7416 end = valoffs + tracememsize;
7417 tracememsize = 0;
7418 }
7419
7420 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF &&
7421 !dtrace_vcanload((void *)(uintptr_t)val,
7422 &dp->dtdo_rtype, NULL, &mstate, vstate))
7423 continue;
7424
7425 dtrace_store_by_ref(dp, tomax, size, &valoffs,
7426 &val, end, act->dta_intuple,
7427 dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ?
7428 DIF_TF_BYREF: DIF_TF_BYUREF);
7429 continue;
7430 }
7431
7432 switch (size) {
7433 case 0:
7434 break;
7435
7436 case sizeof (uint8_t):
7437 DTRACE_STORE(uint8_t, tomax, valoffs, val);
7438 break;
7439 case sizeof (uint16_t):
7440 DTRACE_STORE(uint16_t, tomax, valoffs, val);
7441 break;
7442 case sizeof (uint32_t):
7443 DTRACE_STORE(uint32_t, tomax, valoffs, val);
7444 break;
7445 case sizeof (uint64_t):
7446 DTRACE_STORE(uint64_t, tomax, valoffs, val);
7447 break;
7448 default:
7449 /*
7450 * Any other size should have been returned by
7451 * reference, not by value.
7452 */
7453 ASSERT(0);
7454 break;
7455 }
7456 }
7457
7458 if (*flags & CPU_DTRACE_DROP)
7459 continue;
7460
7461 if (*flags & CPU_DTRACE_FAULT) {
7462 int ndx;
7463 dtrace_action_t *err;
7464
7465 buf->dtb_errors++;
7466
7467 if (probe->dtpr_id == dtrace_probeid_error) {
7468 /*
7469 * There's nothing we can do -- we had an
7470 * error on the error probe. We bump an
7471 * error counter to at least indicate that
7472 * this condition happened.
7473 */
7474 dtrace_error(&state->dts_dblerrors);
7475 continue;
7476 }
7477
7478 if (vtime) {
7479 /*
7480 * Before recursing on dtrace_probe(), we
7481 * need to explicitly clear out our start
7482 * time to prevent it from being accumulated
7483 * into t_dtrace_vtime.
7484 */
7485 curthread->t_dtrace_start = 0;
7486 }
7487
7488 /*
7489 * Iterate over the actions to figure out which action
7490 * we were processing when we experienced the error.
7491 * Note that act points _past_ the faulting action; if
7492 * act is ecb->dte_action, the fault was in the
7493 * predicate, if it's ecb->dte_action->dta_next it's
7494 * in action #1, and so on.
7495 */
7496 for (err = ecb->dte_action, ndx = 0;
7497 err != act; err = err->dta_next, ndx++)
7498 continue;
7499
7500 dtrace_probe_error(state, ecb->dte_epid, ndx,
7501 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7502 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7503 cpu_core[cpuid].cpuc_dtrace_illval);
7504
7505 continue;
7506 }
7507
7508 if (!committed)
7509 buf->dtb_offset = offs + ecb->dte_size;
7510 }
7511
7512 end = dtrace_gethrtime();
7513 if (vtime)
7514 curthread->t_dtrace_start = end;
7515
7516 CPU->cpu_dtrace_nsec += end - now;
7517
7518 dtrace_interrupt_enable(cookie);
7519 }
7520
7521 /*
7522 * DTrace Probe Hashing Functions
7523 *
7524 * The functions in this section (and indeed, the functions in remaining
7525 * sections) are not _called_ from probe context. (Any exceptions to this are
7526 * marked with a "Note:".) Rather, they are called from elsewhere in the
7527 * DTrace framework to look-up probes in, add probes to and remove probes from
7528 * the DTrace probe hashes. (Each probe is hashed by each element of the
7529 * probe tuple -- allowing for fast lookups, regardless of what was
7530 * specified.)
7531 */
7532 static uint_t
7533 dtrace_hash_str(char *p)
7534 {
7535 unsigned int g;
7536 uint_t hval = 0;
7537
7538 while (*p) {
7539 hval = (hval << 4) + *p++;
7540 if ((g = (hval & 0xf0000000)) != 0)
7541 hval ^= g >> 24;
7542 hval &= ~g;
7543 }
7544 return (hval);
7545 }
7546
7547 static dtrace_hash_t *
7548 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
7549 {
7550 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
7551
7552 hash->dth_stroffs = stroffs;
7553 hash->dth_nextoffs = nextoffs;
7554 hash->dth_prevoffs = prevoffs;
7555
7556 hash->dth_size = 1;
7557 hash->dth_mask = hash->dth_size - 1;
7558
7559 hash->dth_tab = kmem_zalloc(hash->dth_size *
7560 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
7561
7562 return (hash);
7563 }
7564
7565 static void
7566 dtrace_hash_destroy(dtrace_hash_t *hash)
7567 {
7568 #ifdef DEBUG
7569 int i;
7570
7571 for (i = 0; i < hash->dth_size; i++)
7572 ASSERT(hash->dth_tab[i] == NULL);
7573 #endif
7574
7575 kmem_free(hash->dth_tab,
7576 hash->dth_size * sizeof (dtrace_hashbucket_t *));
7577 kmem_free(hash, sizeof (dtrace_hash_t));
7578 }
7579
7580 static void
7581 dtrace_hash_resize(dtrace_hash_t *hash)
7582 {
7583 int size = hash->dth_size, i, ndx;
7584 int new_size = hash->dth_size << 1;
7585 int new_mask = new_size - 1;
7586 dtrace_hashbucket_t **new_tab, *bucket, *next;
7587
7588 ASSERT((new_size & new_mask) == 0);
7589
7590 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
7591
7592 for (i = 0; i < size; i++) {
7593 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
7594 dtrace_probe_t *probe = bucket->dthb_chain;
7595
7596 ASSERT(probe != NULL);
7597 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
7598
7599 next = bucket->dthb_next;
7600 bucket->dthb_next = new_tab[ndx];
7601 new_tab[ndx] = bucket;
7602 }
7603 }
7604
7605 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
7606 hash->dth_tab = new_tab;
7607 hash->dth_size = new_size;
7608 hash->dth_mask = new_mask;
7609 }
7610
7611 static void
7612 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
7613 {
7614 int hashval = DTRACE_HASHSTR(hash, new);
7615 int ndx = hashval & hash->dth_mask;
7616 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7617 dtrace_probe_t **nextp, **prevp;
7618
7619 for (; bucket != NULL; bucket = bucket->dthb_next) {
7620 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
7621 goto add;
7622 }
7623
7624 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
7625 dtrace_hash_resize(hash);
7626 dtrace_hash_add(hash, new);
7627 return;
7628 }
7629
7630 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
7631 bucket->dthb_next = hash->dth_tab[ndx];
7632 hash->dth_tab[ndx] = bucket;
7633 hash->dth_nbuckets++;
7634
7635 add:
7636 nextp = DTRACE_HASHNEXT(hash, new);
7637 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
7638 *nextp = bucket->dthb_chain;
7639
7640 if (bucket->dthb_chain != NULL) {
7641 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
7642 ASSERT(*prevp == NULL);
7643 *prevp = new;
7644 }
7645
7646 bucket->dthb_chain = new;
7647 bucket->dthb_len++;
7648 }
7649
7650 static dtrace_probe_t *
7651 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
7652 {
7653 int hashval = DTRACE_HASHSTR(hash, template);
7654 int ndx = hashval & hash->dth_mask;
7655 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7656
7657 for (; bucket != NULL; bucket = bucket->dthb_next) {
7658 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7659 return (bucket->dthb_chain);
7660 }
7661
7662 return (NULL);
7663 }
7664
7665 static int
7666 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
7667 {
7668 int hashval = DTRACE_HASHSTR(hash, template);
7669 int ndx = hashval & hash->dth_mask;
7670 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7671
7672 for (; bucket != NULL; bucket = bucket->dthb_next) {
7673 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7674 return (bucket->dthb_len);
7675 }
7676
7677 return (0);
7678 }
7679
7680 static void
7681 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7682 {
7683 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7684 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7685
7686 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7687 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7688
7689 /*
7690 * Find the bucket that we're removing this probe from.
7691 */
7692 for (; bucket != NULL; bucket = bucket->dthb_next) {
7693 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7694 break;
7695 }
7696
7697 ASSERT(bucket != NULL);
7698
7699 if (*prevp == NULL) {
7700 if (*nextp == NULL) {
7701 /*
7702 * The removed probe was the only probe on this
7703 * bucket; we need to remove the bucket.
7704 */
7705 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7706
7707 ASSERT(bucket->dthb_chain == probe);
7708 ASSERT(b != NULL);
7709
7710 if (b == bucket) {
7711 hash->dth_tab[ndx] = bucket->dthb_next;
7712 } else {
7713 while (b->dthb_next != bucket)
7714 b = b->dthb_next;
7715 b->dthb_next = bucket->dthb_next;
7716 }
7717
7718 ASSERT(hash->dth_nbuckets > 0);
7719 hash->dth_nbuckets--;
7720 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7721 return;
7722 }
7723
7724 bucket->dthb_chain = *nextp;
7725 } else {
7726 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7727 }
7728
7729 if (*nextp != NULL)
7730 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7731 }
7732
7733 /*
7734 * DTrace Utility Functions
7735 *
7736 * These are random utility functions that are _not_ called from probe context.
7737 */
7738 static int
7739 dtrace_badattr(const dtrace_attribute_t *a)
7740 {
7741 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7742 a->dtat_data > DTRACE_STABILITY_MAX ||
7743 a->dtat_class > DTRACE_CLASS_MAX);
7744 }
7745
7746 /*
7747 * Return a duplicate copy of a string. If the specified string is NULL,
7748 * this function returns a zero-length string.
7749 */
7750 static char *
7751 dtrace_strdup(const char *str)
7752 {
7753 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7754
7755 if (str != NULL)
7756 (void) strcpy(new, str);
7757
7758 return (new);
7759 }
7760
7761 #define DTRACE_ISALPHA(c) \
7762 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7763
7764 static int
7765 dtrace_badname(const char *s)
7766 {
7767 char c;
7768
7769 if (s == NULL || (c = *s++) == '\0')
7770 return (0);
7771
7772 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7773 return (1);
7774
7775 while ((c = *s++) != '\0') {
7776 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7777 c != '-' && c != '_' && c != '.' && c != '`')
7778 return (1);
7779 }
7780
7781 return (0);
7782 }
7783
7784 static void
7785 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7786 {
7787 uint32_t priv;
7788
7789 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7790 /*
7791 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7792 */
7793 priv = DTRACE_PRIV_ALL;
7794 } else {
7795 *uidp = crgetuid(cr);
7796 *zoneidp = crgetzonedid(cr);
7797
7798 priv = 0;
7799 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7800 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7801 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7802 priv |= DTRACE_PRIV_USER;
7803 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7804 priv |= DTRACE_PRIV_PROC;
7805 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7806 priv |= DTRACE_PRIV_OWNER;
7807 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7808 priv |= DTRACE_PRIV_ZONEOWNER;
7809 }
7810
7811 *privp = priv;
7812 }
7813
7814 #ifdef DTRACE_ERRDEBUG
7815 static void
7816 dtrace_errdebug(const char *str)
7817 {
7818 int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
7819 int occupied = 0;
7820
7821 mutex_enter(&dtrace_errlock);
7822 dtrace_errlast = str;
7823 dtrace_errthread = curthread;
7824
7825 while (occupied++ < DTRACE_ERRHASHSZ) {
7826 if (dtrace_errhash[hval].dter_msg == str) {
7827 dtrace_errhash[hval].dter_count++;
7828 goto out;
7829 }
7830
7831 if (dtrace_errhash[hval].dter_msg != NULL) {
7832 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7833 continue;
7834 }
7835
7836 dtrace_errhash[hval].dter_msg = str;
7837 dtrace_errhash[hval].dter_count = 1;
7838 goto out;
7839 }
7840
7841 panic("dtrace: undersized error hash");
7842 out:
7843 mutex_exit(&dtrace_errlock);
7844 }
7845 #endif
7846
7847 /*
7848 * DTrace Matching Functions
7849 *
7850 * These functions are used to match groups of probes, given some elements of
7851 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7852 */
7853 static int
7854 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7855 zoneid_t zoneid)
7856 {
7857 if (priv != DTRACE_PRIV_ALL) {
7858 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7859 uint32_t match = priv & ppriv;
7860
7861 /*
7862 * No PRIV_DTRACE_* privileges...
7863 */
7864 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7865 DTRACE_PRIV_KERNEL)) == 0)
7866 return (0);
7867
7868 /*
7869 * No matching bits, but there were bits to match...
7870 */
7871 if (match == 0 && ppriv != 0)
7872 return (0);
7873
7874 /*
7875 * Need to have permissions to the process, but don't...
7876 */
7877 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7878 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7879 return (0);
7880 }
7881
7882 /*
7883 * Need to be in the same zone unless we possess the
7884 * privilege to examine all zones.
7885 */
7886 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7887 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7888 return (0);
7889 }
7890 }
7891
7892 return (1);
7893 }
7894
7895 /*
7896 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7897 * consists of input pattern strings and an ops-vector to evaluate them.
7898 * This function returns >0 for match, 0 for no match, and <0 for error.
7899 */
7900 static int
7901 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7902 uint32_t priv, uid_t uid, zoneid_t zoneid)
7903 {
7904 dtrace_provider_t *pvp = prp->dtpr_provider;
7905 int rv;
7906
7907 if (pvp->dtpv_defunct)
7908 return (0);
7909
7910 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7911 return (rv);
7912
7913 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7914 return (rv);
7915
7916 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7917 return (rv);
7918
7919 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7920 return (rv);
7921
7922 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7923 return (0);
7924
7925 return (rv);
7926 }
7927
7928 /*
7929 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7930 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7931 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7932 * In addition, all of the recursion cases except for '*' matching have been
7933 * unwound. For '*', we still implement recursive evaluation, but a depth
7934 * counter is maintained and matching is aborted if we recurse too deep.
7935 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7936 */
7937 static int
7938 dtrace_match_glob(const char *s, const char *p, int depth)
7939 {
7940 const char *olds;
7941 char s1, c;
7942 int gs;
7943
7944 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7945 return (-1);
7946
7947 if (s == NULL)
7948 s = ""; /* treat NULL as empty string */
7949
7950 top:
7951 olds = s;
7952 s1 = *s++;
7953
7954 if (p == NULL)
7955 return (0);
7956
7957 if ((c = *p++) == '\0')
7958 return (s1 == '\0');
7959
7960 switch (c) {
7961 case '[': {
7962 int ok = 0, notflag = 0;
7963 char lc = '\0';
7964
7965 if (s1 == '\0')
7966 return (0);
7967
7968 if (*p == '!') {
7969 notflag = 1;
7970 p++;
7971 }
7972
7973 if ((c = *p++) == '\0')
7974 return (0);
7975
7976 do {
7977 if (c == '-' && lc != '\0' && *p != ']') {
7978 if ((c = *p++) == '\0')
7979 return (0);
7980 if (c == '\\' && (c = *p++) == '\0')
7981 return (0);
7982
7983 if (notflag) {
7984 if (s1 < lc || s1 > c)
7985 ok++;
7986 else
7987 return (0);
7988 } else if (lc <= s1 && s1 <= c)
7989 ok++;
7990
7991 } else if (c == '\\' && (c = *p++) == '\0')
7992 return (0);
7993
7994 lc = c; /* save left-hand 'c' for next iteration */
7995
7996 if (notflag) {
7997 if (s1 != c)
7998 ok++;
7999 else
8000 return (0);
8001 } else if (s1 == c)
8002 ok++;
8003
8004 if ((c = *p++) == '\0')
8005 return (0);
8006
8007 } while (c != ']');
8008
8009 if (ok)
8010 goto top;
8011
8012 return (0);
8013 }
8014
8015 case '\\':
8016 if ((c = *p++) == '\0')
8017 return (0);
8018 /*FALLTHRU*/
8019
8020 default:
8021 if (c != s1)
8022 return (0);
8023 /*FALLTHRU*/
8024
8025 case '?':
8026 if (s1 != '\0')
8027 goto top;
8028 return (0);
8029
8030 case '*':
8031 while (*p == '*')
8032 p++; /* consecutive *'s are identical to a single one */
8033
8034 if (*p == '\0')
8035 return (1);
8036
8037 for (s = olds; *s != '\0'; s++) {
8038 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
8039 return (gs);
8040 }
8041
8042 return (0);
8043 }
8044 }
8045
8046 /*ARGSUSED*/
8047 static int
8048 dtrace_match_string(const char *s, const char *p, int depth)
8049 {
8050 return (s != NULL && strcmp(s, p) == 0);
8051 }
8052
8053 /*ARGSUSED*/
8054 static int
8055 dtrace_match_nul(const char *s, const char *p, int depth)
8056 {
8057 return (1); /* always match the empty pattern */
8058 }
8059
8060 /*ARGSUSED*/
8061 static int
8062 dtrace_match_nonzero(const char *s, const char *p, int depth)
8063 {
8064 return (s != NULL && s[0] != '\0');
8065 }
8066
8067 static int
8068 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
8069 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
8070 {
8071 dtrace_probe_t template, *probe;
8072 dtrace_hash_t *hash = NULL;
8073 int len, rc, best = INT_MAX, nmatched = 0;
8074 dtrace_id_t i;
8075
8076 ASSERT(MUTEX_HELD(&dtrace_lock));
8077
8078 /*
8079 * If the probe ID is specified in the key, just lookup by ID and
8080 * invoke the match callback once if a matching probe is found.
8081 */
8082 if (pkp->dtpk_id != DTRACE_IDNONE) {
8083 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
8084 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
8085 if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
8086 return (DTRACE_MATCH_FAIL);
8087 nmatched++;
8088 }
8089 return (nmatched);
8090 }
8091
8092 template.dtpr_mod = (char *)pkp->dtpk_mod;
8093 template.dtpr_func = (char *)pkp->dtpk_func;
8094 template.dtpr_name = (char *)pkp->dtpk_name;
8095
8096 /*
8097 * We want to find the most distinct of the module name, function
8098 * name, and name. So for each one that is not a glob pattern or
8099 * empty string, we perform a lookup in the corresponding hash and
8100 * use the hash table with the fewest collisions to do our search.
8101 */
8102 if (pkp->dtpk_mmatch == &dtrace_match_string &&
8103 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
8104 best = len;
8105 hash = dtrace_bymod;
8106 }
8107
8108 if (pkp->dtpk_fmatch == &dtrace_match_string &&
8109 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
8110 best = len;
8111 hash = dtrace_byfunc;
8112 }
8113
8114 if (pkp->dtpk_nmatch == &dtrace_match_string &&
8115 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
8116 best = len;
8117 hash = dtrace_byname;
8118 }
8119
8120 /*
8121 * If we did not select a hash table, iterate over every probe and
8122 * invoke our callback for each one that matches our input probe key.
8123 */
8124 if (hash == NULL) {
8125 for (i = 0; i < dtrace_nprobes; i++) {
8126 if ((probe = dtrace_probes[i]) == NULL ||
8127 dtrace_match_probe(probe, pkp, priv, uid,
8128 zoneid) <= 0)
8129 continue;
8130
8131 nmatched++;
8132
8133 if ((rc = (*matched)(probe, arg)) !=
8134 DTRACE_MATCH_NEXT) {
8135 if (rc == DTRACE_MATCH_FAIL)
8136 return (DTRACE_MATCH_FAIL);
8137 break;
8138 }
8139 }
8140
8141 return (nmatched);
8142 }
8143
8144 /*
8145 * If we selected a hash table, iterate over each probe of the same key
8146 * name and invoke the callback for every probe that matches the other
8147 * attributes of our input probe key.
8148 */
8149 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
8150 probe = *(DTRACE_HASHNEXT(hash, probe))) {
8151
8152 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
8153 continue;
8154
8155 nmatched++;
8156
8157 if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
8158 if (rc == DTRACE_MATCH_FAIL)
8159 return (DTRACE_MATCH_FAIL);
8160 break;
8161 }
8162 }
8163
8164 return (nmatched);
8165 }
8166
8167 /*
8168 * Return the function pointer dtrace_probecmp() should use to compare the
8169 * specified pattern with a string. For NULL or empty patterns, we select
8170 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
8171 * For non-empty non-glob strings, we use dtrace_match_string().
8172 */
8173 static dtrace_probekey_f *
8174 dtrace_probekey_func(const char *p)
8175 {
8176 char c;
8177
8178 if (p == NULL || *p == '\0')
8179 return (&dtrace_match_nul);
8180
8181 while ((c = *p++) != '\0') {
8182 if (c == '[' || c == '?' || c == '*' || c == '\\')
8183 return (&dtrace_match_glob);
8184 }
8185
8186 return (&dtrace_match_string);
8187 }
8188
8189 /*
8190 * Build a probe comparison key for use with dtrace_match_probe() from the
8191 * given probe description. By convention, a null key only matches anchored
8192 * probes: if each field is the empty string, reset dtpk_fmatch to
8193 * dtrace_match_nonzero().
8194 */
8195 static void
8196 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
8197 {
8198 pkp->dtpk_prov = pdp->dtpd_provider;
8199 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
8200
8201 pkp->dtpk_mod = pdp->dtpd_mod;
8202 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
8203
8204 pkp->dtpk_func = pdp->dtpd_func;
8205 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
8206
8207 pkp->dtpk_name = pdp->dtpd_name;
8208 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
8209
8210 pkp->dtpk_id = pdp->dtpd_id;
8211
8212 if (pkp->dtpk_id == DTRACE_IDNONE &&
8213 pkp->dtpk_pmatch == &dtrace_match_nul &&
8214 pkp->dtpk_mmatch == &dtrace_match_nul &&
8215 pkp->dtpk_fmatch == &dtrace_match_nul &&
8216 pkp->dtpk_nmatch == &dtrace_match_nul)
8217 pkp->dtpk_fmatch = &dtrace_match_nonzero;
8218 }
8219
8220 /*
8221 * DTrace Provider-to-Framework API Functions
8222 *
8223 * These functions implement much of the Provider-to-Framework API, as
8224 * described in <sys/dtrace.h>. The parts of the API not in this section are
8225 * the functions in the API for probe management (found below), and
8226 * dtrace_probe() itself (found above).
8227 */
8228
8229 /*
8230 * Register the calling provider with the DTrace framework. This should
8231 * generally be called by DTrace providers in their attach(9E) entry point.
8232 */
8233 int
8234 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
8235 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
8236 {
8237 dtrace_provider_t *provider;
8238
8239 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
8240 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8241 "arguments", name ? name : "<NULL>");
8242 return (EINVAL);
8243 }
8244
8245 if (name[0] == '\0' || dtrace_badname(name)) {
8246 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8247 "provider name", name);
8248 return (EINVAL);
8249 }
8250
8251 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
8252 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
8253 pops->dtps_destroy == NULL ||
8254 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
8255 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8256 "provider ops", name);
8257 return (EINVAL);
8258 }
8259
8260 if (dtrace_badattr(&pap->dtpa_provider) ||
8261 dtrace_badattr(&pap->dtpa_mod) ||
8262 dtrace_badattr(&pap->dtpa_func) ||
8263 dtrace_badattr(&pap->dtpa_name) ||
8264 dtrace_badattr(&pap->dtpa_args)) {
8265 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8266 "provider attributes", name);
8267 return (EINVAL);
8268 }
8269
8270 if (priv & ~DTRACE_PRIV_ALL) {
8271 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8272 "privilege attributes", name);
8273 return (EINVAL);
8274 }
8275
8276 if ((priv & DTRACE_PRIV_KERNEL) &&
8277 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
8278 pops->dtps_mode == NULL) {
8279 cmn_err(CE_WARN, "failed to register provider '%s': need "
8280 "dtps_mode() op for given privilege attributes", name);
8281 return (EINVAL);
8282 }
8283
8284 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
8285 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8286 (void) strcpy(provider->dtpv_name, name);
8287
8288 provider->dtpv_attr = *pap;
8289 provider->dtpv_priv.dtpp_flags = priv;
8290 if (cr != NULL) {
8291 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
8292 provider->dtpv_priv.dtpp_zoneid = crgetzonedid(cr);
8293 }
8294 provider->dtpv_pops = *pops;
8295
8296 if (pops->dtps_provide == NULL) {
8297 ASSERT(pops->dtps_provide_module != NULL);
8298 provider->dtpv_pops.dtps_provide = dtrace_nullop_provide;
8299 }
8300
8301 if (pops->dtps_provide_module == NULL) {
8302 ASSERT(pops->dtps_provide != NULL);
8303 provider->dtpv_pops.dtps_provide_module = dtrace_nullop_module;
8304 }
8305
8306 if (pops->dtps_suspend == NULL) {
8307 ASSERT(pops->dtps_resume == NULL);
8308 provider->dtpv_pops.dtps_suspend = dtrace_nullop;
8309 provider->dtpv_pops.dtps_resume = dtrace_nullop;
8310 }
8311
8312 provider->dtpv_arg = arg;
8313 *idp = (dtrace_provider_id_t)provider;
8314
8315 if (pops == &dtrace_provider_ops) {
8316 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8317 ASSERT(MUTEX_HELD(&dtrace_lock));
8318 ASSERT(dtrace_anon.dta_enabling == NULL);
8319
8320 /*
8321 * We make sure that the DTrace provider is at the head of
8322 * the provider chain.
8323 */
8324 provider->dtpv_next = dtrace_provider;
8325 dtrace_provider = provider;
8326 return (0);
8327 }
8328
8329 mutex_enter(&dtrace_provider_lock);
8330 mutex_enter(&dtrace_lock);
8331
8332 /*
8333 * If there is at least one provider registered, we'll add this
8334 * provider after the first provider.
8335 */
8336 if (dtrace_provider != NULL) {
8337 provider->dtpv_next = dtrace_provider->dtpv_next;
8338 dtrace_provider->dtpv_next = provider;
8339 } else {
8340 dtrace_provider = provider;
8341 }
8342
8343 if (dtrace_retained != NULL) {
8344 dtrace_enabling_provide(provider);
8345
8346 /*
8347 * Now we need to call dtrace_enabling_matchall() -- which
8348 * will acquire cpu_lock and dtrace_lock. We therefore need
8349 * to drop all of our locks before calling into it...
8350 */
8351 mutex_exit(&dtrace_lock);
8352 mutex_exit(&dtrace_provider_lock);
8353 dtrace_enabling_matchall();
8354
8355 return (0);
8356 }
8357
8358 mutex_exit(&dtrace_lock);
8359 mutex_exit(&dtrace_provider_lock);
8360
8361 return (0);
8362 }
8363
8364 /*
8365 * Unregister the specified provider from the DTrace framework. This should
8366 * generally be called by DTrace providers in their detach(9E) entry point.
8367 */
8368 int
8369 dtrace_unregister(dtrace_provider_id_t id)
8370 {
8371 dtrace_provider_t *old = (dtrace_provider_t *)id;
8372 dtrace_provider_t *prev = NULL;
8373 int i, self = 0, noreap = 0;
8374 dtrace_probe_t *probe, *first = NULL;
8375
8376 if (old->dtpv_pops.dtps_enable == dtrace_enable_nullop) {
8377 /*
8378 * If DTrace itself is the provider, we're called with locks
8379 * already held.
8380 */
8381 ASSERT(old == dtrace_provider);
8382 ASSERT(dtrace_devi != NULL);
8383 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8384 ASSERT(MUTEX_HELD(&dtrace_lock));
8385 self = 1;
8386
8387 if (dtrace_provider->dtpv_next != NULL) {
8388 /*
8389 * There's another provider here; return failure.
8390 */
8391 return (EBUSY);
8392 }
8393 } else {
8394 mutex_enter(&dtrace_provider_lock);
8395 mutex_enter(&mod_lock);
8396 mutex_enter(&dtrace_lock);
8397 }
8398
8399 /*
8400 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8401 * probes, we refuse to let providers slither away, unless this
8402 * provider has already been explicitly invalidated.
8403 */
8404 if (!old->dtpv_defunct &&
8405 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8406 dtrace_anon.dta_state->dts_necbs > 0))) {
8407 if (!self) {
8408 mutex_exit(&dtrace_lock);
8409 mutex_exit(&mod_lock);
8410 mutex_exit(&dtrace_provider_lock);
8411 }
8412 return (EBUSY);
8413 }
8414
8415 /*
8416 * Attempt to destroy the probes associated with this provider.
8417 */
8418 for (i = 0; i < dtrace_nprobes; i++) {
8419 if ((probe = dtrace_probes[i]) == NULL)
8420 continue;
8421
8422 if (probe->dtpr_provider != old)
8423 continue;
8424
8425 if (probe->dtpr_ecb == NULL)
8426 continue;
8427
8428 /*
8429 * If we are trying to unregister a defunct provider, and the
8430 * provider was made defunct within the interval dictated by
8431 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8432 * attempt to reap our enablings. To denote that the provider
8433 * should reattempt to unregister itself at some point in the
8434 * future, we will return a differentiable error code (EAGAIN
8435 * instead of EBUSY) in this case.
8436 */
8437 if (dtrace_gethrtime() - old->dtpv_defunct >
8438 dtrace_unregister_defunct_reap)
8439 noreap = 1;
8440
8441 if (!self) {
8442 mutex_exit(&dtrace_lock);
8443 mutex_exit(&mod_lock);
8444 mutex_exit(&dtrace_provider_lock);
8445 }
8446
8447 if (noreap)
8448 return (EBUSY);
8449
8450 (void) taskq_dispatch(dtrace_taskq,
8451 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8452
8453 return (EAGAIN);
8454 }
8455
8456 /*
8457 * All of the probes for this provider are disabled; we can safely
8458 * remove all of them from their hash chains and from the probe array.
8459 */
8460 for (i = 0; i < dtrace_nprobes; i++) {
8461 if ((probe = dtrace_probes[i]) == NULL)
8462 continue;
8463
8464 if (probe->dtpr_provider != old)
8465 continue;
8466
8467 dtrace_probes[i] = NULL;
8468
8469 dtrace_hash_remove(dtrace_bymod, probe);
8470 dtrace_hash_remove(dtrace_byfunc, probe);
8471 dtrace_hash_remove(dtrace_byname, probe);
8472
8473 if (first == NULL) {
8474 first = probe;
8475 probe->dtpr_nextmod = NULL;
8476 } else {
8477 probe->dtpr_nextmod = first;
8478 first = probe;
8479 }
8480 }
8481
8482 /*
8483 * The provider's probes have been removed from the hash chains and
8484 * from the probe array. Now issue a dtrace_sync() to be sure that
8485 * everyone has cleared out from any probe array processing.
8486 */
8487 dtrace_sync();
8488
8489 for (probe = first; probe != NULL; probe = first) {
8490 first = probe->dtpr_nextmod;
8491
8492 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8493 probe->dtpr_arg);
8494 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8495 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8496 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8497 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8498 kmem_free(probe, sizeof (dtrace_probe_t));
8499 }
8500
8501 if ((prev = dtrace_provider) == old) {
8502 ASSERT(self || dtrace_devi == NULL);
8503 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8504 dtrace_provider = old->dtpv_next;
8505 } else {
8506 while (prev != NULL && prev->dtpv_next != old)
8507 prev = prev->dtpv_next;
8508
8509 if (prev == NULL) {
8510 panic("attempt to unregister non-existent "
8511 "dtrace provider %p\n", (void *)id);
8512 }
8513
8514 prev->dtpv_next = old->dtpv_next;
8515 }
8516
8517 if (!self) {
8518 mutex_exit(&dtrace_lock);
8519 mutex_exit(&mod_lock);
8520 mutex_exit(&dtrace_provider_lock);
8521 }
8522
8523 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
8524 kmem_free(old, sizeof (dtrace_provider_t));
8525
8526 return (0);
8527 }
8528
8529 /*
8530 * Invalidate the specified provider. All subsequent probe lookups for the
8531 * specified provider will fail, but its probes will not be removed.
8532 */
8533 void
8534 dtrace_invalidate(dtrace_provider_id_t id)
8535 {
8536 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
8537
8538 ASSERT(pvp->dtpv_pops.dtps_enable != dtrace_enable_nullop);
8539
8540 mutex_enter(&dtrace_provider_lock);
8541 mutex_enter(&dtrace_lock);
8542
8543 pvp->dtpv_defunct = dtrace_gethrtime();
8544
8545 mutex_exit(&dtrace_lock);
8546 mutex_exit(&dtrace_provider_lock);
8547 }
8548
8549 /*
8550 * Indicate whether or not DTrace has attached.
8551 */
8552 int
8553 dtrace_attached(void)
8554 {
8555 /*
8556 * dtrace_provider will be non-NULL iff the DTrace driver has
8557 * attached. (It's non-NULL because DTrace is always itself a
8558 * provider.)
8559 */
8560 return (dtrace_provider != NULL);
8561 }
8562
8563 /*
8564 * Remove all the unenabled probes for the given provider. This function is
8565 * not unlike dtrace_unregister(), except that it doesn't remove the provider
8566 * -- just as many of its associated probes as it can.
8567 */
8568 int
8569 dtrace_condense(dtrace_provider_id_t id)
8570 {
8571 dtrace_provider_t *prov = (dtrace_provider_t *)id;
8572 int i;
8573 dtrace_probe_t *probe;
8574
8575 /*
8576 * Make sure this isn't the dtrace provider itself.
8577 */
8578 ASSERT(prov->dtpv_pops.dtps_enable != dtrace_enable_nullop);
8579
8580 mutex_enter(&dtrace_provider_lock);
8581 mutex_enter(&dtrace_lock);
8582
8583 /*
8584 * Attempt to destroy the probes associated with this provider.
8585 */
8586 for (i = 0; i < dtrace_nprobes; i++) {
8587 if ((probe = dtrace_probes[i]) == NULL)
8588 continue;
8589
8590 if (probe->dtpr_provider != prov)
8591 continue;
8592
8593 if (probe->dtpr_ecb != NULL)
8594 continue;
8595
8596 dtrace_probes[i] = NULL;
8597
8598 dtrace_hash_remove(dtrace_bymod, probe);
8599 dtrace_hash_remove(dtrace_byfunc, probe);
8600 dtrace_hash_remove(dtrace_byname, probe);
8601
8602 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
8603 probe->dtpr_arg);
8604 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8605 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8606 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8607 kmem_free(probe, sizeof (dtrace_probe_t));
8608 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
8609 }
8610
8611 mutex_exit(&dtrace_lock);
8612 mutex_exit(&dtrace_provider_lock);
8613
8614 return (0);
8615 }
8616
8617 /*
8618 * DTrace Probe Management Functions
8619 *
8620 * The functions in this section perform the DTrace probe management,
8621 * including functions to create probes, look-up probes, and call into the
8622 * providers to request that probes be provided. Some of these functions are
8623 * in the Provider-to-Framework API; these functions can be identified by the
8624 * fact that they are not declared "static".
8625 */
8626
8627 /*
8628 * Create a probe with the specified module name, function name, and name.
8629 */
8630 dtrace_id_t
8631 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
8632 const char *func, const char *name, int aframes, void *arg)
8633 {
8634 dtrace_probe_t *probe, **probes;
8635 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
8636 dtrace_id_t id;
8637
8638 if (provider == dtrace_provider) {
8639 ASSERT(MUTEX_HELD(&dtrace_lock));
8640 } else {
8641 mutex_enter(&dtrace_lock);
8642 }
8643
8644 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
8645 VM_BESTFIT | VM_SLEEP);
8646 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8647
8648 probe->dtpr_id = id;
8649 probe->dtpr_gen = dtrace_probegen++;
8650 probe->dtpr_mod = dtrace_strdup(mod);
8651 probe->dtpr_func = dtrace_strdup(func);
8652 probe->dtpr_name = dtrace_strdup(name);
8653 probe->dtpr_arg = arg;
8654 probe->dtpr_aframes = aframes;
8655 probe->dtpr_provider = provider;
8656
8657 dtrace_hash_add(dtrace_bymod, probe);
8658 dtrace_hash_add(dtrace_byfunc, probe);
8659 dtrace_hash_add(dtrace_byname, probe);
8660
8661 if (id - 1 >= dtrace_nprobes) {
8662 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8663 size_t nsize = osize << 1;
8664
8665 if (nsize == 0) {
8666 ASSERT(osize == 0);
8667 ASSERT(dtrace_probes == NULL);
8668 nsize = sizeof (dtrace_probe_t *);
8669 }
8670
8671 probes = kmem_zalloc(nsize, KM_SLEEP);
8672
8673 if (dtrace_probes == NULL) {
8674 ASSERT(osize == 0);
8675 dtrace_probes = probes;
8676 dtrace_nprobes = 1;
8677 } else {
8678 dtrace_probe_t **oprobes = dtrace_probes;
8679
8680 bcopy(oprobes, probes, osize);
8681 dtrace_membar_producer();
8682 dtrace_probes = probes;
8683
8684 dtrace_sync();
8685
8686 /*
8687 * All CPUs are now seeing the new probes array; we can
8688 * safely free the old array.
8689 */
8690 kmem_free(oprobes, osize);
8691 dtrace_nprobes <<= 1;
8692 }
8693
8694 ASSERT(id - 1 < dtrace_nprobes);
8695 }
8696
8697 ASSERT(dtrace_probes[id - 1] == NULL);
8698 dtrace_probes[id - 1] = probe;
8699
8700 if (provider != dtrace_provider)
8701 mutex_exit(&dtrace_lock);
8702
8703 return (id);
8704 }
8705
8706 static dtrace_probe_t *
8707 dtrace_probe_lookup_id(dtrace_id_t id)
8708 {
8709 ASSERT(MUTEX_HELD(&dtrace_lock));
8710
8711 if (id == 0 || id > dtrace_nprobes)
8712 return (NULL);
8713
8714 return (dtrace_probes[id - 1]);
8715 }
8716
8717 static int
8718 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8719 {
8720 *((dtrace_id_t *)arg) = probe->dtpr_id;
8721
8722 return (DTRACE_MATCH_DONE);
8723 }
8724
8725 /*
8726 * Look up a probe based on provider and one or more of module name, function
8727 * name and probe name.
8728 */
8729 dtrace_id_t
8730 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
8731 const char *func, const char *name)
8732 {
8733 dtrace_probekey_t pkey;
8734 dtrace_id_t id;
8735 int match;
8736
8737 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8738 pkey.dtpk_pmatch = &dtrace_match_string;
8739 pkey.dtpk_mod = mod;
8740 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8741 pkey.dtpk_func = func;
8742 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8743 pkey.dtpk_name = name;
8744 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8745 pkey.dtpk_id = DTRACE_IDNONE;
8746
8747 mutex_enter(&dtrace_lock);
8748 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8749 dtrace_probe_lookup_match, &id);
8750 mutex_exit(&dtrace_lock);
8751
8752 ASSERT(match == 1 || match == 0);
8753 return (match ? id : 0);
8754 }
8755
8756 /*
8757 * Returns the probe argument associated with the specified probe.
8758 */
8759 void *
8760 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8761 {
8762 dtrace_probe_t *probe;
8763 void *rval = NULL;
8764
8765 mutex_enter(&dtrace_lock);
8766
8767 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8768 probe->dtpr_provider == (dtrace_provider_t *)id)
8769 rval = probe->dtpr_arg;
8770
8771 mutex_exit(&dtrace_lock);
8772
8773 return (rval);
8774 }
8775
8776 /*
8777 * Copy a probe into a probe description.
8778 */
8779 static void
8780 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8781 {
8782 bzero(pdp, sizeof (dtrace_probedesc_t));
8783 pdp->dtpd_id = prp->dtpr_id;
8784
8785 (void) strncpy(pdp->dtpd_provider,
8786 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8787
8788 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8789 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8790 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8791 }
8792
8793 /*
8794 * Called to indicate that a probe -- or probes -- should be provided by a
8795 * specfied provider. If the specified description is NULL, the provider will
8796 * be told to provide all of its probes. (This is done whenever a new
8797 * consumer comes along, or whenever a retained enabling is to be matched.) If
8798 * the specified description is non-NULL, the provider is given the
8799 * opportunity to dynamically provide the specified probe, allowing providers
8800 * to support the creation of probes on-the-fly. (So-called _autocreated_
8801 * probes.) If the provider is NULL, the operations will be applied to all
8802 * providers; if the provider is non-NULL the operations will only be applied
8803 * to the specified provider. The dtrace_provider_lock must be held, and the
8804 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8805 * will need to grab the dtrace_lock when it reenters the framework through
8806 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8807 */
8808 static void
8809 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8810 {
8811 struct modctl *ctl;
8812 int all = 0;
8813
8814 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8815
8816 if (prv == NULL) {
8817 all = 1;
8818 prv = dtrace_provider;
8819 }
8820
8821 do {
8822 /*
8823 * First, call the blanket provide operation.
8824 */
8825 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8826
8827 /*
8828 * Now call the per-module provide operation. We will grab
8829 * mod_lock to prevent the list from being modified. Note
8830 * that this also prevents the mod_busy bits from changing.
8831 * (mod_busy can only be changed with mod_lock held.)
8832 */
8833 mutex_enter(&mod_lock);
8834
8835 ctl = &modules;
8836 do {
8837 if (ctl->mod_busy || ctl->mod_mp == NULL)
8838 continue;
8839
8840 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8841
8842 } while ((ctl = ctl->mod_next) != &modules);
8843
8844 mutex_exit(&mod_lock);
8845 } while (all && (prv = prv->dtpv_next) != NULL);
8846 }
8847
8848 /*
8849 * Iterate over each probe, and call the Framework-to-Provider API function
8850 * denoted by offs.
8851 */
8852 static void
8853 dtrace_probe_foreach(uintptr_t offs)
8854 {
8855 dtrace_provider_t *prov;
8856 void (*func)(void *, dtrace_id_t, void *);
8857 dtrace_probe_t *probe;
8858 dtrace_icookie_t cookie;
8859 int i;
8860
8861 /*
8862 * We disable interrupts to walk through the probe array. This is
8863 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8864 * won't see stale data.
8865 */
8866 cookie = dtrace_interrupt_disable();
8867
8868 for (i = 0; i < dtrace_nprobes; i++) {
8869 if ((probe = dtrace_probes[i]) == NULL)
8870 continue;
8871
8872 if (probe->dtpr_ecb == NULL) {
8873 /*
8874 * This probe isn't enabled -- don't call the function.
8875 */
8876 continue;
8877 }
8878
8879 prov = probe->dtpr_provider;
8880 func = *((void(**)(void *, dtrace_id_t, void *))
8881 ((uintptr_t)&prov->dtpv_pops + offs));
8882
8883 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8884 }
8885
8886 dtrace_interrupt_enable(cookie);
8887 }
8888
8889 static int
8890 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8891 {
8892 dtrace_probekey_t pkey;
8893 uint32_t priv;
8894 uid_t uid;
8895 zoneid_t zoneid;
8896 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
8897
8898 ASSERT(MUTEX_HELD(&dtrace_lock));
8899 dtrace_ecb_create_cache = NULL;
8900
8901 if (desc == NULL) {
8902 /*
8903 * If we're passed a NULL description, we're being asked to
8904 * create an ECB with a NULL probe.
8905 */
8906 (void) dtrace_ecb_create_enable(NULL, enab);
8907 return (0);
8908 }
8909
8910 dtrace_probekey(desc, &pkey);
8911 dtrace_cred2priv(state->dts_cred.dcr_cred, &priv, &uid, &zoneid);
8912
8913 if ((priv & DTRACE_PRIV_ZONEOWNER) &&
8914 state->dts_options[DTRACEOPT_ZONE] != DTRACEOPT_UNSET) {
8915 /*
8916 * If we have the privilege of instrumenting all zones but we
8917 * have been told to instrument but one, we will spoof this up
8918 * depriving ourselves of DTRACE_PRIV_ZONEOWNER for purposes
8919 * of dtrace_match(). (Note that DTRACEOPT_ZONE is not for
8920 * security but rather for performance: it allows the global
8921 * zone to instrument USDT probes in a local zone without
8922 * requiring all zones to be instrumented.)
8923 */
8924 priv &= ~DTRACE_PRIV_ZONEOWNER;
8925 zoneid = state->dts_options[DTRACEOPT_ZONE];
8926 }
8927
8928 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8929 enab));
8930 }
8931
8932 /*
8933 * DTrace Helper Provider Functions
8934 */
8935 static void
8936 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8937 {
8938 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8939 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8940 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8941 }
8942
8943 static void
8944 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8945 const dof_provider_t *dofprov, char *strtab)
8946 {
8947 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8948 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8949 dofprov->dofpv_provattr);
8950 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8951 dofprov->dofpv_modattr);
8952 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8953 dofprov->dofpv_funcattr);
8954 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8955 dofprov->dofpv_nameattr);
8956 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8957 dofprov->dofpv_argsattr);
8958 }
8959
8960 static void
8961 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8962 {
8963 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8964 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8965 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8966 dof_provider_t *provider;
8967 dof_probe_t *probe;
8968 uint32_t *off, *enoff;
8969 uint8_t *arg;
8970 char *strtab;
8971 uint_t i, nprobes;
8972 dtrace_helper_provdesc_t dhpv;
8973 dtrace_helper_probedesc_t dhpb;
8974 dtrace_meta_t *meta = dtrace_meta_pid;
8975 dtrace_mops_t *mops = &meta->dtm_mops;
8976 void *parg;
8977
8978 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8979 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8980 provider->dofpv_strtab * dof->dofh_secsize);
8981 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8982 provider->dofpv_probes * dof->dofh_secsize);
8983 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8984 provider->dofpv_prargs * dof->dofh_secsize);
8985 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8986 provider->dofpv_proffs * dof->dofh_secsize);
8987
8988 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8989 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8990 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8991 enoff = NULL;
8992
8993 /*
8994 * See dtrace_helper_provider_validate().
8995 */
8996 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8997 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8998 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8999 provider->dofpv_prenoffs * dof->dofh_secsize);
9000 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
9001 }
9002
9003 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
9004
9005 /*
9006 * Create the provider.
9007 */
9008 dtrace_dofprov2hprov(&dhpv, provider, strtab);
9009
9010 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
9011 return;
9012
9013 meta->dtm_count++;
9014
9015 /*
9016 * Create the probes.
9017 */
9018 for (i = 0; i < nprobes; i++) {
9019 probe = (dof_probe_t *)(uintptr_t)(daddr +
9020 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
9021
9022 dhpb.dthpb_mod = dhp->dofhp_mod;
9023 dhpb.dthpb_func = strtab + probe->dofpr_func;
9024 dhpb.dthpb_name = strtab + probe->dofpr_name;
9025 dhpb.dthpb_base = probe->dofpr_addr;
9026 dhpb.dthpb_offs = off + probe->dofpr_offidx;
9027 dhpb.dthpb_noffs = probe->dofpr_noffs;
9028 if (enoff != NULL) {
9029 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
9030 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
9031 } else {
9032 dhpb.dthpb_enoffs = NULL;
9033 dhpb.dthpb_nenoffs = 0;
9034 }
9035 dhpb.dthpb_args = arg + probe->dofpr_argidx;
9036 dhpb.dthpb_nargc = probe->dofpr_nargc;
9037 dhpb.dthpb_xargc = probe->dofpr_xargc;
9038 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
9039 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
9040
9041 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
9042 }
9043 }
9044
9045 static void
9046 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
9047 {
9048 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9049 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9050 int i;
9051
9052 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9053
9054 for (i = 0; i < dof->dofh_secnum; i++) {
9055 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9056 dof->dofh_secoff + i * dof->dofh_secsize);
9057
9058 if (sec->dofs_type != DOF_SECT_PROVIDER)
9059 continue;
9060
9061 dtrace_helper_provide_one(dhp, sec, pid);
9062 }
9063
9064 /*
9065 * We may have just created probes, so we must now rematch against
9066 * any retained enablings. Note that this call will acquire both
9067 * cpu_lock and dtrace_lock; the fact that we are holding
9068 * dtrace_meta_lock now is what defines the ordering with respect to
9069 * these three locks.
9070 */
9071 dtrace_enabling_matchall();
9072 }
9073
9074 static void
9075 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9076 {
9077 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9078 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9079 dof_sec_t *str_sec;
9080 dof_provider_t *provider;
9081 char *strtab;
9082 dtrace_helper_provdesc_t dhpv;
9083 dtrace_meta_t *meta = dtrace_meta_pid;
9084 dtrace_mops_t *mops = &meta->dtm_mops;
9085
9086 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9087 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9088 provider->dofpv_strtab * dof->dofh_secsize);
9089
9090 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9091
9092 /*
9093 * Create the provider.
9094 */
9095 dtrace_dofprov2hprov(&dhpv, provider, strtab);
9096
9097 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
9098
9099 meta->dtm_count--;
9100 }
9101
9102 static void
9103 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
9104 {
9105 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9106 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9107 int i;
9108
9109 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9110
9111 for (i = 0; i < dof->dofh_secnum; i++) {
9112 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9113 dof->dofh_secoff + i * dof->dofh_secsize);
9114
9115 if (sec->dofs_type != DOF_SECT_PROVIDER)
9116 continue;
9117
9118 dtrace_helper_provider_remove_one(dhp, sec, pid);
9119 }
9120 }
9121
9122 /*
9123 * DTrace Meta Provider-to-Framework API Functions
9124 *
9125 * These functions implement the Meta Provider-to-Framework API, as described
9126 * in <sys/dtrace.h>.
9127 */
9128 int
9129 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
9130 dtrace_meta_provider_id_t *idp)
9131 {
9132 dtrace_meta_t *meta;
9133 dtrace_helpers_t *help, *next;
9134 int i;
9135
9136 *idp = DTRACE_METAPROVNONE;
9137
9138 /*
9139 * We strictly don't need the name, but we hold onto it for
9140 * debuggability. All hail error queues!
9141 */
9142 if (name == NULL) {
9143 cmn_err(CE_WARN, "failed to register meta-provider: "
9144 "invalid name");
9145 return (EINVAL);
9146 }
9147
9148 if (mops == NULL ||
9149 mops->dtms_create_probe == NULL ||
9150 mops->dtms_provide_pid == NULL ||
9151 mops->dtms_remove_pid == NULL) {
9152 cmn_err(CE_WARN, "failed to register meta-register %s: "
9153 "invalid ops", name);
9154 return (EINVAL);
9155 }
9156
9157 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
9158 meta->dtm_mops = *mops;
9159 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
9160 (void) strcpy(meta->dtm_name, name);
9161 meta->dtm_arg = arg;
9162
9163 mutex_enter(&dtrace_meta_lock);
9164 mutex_enter(&dtrace_lock);
9165
9166 if (dtrace_meta_pid != NULL) {
9167 mutex_exit(&dtrace_lock);
9168 mutex_exit(&dtrace_meta_lock);
9169 cmn_err(CE_WARN, "failed to register meta-register %s: "
9170 "user-land meta-provider exists", name);
9171 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
9172 kmem_free(meta, sizeof (dtrace_meta_t));
9173 return (EINVAL);
9174 }
9175
9176 dtrace_meta_pid = meta;
9177 *idp = (dtrace_meta_provider_id_t)meta;
9178
9179 /*
9180 * If there are providers and probes ready to go, pass them
9181 * off to the new meta provider now.
9182 */
9183
9184 help = dtrace_deferred_pid;
9185 dtrace_deferred_pid = NULL;
9186
9187 mutex_exit(&dtrace_lock);
9188
9189 while (help != NULL) {
9190 for (i = 0; i < help->dthps_nprovs; i++) {
9191 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
9192 help->dthps_pid);
9193 }
9194
9195 next = help->dthps_next;
9196 help->dthps_next = NULL;
9197 help->dthps_prev = NULL;
9198 help->dthps_deferred = 0;
9199 help = next;
9200 }
9201
9202 mutex_exit(&dtrace_meta_lock);
9203
9204 return (0);
9205 }
9206
9207 int
9208 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
9209 {
9210 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
9211
9212 mutex_enter(&dtrace_meta_lock);
9213 mutex_enter(&dtrace_lock);
9214
9215 if (old == dtrace_meta_pid) {
9216 pp = &dtrace_meta_pid;
9217 } else {
9218 panic("attempt to unregister non-existent "
9219 "dtrace meta-provider %p\n", (void *)old);
9220 }
9221
9222 if (old->dtm_count != 0) {
9223 mutex_exit(&dtrace_lock);
9224 mutex_exit(&dtrace_meta_lock);
9225 return (EBUSY);
9226 }
9227
9228 *pp = NULL;
9229
9230 mutex_exit(&dtrace_lock);
9231 mutex_exit(&dtrace_meta_lock);
9232
9233 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
9234 kmem_free(old, sizeof (dtrace_meta_t));
9235
9236 return (0);
9237 }
9238
9239
9240 /*
9241 * DTrace DIF Object Functions
9242 */
9243 static int
9244 dtrace_difo_err(uint_t pc, const char *format, ...)
9245 {
9246 if (dtrace_err_verbose) {
9247 va_list alist;
9248
9249 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
9250 va_start(alist, format);
9251 (void) vuprintf(format, alist);
9252 va_end(alist);
9253 }
9254
9255 #ifdef DTRACE_ERRDEBUG
9256 dtrace_errdebug(format);
9257 #endif
9258 return (1);
9259 }
9260
9261 /*
9262 * Validate a DTrace DIF object by checking the IR instructions. The following
9263 * rules are currently enforced by dtrace_difo_validate():
9264 *
9265 * 1. Each instruction must have a valid opcode
9266 * 2. Each register, string, variable, or subroutine reference must be valid
9267 * 3. No instruction can modify register %r0 (must be zero)
9268 * 4. All instruction reserved bits must be set to zero
9269 * 5. The last instruction must be a "ret" instruction
9270 * 6. All branch targets must reference a valid instruction _after_ the branch
9271 */
9272 static int
9273 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
9274 cred_t *cr)
9275 {
9276 int err = 0, i;
9277 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9278 int kcheckload;
9279 uint_t pc;
9280 int maxglobal = -1, maxlocal = -1, maxtlocal = -1;
9281
9282 kcheckload = cr == NULL ||
9283 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
9284
9285 dp->dtdo_destructive = 0;
9286
9287 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9288 dif_instr_t instr = dp->dtdo_buf[pc];
9289
9290 uint_t r1 = DIF_INSTR_R1(instr);
9291 uint_t r2 = DIF_INSTR_R2(instr);
9292 uint_t rd = DIF_INSTR_RD(instr);
9293 uint_t rs = DIF_INSTR_RS(instr);
9294 uint_t label = DIF_INSTR_LABEL(instr);
9295 uint_t v = DIF_INSTR_VAR(instr);
9296 uint_t subr = DIF_INSTR_SUBR(instr);
9297 uint_t type = DIF_INSTR_TYPE(instr);
9298 uint_t op = DIF_INSTR_OP(instr);
9299
9300 switch (op) {
9301 case DIF_OP_OR:
9302 case DIF_OP_XOR:
9303 case DIF_OP_AND:
9304 case DIF_OP_SLL:
9305 case DIF_OP_SRL:
9306 case DIF_OP_SRA:
9307 case DIF_OP_SUB:
9308 case DIF_OP_ADD:
9309 case DIF_OP_MUL:
9310 case DIF_OP_SDIV:
9311 case DIF_OP_UDIV:
9312 case DIF_OP_SREM:
9313 case DIF_OP_UREM:
9314 case DIF_OP_COPYS:
9315 if (r1 >= nregs)
9316 err += efunc(pc, "invalid register %u\n", r1);
9317 if (r2 >= nregs)
9318 err += efunc(pc, "invalid register %u\n", r2);
9319 if (rd >= nregs)
9320 err += efunc(pc, "invalid register %u\n", rd);
9321 if (rd == 0)
9322 err += efunc(pc, "cannot write to %r0\n");
9323 break;
9324 case DIF_OP_NOT:
9325 case DIF_OP_MOV:
9326 case DIF_OP_ALLOCS:
9327 if (r1 >= nregs)
9328 err += efunc(pc, "invalid register %u\n", r1);
9329 if (r2 != 0)
9330 err += efunc(pc, "non-zero reserved bits\n");
9331 if (rd >= nregs)
9332 err += efunc(pc, "invalid register %u\n", rd);
9333 if (rd == 0)
9334 err += efunc(pc, "cannot write to %r0\n");
9335 break;
9336 case DIF_OP_LDSB:
9337 case DIF_OP_LDSH:
9338 case DIF_OP_LDSW:
9339 case DIF_OP_LDUB:
9340 case DIF_OP_LDUH:
9341 case DIF_OP_LDUW:
9342 case DIF_OP_LDX:
9343 if (r1 >= nregs)
9344 err += efunc(pc, "invalid register %u\n", r1);
9345 if (r2 != 0)
9346 err += efunc(pc, "non-zero reserved bits\n");
9347 if (rd >= nregs)
9348 err += efunc(pc, "invalid register %u\n", rd);
9349 if (rd == 0)
9350 err += efunc(pc, "cannot write to %r0\n");
9351 if (kcheckload)
9352 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9353 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9354 break;
9355 case DIF_OP_RLDSB:
9356 case DIF_OP_RLDSH:
9357 case DIF_OP_RLDSW:
9358 case DIF_OP_RLDUB:
9359 case DIF_OP_RLDUH:
9360 case DIF_OP_RLDUW:
9361 case DIF_OP_RLDX:
9362 if (r1 >= nregs)
9363 err += efunc(pc, "invalid register %u\n", r1);
9364 if (r2 != 0)
9365 err += efunc(pc, "non-zero reserved bits\n");
9366 if (rd >= nregs)
9367 err += efunc(pc, "invalid register %u\n", rd);
9368 if (rd == 0)
9369 err += efunc(pc, "cannot write to %r0\n");
9370 break;
9371 case DIF_OP_ULDSB:
9372 case DIF_OP_ULDSH:
9373 case DIF_OP_ULDSW:
9374 case DIF_OP_ULDUB:
9375 case DIF_OP_ULDUH:
9376 case DIF_OP_ULDUW:
9377 case DIF_OP_ULDX:
9378 if (r1 >= nregs)
9379 err += efunc(pc, "invalid register %u\n", r1);
9380 if (r2 != 0)
9381 err += efunc(pc, "non-zero reserved bits\n");
9382 if (rd >= nregs)
9383 err += efunc(pc, "invalid register %u\n", rd);
9384 if (rd == 0)
9385 err += efunc(pc, "cannot write to %r0\n");
9386 break;
9387 case DIF_OP_STB:
9388 case DIF_OP_STH:
9389 case DIF_OP_STW:
9390 case DIF_OP_STX:
9391 if (r1 >= nregs)
9392 err += efunc(pc, "invalid register %u\n", r1);
9393 if (r2 != 0)
9394 err += efunc(pc, "non-zero reserved bits\n");
9395 if (rd >= nregs)
9396 err += efunc(pc, "invalid register %u\n", rd);
9397 if (rd == 0)
9398 err += efunc(pc, "cannot write to 0 address\n");
9399 break;
9400 case DIF_OP_CMP:
9401 case DIF_OP_SCMP:
9402 if (r1 >= nregs)
9403 err += efunc(pc, "invalid register %u\n", r1);
9404 if (r2 >= nregs)
9405 err += efunc(pc, "invalid register %u\n", r2);
9406 if (rd != 0)
9407 err += efunc(pc, "non-zero reserved bits\n");
9408 break;
9409 case DIF_OP_TST:
9410 if (r1 >= nregs)
9411 err += efunc(pc, "invalid register %u\n", r1);
9412 if (r2 != 0 || rd != 0)
9413 err += efunc(pc, "non-zero reserved bits\n");
9414 break;
9415 case DIF_OP_BA:
9416 case DIF_OP_BE:
9417 case DIF_OP_BNE:
9418 case DIF_OP_BG:
9419 case DIF_OP_BGU:
9420 case DIF_OP_BGE:
9421 case DIF_OP_BGEU:
9422 case DIF_OP_BL:
9423 case DIF_OP_BLU:
9424 case DIF_OP_BLE:
9425 case DIF_OP_BLEU:
9426 if (label >= dp->dtdo_len) {
9427 err += efunc(pc, "invalid branch target %u\n",
9428 label);
9429 }
9430 if (label <= pc) {
9431 err += efunc(pc, "backward branch to %u\n",
9432 label);
9433 }
9434 break;
9435 case DIF_OP_RET:
9436 if (r1 != 0 || r2 != 0)
9437 err += efunc(pc, "non-zero reserved bits\n");
9438 if (rd >= nregs)
9439 err += efunc(pc, "invalid register %u\n", rd);
9440 break;
9441 case DIF_OP_NOP:
9442 case DIF_OP_POPTS:
9443 case DIF_OP_FLUSHTS:
9444 if (r1 != 0 || r2 != 0 || rd != 0)
9445 err += efunc(pc, "non-zero reserved bits\n");
9446 break;
9447 case DIF_OP_SETX:
9448 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9449 err += efunc(pc, "invalid integer ref %u\n",
9450 DIF_INSTR_INTEGER(instr));
9451 }
9452 if (rd >= nregs)
9453 err += efunc(pc, "invalid register %u\n", rd);
9454 if (rd == 0)
9455 err += efunc(pc, "cannot write to %r0\n");
9456 break;
9457 case DIF_OP_SETS:
9458 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9459 err += efunc(pc, "invalid string ref %u\n",
9460 DIF_INSTR_STRING(instr));
9461 }
9462 if (rd >= nregs)
9463 err += efunc(pc, "invalid register %u\n", rd);
9464 if (rd == 0)
9465 err += efunc(pc, "cannot write to %r0\n");
9466 break;
9467 case DIF_OP_LDGA:
9468 case DIF_OP_LDTA:
9469 if (r1 > DIF_VAR_ARRAY_MAX)
9470 err += efunc(pc, "invalid array %u\n", r1);
9471 if (r2 >= nregs)
9472 err += efunc(pc, "invalid register %u\n", r2);
9473 if (rd >= nregs)
9474 err += efunc(pc, "invalid register %u\n", rd);
9475 if (rd == 0)
9476 err += efunc(pc, "cannot write to %r0\n");
9477 break;
9478 case DIF_OP_STGA:
9479 if (r1 > DIF_VAR_ARRAY_MAX)
9480 err += efunc(pc, "invalid array %u\n", r1);
9481 if (r2 >= nregs)
9482 err += efunc(pc, "invalid register %u\n", r2);
9483 if (rd >= nregs)
9484 err += efunc(pc, "invalid register %u\n", rd);
9485 dp->dtdo_destructive = 1;
9486 break;
9487 case DIF_OP_LDGS:
9488 case DIF_OP_LDTS:
9489 case DIF_OP_LDLS:
9490 case DIF_OP_LDGAA:
9491 case DIF_OP_LDTAA:
9492 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9493 err += efunc(pc, "invalid variable %u\n", v);
9494 if (rd >= nregs)
9495 err += efunc(pc, "invalid register %u\n", rd);
9496 if (rd == 0)
9497 err += efunc(pc, "cannot write to %r0\n");
9498 break;
9499 case DIF_OP_STGS:
9500 case DIF_OP_STTS:
9501 case DIF_OP_STLS:
9502 case DIF_OP_STGAA:
9503 case DIF_OP_STTAA:
9504 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9505 err += efunc(pc, "invalid variable %u\n", v);
9506 if (rs >= nregs)
9507 err += efunc(pc, "invalid register %u\n", rd);
9508 break;
9509 case DIF_OP_CALL:
9510 if (subr > DIF_SUBR_MAX)
9511 err += efunc(pc, "invalid subr %u\n", subr);
9512 if (rd >= nregs)
9513 err += efunc(pc, "invalid register %u\n", rd);
9514 if (rd == 0)
9515 err += efunc(pc, "cannot write to %r0\n");
9516
9517 if (subr == DIF_SUBR_COPYOUT ||
9518 subr == DIF_SUBR_COPYOUTSTR) {
9519 dp->dtdo_destructive = 1;
9520 }
9521
9522 if (subr == DIF_SUBR_GETF) {
9523 /*
9524 * If we have a getf() we need to record that
9525 * in our state. Note that our state can be
9526 * NULL if this is a helper -- but in that
9527 * case, the call to getf() is itself illegal,
9528 * and will be caught (slightly later) when
9529 * the helper is validated.
9530 */
9531 if (vstate->dtvs_state != NULL)
9532 vstate->dtvs_state->dts_getf++;
9533 }
9534
9535 break;
9536 case DIF_OP_PUSHTR:
9537 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
9538 err += efunc(pc, "invalid ref type %u\n", type);
9539 if (r2 >= nregs)
9540 err += efunc(pc, "invalid register %u\n", r2);
9541 if (rs >= nregs)
9542 err += efunc(pc, "invalid register %u\n", rs);
9543 break;
9544 case DIF_OP_PUSHTV:
9545 if (type != DIF_TYPE_CTF)
9546 err += efunc(pc, "invalid val type %u\n", type);
9547 if (r2 >= nregs)
9548 err += efunc(pc, "invalid register %u\n", r2);
9549 if (rs >= nregs)
9550 err += efunc(pc, "invalid register %u\n", rs);
9551 break;
9552 default:
9553 err += efunc(pc, "invalid opcode %u\n",
9554 DIF_INSTR_OP(instr));
9555 }
9556 }
9557
9558 if (dp->dtdo_len != 0 &&
9559 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
9560 err += efunc(dp->dtdo_len - 1,
9561 "expected 'ret' as last DIF instruction\n");
9562 }
9563
9564 if (!(dp->dtdo_rtype.dtdt_flags & (DIF_TF_BYREF | DIF_TF_BYUREF))) {
9565 /*
9566 * If we're not returning by reference, the size must be either
9567 * 0 or the size of one of the base types.
9568 */
9569 switch (dp->dtdo_rtype.dtdt_size) {
9570 case 0:
9571 case sizeof (uint8_t):
9572 case sizeof (uint16_t):
9573 case sizeof (uint32_t):
9574 case sizeof (uint64_t):
9575 break;
9576
9577 default:
9578 err += efunc(dp->dtdo_len - 1, "bad return size\n");
9579 }
9580 }
9581
9582 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
9583 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
9584 dtrace_diftype_t *vt, *et;
9585 uint_t id, ndx;
9586
9587 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
9588 v->dtdv_scope != DIFV_SCOPE_THREAD &&
9589 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
9590 err += efunc(i, "unrecognized variable scope %d\n",
9591 v->dtdv_scope);
9592 break;
9593 }
9594
9595 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
9596 v->dtdv_kind != DIFV_KIND_SCALAR) {
9597 err += efunc(i, "unrecognized variable type %d\n",
9598 v->dtdv_kind);
9599 break;
9600 }
9601
9602 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
9603 err += efunc(i, "%d exceeds variable id limit\n", id);
9604 break;
9605 }
9606
9607 if (id < DIF_VAR_OTHER_UBASE)
9608 continue;
9609
9610 /*
9611 * For user-defined variables, we need to check that this
9612 * definition is identical to any previous definition that we
9613 * encountered.
9614 */
9615 ndx = id - DIF_VAR_OTHER_UBASE;
9616
9617 switch (v->dtdv_scope) {
9618 case DIFV_SCOPE_GLOBAL:
9619 if (maxglobal == -1 || ndx > maxglobal)
9620 maxglobal = ndx;
9621
9622 if (ndx < vstate->dtvs_nglobals) {
9623 dtrace_statvar_t *svar;
9624
9625 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
9626 existing = &svar->dtsv_var;
9627 }
9628
9629 break;
9630
9631 case DIFV_SCOPE_THREAD:
9632 if (maxtlocal == -1 || ndx > maxtlocal)
9633 maxtlocal = ndx;
9634
9635 if (ndx < vstate->dtvs_ntlocals)
9636 existing = &vstate->dtvs_tlocals[ndx];
9637 break;
9638
9639 case DIFV_SCOPE_LOCAL:
9640 if (maxlocal == -1 || ndx > maxlocal)
9641 maxlocal = ndx;
9642
9643 if (ndx < vstate->dtvs_nlocals) {
9644 dtrace_statvar_t *svar;
9645
9646 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
9647 existing = &svar->dtsv_var;
9648 }
9649
9650 break;
9651 }
9652
9653 vt = &v->dtdv_type;
9654
9655 if (vt->dtdt_flags & DIF_TF_BYREF) {
9656 if (vt->dtdt_size == 0) {
9657 err += efunc(i, "zero-sized variable\n");
9658 break;
9659 }
9660
9661 if ((v->dtdv_scope == DIFV_SCOPE_GLOBAL ||
9662 v->dtdv_scope == DIFV_SCOPE_LOCAL) &&
9663 vt->dtdt_size > dtrace_statvar_maxsize) {
9664 err += efunc(i, "oversized by-ref static\n");
9665 break;
9666 }
9667 }
9668
9669 if (existing == NULL || existing->dtdv_id == 0)
9670 continue;
9671
9672 ASSERT(existing->dtdv_id == v->dtdv_id);
9673 ASSERT(existing->dtdv_scope == v->dtdv_scope);
9674
9675 if (existing->dtdv_kind != v->dtdv_kind)
9676 err += efunc(i, "%d changed variable kind\n", id);
9677
9678 et = &existing->dtdv_type;
9679
9680 if (vt->dtdt_flags != et->dtdt_flags) {
9681 err += efunc(i, "%d changed variable type flags\n", id);
9682 break;
9683 }
9684
9685 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
9686 err += efunc(i, "%d changed variable type size\n", id);
9687 break;
9688 }
9689 }
9690
9691 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9692 dif_instr_t instr = dp->dtdo_buf[pc];
9693
9694 uint_t v = DIF_INSTR_VAR(instr);
9695 uint_t op = DIF_INSTR_OP(instr);
9696
9697 switch (op) {
9698 case DIF_OP_LDGS:
9699 case DIF_OP_LDGAA:
9700 case DIF_OP_STGS:
9701 case DIF_OP_STGAA:
9702 if (v > DIF_VAR_OTHER_UBASE + maxglobal)
9703 err += efunc(pc, "invalid variable %u\n", v);
9704 break;
9705 case DIF_OP_LDTS:
9706 case DIF_OP_LDTAA:
9707 case DIF_OP_STTS:
9708 case DIF_OP_STTAA:
9709 if (v > DIF_VAR_OTHER_UBASE + maxtlocal)
9710 err += efunc(pc, "invalid variable %u\n", v);
9711 break;
9712 case DIF_OP_LDLS:
9713 case DIF_OP_STLS:
9714 if (v > DIF_VAR_OTHER_UBASE + maxlocal)
9715 err += efunc(pc, "invalid variable %u\n", v);
9716 break;
9717 default:
9718 break;
9719 }
9720 }
9721
9722 return (err);
9723 }
9724
9725 /*
9726 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
9727 * are much more constrained than normal DIFOs. Specifically, they may
9728 * not:
9729 *
9730 * 1. Make calls to subroutines other than copyin(), copyinstr() or
9731 * miscellaneous string routines
9732 * 2. Access DTrace variables other than the args[] array, and the
9733 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9734 * 3. Have thread-local variables.
9735 * 4. Have dynamic variables.
9736 */
9737 static int
9738 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9739 {
9740 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9741 int err = 0;
9742 uint_t pc;
9743
9744 for (pc = 0; pc < dp->dtdo_len; pc++) {
9745 dif_instr_t instr = dp->dtdo_buf[pc];
9746
9747 uint_t v = DIF_INSTR_VAR(instr);
9748 uint_t subr = DIF_INSTR_SUBR(instr);
9749 uint_t op = DIF_INSTR_OP(instr);
9750
9751 switch (op) {
9752 case DIF_OP_OR:
9753 case DIF_OP_XOR:
9754 case DIF_OP_AND:
9755 case DIF_OP_SLL:
9756 case DIF_OP_SRL:
9757 case DIF_OP_SRA:
9758 case DIF_OP_SUB:
9759 case DIF_OP_ADD:
9760 case DIF_OP_MUL:
9761 case DIF_OP_SDIV:
9762 case DIF_OP_UDIV:
9763 case DIF_OP_SREM:
9764 case DIF_OP_UREM:
9765 case DIF_OP_COPYS:
9766 case DIF_OP_NOT:
9767 case DIF_OP_MOV:
9768 case DIF_OP_RLDSB:
9769 case DIF_OP_RLDSH:
9770 case DIF_OP_RLDSW:
9771 case DIF_OP_RLDUB:
9772 case DIF_OP_RLDUH:
9773 case DIF_OP_RLDUW:
9774 case DIF_OP_RLDX:
9775 case DIF_OP_ULDSB:
9776 case DIF_OP_ULDSH:
9777 case DIF_OP_ULDSW:
9778 case DIF_OP_ULDUB:
9779 case DIF_OP_ULDUH:
9780 case DIF_OP_ULDUW:
9781 case DIF_OP_ULDX:
9782 case DIF_OP_STB:
9783 case DIF_OP_STH:
9784 case DIF_OP_STW:
9785 case DIF_OP_STX:
9786 case DIF_OP_ALLOCS:
9787 case DIF_OP_CMP:
9788 case DIF_OP_SCMP:
9789 case DIF_OP_TST:
9790 case DIF_OP_BA:
9791 case DIF_OP_BE:
9792 case DIF_OP_BNE:
9793 case DIF_OP_BG:
9794 case DIF_OP_BGU:
9795 case DIF_OP_BGE:
9796 case DIF_OP_BGEU:
9797 case DIF_OP_BL:
9798 case DIF_OP_BLU:
9799 case DIF_OP_BLE:
9800 case DIF_OP_BLEU:
9801 case DIF_OP_RET:
9802 case DIF_OP_NOP:
9803 case DIF_OP_POPTS:
9804 case DIF_OP_FLUSHTS:
9805 case DIF_OP_SETX:
9806 case DIF_OP_SETS:
9807 case DIF_OP_LDGA:
9808 case DIF_OP_LDLS:
9809 case DIF_OP_STGS:
9810 case DIF_OP_STLS:
9811 case DIF_OP_PUSHTR:
9812 case DIF_OP_PUSHTV:
9813 break;
9814
9815 case DIF_OP_LDGS:
9816 if (v >= DIF_VAR_OTHER_UBASE)
9817 break;
9818
9819 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9820 break;
9821
9822 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9823 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9824 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9825 v == DIF_VAR_UID || v == DIF_VAR_GID)
9826 break;
9827
9828 err += efunc(pc, "illegal variable %u\n", v);
9829 break;
9830
9831 case DIF_OP_LDTA:
9832 if (v < DIF_VAR_OTHER_UBASE) {
9833 err += efunc(pc, "illegal variable load\n");
9834 break;
9835 }
9836 /* FALLTHROUGH */
9837 case DIF_OP_LDTS:
9838 case DIF_OP_LDGAA:
9839 case DIF_OP_LDTAA:
9840 err += efunc(pc, "illegal dynamic variable load\n");
9841 break;
9842
9843 case DIF_OP_STGA:
9844 if (v < DIF_VAR_OTHER_UBASE) {
9845 err += efunc(pc, "illegal variable store\n");
9846 break;
9847 }
9848 /* FALLTHROUGH */
9849 case DIF_OP_STTS:
9850 case DIF_OP_STGAA:
9851 case DIF_OP_STTAA:
9852 err += efunc(pc, "illegal dynamic variable store\n");
9853 break;
9854
9855 case DIF_OP_CALL:
9856 if (subr == DIF_SUBR_ALLOCA ||
9857 subr == DIF_SUBR_BCOPY ||
9858 subr == DIF_SUBR_COPYIN ||
9859 subr == DIF_SUBR_COPYINTO ||
9860 subr == DIF_SUBR_COPYINSTR ||
9861 subr == DIF_SUBR_INDEX ||
9862 subr == DIF_SUBR_INET_NTOA ||
9863 subr == DIF_SUBR_INET_NTOA6 ||
9864 subr == DIF_SUBR_INET_NTOP ||
9865 subr == DIF_SUBR_JSON ||
9866 subr == DIF_SUBR_LLTOSTR ||
9867 subr == DIF_SUBR_STRTOLL ||
9868 subr == DIF_SUBR_RINDEX ||
9869 subr == DIF_SUBR_STRCHR ||
9870 subr == DIF_SUBR_STRJOIN ||
9871 subr == DIF_SUBR_STRRCHR ||
9872 subr == DIF_SUBR_STRSTR ||
9873 subr == DIF_SUBR_HTONS ||
9874 subr == DIF_SUBR_HTONL ||
9875 subr == DIF_SUBR_HTONLL ||
9876 subr == DIF_SUBR_NTOHS ||
9877 subr == DIF_SUBR_NTOHL ||
9878 subr == DIF_SUBR_NTOHLL)
9879 break;
9880
9881 err += efunc(pc, "invalid subr %u\n", subr);
9882 break;
9883
9884 default:
9885 err += efunc(pc, "invalid opcode %u\n",
9886 DIF_INSTR_OP(instr));
9887 }
9888 }
9889
9890 return (err);
9891 }
9892
9893 /*
9894 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9895 * basis; 0 if not.
9896 */
9897 static int
9898 dtrace_difo_cacheable(dtrace_difo_t *dp)
9899 {
9900 int i;
9901
9902 if (dp == NULL)
9903 return (0);
9904
9905 for (i = 0; i < dp->dtdo_varlen; i++) {
9906 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9907
9908 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9909 continue;
9910
9911 switch (v->dtdv_id) {
9912 case DIF_VAR_CURTHREAD:
9913 case DIF_VAR_PID:
9914 case DIF_VAR_TID:
9915 case DIF_VAR_EXECNAME:
9916 case DIF_VAR_ZONENAME:
9917 break;
9918
9919 default:
9920 return (0);
9921 }
9922 }
9923
9924 /*
9925 * This DIF object may be cacheable. Now we need to look for any
9926 * array loading instructions, any memory loading instructions, or
9927 * any stores to thread-local variables.
9928 */
9929 for (i = 0; i < dp->dtdo_len; i++) {
9930 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9931
9932 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9933 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9934 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9935 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9936 return (0);
9937 }
9938
9939 return (1);
9940 }
9941
9942 static void
9943 dtrace_difo_hold(dtrace_difo_t *dp)
9944 {
9945 int i;
9946
9947 ASSERT(MUTEX_HELD(&dtrace_lock));
9948
9949 dp->dtdo_refcnt++;
9950 ASSERT(dp->dtdo_refcnt != 0);
9951
9952 /*
9953 * We need to check this DIF object for references to the variable
9954 * DIF_VAR_VTIMESTAMP.
9955 */
9956 for (i = 0; i < dp->dtdo_varlen; i++) {
9957 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9958
9959 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9960 continue;
9961
9962 if (dtrace_vtime_references++ == 0)
9963 dtrace_vtime_enable();
9964 }
9965 }
9966
9967 /*
9968 * This routine calculates the dynamic variable chunksize for a given DIF
9969 * object. The calculation is not fool-proof, and can probably be tricked by
9970 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9971 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9972 * if a dynamic variable size exceeds the chunksize.
9973 */
9974 static void
9975 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9976 {
9977 uint64_t sval;
9978 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9979 const dif_instr_t *text = dp->dtdo_buf;
9980 uint_t pc, srd = 0;
9981 uint_t ttop = 0;
9982 size_t size, ksize;
9983 uint_t id, i;
9984
9985 for (pc = 0; pc < dp->dtdo_len; pc++) {
9986 dif_instr_t instr = text[pc];
9987 uint_t op = DIF_INSTR_OP(instr);
9988 uint_t rd = DIF_INSTR_RD(instr);
9989 uint_t r1 = DIF_INSTR_R1(instr);
9990 uint_t nkeys = 0;
9991 uchar_t scope;
9992
9993 dtrace_key_t *key = tupregs;
9994
9995 switch (op) {
9996 case DIF_OP_SETX:
9997 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9998 srd = rd;
9999 continue;
10000
10001 case DIF_OP_STTS:
10002 key = &tupregs[DIF_DTR_NREGS];
10003 key[0].dttk_size = 0;
10004 key[1].dttk_size = 0;
10005 nkeys = 2;
10006 scope = DIFV_SCOPE_THREAD;
10007 break;
10008
10009 case DIF_OP_STGAA:
10010 case DIF_OP_STTAA:
10011 nkeys = ttop;
10012
10013 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
10014 key[nkeys++].dttk_size = 0;
10015
10016 key[nkeys++].dttk_size = 0;
10017
10018 if (op == DIF_OP_STTAA) {
10019 scope = DIFV_SCOPE_THREAD;
10020 } else {
10021 scope = DIFV_SCOPE_GLOBAL;
10022 }
10023
10024 break;
10025
10026 case DIF_OP_PUSHTR:
10027 if (ttop == DIF_DTR_NREGS)
10028 return;
10029
10030 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
10031 /*
10032 * If the register for the size of the "pushtr"
10033 * is %r0 (or the value is 0) and the type is
10034 * a string, we'll use the system-wide default
10035 * string size.
10036 */
10037 tupregs[ttop++].dttk_size =
10038 dtrace_strsize_default;
10039 } else {
10040 if (srd == 0)
10041 return;
10042
10043 if (sval > LONG_MAX)
10044 return;
10045
10046 tupregs[ttop++].dttk_size = sval;
10047 }
10048
10049 break;
10050
10051 case DIF_OP_PUSHTV:
10052 if (ttop == DIF_DTR_NREGS)
10053 return;
10054
10055 tupregs[ttop++].dttk_size = 0;
10056 break;
10057
10058 case DIF_OP_FLUSHTS:
10059 ttop = 0;
10060 break;
10061
10062 case DIF_OP_POPTS:
10063 if (ttop != 0)
10064 ttop--;
10065 break;
10066 }
10067
10068 sval = 0;
10069 srd = 0;
10070
10071 if (nkeys == 0)
10072 continue;
10073
10074 /*
10075 * We have a dynamic variable allocation; calculate its size.
10076 */
10077 for (ksize = 0, i = 0; i < nkeys; i++)
10078 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
10079
10080 size = sizeof (dtrace_dynvar_t);
10081 size += sizeof (dtrace_key_t) * (nkeys - 1);
10082 size += ksize;
10083
10084 /*
10085 * Now we need to determine the size of the stored data.
10086 */
10087 id = DIF_INSTR_VAR(instr);
10088
10089 for (i = 0; i < dp->dtdo_varlen; i++) {
10090 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10091
10092 if (v->dtdv_id == id && v->dtdv_scope == scope) {
10093 size += v->dtdv_type.dtdt_size;
10094 break;
10095 }
10096 }
10097
10098 if (i == dp->dtdo_varlen)
10099 return;
10100
10101 /*
10102 * We have the size. If this is larger than the chunk size
10103 * for our dynamic variable state, reset the chunk size.
10104 */
10105 size = P2ROUNDUP(size, sizeof (uint64_t));
10106
10107 /*
10108 * Before setting the chunk size, check that we're not going
10109 * to set it to a negative value...
10110 */
10111 if (size > LONG_MAX)
10112 return;
10113
10114 /*
10115 * ...and make certain that we didn't badly overflow.
10116 */
10117 if (size < ksize || size < sizeof (dtrace_dynvar_t))
10118 return;
10119
10120 if (size > vstate->dtvs_dynvars.dtds_chunksize)
10121 vstate->dtvs_dynvars.dtds_chunksize = size;
10122 }
10123 }
10124
10125 static void
10126 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10127 {
10128 int i, oldsvars, osz, nsz, otlocals, ntlocals;
10129 uint_t id;
10130
10131 ASSERT(MUTEX_HELD(&dtrace_lock));
10132 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
10133
10134 for (i = 0; i < dp->dtdo_varlen; i++) {
10135 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10136 dtrace_statvar_t *svar, ***svarp;
10137 size_t dsize = 0;
10138 uint8_t scope = v->dtdv_scope;
10139 int *np;
10140
10141 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10142 continue;
10143
10144 id -= DIF_VAR_OTHER_UBASE;
10145
10146 switch (scope) {
10147 case DIFV_SCOPE_THREAD:
10148 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
10149 dtrace_difv_t *tlocals;
10150
10151 if ((ntlocals = (otlocals << 1)) == 0)
10152 ntlocals = 1;
10153
10154 osz = otlocals * sizeof (dtrace_difv_t);
10155 nsz = ntlocals * sizeof (dtrace_difv_t);
10156
10157 tlocals = kmem_zalloc(nsz, KM_SLEEP);
10158
10159 if (osz != 0) {
10160 bcopy(vstate->dtvs_tlocals,
10161 tlocals, osz);
10162 kmem_free(vstate->dtvs_tlocals, osz);
10163 }
10164
10165 vstate->dtvs_tlocals = tlocals;
10166 vstate->dtvs_ntlocals = ntlocals;
10167 }
10168
10169 vstate->dtvs_tlocals[id] = *v;
10170 continue;
10171
10172 case DIFV_SCOPE_LOCAL:
10173 np = &vstate->dtvs_nlocals;
10174 svarp = &vstate->dtvs_locals;
10175
10176 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10177 dsize = NCPU * (v->dtdv_type.dtdt_size +
10178 sizeof (uint64_t));
10179 else
10180 dsize = NCPU * sizeof (uint64_t);
10181
10182 break;
10183
10184 case DIFV_SCOPE_GLOBAL:
10185 np = &vstate->dtvs_nglobals;
10186 svarp = &vstate->dtvs_globals;
10187
10188 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10189 dsize = v->dtdv_type.dtdt_size +
10190 sizeof (uint64_t);
10191
10192 break;
10193
10194 default:
10195 ASSERT(0);
10196 }
10197
10198 while (id >= (oldsvars = *np)) {
10199 dtrace_statvar_t **statics;
10200 int newsvars, oldsize, newsize;
10201
10202 if ((newsvars = (oldsvars << 1)) == 0)
10203 newsvars = 1;
10204
10205 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
10206 newsize = newsvars * sizeof (dtrace_statvar_t *);
10207
10208 statics = kmem_zalloc(newsize, KM_SLEEP);
10209
10210 if (oldsize != 0) {
10211 bcopy(*svarp, statics, oldsize);
10212 kmem_free(*svarp, oldsize);
10213 }
10214
10215 *svarp = statics;
10216 *np = newsvars;
10217 }
10218
10219 if ((svar = (*svarp)[id]) == NULL) {
10220 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
10221 svar->dtsv_var = *v;
10222
10223 if ((svar->dtsv_size = dsize) != 0) {
10224 svar->dtsv_data = (uint64_t)(uintptr_t)
10225 kmem_zalloc(dsize, KM_SLEEP);
10226 }
10227
10228 (*svarp)[id] = svar;
10229 }
10230
10231 svar->dtsv_refcnt++;
10232 }
10233
10234 dtrace_difo_chunksize(dp, vstate);
10235 dtrace_difo_hold(dp);
10236 }
10237
10238 static dtrace_difo_t *
10239 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10240 {
10241 dtrace_difo_t *new;
10242 size_t sz;
10243
10244 ASSERT(dp->dtdo_buf != NULL);
10245 ASSERT(dp->dtdo_refcnt != 0);
10246
10247 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
10248
10249 ASSERT(dp->dtdo_buf != NULL);
10250 sz = dp->dtdo_len * sizeof (dif_instr_t);
10251 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
10252 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
10253 new->dtdo_len = dp->dtdo_len;
10254
10255 if (dp->dtdo_strtab != NULL) {
10256 ASSERT(dp->dtdo_strlen != 0);
10257 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
10258 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
10259 new->dtdo_strlen = dp->dtdo_strlen;
10260 }
10261
10262 if (dp->dtdo_inttab != NULL) {
10263 ASSERT(dp->dtdo_intlen != 0);
10264 sz = dp->dtdo_intlen * sizeof (uint64_t);
10265 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
10266 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
10267 new->dtdo_intlen = dp->dtdo_intlen;
10268 }
10269
10270 if (dp->dtdo_vartab != NULL) {
10271 ASSERT(dp->dtdo_varlen != 0);
10272 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
10273 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
10274 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
10275 new->dtdo_varlen = dp->dtdo_varlen;
10276 }
10277
10278 dtrace_difo_init(new, vstate);
10279 return (new);
10280 }
10281
10282 static void
10283 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10284 {
10285 int i;
10286
10287 ASSERT(dp->dtdo_refcnt == 0);
10288
10289 for (i = 0; i < dp->dtdo_varlen; i++) {
10290 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10291 dtrace_statvar_t *svar, **svarp;
10292 uint_t id;
10293 uint8_t scope = v->dtdv_scope;
10294 int *np;
10295
10296 switch (scope) {
10297 case DIFV_SCOPE_THREAD:
10298 continue;
10299
10300 case DIFV_SCOPE_LOCAL:
10301 np = &vstate->dtvs_nlocals;
10302 svarp = vstate->dtvs_locals;
10303 break;
10304
10305 case DIFV_SCOPE_GLOBAL:
10306 np = &vstate->dtvs_nglobals;
10307 svarp = vstate->dtvs_globals;
10308 break;
10309
10310 default:
10311 ASSERT(0);
10312 }
10313
10314 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10315 continue;
10316
10317 id -= DIF_VAR_OTHER_UBASE;
10318 ASSERT(id < *np);
10319
10320 svar = svarp[id];
10321 ASSERT(svar != NULL);
10322 ASSERT(svar->dtsv_refcnt > 0);
10323
10324 if (--svar->dtsv_refcnt > 0)
10325 continue;
10326
10327 if (svar->dtsv_size != 0) {
10328 ASSERT(svar->dtsv_data != 0);
10329 kmem_free((void *)(uintptr_t)svar->dtsv_data,
10330 svar->dtsv_size);
10331 }
10332
10333 kmem_free(svar, sizeof (dtrace_statvar_t));
10334 svarp[id] = NULL;
10335 }
10336
10337 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
10338 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
10339 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
10340 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
10341
10342 kmem_free(dp, sizeof (dtrace_difo_t));
10343 }
10344
10345 static void
10346 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10347 {
10348 int i;
10349
10350 ASSERT(MUTEX_HELD(&dtrace_lock));
10351 ASSERT(dp->dtdo_refcnt != 0);
10352
10353 for (i = 0; i < dp->dtdo_varlen; i++) {
10354 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10355
10356 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10357 continue;
10358
10359 ASSERT(dtrace_vtime_references > 0);
10360 if (--dtrace_vtime_references == 0)
10361 dtrace_vtime_disable();
10362 }
10363
10364 if (--dp->dtdo_refcnt == 0)
10365 dtrace_difo_destroy(dp, vstate);
10366 }
10367
10368 /*
10369 * DTrace Format Functions
10370 */
10371 static uint16_t
10372 dtrace_format_add(dtrace_state_t *state, char *str)
10373 {
10374 char *fmt, **new;
10375 uint16_t ndx, len = strlen(str) + 1;
10376
10377 fmt = kmem_zalloc(len, KM_SLEEP);
10378 bcopy(str, fmt, len);
10379
10380 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
10381 if (state->dts_formats[ndx] == NULL) {
10382 state->dts_formats[ndx] = fmt;
10383 return (ndx + 1);
10384 }
10385 }
10386
10387 if (state->dts_nformats == USHRT_MAX) {
10388 /*
10389 * This is only likely if a denial-of-service attack is being
10390 * attempted. As such, it's okay to fail silently here.
10391 */
10392 kmem_free(fmt, len);
10393 return (0);
10394 }
10395
10396 /*
10397 * For simplicity, we always resize the formats array to be exactly the
10398 * number of formats.
10399 */
10400 ndx = state->dts_nformats++;
10401 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10402
10403 if (state->dts_formats != NULL) {
10404 ASSERT(ndx != 0);
10405 bcopy(state->dts_formats, new, ndx * sizeof (char *));
10406 kmem_free(state->dts_formats, ndx * sizeof (char *));
10407 }
10408
10409 state->dts_formats = new;
10410 state->dts_formats[ndx] = fmt;
10411
10412 return (ndx + 1);
10413 }
10414
10415 static void
10416 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10417 {
10418 char *fmt;
10419
10420 ASSERT(state->dts_formats != NULL);
10421 ASSERT(format <= state->dts_nformats);
10422 ASSERT(state->dts_formats[format - 1] != NULL);
10423
10424 fmt = state->dts_formats[format - 1];
10425 kmem_free(fmt, strlen(fmt) + 1);
10426 state->dts_formats[format - 1] = NULL;
10427 }
10428
10429 static void
10430 dtrace_format_destroy(dtrace_state_t *state)
10431 {
10432 int i;
10433
10434 if (state->dts_nformats == 0) {
10435 ASSERT(state->dts_formats == NULL);
10436 return;
10437 }
10438
10439 ASSERT(state->dts_formats != NULL);
10440
10441 for (i = 0; i < state->dts_nformats; i++) {
10442 char *fmt = state->dts_formats[i];
10443
10444 if (fmt == NULL)
10445 continue;
10446
10447 kmem_free(fmt, strlen(fmt) + 1);
10448 }
10449
10450 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10451 state->dts_nformats = 0;
10452 state->dts_formats = NULL;
10453 }
10454
10455 /*
10456 * DTrace Predicate Functions
10457 */
10458 static dtrace_predicate_t *
10459 dtrace_predicate_create(dtrace_difo_t *dp)
10460 {
10461 dtrace_predicate_t *pred;
10462
10463 ASSERT(MUTEX_HELD(&dtrace_lock));
10464 ASSERT(dp->dtdo_refcnt != 0);
10465
10466 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10467 pred->dtp_difo = dp;
10468 pred->dtp_refcnt = 1;
10469
10470 if (!dtrace_difo_cacheable(dp))
10471 return (pred);
10472
10473 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10474 /*
10475 * This is only theoretically possible -- we have had 2^32
10476 * cacheable predicates on this machine. We cannot allow any
10477 * more predicates to become cacheable: as unlikely as it is,
10478 * there may be a thread caching a (now stale) predicate cache
10479 * ID. (N.B.: the temptation is being successfully resisted to
10480 * have this cmn_err() "Holy shit -- we executed this code!")
10481 */
10482 return (pred);
10483 }
10484
10485 pred->dtp_cacheid = dtrace_predcache_id++;
10486
10487 return (pred);
10488 }
10489
10490 static void
10491 dtrace_predicate_hold(dtrace_predicate_t *pred)
10492 {
10493 ASSERT(MUTEX_HELD(&dtrace_lock));
10494 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10495 ASSERT(pred->dtp_refcnt > 0);
10496
10497 pred->dtp_refcnt++;
10498 }
10499
10500 static void
10501 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10502 {
10503 dtrace_difo_t *dp = pred->dtp_difo;
10504
10505 ASSERT(MUTEX_HELD(&dtrace_lock));
10506 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
10507 ASSERT(pred->dtp_refcnt > 0);
10508
10509 if (--pred->dtp_refcnt == 0) {
10510 dtrace_difo_release(pred->dtp_difo, vstate);
10511 kmem_free(pred, sizeof (dtrace_predicate_t));
10512 }
10513 }
10514
10515 /*
10516 * DTrace Action Description Functions
10517 */
10518 static dtrace_actdesc_t *
10519 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
10520 uint64_t uarg, uint64_t arg)
10521 {
10522 dtrace_actdesc_t *act;
10523
10524 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != 0 &&
10525 arg >= KERNELBASE) || (arg == 0 && kind == DTRACEACT_PRINTA));
10526
10527 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
10528 act->dtad_kind = kind;
10529 act->dtad_ntuple = ntuple;
10530 act->dtad_uarg = uarg;
10531 act->dtad_arg = arg;
10532 act->dtad_refcnt = 1;
10533
10534 return (act);
10535 }
10536
10537 static void
10538 dtrace_actdesc_hold(dtrace_actdesc_t *act)
10539 {
10540 ASSERT(act->dtad_refcnt >= 1);
10541 act->dtad_refcnt++;
10542 }
10543
10544 static void
10545 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
10546 {
10547 dtrace_actkind_t kind = act->dtad_kind;
10548 dtrace_difo_t *dp;
10549
10550 ASSERT(act->dtad_refcnt >= 1);
10551
10552 if (--act->dtad_refcnt != 0)
10553 return;
10554
10555 if ((dp = act->dtad_difo) != NULL)
10556 dtrace_difo_release(dp, vstate);
10557
10558 if (DTRACEACT_ISPRINTFLIKE(kind)) {
10559 char *str = (char *)(uintptr_t)act->dtad_arg;
10560
10561 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
10562 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
10563
10564 if (str != NULL)
10565 kmem_free(str, strlen(str) + 1);
10566 }
10567
10568 kmem_free(act, sizeof (dtrace_actdesc_t));
10569 }
10570
10571 /*
10572 * DTrace ECB Functions
10573 */
10574 static dtrace_ecb_t *
10575 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
10576 {
10577 dtrace_ecb_t *ecb;
10578 dtrace_epid_t epid;
10579
10580 ASSERT(MUTEX_HELD(&dtrace_lock));
10581
10582 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
10583 ecb->dte_predicate = NULL;
10584 ecb->dte_probe = probe;
10585
10586 /*
10587 * The default size is the size of the default action: recording
10588 * the header.
10589 */
10590 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
10591 ecb->dte_alignment = sizeof (dtrace_epid_t);
10592
10593 epid = state->dts_epid++;
10594
10595 if (epid - 1 >= state->dts_necbs) {
10596 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
10597 int necbs = state->dts_necbs << 1;
10598
10599 ASSERT(epid == state->dts_necbs + 1);
10600
10601 if (necbs == 0) {
10602 ASSERT(oecbs == NULL);
10603 necbs = 1;
10604 }
10605
10606 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
10607
10608 if (oecbs != NULL)
10609 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
10610
10611 dtrace_membar_producer();
10612 state->dts_ecbs = ecbs;
10613
10614 if (oecbs != NULL) {
10615 /*
10616 * If this state is active, we must dtrace_sync()
10617 * before we can free the old dts_ecbs array: we're
10618 * coming in hot, and there may be active ring
10619 * buffer processing (which indexes into the dts_ecbs
10620 * array) on another CPU.
10621 */
10622 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
10623 dtrace_sync();
10624
10625 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
10626 }
10627
10628 dtrace_membar_producer();
10629 state->dts_necbs = necbs;
10630 }
10631
10632 ecb->dte_state = state;
10633
10634 ASSERT(state->dts_ecbs[epid - 1] == NULL);
10635 dtrace_membar_producer();
10636 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
10637
10638 return (ecb);
10639 }
10640
10641 static int
10642 dtrace_ecb_enable(dtrace_ecb_t *ecb)
10643 {
10644 dtrace_probe_t *probe = ecb->dte_probe;
10645
10646 ASSERT(MUTEX_HELD(&cpu_lock));
10647 ASSERT(MUTEX_HELD(&dtrace_lock));
10648 ASSERT(ecb->dte_next == NULL);
10649
10650 if (probe == NULL) {
10651 /*
10652 * This is the NULL probe -- there's nothing to do.
10653 */
10654 return (0);
10655 }
10656
10657 if (probe->dtpr_ecb == NULL) {
10658 dtrace_provider_t *prov = probe->dtpr_provider;
10659
10660 /*
10661 * We're the first ECB on this probe.
10662 */
10663 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
10664
10665 if (ecb->dte_predicate != NULL)
10666 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
10667
10668 return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
10669 probe->dtpr_id, probe->dtpr_arg));
10670 } else {
10671 /*
10672 * This probe is already active. Swing the last pointer to
10673 * point to the new ECB, and issue a dtrace_sync() to assure
10674 * that all CPUs have seen the change.
10675 */
10676 ASSERT(probe->dtpr_ecb_last != NULL);
10677 probe->dtpr_ecb_last->dte_next = ecb;
10678 probe->dtpr_ecb_last = ecb;
10679 probe->dtpr_predcache = 0;
10680
10681 dtrace_sync();
10682 return (0);
10683 }
10684 }
10685
10686 static int
10687 dtrace_ecb_resize(dtrace_ecb_t *ecb)
10688 {
10689 dtrace_action_t *act;
10690 uint32_t curneeded = UINT32_MAX;
10691 uint32_t aggbase = UINT32_MAX;
10692
10693 /*
10694 * If we record anything, we always record the dtrace_rechdr_t. (And
10695 * we always record it first.)
10696 */
10697 ecb->dte_size = sizeof (dtrace_rechdr_t);
10698 ecb->dte_alignment = sizeof (dtrace_epid_t);
10699
10700 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10701 dtrace_recdesc_t *rec = &act->dta_rec;
10702 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
10703
10704 ecb->dte_alignment = MAX(ecb->dte_alignment,
10705 rec->dtrd_alignment);
10706
10707 if (DTRACEACT_ISAGG(act->dta_kind)) {
10708 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10709
10710 ASSERT(rec->dtrd_size != 0);
10711 ASSERT(agg->dtag_first != NULL);
10712 ASSERT(act->dta_prev->dta_intuple);
10713 ASSERT(aggbase != UINT32_MAX);
10714 ASSERT(curneeded != UINT32_MAX);
10715
10716 agg->dtag_base = aggbase;
10717
10718 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10719 rec->dtrd_offset = curneeded;
10720 if (curneeded + rec->dtrd_size < curneeded)
10721 return (EINVAL);
10722 curneeded += rec->dtrd_size;
10723 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
10724
10725 aggbase = UINT32_MAX;
10726 curneeded = UINT32_MAX;
10727 } else if (act->dta_intuple) {
10728 if (curneeded == UINT32_MAX) {
10729 /*
10730 * This is the first record in a tuple. Align
10731 * curneeded to be at offset 4 in an 8-byte
10732 * aligned block.
10733 */
10734 ASSERT(act->dta_prev == NULL ||
10735 !act->dta_prev->dta_intuple);
10736 ASSERT3U(aggbase, ==, UINT32_MAX);
10737 curneeded = P2PHASEUP(ecb->dte_size,
10738 sizeof (uint64_t), sizeof (dtrace_aggid_t));
10739
10740 aggbase = curneeded - sizeof (dtrace_aggid_t);
10741 ASSERT(IS_P2ALIGNED(aggbase,
10742 sizeof (uint64_t)));
10743 }
10744 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10745 rec->dtrd_offset = curneeded;
10746 if (curneeded + rec->dtrd_size < curneeded)
10747 return (EINVAL);
10748 curneeded += rec->dtrd_size;
10749 } else {
10750 /* tuples must be followed by an aggregation */
10751 ASSERT(act->dta_prev == NULL ||
10752 !act->dta_prev->dta_intuple);
10753
10754 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10755 rec->dtrd_alignment);
10756 rec->dtrd_offset = ecb->dte_size;
10757 if (ecb->dte_size + rec->dtrd_size < ecb->dte_size)
10758 return (EINVAL);
10759 ecb->dte_size += rec->dtrd_size;
10760 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10761 }
10762 }
10763
10764 if ((act = ecb->dte_action) != NULL &&
10765 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10766 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10767 /*
10768 * If the size is still sizeof (dtrace_rechdr_t), then all
10769 * actions store no data; set the size to 0.
10770 */
10771 ecb->dte_size = 0;
10772 }
10773
10774 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10775 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10776 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10777 ecb->dte_needed);
10778 return (0);
10779 }
10780
10781 static dtrace_action_t *
10782 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10783 {
10784 dtrace_aggregation_t *agg;
10785 size_t size = sizeof (uint64_t);
10786 int ntuple = desc->dtad_ntuple;
10787 dtrace_action_t *act;
10788 dtrace_recdesc_t *frec;
10789 dtrace_aggid_t aggid;
10790 dtrace_state_t *state = ecb->dte_state;
10791
10792 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10793 agg->dtag_ecb = ecb;
10794
10795 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10796
10797 switch (desc->dtad_kind) {
10798 case DTRACEAGG_MIN:
10799 agg->dtag_initial = INT64_MAX;
10800 agg->dtag_aggregate = dtrace_aggregate_min;
10801 break;
10802
10803 case DTRACEAGG_MAX:
10804 agg->dtag_initial = INT64_MIN;
10805 agg->dtag_aggregate = dtrace_aggregate_max;
10806 break;
10807
10808 case DTRACEAGG_COUNT:
10809 agg->dtag_aggregate = dtrace_aggregate_count;
10810 break;
10811
10812 case DTRACEAGG_QUANTIZE:
10813 agg->dtag_aggregate = dtrace_aggregate_quantize;
10814 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10815 sizeof (uint64_t);
10816 break;
10817
10818 case DTRACEAGG_LQUANTIZE: {
10819 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10820 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10821
10822 agg->dtag_initial = desc->dtad_arg;
10823 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10824
10825 if (step == 0 || levels == 0)
10826 goto err;
10827
10828 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10829 break;
10830 }
10831
10832 case DTRACEAGG_LLQUANTIZE: {
10833 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10834 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10835 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10836 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10837 int64_t v;
10838
10839 agg->dtag_initial = desc->dtad_arg;
10840 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10841
10842 if (factor < 2 || low >= high || nsteps < factor)
10843 goto err;
10844
10845 /*
10846 * Now check that the number of steps evenly divides a power
10847 * of the factor. (This assures both integer bucket size and
10848 * linearity within each magnitude.)
10849 */
10850 for (v = factor; v < nsteps; v *= factor)
10851 continue;
10852
10853 if ((v % nsteps) || (nsteps % factor))
10854 goto err;
10855
10856 size = (dtrace_aggregate_llquantize_bucket(factor,
10857 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10858 break;
10859 }
10860
10861 case DTRACEAGG_AVG:
10862 agg->dtag_aggregate = dtrace_aggregate_avg;
10863 size = sizeof (uint64_t) * 2;
10864 break;
10865
10866 case DTRACEAGG_STDDEV:
10867 agg->dtag_aggregate = dtrace_aggregate_stddev;
10868 size = sizeof (uint64_t) * 4;
10869 break;
10870
10871 case DTRACEAGG_SUM:
10872 agg->dtag_aggregate = dtrace_aggregate_sum;
10873 break;
10874
10875 default:
10876 goto err;
10877 }
10878
10879 agg->dtag_action.dta_rec.dtrd_size = size;
10880
10881 if (ntuple == 0)
10882 goto err;
10883
10884 /*
10885 * We must make sure that we have enough actions for the n-tuple.
10886 */
10887 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10888 if (DTRACEACT_ISAGG(act->dta_kind))
10889 break;
10890
10891 if (--ntuple == 0) {
10892 /*
10893 * This is the action with which our n-tuple begins.
10894 */
10895 agg->dtag_first = act;
10896 goto success;
10897 }
10898 }
10899
10900 /*
10901 * This n-tuple is short by ntuple elements. Return failure.
10902 */
10903 ASSERT(ntuple != 0);
10904 err:
10905 kmem_free(agg, sizeof (dtrace_aggregation_t));
10906 return (NULL);
10907
10908 success:
10909 /*
10910 * If the last action in the tuple has a size of zero, it's actually
10911 * an expression argument for the aggregating action.
10912 */
10913 ASSERT(ecb->dte_action_last != NULL);
10914 act = ecb->dte_action_last;
10915
10916 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10917 ASSERT(act->dta_difo != NULL);
10918
10919 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10920 agg->dtag_hasarg = 1;
10921 }
10922
10923 /*
10924 * We need to allocate an id for this aggregation.
10925 */
10926 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10927 VM_BESTFIT | VM_SLEEP);
10928
10929 if (aggid - 1 >= state->dts_naggregations) {
10930 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10931 dtrace_aggregation_t **aggs;
10932 int naggs = state->dts_naggregations << 1;
10933 int onaggs = state->dts_naggregations;
10934
10935 ASSERT(aggid == state->dts_naggregations + 1);
10936
10937 if (naggs == 0) {
10938 ASSERT(oaggs == NULL);
10939 naggs = 1;
10940 }
10941
10942 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10943
10944 if (oaggs != NULL) {
10945 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10946 kmem_free(oaggs, onaggs * sizeof (*aggs));
10947 }
10948
10949 state->dts_aggregations = aggs;
10950 state->dts_naggregations = naggs;
10951 }
10952
10953 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10954 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10955
10956 frec = &agg->dtag_first->dta_rec;
10957 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10958 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10959
10960 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10961 ASSERT(!act->dta_intuple);
10962 act->dta_intuple = 1;
10963 }
10964
10965 return (&agg->dtag_action);
10966 }
10967
10968 static void
10969 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10970 {
10971 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10972 dtrace_state_t *state = ecb->dte_state;
10973 dtrace_aggid_t aggid = agg->dtag_id;
10974
10975 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10976 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10977
10978 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10979 state->dts_aggregations[aggid - 1] = NULL;
10980
10981 kmem_free(agg, sizeof (dtrace_aggregation_t));
10982 }
10983
10984 static int
10985 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10986 {
10987 dtrace_action_t *action, *last;
10988 dtrace_difo_t *dp = desc->dtad_difo;
10989 uint32_t size = 0, align = sizeof (uint8_t), mask;
10990 uint16_t format = 0;
10991 dtrace_recdesc_t *rec;
10992 dtrace_state_t *state = ecb->dte_state;
10993 dtrace_optval_t *opt = state->dts_options, nframes, strsize;
10994 uint64_t arg = desc->dtad_arg;
10995
10996 ASSERT(MUTEX_HELD(&dtrace_lock));
10997 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10998
10999 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
11000 /*
11001 * If this is an aggregating action, there must be neither
11002 * a speculate nor a commit on the action chain.
11003 */
11004 dtrace_action_t *act;
11005
11006 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11007 if (act->dta_kind == DTRACEACT_COMMIT)
11008 return (EINVAL);
11009
11010 if (act->dta_kind == DTRACEACT_SPECULATE)
11011 return (EINVAL);
11012 }
11013
11014 action = dtrace_ecb_aggregation_create(ecb, desc);
11015
11016 if (action == NULL)
11017 return (EINVAL);
11018 } else {
11019 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
11020 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
11021 dp != NULL && dp->dtdo_destructive)) {
11022 state->dts_destructive = 1;
11023 }
11024
11025 switch (desc->dtad_kind) {
11026 case DTRACEACT_PRINTF:
11027 case DTRACEACT_PRINTA:
11028 case DTRACEACT_SYSTEM:
11029 case DTRACEACT_FREOPEN:
11030 case DTRACEACT_DIFEXPR:
11031 /*
11032 * We know that our arg is a string -- turn it into a
11033 * format.
11034 */
11035 if (arg == 0) {
11036 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
11037 desc->dtad_kind == DTRACEACT_DIFEXPR);
11038 format = 0;
11039 } else {
11040 ASSERT(arg != 0);
11041 ASSERT(arg > KERNELBASE);
11042 format = dtrace_format_add(state,
11043 (char *)(uintptr_t)arg);
11044 }
11045
11046 /*FALLTHROUGH*/
11047 case DTRACEACT_LIBACT:
11048 case DTRACEACT_TRACEMEM:
11049 case DTRACEACT_TRACEMEM_DYNSIZE:
11050 if (dp == NULL)
11051 return (EINVAL);
11052
11053 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
11054 break;
11055
11056 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
11057 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11058 return (EINVAL);
11059
11060 size = opt[DTRACEOPT_STRSIZE];
11061 }
11062
11063 break;
11064
11065 case DTRACEACT_STACK:
11066 if ((nframes = arg) == 0) {
11067 nframes = opt[DTRACEOPT_STACKFRAMES];
11068 ASSERT(nframes > 0);
11069 arg = nframes;
11070 }
11071
11072 size = nframes * sizeof (pc_t);
11073 break;
11074
11075 case DTRACEACT_JSTACK:
11076 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
11077 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
11078
11079 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
11080 nframes = opt[DTRACEOPT_JSTACKFRAMES];
11081
11082 arg = DTRACE_USTACK_ARG(nframes, strsize);
11083
11084 /*FALLTHROUGH*/
11085 case DTRACEACT_USTACK:
11086 if (desc->dtad_kind != DTRACEACT_JSTACK &&
11087 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
11088 strsize = DTRACE_USTACK_STRSIZE(arg);
11089 nframes = opt[DTRACEOPT_USTACKFRAMES];
11090 ASSERT(nframes > 0);
11091 arg = DTRACE_USTACK_ARG(nframes, strsize);
11092 }
11093
11094 /*
11095 * Save a slot for the pid.
11096 */
11097 size = (nframes + 1) * sizeof (uint64_t);
11098 size += DTRACE_USTACK_STRSIZE(arg);
11099 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
11100
11101 break;
11102
11103 case DTRACEACT_SYM:
11104 case DTRACEACT_MOD:
11105 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
11106 sizeof (uint64_t)) ||
11107 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11108 return (EINVAL);
11109 break;
11110
11111 case DTRACEACT_USYM:
11112 case DTRACEACT_UMOD:
11113 case DTRACEACT_UADDR:
11114 if (dp == NULL ||
11115 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
11116 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11117 return (EINVAL);
11118
11119 /*
11120 * We have a slot for the pid, plus a slot for the
11121 * argument. To keep things simple (aligned with
11122 * bitness-neutral sizing), we store each as a 64-bit
11123 * quantity.
11124 */
11125 size = 2 * sizeof (uint64_t);
11126 break;
11127
11128 case DTRACEACT_STOP:
11129 case DTRACEACT_BREAKPOINT:
11130 case DTRACEACT_PANIC:
11131 break;
11132
11133 case DTRACEACT_CHILL:
11134 case DTRACEACT_DISCARD:
11135 case DTRACEACT_RAISE:
11136 if (dp == NULL)
11137 return (EINVAL);
11138 break;
11139
11140 case DTRACEACT_EXIT:
11141 if (dp == NULL ||
11142 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
11143 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11144 return (EINVAL);
11145 break;
11146
11147 case DTRACEACT_SPECULATE:
11148 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
11149 return (EINVAL);
11150
11151 if (dp == NULL)
11152 return (EINVAL);
11153
11154 state->dts_speculates = 1;
11155 break;
11156
11157 case DTRACEACT_COMMIT: {
11158 dtrace_action_t *act = ecb->dte_action;
11159
11160 for (; act != NULL; act = act->dta_next) {
11161 if (act->dta_kind == DTRACEACT_COMMIT)
11162 return (EINVAL);
11163 }
11164
11165 if (dp == NULL)
11166 return (EINVAL);
11167 break;
11168 }
11169
11170 default:
11171 return (EINVAL);
11172 }
11173
11174 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
11175 /*
11176 * If this is a data-storing action or a speculate,
11177 * we must be sure that there isn't a commit on the
11178 * action chain.
11179 */
11180 dtrace_action_t *act = ecb->dte_action;
11181
11182 for (; act != NULL; act = act->dta_next) {
11183 if (act->dta_kind == DTRACEACT_COMMIT)
11184 return (EINVAL);
11185 }
11186 }
11187
11188 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
11189 action->dta_rec.dtrd_size = size;
11190 }
11191
11192 action->dta_refcnt = 1;
11193 rec = &action->dta_rec;
11194 size = rec->dtrd_size;
11195
11196 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
11197 if (!(size & mask)) {
11198 align = mask + 1;
11199 break;
11200 }
11201 }
11202
11203 action->dta_kind = desc->dtad_kind;
11204
11205 if ((action->dta_difo = dp) != NULL)
11206 dtrace_difo_hold(dp);
11207
11208 rec->dtrd_action = action->dta_kind;
11209 rec->dtrd_arg = arg;
11210 rec->dtrd_uarg = desc->dtad_uarg;
11211 rec->dtrd_alignment = (uint16_t)align;
11212 rec->dtrd_format = format;
11213
11214 if ((last = ecb->dte_action_last) != NULL) {
11215 ASSERT(ecb->dte_action != NULL);
11216 action->dta_prev = last;
11217 last->dta_next = action;
11218 } else {
11219 ASSERT(ecb->dte_action == NULL);
11220 ecb->dte_action = action;
11221 }
11222
11223 ecb->dte_action_last = action;
11224
11225 return (0);
11226 }
11227
11228 static void
11229 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
11230 {
11231 dtrace_action_t *act = ecb->dte_action, *next;
11232 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
11233 dtrace_difo_t *dp;
11234 uint16_t format;
11235
11236 if (act != NULL && act->dta_refcnt > 1) {
11237 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
11238 act->dta_refcnt--;
11239 } else {
11240 for (; act != NULL; act = next) {
11241 next = act->dta_next;
11242 ASSERT(next != NULL || act == ecb->dte_action_last);
11243 ASSERT(act->dta_refcnt == 1);
11244
11245 if ((format = act->dta_rec.dtrd_format) != 0)
11246 dtrace_format_remove(ecb->dte_state, format);
11247
11248 if ((dp = act->dta_difo) != NULL)
11249 dtrace_difo_release(dp, vstate);
11250
11251 if (DTRACEACT_ISAGG(act->dta_kind)) {
11252 dtrace_ecb_aggregation_destroy(ecb, act);
11253 } else {
11254 kmem_free(act, sizeof (dtrace_action_t));
11255 }
11256 }
11257 }
11258
11259 ecb->dte_action = NULL;
11260 ecb->dte_action_last = NULL;
11261 ecb->dte_size = 0;
11262 }
11263
11264 static void
11265 dtrace_ecb_disable(dtrace_ecb_t *ecb)
11266 {
11267 /*
11268 * We disable the ECB by removing it from its probe.
11269 */
11270 dtrace_ecb_t *pecb, *prev = NULL;
11271 dtrace_probe_t *probe = ecb->dte_probe;
11272
11273 ASSERT(MUTEX_HELD(&dtrace_lock));
11274
11275 if (probe == NULL) {
11276 /*
11277 * This is the NULL probe; there is nothing to disable.
11278 */
11279 return;
11280 }
11281
11282 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
11283 if (pecb == ecb)
11284 break;
11285 prev = pecb;
11286 }
11287
11288 ASSERT(pecb != NULL);
11289
11290 if (prev == NULL) {
11291 probe->dtpr_ecb = ecb->dte_next;
11292 } else {
11293 prev->dte_next = ecb->dte_next;
11294 }
11295
11296 if (ecb == probe->dtpr_ecb_last) {
11297 ASSERT(ecb->dte_next == NULL);
11298 probe->dtpr_ecb_last = prev;
11299 }
11300
11301 /*
11302 * The ECB has been disconnected from the probe; now sync to assure
11303 * that all CPUs have seen the change before returning.
11304 */
11305 dtrace_sync();
11306
11307 if (probe->dtpr_ecb == NULL) {
11308 /*
11309 * That was the last ECB on the probe; clear the predicate
11310 * cache ID for the probe, disable it and sync one more time
11311 * to assure that we'll never hit it again.
11312 */
11313 dtrace_provider_t *prov = probe->dtpr_provider;
11314
11315 ASSERT(ecb->dte_next == NULL);
11316 ASSERT(probe->dtpr_ecb_last == NULL);
11317 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
11318 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
11319 probe->dtpr_id, probe->dtpr_arg);
11320 dtrace_sync();
11321 } else {
11322 /*
11323 * There is at least one ECB remaining on the probe. If there
11324 * is _exactly_ one, set the probe's predicate cache ID to be
11325 * the predicate cache ID of the remaining ECB.
11326 */
11327 ASSERT(probe->dtpr_ecb_last != NULL);
11328 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
11329
11330 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
11331 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
11332
11333 ASSERT(probe->dtpr_ecb->dte_next == NULL);
11334
11335 if (p != NULL)
11336 probe->dtpr_predcache = p->dtp_cacheid;
11337 }
11338
11339 ecb->dte_next = NULL;
11340 }
11341 }
11342
11343 static void
11344 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
11345 {
11346 dtrace_state_t *state = ecb->dte_state;
11347 dtrace_vstate_t *vstate = &state->dts_vstate;
11348 dtrace_predicate_t *pred;
11349 dtrace_epid_t epid = ecb->dte_epid;
11350
11351 ASSERT(MUTEX_HELD(&dtrace_lock));
11352 ASSERT(ecb->dte_next == NULL);
11353 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
11354
11355 if ((pred = ecb->dte_predicate) != NULL)
11356 dtrace_predicate_release(pred, vstate);
11357
11358 dtrace_ecb_action_remove(ecb);
11359
11360 ASSERT(state->dts_ecbs[epid - 1] == ecb);
11361 state->dts_ecbs[epid - 1] = NULL;
11362
11363 kmem_free(ecb, sizeof (dtrace_ecb_t));
11364 }
11365
11366 static dtrace_ecb_t *
11367 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
11368 dtrace_enabling_t *enab)
11369 {
11370 dtrace_ecb_t *ecb;
11371 dtrace_predicate_t *pred;
11372 dtrace_actdesc_t *act;
11373 dtrace_provider_t *prov;
11374 dtrace_ecbdesc_t *desc = enab->dten_current;
11375
11376 ASSERT(MUTEX_HELD(&dtrace_lock));
11377 ASSERT(state != NULL);
11378
11379 ecb = dtrace_ecb_add(state, probe);
11380 ecb->dte_uarg = desc->dted_uarg;
11381
11382 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
11383 dtrace_predicate_hold(pred);
11384 ecb->dte_predicate = pred;
11385 }
11386
11387 if (probe != NULL) {
11388 /*
11389 * If the provider shows more leg than the consumer is old
11390 * enough to see, we need to enable the appropriate implicit
11391 * predicate bits to prevent the ecb from activating at
11392 * revealing times.
11393 *
11394 * Providers specifying DTRACE_PRIV_USER at register time
11395 * are stating that they need the /proc-style privilege
11396 * model to be enforced, and this is what DTRACE_COND_OWNER
11397 * and DTRACE_COND_ZONEOWNER will then do at probe time.
11398 */
11399 prov = probe->dtpr_provider;
11400 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
11401 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11402 ecb->dte_cond |= DTRACE_COND_OWNER;
11403
11404 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11405 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11406 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11407
11408 /*
11409 * If the provider shows us kernel innards and the user
11410 * is lacking sufficient privilege, enable the
11411 * DTRACE_COND_USERMODE implicit predicate.
11412 */
11413 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11414 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11415 ecb->dte_cond |= DTRACE_COND_USERMODE;
11416 }
11417
11418 if (dtrace_ecb_create_cache != NULL) {
11419 /*
11420 * If we have a cached ecb, we'll use its action list instead
11421 * of creating our own (saving both time and space).
11422 */
11423 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11424 dtrace_action_t *act = cached->dte_action;
11425
11426 if (act != NULL) {
11427 ASSERT(act->dta_refcnt > 0);
11428 act->dta_refcnt++;
11429 ecb->dte_action = act;
11430 ecb->dte_action_last = cached->dte_action_last;
11431 ecb->dte_needed = cached->dte_needed;
11432 ecb->dte_size = cached->dte_size;
11433 ecb->dte_alignment = cached->dte_alignment;
11434 }
11435
11436 return (ecb);
11437 }
11438
11439 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11440 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11441 dtrace_ecb_destroy(ecb);
11442 return (NULL);
11443 }
11444 }
11445
11446 if ((enab->dten_error = dtrace_ecb_resize(ecb)) != 0) {
11447 dtrace_ecb_destroy(ecb);
11448 return (NULL);
11449 }
11450
11451 return (dtrace_ecb_create_cache = ecb);
11452 }
11453
11454 static int
11455 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11456 {
11457 dtrace_ecb_t *ecb;
11458 dtrace_enabling_t *enab = arg;
11459 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11460
11461 ASSERT(state != NULL);
11462
11463 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11464 /*
11465 * This probe was created in a generation for which this
11466 * enabling has previously created ECBs; we don't want to
11467 * enable it again, so just kick out.
11468 */
11469 return (DTRACE_MATCH_NEXT);
11470 }
11471
11472 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11473 return (DTRACE_MATCH_DONE);
11474
11475 if (dtrace_ecb_enable(ecb) < 0)
11476 return (DTRACE_MATCH_FAIL);
11477
11478 return (DTRACE_MATCH_NEXT);
11479 }
11480
11481 static dtrace_ecb_t *
11482 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11483 {
11484 dtrace_ecb_t *ecb;
11485
11486 ASSERT(MUTEX_HELD(&dtrace_lock));
11487
11488 if (id == 0 || id > state->dts_necbs)
11489 return (NULL);
11490
11491 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
11492 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
11493
11494 return (state->dts_ecbs[id - 1]);
11495 }
11496
11497 static dtrace_aggregation_t *
11498 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
11499 {
11500 dtrace_aggregation_t *agg;
11501
11502 ASSERT(MUTEX_HELD(&dtrace_lock));
11503
11504 if (id == 0 || id > state->dts_naggregations)
11505 return (NULL);
11506
11507 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
11508 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
11509 agg->dtag_id == id);
11510
11511 return (state->dts_aggregations[id - 1]);
11512 }
11513
11514 /*
11515 * DTrace Buffer Functions
11516 *
11517 * The following functions manipulate DTrace buffers. Most of these functions
11518 * are called in the context of establishing or processing consumer state;
11519 * exceptions are explicitly noted.
11520 */
11521
11522 /*
11523 * Note: called from cross call context. This function switches the two
11524 * buffers on a given CPU. The atomicity of this operation is assured by
11525 * disabling interrupts while the actual switch takes place; the disabling of
11526 * interrupts serializes the execution with any execution of dtrace_probe() on
11527 * the same CPU.
11528 */
11529 static void
11530 dtrace_buffer_switch(dtrace_buffer_t *buf)
11531 {
11532 caddr_t tomax = buf->dtb_tomax;
11533 caddr_t xamot = buf->dtb_xamot;
11534 dtrace_icookie_t cookie;
11535 hrtime_t now;
11536
11537 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11538 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
11539
11540 cookie = dtrace_interrupt_disable();
11541 now = dtrace_gethrtime();
11542 buf->dtb_tomax = xamot;
11543 buf->dtb_xamot = tomax;
11544 buf->dtb_xamot_drops = buf->dtb_drops;
11545 buf->dtb_xamot_offset = buf->dtb_offset;
11546 buf->dtb_xamot_errors = buf->dtb_errors;
11547 buf->dtb_xamot_flags = buf->dtb_flags;
11548 buf->dtb_offset = 0;
11549 buf->dtb_drops = 0;
11550 buf->dtb_errors = 0;
11551 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
11552 buf->dtb_interval = now - buf->dtb_switched;
11553 buf->dtb_switched = now;
11554 dtrace_interrupt_enable(cookie);
11555 }
11556
11557 /*
11558 * Note: called from cross call context. This function activates a buffer
11559 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
11560 * is guaranteed by the disabling of interrupts.
11561 */
11562 static void
11563 dtrace_buffer_activate(dtrace_state_t *state)
11564 {
11565 dtrace_buffer_t *buf;
11566 dtrace_icookie_t cookie = dtrace_interrupt_disable();
11567
11568 buf = &state->dts_buffer[CPU->cpu_id];
11569
11570 if (buf->dtb_tomax != NULL) {
11571 /*
11572 * We might like to assert that the buffer is marked inactive,
11573 * but this isn't necessarily true: the buffer for the CPU
11574 * that processes the BEGIN probe has its buffer activated
11575 * manually. In this case, we take the (harmless) action
11576 * re-clearing the bit INACTIVE bit.
11577 */
11578 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
11579 }
11580
11581 dtrace_interrupt_enable(cookie);
11582 }
11583
11584 static int
11585 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
11586 processorid_t cpu, int *factor)
11587 {
11588 cpu_t *cp;
11589 dtrace_buffer_t *buf;
11590 int allocated = 0, desired = 0;
11591
11592 ASSERT(MUTEX_HELD(&cpu_lock));
11593 ASSERT(MUTEX_HELD(&dtrace_lock));
11594
11595 *factor = 1;
11596
11597 if (size > dtrace_nonroot_maxsize &&
11598 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
11599 return (EFBIG);
11600
11601 cp = cpu_list;
11602
11603 do {
11604 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11605 continue;
11606
11607 buf = &bufs[cp->cpu_id];
11608
11609 /*
11610 * If there is already a buffer allocated for this CPU, it
11611 * is only possible that this is a DR event. In this case,
11612 * the buffer size must match our specified size.
11613 */
11614 if (buf->dtb_tomax != NULL) {
11615 ASSERT(buf->dtb_size == size);
11616 continue;
11617 }
11618
11619 ASSERT(buf->dtb_xamot == NULL);
11620
11621 if ((buf->dtb_tomax = kmem_zalloc(size,
11622 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11623 goto err;
11624
11625 buf->dtb_size = size;
11626 buf->dtb_flags = flags;
11627 buf->dtb_offset = 0;
11628 buf->dtb_drops = 0;
11629
11630 if (flags & DTRACEBUF_NOSWITCH)
11631 continue;
11632
11633 if ((buf->dtb_xamot = kmem_zalloc(size,
11634 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11635 goto err;
11636 } while ((cp = cp->cpu_next) != cpu_list);
11637
11638 return (0);
11639
11640 err:
11641 cp = cpu_list;
11642
11643 do {
11644 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11645 continue;
11646
11647 buf = &bufs[cp->cpu_id];
11648 desired += 2;
11649
11650 if (buf->dtb_xamot != NULL) {
11651 ASSERT(buf->dtb_tomax != NULL);
11652 ASSERT(buf->dtb_size == size);
11653 kmem_free(buf->dtb_xamot, size);
11654 allocated++;
11655 }
11656
11657 if (buf->dtb_tomax != NULL) {
11658 ASSERT(buf->dtb_size == size);
11659 kmem_free(buf->dtb_tomax, size);
11660 allocated++;
11661 }
11662
11663 buf->dtb_tomax = NULL;
11664 buf->dtb_xamot = NULL;
11665 buf->dtb_size = 0;
11666 } while ((cp = cp->cpu_next) != cpu_list);
11667
11668 *factor = desired / (allocated > 0 ? allocated : 1);
11669
11670 return (ENOMEM);
11671 }
11672
11673 /*
11674 * Note: called from probe context. This function just increments the drop
11675 * count on a buffer. It has been made a function to allow for the
11676 * possibility of understanding the source of mysterious drop counts. (A
11677 * problem for which one may be particularly disappointed that DTrace cannot
11678 * be used to understand DTrace.)
11679 */
11680 static void
11681 dtrace_buffer_drop(dtrace_buffer_t *buf)
11682 {
11683 buf->dtb_drops++;
11684 }
11685
11686 /*
11687 * Note: called from probe context. This function is called to reserve space
11688 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
11689 * mstate. Returns the new offset in the buffer, or a negative value if an
11690 * error has occurred.
11691 */
11692 static intptr_t
11693 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11694 dtrace_state_t *state, dtrace_mstate_t *mstate)
11695 {
11696 intptr_t offs = buf->dtb_offset, soffs;
11697 intptr_t woffs;
11698 caddr_t tomax;
11699 size_t total;
11700
11701 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11702 return (-1);
11703
11704 if ((tomax = buf->dtb_tomax) == NULL) {
11705 dtrace_buffer_drop(buf);
11706 return (-1);
11707 }
11708
11709 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11710 while (offs & (align - 1)) {
11711 /*
11712 * Assert that our alignment is off by a number which
11713 * is itself sizeof (uint32_t) aligned.
11714 */
11715 ASSERT(!((align - (offs & (align - 1))) &
11716 (sizeof (uint32_t) - 1)));
11717 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11718 offs += sizeof (uint32_t);
11719 }
11720
11721 if ((soffs = offs + needed) > buf->dtb_size) {
11722 dtrace_buffer_drop(buf);
11723 return (-1);
11724 }
11725
11726 if (mstate == NULL)
11727 return (offs);
11728
11729 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11730 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11731 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11732
11733 return (offs);
11734 }
11735
11736 if (buf->dtb_flags & DTRACEBUF_FILL) {
11737 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11738 (buf->dtb_flags & DTRACEBUF_FULL))
11739 return (-1);
11740 goto out;
11741 }
11742
11743 total = needed + (offs & (align - 1));
11744
11745 /*
11746 * For a ring buffer, life is quite a bit more complicated. Before
11747 * we can store any padding, we need to adjust our wrapping offset.
11748 * (If we've never before wrapped or we're not about to, no adjustment
11749 * is required.)
11750 */
11751 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11752 offs + total > buf->dtb_size) {
11753 woffs = buf->dtb_xamot_offset;
11754
11755 if (offs + total > buf->dtb_size) {
11756 /*
11757 * We can't fit in the end of the buffer. First, a
11758 * sanity check that we can fit in the buffer at all.
11759 */
11760 if (total > buf->dtb_size) {
11761 dtrace_buffer_drop(buf);
11762 return (-1);
11763 }
11764
11765 /*
11766 * We're going to be storing at the top of the buffer,
11767 * so now we need to deal with the wrapped offset. We
11768 * only reset our wrapped offset to 0 if it is
11769 * currently greater than the current offset. If it
11770 * is less than the current offset, it is because a
11771 * previous allocation induced a wrap -- but the
11772 * allocation didn't subsequently take the space due
11773 * to an error or false predicate evaluation. In this
11774 * case, we'll just leave the wrapped offset alone: if
11775 * the wrapped offset hasn't been advanced far enough
11776 * for this allocation, it will be adjusted in the
11777 * lower loop.
11778 */
11779 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11780 if (woffs >= offs)
11781 woffs = 0;
11782 } else {
11783 woffs = 0;
11784 }
11785
11786 /*
11787 * Now we know that we're going to be storing to the
11788 * top of the buffer and that there is room for us
11789 * there. We need to clear the buffer from the current
11790 * offset to the end (there may be old gunk there).
11791 */
11792 while (offs < buf->dtb_size)
11793 tomax[offs++] = 0;
11794
11795 /*
11796 * We need to set our offset to zero. And because we
11797 * are wrapping, we need to set the bit indicating as
11798 * much. We can also adjust our needed space back
11799 * down to the space required by the ECB -- we know
11800 * that the top of the buffer is aligned.
11801 */
11802 offs = 0;
11803 total = needed;
11804 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11805 } else {
11806 /*
11807 * There is room for us in the buffer, so we simply
11808 * need to check the wrapped offset.
11809 */
11810 if (woffs < offs) {
11811 /*
11812 * The wrapped offset is less than the offset.
11813 * This can happen if we allocated buffer space
11814 * that induced a wrap, but then we didn't
11815 * subsequently take the space due to an error
11816 * or false predicate evaluation. This is
11817 * okay; we know that _this_ allocation isn't
11818 * going to induce a wrap. We still can't
11819 * reset the wrapped offset to be zero,
11820 * however: the space may have been trashed in
11821 * the previous failed probe attempt. But at
11822 * least the wrapped offset doesn't need to
11823 * be adjusted at all...
11824 */
11825 goto out;
11826 }
11827 }
11828
11829 while (offs + total > woffs) {
11830 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11831 size_t size;
11832
11833 if (epid == DTRACE_EPIDNONE) {
11834 size = sizeof (uint32_t);
11835 } else {
11836 ASSERT3U(epid, <=, state->dts_necbs);
11837 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11838
11839 size = state->dts_ecbs[epid - 1]->dte_size;
11840 }
11841
11842 ASSERT(woffs + size <= buf->dtb_size);
11843 ASSERT(size != 0);
11844
11845 if (woffs + size == buf->dtb_size) {
11846 /*
11847 * We've reached the end of the buffer; we want
11848 * to set the wrapped offset to 0 and break
11849 * out. However, if the offs is 0, then we're
11850 * in a strange edge-condition: the amount of
11851 * space that we want to reserve plus the size
11852 * of the record that we're overwriting is
11853 * greater than the size of the buffer. This
11854 * is problematic because if we reserve the
11855 * space but subsequently don't consume it (due
11856 * to a failed predicate or error) the wrapped
11857 * offset will be 0 -- yet the EPID at offset 0
11858 * will not be committed. This situation is
11859 * relatively easy to deal with: if we're in
11860 * this case, the buffer is indistinguishable
11861 * from one that hasn't wrapped; we need only
11862 * finish the job by clearing the wrapped bit,
11863 * explicitly setting the offset to be 0, and
11864 * zero'ing out the old data in the buffer.
11865 */
11866 if (offs == 0) {
11867 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11868 buf->dtb_offset = 0;
11869 woffs = total;
11870
11871 while (woffs < buf->dtb_size)
11872 tomax[woffs++] = 0;
11873 }
11874
11875 woffs = 0;
11876 break;
11877 }
11878
11879 woffs += size;
11880 }
11881
11882 /*
11883 * We have a wrapped offset. It may be that the wrapped offset
11884 * has become zero -- that's okay.
11885 */
11886 buf->dtb_xamot_offset = woffs;
11887 }
11888
11889 out:
11890 /*
11891 * Now we can plow the buffer with any necessary padding.
11892 */
11893 while (offs & (align - 1)) {
11894 /*
11895 * Assert that our alignment is off by a number which
11896 * is itself sizeof (uint32_t) aligned.
11897 */
11898 ASSERT(!((align - (offs & (align - 1))) &
11899 (sizeof (uint32_t) - 1)));
11900 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11901 offs += sizeof (uint32_t);
11902 }
11903
11904 if (buf->dtb_flags & DTRACEBUF_FILL) {
11905 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11906 buf->dtb_flags |= DTRACEBUF_FULL;
11907 return (-1);
11908 }
11909 }
11910
11911 if (mstate == NULL)
11912 return (offs);
11913
11914 /*
11915 * For ring buffers and fill buffers, the scratch space is always
11916 * the inactive buffer.
11917 */
11918 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11919 mstate->dtms_scratch_size = buf->dtb_size;
11920 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11921
11922 return (offs);
11923 }
11924
11925 static void
11926 dtrace_buffer_polish(dtrace_buffer_t *buf)
11927 {
11928 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11929 ASSERT(MUTEX_HELD(&dtrace_lock));
11930
11931 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11932 return;
11933
11934 /*
11935 * We need to polish the ring buffer. There are three cases:
11936 *
11937 * - The first (and presumably most common) is that there is no gap
11938 * between the buffer offset and the wrapped offset. In this case,
11939 * there is nothing in the buffer that isn't valid data; we can
11940 * mark the buffer as polished and return.
11941 *
11942 * - The second (less common than the first but still more common
11943 * than the third) is that there is a gap between the buffer offset
11944 * and the wrapped offset, and the wrapped offset is larger than the
11945 * buffer offset. This can happen because of an alignment issue, or
11946 * can happen because of a call to dtrace_buffer_reserve() that
11947 * didn't subsequently consume the buffer space. In this case,
11948 * we need to zero the data from the buffer offset to the wrapped
11949 * offset.
11950 *
11951 * - The third (and least common) is that there is a gap between the
11952 * buffer offset and the wrapped offset, but the wrapped offset is
11953 * _less_ than the buffer offset. This can only happen because a
11954 * call to dtrace_buffer_reserve() induced a wrap, but the space
11955 * was not subsequently consumed. In this case, we need to zero the
11956 * space from the offset to the end of the buffer _and_ from the
11957 * top of the buffer to the wrapped offset.
11958 */
11959 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11960 bzero(buf->dtb_tomax + buf->dtb_offset,
11961 buf->dtb_xamot_offset - buf->dtb_offset);
11962 }
11963
11964 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11965 bzero(buf->dtb_tomax + buf->dtb_offset,
11966 buf->dtb_size - buf->dtb_offset);
11967 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11968 }
11969 }
11970
11971 /*
11972 * This routine determines if data generated at the specified time has likely
11973 * been entirely consumed at user-level. This routine is called to determine
11974 * if an ECB on a defunct probe (but for an active enabling) can be safely
11975 * disabled and destroyed.
11976 */
11977 static int
11978 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11979 {
11980 int i;
11981
11982 for (i = 0; i < NCPU; i++) {
11983 dtrace_buffer_t *buf = &bufs[i];
11984
11985 if (buf->dtb_size == 0)
11986 continue;
11987
11988 if (buf->dtb_flags & DTRACEBUF_RING)
11989 return (0);
11990
11991 if (!buf->dtb_switched && buf->dtb_offset != 0)
11992 return (0);
11993
11994 if (buf->dtb_switched - buf->dtb_interval < when)
11995 return (0);
11996 }
11997
11998 return (1);
11999 }
12000
12001 static void
12002 dtrace_buffer_free(dtrace_buffer_t *bufs)
12003 {
12004 int i;
12005
12006 for (i = 0; i < NCPU; i++) {
12007 dtrace_buffer_t *buf = &bufs[i];
12008
12009 if (buf->dtb_tomax == NULL) {
12010 ASSERT(buf->dtb_xamot == NULL);
12011 ASSERT(buf->dtb_size == 0);
12012 continue;
12013 }
12014
12015 if (buf->dtb_xamot != NULL) {
12016 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
12017 kmem_free(buf->dtb_xamot, buf->dtb_size);
12018 }
12019
12020 kmem_free(buf->dtb_tomax, buf->dtb_size);
12021 buf->dtb_size = 0;
12022 buf->dtb_tomax = NULL;
12023 buf->dtb_xamot = NULL;
12024 }
12025 }
12026
12027 /*
12028 * DTrace Enabling Functions
12029 */
12030 static dtrace_enabling_t *
12031 dtrace_enabling_create(dtrace_vstate_t *vstate)
12032 {
12033 dtrace_enabling_t *enab;
12034
12035 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
12036 enab->dten_vstate = vstate;
12037
12038 return (enab);
12039 }
12040
12041 static void
12042 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
12043 {
12044 dtrace_ecbdesc_t **ndesc;
12045 size_t osize, nsize;
12046
12047 /*
12048 * We can't add to enablings after we've enabled them, or after we've
12049 * retained them.
12050 */
12051 ASSERT(enab->dten_probegen == 0);
12052 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12053
12054 if (enab->dten_ndesc < enab->dten_maxdesc) {
12055 enab->dten_desc[enab->dten_ndesc++] = ecb;
12056 return;
12057 }
12058
12059 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12060
12061 if (enab->dten_maxdesc == 0) {
12062 enab->dten_maxdesc = 1;
12063 } else {
12064 enab->dten_maxdesc <<= 1;
12065 }
12066
12067 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
12068
12069 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12070 ndesc = kmem_zalloc(nsize, KM_SLEEP);
12071 bcopy(enab->dten_desc, ndesc, osize);
12072 kmem_free(enab->dten_desc, osize);
12073
12074 enab->dten_desc = ndesc;
12075 enab->dten_desc[enab->dten_ndesc++] = ecb;
12076 }
12077
12078 static void
12079 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
12080 dtrace_probedesc_t *pd)
12081 {
12082 dtrace_ecbdesc_t *new;
12083 dtrace_predicate_t *pred;
12084 dtrace_actdesc_t *act;
12085
12086 /*
12087 * We're going to create a new ECB description that matches the
12088 * specified ECB in every way, but has the specified probe description.
12089 */
12090 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12091
12092 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
12093 dtrace_predicate_hold(pred);
12094
12095 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
12096 dtrace_actdesc_hold(act);
12097
12098 new->dted_action = ecb->dted_action;
12099 new->dted_pred = ecb->dted_pred;
12100 new->dted_probe = *pd;
12101 new->dted_uarg = ecb->dted_uarg;
12102
12103 dtrace_enabling_add(enab, new);
12104 }
12105
12106 static void
12107 dtrace_enabling_dump(dtrace_enabling_t *enab)
12108 {
12109 int i;
12110
12111 for (i = 0; i < enab->dten_ndesc; i++) {
12112 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
12113
12114 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
12115 desc->dtpd_provider, desc->dtpd_mod,
12116 desc->dtpd_func, desc->dtpd_name);
12117 }
12118 }
12119
12120 static void
12121 dtrace_enabling_destroy(dtrace_enabling_t *enab)
12122 {
12123 int i;
12124 dtrace_ecbdesc_t *ep;
12125 dtrace_vstate_t *vstate = enab->dten_vstate;
12126
12127 ASSERT(MUTEX_HELD(&dtrace_lock));
12128
12129 for (i = 0; i < enab->dten_ndesc; i++) {
12130 dtrace_actdesc_t *act, *next;
12131 dtrace_predicate_t *pred;
12132
12133 ep = enab->dten_desc[i];
12134
12135 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
12136 dtrace_predicate_release(pred, vstate);
12137
12138 for (act = ep->dted_action; act != NULL; act = next) {
12139 next = act->dtad_next;
12140 dtrace_actdesc_release(act, vstate);
12141 }
12142
12143 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12144 }
12145
12146 kmem_free(enab->dten_desc,
12147 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
12148
12149 /*
12150 * If this was a retained enabling, decrement the dts_nretained count
12151 * and take it off of the dtrace_retained list.
12152 */
12153 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
12154 dtrace_retained == enab) {
12155 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12156 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
12157 enab->dten_vstate->dtvs_state->dts_nretained--;
12158 dtrace_retained_gen++;
12159 }
12160
12161 if (enab->dten_prev == NULL) {
12162 if (dtrace_retained == enab) {
12163 dtrace_retained = enab->dten_next;
12164
12165 if (dtrace_retained != NULL)
12166 dtrace_retained->dten_prev = NULL;
12167 }
12168 } else {
12169 ASSERT(enab != dtrace_retained);
12170 ASSERT(dtrace_retained != NULL);
12171 enab->dten_prev->dten_next = enab->dten_next;
12172 }
12173
12174 if (enab->dten_next != NULL) {
12175 ASSERT(dtrace_retained != NULL);
12176 enab->dten_next->dten_prev = enab->dten_prev;
12177 }
12178
12179 kmem_free(enab, sizeof (dtrace_enabling_t));
12180 }
12181
12182 static int
12183 dtrace_enabling_retain(dtrace_enabling_t *enab)
12184 {
12185 dtrace_state_t *state;
12186
12187 ASSERT(MUTEX_HELD(&dtrace_lock));
12188 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12189 ASSERT(enab->dten_vstate != NULL);
12190
12191 state = enab->dten_vstate->dtvs_state;
12192 ASSERT(state != NULL);
12193
12194 /*
12195 * We only allow each state to retain dtrace_retain_max enablings.
12196 */
12197 if (state->dts_nretained >= dtrace_retain_max)
12198 return (ENOSPC);
12199
12200 state->dts_nretained++;
12201 dtrace_retained_gen++;
12202
12203 if (dtrace_retained == NULL) {
12204 dtrace_retained = enab;
12205 return (0);
12206 }
12207
12208 enab->dten_next = dtrace_retained;
12209 dtrace_retained->dten_prev = enab;
12210 dtrace_retained = enab;
12211
12212 return (0);
12213 }
12214
12215 static int
12216 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
12217 dtrace_probedesc_t *create)
12218 {
12219 dtrace_enabling_t *new, *enab;
12220 int found = 0, err = ENOENT;
12221
12222 ASSERT(MUTEX_HELD(&dtrace_lock));
12223 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
12224 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
12225 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
12226 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
12227
12228 new = dtrace_enabling_create(&state->dts_vstate);
12229
12230 /*
12231 * Iterate over all retained enablings, looking for enablings that
12232 * match the specified state.
12233 */
12234 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12235 int i;
12236
12237 /*
12238 * dtvs_state can only be NULL for helper enablings -- and
12239 * helper enablings can't be retained.
12240 */
12241 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12242
12243 if (enab->dten_vstate->dtvs_state != state)
12244 continue;
12245
12246 /*
12247 * Now iterate over each probe description; we're looking for
12248 * an exact match to the specified probe description.
12249 */
12250 for (i = 0; i < enab->dten_ndesc; i++) {
12251 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12252 dtrace_probedesc_t *pd = &ep->dted_probe;
12253
12254 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
12255 continue;
12256
12257 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
12258 continue;
12259
12260 if (strcmp(pd->dtpd_func, match->dtpd_func))
12261 continue;
12262
12263 if (strcmp(pd->dtpd_name, match->dtpd_name))
12264 continue;
12265
12266 /*
12267 * We have a winning probe! Add it to our growing
12268 * enabling.
12269 */
12270 found = 1;
12271 dtrace_enabling_addlike(new, ep, create);
12272 }
12273 }
12274
12275 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
12276 dtrace_enabling_destroy(new);
12277 return (err);
12278 }
12279
12280 return (0);
12281 }
12282
12283 static void
12284 dtrace_enabling_retract(dtrace_state_t *state)
12285 {
12286 dtrace_enabling_t *enab, *next;
12287
12288 ASSERT(MUTEX_HELD(&dtrace_lock));
12289
12290 /*
12291 * Iterate over all retained enablings, destroy the enablings retained
12292 * for the specified state.
12293 */
12294 for (enab = dtrace_retained; enab != NULL; enab = next) {
12295 next = enab->dten_next;
12296
12297 /*
12298 * dtvs_state can only be NULL for helper enablings -- and
12299 * helper enablings can't be retained.
12300 */
12301 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12302
12303 if (enab->dten_vstate->dtvs_state == state) {
12304 ASSERT(state->dts_nretained > 0);
12305 dtrace_enabling_destroy(enab);
12306 }
12307 }
12308
12309 ASSERT(state->dts_nretained == 0);
12310 }
12311
12312 static int
12313 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
12314 {
12315 int i = 0;
12316 int total_matched = 0, matched = 0;
12317
12318 ASSERT(MUTEX_HELD(&cpu_lock));
12319 ASSERT(MUTEX_HELD(&dtrace_lock));
12320
12321 for (i = 0; i < enab->dten_ndesc; i++) {
12322 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12323
12324 enab->dten_current = ep;
12325 enab->dten_error = 0;
12326
12327 /*
12328 * If a provider failed to enable a probe then get out and
12329 * let the consumer know we failed.
12330 */
12331 if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
12332 return (EBUSY);
12333
12334 total_matched += matched;
12335
12336 if (enab->dten_error != 0) {
12337 /*
12338 * If we get an error half-way through enabling the
12339 * probes, we kick out -- perhaps with some number of
12340 * them enabled. Leaving enabled probes enabled may
12341 * be slightly confusing for user-level, but we expect
12342 * that no one will attempt to actually drive on in
12343 * the face of such errors. If this is an anonymous
12344 * enabling (indicated with a NULL nmatched pointer),
12345 * we cmn_err() a message. We aren't expecting to
12346 * get such an error -- such as it can exist at all,
12347 * it would be a result of corrupted DOF in the driver
12348 * properties.
12349 */
12350 if (nmatched == NULL) {
12351 cmn_err(CE_WARN, "dtrace_enabling_match() "
12352 "error on %p: %d", (void *)ep,
12353 enab->dten_error);
12354 }
12355
12356 return (enab->dten_error);
12357 }
12358 }
12359
12360 enab->dten_probegen = dtrace_probegen;
12361 if (nmatched != NULL)
12362 *nmatched = total_matched;
12363
12364 return (0);
12365 }
12366
12367 static void
12368 dtrace_enabling_matchall(void)
12369 {
12370 dtrace_enabling_t *enab;
12371
12372 mutex_enter(&cpu_lock);
12373 mutex_enter(&dtrace_lock);
12374
12375 /*
12376 * Iterate over all retained enablings to see if any probes match
12377 * against them. We only perform this operation on enablings for which
12378 * we have sufficient permissions by virtue of being in the global zone
12379 * or in the same zone as the DTrace client. Because we can be called
12380 * after dtrace_detach() has been called, we cannot assert that there
12381 * are retained enablings. We can safely load from dtrace_retained,
12382 * however: the taskq_destroy() at the end of dtrace_detach() will
12383 * block pending our completion.
12384 */
12385 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12386 dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
12387 cred_t *cr = dcr->dcr_cred;
12388 zoneid_t zone = cr != NULL ? crgetzonedid(cr) : 0;
12389
12390 if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
12391 (zone == GLOBAL_ZONEID || getzonedid() == zone)))
12392 (void) dtrace_enabling_match(enab, NULL);
12393 }
12394
12395 mutex_exit(&dtrace_lock);
12396 mutex_exit(&cpu_lock);
12397 }
12398
12399 /*
12400 * If an enabling is to be enabled without having matched probes (that is, if
12401 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
12402 * enabling must be _primed_ by creating an ECB for every ECB description.
12403 * This must be done to assure that we know the number of speculations, the
12404 * number of aggregations, the minimum buffer size needed, etc. before we
12405 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
12406 * enabling any probes, we create ECBs for every ECB decription, but with a
12407 * NULL probe -- which is exactly what this function does.
12408 */
12409 static void
12410 dtrace_enabling_prime(dtrace_state_t *state)
12411 {
12412 dtrace_enabling_t *enab;
12413 int i;
12414
12415 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12416 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12417
12418 if (enab->dten_vstate->dtvs_state != state)
12419 continue;
12420
12421 /*
12422 * We don't want to prime an enabling more than once, lest
12423 * we allow a malicious user to induce resource exhaustion.
12424 * (The ECBs that result from priming an enabling aren't
12425 * leaked -- but they also aren't deallocated until the
12426 * consumer state is destroyed.)
12427 */
12428 if (enab->dten_primed)
12429 continue;
12430
12431 for (i = 0; i < enab->dten_ndesc; i++) {
12432 enab->dten_current = enab->dten_desc[i];
12433 (void) dtrace_probe_enable(NULL, enab);
12434 }
12435
12436 enab->dten_primed = 1;
12437 }
12438 }
12439
12440 /*
12441 * Called to indicate that probes should be provided due to retained
12442 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
12443 * must take an initial lap through the enabling calling the dtps_provide()
12444 * entry point explicitly to allow for autocreated probes.
12445 */
12446 static void
12447 dtrace_enabling_provide(dtrace_provider_t *prv)
12448 {
12449 int i, all = 0;
12450 dtrace_probedesc_t desc;
12451 dtrace_genid_t gen;
12452
12453 ASSERT(MUTEX_HELD(&dtrace_lock));
12454 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
12455
12456 if (prv == NULL) {
12457 all = 1;
12458 prv = dtrace_provider;
12459 }
12460
12461 do {
12462 dtrace_enabling_t *enab;
12463 void *parg = prv->dtpv_arg;
12464
12465 retry:
12466 gen = dtrace_retained_gen;
12467 for (enab = dtrace_retained; enab != NULL;
12468 enab = enab->dten_next) {
12469 for (i = 0; i < enab->dten_ndesc; i++) {
12470 desc = enab->dten_desc[i]->dted_probe;
12471 mutex_exit(&dtrace_lock);
12472 prv->dtpv_pops.dtps_provide(parg, &desc);
12473 mutex_enter(&dtrace_lock);
12474 /*
12475 * Process the retained enablings again if
12476 * they have changed while we weren't holding
12477 * dtrace_lock.
12478 */
12479 if (gen != dtrace_retained_gen)
12480 goto retry;
12481 }
12482 }
12483 } while (all && (prv = prv->dtpv_next) != NULL);
12484
12485 mutex_exit(&dtrace_lock);
12486 dtrace_probe_provide(NULL, all ? NULL : prv);
12487 mutex_enter(&dtrace_lock);
12488 }
12489
12490 /*
12491 * Called to reap ECBs that are attached to probes from defunct providers.
12492 */
12493 static void
12494 dtrace_enabling_reap(void)
12495 {
12496 dtrace_provider_t *prov;
12497 dtrace_probe_t *probe;
12498 dtrace_ecb_t *ecb;
12499 hrtime_t when;
12500 int i;
12501
12502 mutex_enter(&cpu_lock);
12503 mutex_enter(&dtrace_lock);
12504
12505 for (i = 0; i < dtrace_nprobes; i++) {
12506 if ((probe = dtrace_probes[i]) == NULL)
12507 continue;
12508
12509 if (probe->dtpr_ecb == NULL)
12510 continue;
12511
12512 prov = probe->dtpr_provider;
12513
12514 if ((when = prov->dtpv_defunct) == 0)
12515 continue;
12516
12517 /*
12518 * We have ECBs on a defunct provider: we want to reap these
12519 * ECBs to allow the provider to unregister. The destruction
12520 * of these ECBs must be done carefully: if we destroy the ECB
12521 * and the consumer later wishes to consume an EPID that
12522 * corresponds to the destroyed ECB (and if the EPID metadata
12523 * has not been previously consumed), the consumer will abort
12524 * processing on the unknown EPID. To reduce (but not, sadly,
12525 * eliminate) the possibility of this, we will only destroy an
12526 * ECB for a defunct provider if, for the state that
12527 * corresponds to the ECB:
12528 *
12529 * (a) There is no speculative tracing (which can effectively
12530 * cache an EPID for an arbitrary amount of time).
12531 *
12532 * (b) The principal buffers have been switched twice since the
12533 * provider became defunct.
12534 *
12535 * (c) The aggregation buffers are of zero size or have been
12536 * switched twice since the provider became defunct.
12537 *
12538 * We use dts_speculates to determine (a) and call a function
12539 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
12540 * that as soon as we've been unable to destroy one of the ECBs
12541 * associated with the probe, we quit trying -- reaping is only
12542 * fruitful in as much as we can destroy all ECBs associated
12543 * with the defunct provider's probes.
12544 */
12545 while ((ecb = probe->dtpr_ecb) != NULL) {
12546 dtrace_state_t *state = ecb->dte_state;
12547 dtrace_buffer_t *buf = state->dts_buffer;
12548 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
12549
12550 if (state->dts_speculates)
12551 break;
12552
12553 if (!dtrace_buffer_consumed(buf, when))
12554 break;
12555
12556 if (!dtrace_buffer_consumed(aggbuf, when))
12557 break;
12558
12559 dtrace_ecb_disable(ecb);
12560 ASSERT(probe->dtpr_ecb != ecb);
12561 dtrace_ecb_destroy(ecb);
12562 }
12563 }
12564
12565 mutex_exit(&dtrace_lock);
12566 mutex_exit(&cpu_lock);
12567 }
12568
12569 /*
12570 * DTrace DOF Functions
12571 */
12572 /*ARGSUSED*/
12573 static void
12574 dtrace_dof_error(dof_hdr_t *dof, const char *str)
12575 {
12576 if (dtrace_err_verbose)
12577 cmn_err(CE_WARN, "failed to process DOF: %s", str);
12578
12579 #ifdef DTRACE_ERRDEBUG
12580 dtrace_errdebug(str);
12581 #endif
12582 }
12583
12584 /*
12585 * Create DOF out of a currently enabled state. Right now, we only create
12586 * DOF containing the run-time options -- but this could be expanded to create
12587 * complete DOF representing the enabled state.
12588 */
12589 static dof_hdr_t *
12590 dtrace_dof_create(dtrace_state_t *state)
12591 {
12592 dof_hdr_t *dof;
12593 dof_sec_t *sec;
12594 dof_optdesc_t *opt;
12595 int i, len = sizeof (dof_hdr_t) +
12596 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
12597 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12598
12599 ASSERT(MUTEX_HELD(&dtrace_lock));
12600
12601 dof = kmem_zalloc(len, KM_SLEEP);
12602 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
12603 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
12604 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
12605 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
12606
12607 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
12608 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
12609 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
12610 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
12611 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
12612 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
12613
12614 dof->dofh_flags = 0;
12615 dof->dofh_hdrsize = sizeof (dof_hdr_t);
12616 dof->dofh_secsize = sizeof (dof_sec_t);
12617 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
12618 dof->dofh_secoff = sizeof (dof_hdr_t);
12619 dof->dofh_loadsz = len;
12620 dof->dofh_filesz = len;
12621 dof->dofh_pad = 0;
12622
12623 /*
12624 * Fill in the option section header...
12625 */
12626 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
12627 sec->dofs_type = DOF_SECT_OPTDESC;
12628 sec->dofs_align = sizeof (uint64_t);
12629 sec->dofs_flags = DOF_SECF_LOAD;
12630 sec->dofs_entsize = sizeof (dof_optdesc_t);
12631
12632 opt = (dof_optdesc_t *)((uintptr_t)sec +
12633 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
12634
12635 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
12636 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12637
12638 for (i = 0; i < DTRACEOPT_MAX; i++) {
12639 opt[i].dofo_option = i;
12640 opt[i].dofo_strtab = DOF_SECIDX_NONE;
12641 opt[i].dofo_value = state->dts_options[i];
12642 }
12643
12644 return (dof);
12645 }
12646
12647 static dof_hdr_t *
12648 dtrace_dof_copyin(uintptr_t uarg, int *errp)
12649 {
12650 dof_hdr_t hdr, *dof;
12651
12652 ASSERT(!MUTEX_HELD(&dtrace_lock));
12653
12654 /*
12655 * First, we're going to copyin() the sizeof (dof_hdr_t).
12656 */
12657 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
12658 dtrace_dof_error(NULL, "failed to copyin DOF header");
12659 *errp = EFAULT;
12660 return (NULL);
12661 }
12662
12663 /*
12664 * Now we'll allocate the entire DOF and copy it in -- provided
12665 * that the length isn't outrageous.
12666 */
12667 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
12668 dtrace_dof_error(&hdr, "load size exceeds maximum");
12669 *errp = E2BIG;
12670 return (NULL);
12671 }
12672
12673 if (hdr.dofh_loadsz < sizeof (hdr)) {
12674 dtrace_dof_error(&hdr, "invalid load size");
12675 *errp = EINVAL;
12676 return (NULL);
12677 }
12678
12679 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12680
12681 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
12682 dof->dofh_loadsz != hdr.dofh_loadsz) {
12683 kmem_free(dof, hdr.dofh_loadsz);
12684 *errp = EFAULT;
12685 return (NULL);
12686 }
12687
12688 return (dof);
12689 }
12690
12691 static dof_hdr_t *
12692 dtrace_dof_property(const char *name)
12693 {
12694 uchar_t *buf;
12695 uint64_t loadsz;
12696 unsigned int len, i;
12697 dof_hdr_t *dof;
12698
12699 /*
12700 * Unfortunately, array of values in .conf files are always (and
12701 * only) interpreted to be integer arrays. We must read our DOF
12702 * as an integer array, and then squeeze it into a byte array.
12703 */
12704 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12705 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12706 return (NULL);
12707
12708 for (i = 0; i < len; i++)
12709 buf[i] = (uchar_t)(((int *)buf)[i]);
12710
12711 if (len < sizeof (dof_hdr_t)) {
12712 ddi_prop_free(buf);
12713 dtrace_dof_error(NULL, "truncated header");
12714 return (NULL);
12715 }
12716
12717 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12718 ddi_prop_free(buf);
12719 dtrace_dof_error(NULL, "truncated DOF");
12720 return (NULL);
12721 }
12722
12723 if (loadsz >= dtrace_dof_maxsize) {
12724 ddi_prop_free(buf);
12725 dtrace_dof_error(NULL, "oversized DOF");
12726 return (NULL);
12727 }
12728
12729 dof = kmem_alloc(loadsz, KM_SLEEP);
12730 bcopy(buf, dof, loadsz);
12731 ddi_prop_free(buf);
12732
12733 return (dof);
12734 }
12735
12736 static void
12737 dtrace_dof_destroy(dof_hdr_t *dof)
12738 {
12739 kmem_free(dof, dof->dofh_loadsz);
12740 }
12741
12742 /*
12743 * Return the dof_sec_t pointer corresponding to a given section index. If the
12744 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12745 * a type other than DOF_SECT_NONE is specified, the header is checked against
12746 * this type and NULL is returned if the types do not match.
12747 */
12748 static dof_sec_t *
12749 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12750 {
12751 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12752 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12753
12754 if (i >= dof->dofh_secnum) {
12755 dtrace_dof_error(dof, "referenced section index is invalid");
12756 return (NULL);
12757 }
12758
12759 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12760 dtrace_dof_error(dof, "referenced section is not loadable");
12761 return (NULL);
12762 }
12763
12764 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12765 dtrace_dof_error(dof, "referenced section is the wrong type");
12766 return (NULL);
12767 }
12768
12769 return (sec);
12770 }
12771
12772 static dtrace_probedesc_t *
12773 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12774 {
12775 dof_probedesc_t *probe;
12776 dof_sec_t *strtab;
12777 uintptr_t daddr = (uintptr_t)dof;
12778 uintptr_t str;
12779 size_t size;
12780
12781 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12782 dtrace_dof_error(dof, "invalid probe section");
12783 return (NULL);
12784 }
12785
12786 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12787 dtrace_dof_error(dof, "bad alignment in probe description");
12788 return (NULL);
12789 }
12790
12791 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12792 dtrace_dof_error(dof, "truncated probe description");
12793 return (NULL);
12794 }
12795
12796 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12797 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12798
12799 if (strtab == NULL)
12800 return (NULL);
12801
12802 str = daddr + strtab->dofs_offset;
12803 size = strtab->dofs_size;
12804
12805 if (probe->dofp_provider >= strtab->dofs_size) {
12806 dtrace_dof_error(dof, "corrupt probe provider");
12807 return (NULL);
12808 }
12809
12810 (void) strncpy(desc->dtpd_provider,
12811 (char *)(str + probe->dofp_provider),
12812 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12813
12814 if (probe->dofp_mod >= strtab->dofs_size) {
12815 dtrace_dof_error(dof, "corrupt probe module");
12816 return (NULL);
12817 }
12818
12819 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12820 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12821
12822 if (probe->dofp_func >= strtab->dofs_size) {
12823 dtrace_dof_error(dof, "corrupt probe function");
12824 return (NULL);
12825 }
12826
12827 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12828 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12829
12830 if (probe->dofp_name >= strtab->dofs_size) {
12831 dtrace_dof_error(dof, "corrupt probe name");
12832 return (NULL);
12833 }
12834
12835 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12836 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12837
12838 return (desc);
12839 }
12840
12841 static dtrace_difo_t *
12842 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12843 cred_t *cr)
12844 {
12845 dtrace_difo_t *dp;
12846 size_t ttl = 0;
12847 dof_difohdr_t *dofd;
12848 uintptr_t daddr = (uintptr_t)dof;
12849 size_t max = dtrace_difo_maxsize;
12850 int i, l, n;
12851
12852 static const struct {
12853 int section;
12854 int bufoffs;
12855 int lenoffs;
12856 int entsize;
12857 int align;
12858 const char *msg;
12859 } difo[] = {
12860 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12861 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12862 sizeof (dif_instr_t), "multiple DIF sections" },
12863
12864 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12865 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12866 sizeof (uint64_t), "multiple integer tables" },
12867
12868 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12869 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12870 sizeof (char), "multiple string tables" },
12871
12872 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12873 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12874 sizeof (uint_t), "multiple variable tables" },
12875
12876 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
12877 };
12878
12879 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12880 dtrace_dof_error(dof, "invalid DIFO header section");
12881 return (NULL);
12882 }
12883
12884 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12885 dtrace_dof_error(dof, "bad alignment in DIFO header");
12886 return (NULL);
12887 }
12888
12889 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12890 sec->dofs_size % sizeof (dof_secidx_t)) {
12891 dtrace_dof_error(dof, "bad size in DIFO header");
12892 return (NULL);
12893 }
12894
12895 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12896 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12897
12898 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12899 dp->dtdo_rtype = dofd->dofd_rtype;
12900
12901 for (l = 0; l < n; l++) {
12902 dof_sec_t *subsec;
12903 void **bufp;
12904 uint32_t *lenp;
12905
12906 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12907 dofd->dofd_links[l])) == NULL)
12908 goto err; /* invalid section link */
12909
12910 if (ttl + subsec->dofs_size > max) {
12911 dtrace_dof_error(dof, "exceeds maximum size");
12912 goto err;
12913 }
12914
12915 ttl += subsec->dofs_size;
12916
12917 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12918 if (subsec->dofs_type != difo[i].section)
12919 continue;
12920
12921 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12922 dtrace_dof_error(dof, "section not loaded");
12923 goto err;
12924 }
12925
12926 if (subsec->dofs_align != difo[i].align) {
12927 dtrace_dof_error(dof, "bad alignment");
12928 goto err;
12929 }
12930
12931 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12932 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12933
12934 if (*bufp != NULL) {
12935 dtrace_dof_error(dof, difo[i].msg);
12936 goto err;
12937 }
12938
12939 if (difo[i].entsize != subsec->dofs_entsize) {
12940 dtrace_dof_error(dof, "entry size mismatch");
12941 goto err;
12942 }
12943
12944 if (subsec->dofs_entsize != 0 &&
12945 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12946 dtrace_dof_error(dof, "corrupt entry size");
12947 goto err;
12948 }
12949
12950 *lenp = subsec->dofs_size;
12951 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12952 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12953 *bufp, subsec->dofs_size);
12954
12955 if (subsec->dofs_entsize != 0)
12956 *lenp /= subsec->dofs_entsize;
12957
12958 break;
12959 }
12960
12961 /*
12962 * If we encounter a loadable DIFO sub-section that is not
12963 * known to us, assume this is a broken program and fail.
12964 */
12965 if (difo[i].section == DOF_SECT_NONE &&
12966 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12967 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12968 goto err;
12969 }
12970 }
12971
12972 if (dp->dtdo_buf == NULL) {
12973 /*
12974 * We can't have a DIF object without DIF text.
12975 */
12976 dtrace_dof_error(dof, "missing DIF text");
12977 goto err;
12978 }
12979
12980 /*
12981 * Before we validate the DIF object, run through the variable table
12982 * looking for the strings -- if any of their size are under, we'll set
12983 * their size to be the system-wide default string size. Note that
12984 * this should _not_ happen if the "strsize" option has been set --
12985 * in this case, the compiler should have set the size to reflect the
12986 * setting of the option.
12987 */
12988 for (i = 0; i < dp->dtdo_varlen; i++) {
12989 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12990 dtrace_diftype_t *t = &v->dtdv_type;
12991
12992 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12993 continue;
12994
12995 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12996 t->dtdt_size = dtrace_strsize_default;
12997 }
12998
12999 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
13000 goto err;
13001
13002 dtrace_difo_init(dp, vstate);
13003 return (dp);
13004
13005 err:
13006 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
13007 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
13008 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
13009 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
13010
13011 kmem_free(dp, sizeof (dtrace_difo_t));
13012 return (NULL);
13013 }
13014
13015 static dtrace_predicate_t *
13016 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13017 cred_t *cr)
13018 {
13019 dtrace_difo_t *dp;
13020
13021 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
13022 return (NULL);
13023
13024 return (dtrace_predicate_create(dp));
13025 }
13026
13027 static dtrace_actdesc_t *
13028 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13029 cred_t *cr)
13030 {
13031 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
13032 dof_actdesc_t *desc;
13033 dof_sec_t *difosec;
13034 size_t offs;
13035 uintptr_t daddr = (uintptr_t)dof;
13036 uint64_t arg;
13037 dtrace_actkind_t kind;
13038
13039 if (sec->dofs_type != DOF_SECT_ACTDESC) {
13040 dtrace_dof_error(dof, "invalid action section");
13041 return (NULL);
13042 }
13043
13044 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
13045 dtrace_dof_error(dof, "truncated action description");
13046 return (NULL);
13047 }
13048
13049 if (sec->dofs_align != sizeof (uint64_t)) {
13050 dtrace_dof_error(dof, "bad alignment in action description");
13051 return (NULL);
13052 }
13053
13054 if (sec->dofs_size < sec->dofs_entsize) {
13055 dtrace_dof_error(dof, "section entry size exceeds total size");
13056 return (NULL);
13057 }
13058
13059 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
13060 dtrace_dof_error(dof, "bad entry size in action description");
13061 return (NULL);
13062 }
13063
13064 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
13065 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
13066 return (NULL);
13067 }
13068
13069 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
13070 desc = (dof_actdesc_t *)(daddr +
13071 (uintptr_t)sec->dofs_offset + offs);
13072 kind = (dtrace_actkind_t)desc->dofa_kind;
13073
13074 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
13075 (kind != DTRACEACT_PRINTA ||
13076 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
13077 (kind == DTRACEACT_DIFEXPR &&
13078 desc->dofa_strtab != DOF_SECIDX_NONE)) {
13079 dof_sec_t *strtab;
13080 char *str, *fmt;
13081 uint64_t i;
13082
13083 /*
13084 * The argument to these actions is an index into the
13085 * DOF string table. For printf()-like actions, this
13086 * is the format string. For print(), this is the
13087 * CTF type of the expression result.
13088 */
13089 if ((strtab = dtrace_dof_sect(dof,
13090 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
13091 goto err;
13092
13093 str = (char *)((uintptr_t)dof +
13094 (uintptr_t)strtab->dofs_offset);
13095
13096 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
13097 if (str[i] == '\0')
13098 break;
13099 }
13100
13101 if (i >= strtab->dofs_size) {
13102 dtrace_dof_error(dof, "bogus format string");
13103 goto err;
13104 }
13105
13106 if (i == desc->dofa_arg) {
13107 dtrace_dof_error(dof, "empty format string");
13108 goto err;
13109 }
13110
13111 i -= desc->dofa_arg;
13112 fmt = kmem_alloc(i + 1, KM_SLEEP);
13113 bcopy(&str[desc->dofa_arg], fmt, i + 1);
13114 arg = (uint64_t)(uintptr_t)fmt;
13115 } else {
13116 if (kind == DTRACEACT_PRINTA) {
13117 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
13118 arg = 0;
13119 } else {
13120 arg = desc->dofa_arg;
13121 }
13122 }
13123
13124 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
13125 desc->dofa_uarg, arg);
13126
13127 if (last != NULL) {
13128 last->dtad_next = act;
13129 } else {
13130 first = act;
13131 }
13132
13133 last = act;
13134
13135 if (desc->dofa_difo == DOF_SECIDX_NONE)
13136 continue;
13137
13138 if ((difosec = dtrace_dof_sect(dof,
13139 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
13140 goto err;
13141
13142 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
13143
13144 if (act->dtad_difo == NULL)
13145 goto err;
13146 }
13147
13148 ASSERT(first != NULL);
13149 return (first);
13150
13151 err:
13152 for (act = first; act != NULL; act = next) {
13153 next = act->dtad_next;
13154 dtrace_actdesc_release(act, vstate);
13155 }
13156
13157 return (NULL);
13158 }
13159
13160 static dtrace_ecbdesc_t *
13161 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13162 cred_t *cr)
13163 {
13164 dtrace_ecbdesc_t *ep;
13165 dof_ecbdesc_t *ecb;
13166 dtrace_probedesc_t *desc;
13167 dtrace_predicate_t *pred = NULL;
13168
13169 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
13170 dtrace_dof_error(dof, "truncated ECB description");
13171 return (NULL);
13172 }
13173
13174 if (sec->dofs_align != sizeof (uint64_t)) {
13175 dtrace_dof_error(dof, "bad alignment in ECB description");
13176 return (NULL);
13177 }
13178
13179 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
13180 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
13181
13182 if (sec == NULL)
13183 return (NULL);
13184
13185 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
13186 ep->dted_uarg = ecb->dofe_uarg;
13187 desc = &ep->dted_probe;
13188
13189 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
13190 goto err;
13191
13192 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
13193 if ((sec = dtrace_dof_sect(dof,
13194 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
13195 goto err;
13196
13197 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
13198 goto err;
13199
13200 ep->dted_pred.dtpdd_predicate = pred;
13201 }
13202
13203 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
13204 if ((sec = dtrace_dof_sect(dof,
13205 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
13206 goto err;
13207
13208 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
13209
13210 if (ep->dted_action == NULL)
13211 goto err;
13212 }
13213
13214 return (ep);
13215
13216 err:
13217 if (pred != NULL)
13218 dtrace_predicate_release(pred, vstate);
13219 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
13220 return (NULL);
13221 }
13222
13223 /*
13224 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
13225 * specified DOF. At present, this amounts to simply adding 'ubase' to the
13226 * site of any user SETX relocations to account for load object base address.
13227 * In the future, if we need other relocations, this function can be extended.
13228 */
13229 static int
13230 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
13231 {
13232 uintptr_t daddr = (uintptr_t)dof;
13233 uintptr_t ts_end;
13234 dof_relohdr_t *dofr =
13235 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13236 dof_sec_t *ss, *rs, *ts;
13237 dof_relodesc_t *r;
13238 uint_t i, n;
13239
13240 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
13241 sec->dofs_align != sizeof (dof_secidx_t)) {
13242 dtrace_dof_error(dof, "invalid relocation header");
13243 return (-1);
13244 }
13245
13246 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
13247 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
13248 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
13249 ts_end = (uintptr_t)ts + sizeof (dof_sec_t);
13250
13251 if (ss == NULL || rs == NULL || ts == NULL)
13252 return (-1); /* dtrace_dof_error() has been called already */
13253
13254 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
13255 rs->dofs_align != sizeof (uint64_t)) {
13256 dtrace_dof_error(dof, "invalid relocation section");
13257 return (-1);
13258 }
13259
13260 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
13261 n = rs->dofs_size / rs->dofs_entsize;
13262
13263 for (i = 0; i < n; i++) {
13264 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
13265
13266 switch (r->dofr_type) {
13267 case DOF_RELO_NONE:
13268 break;
13269 case DOF_RELO_SETX:
13270 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
13271 sizeof (uint64_t) > ts->dofs_size) {
13272 dtrace_dof_error(dof, "bad relocation offset");
13273 return (-1);
13274 }
13275
13276 if (taddr >= (uintptr_t)ts && taddr < ts_end) {
13277 dtrace_dof_error(dof, "bad relocation offset");
13278 return (-1);
13279 }
13280
13281 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
13282 dtrace_dof_error(dof, "misaligned setx relo");
13283 return (-1);
13284 }
13285
13286 *(uint64_t *)taddr += ubase;
13287 break;
13288 default:
13289 dtrace_dof_error(dof, "invalid relocation type");
13290 return (-1);
13291 }
13292
13293 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
13294 }
13295
13296 return (0);
13297 }
13298
13299 /*
13300 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
13301 * header: it should be at the front of a memory region that is at least
13302 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
13303 * size. It need not be validated in any other way.
13304 */
13305 static int
13306 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
13307 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
13308 {
13309 uint64_t len = dof->dofh_loadsz, seclen;
13310 uintptr_t daddr = (uintptr_t)dof;
13311 dtrace_ecbdesc_t *ep;
13312 dtrace_enabling_t *enab;
13313 uint_t i;
13314
13315 ASSERT(MUTEX_HELD(&dtrace_lock));
13316 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
13317
13318 /*
13319 * Check the DOF header identification bytes. In addition to checking
13320 * valid settings, we also verify that unused bits/bytes are zeroed so
13321 * we can use them later without fear of regressing existing binaries.
13322 */
13323 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
13324 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
13325 dtrace_dof_error(dof, "DOF magic string mismatch");
13326 return (-1);
13327 }
13328
13329 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
13330 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
13331 dtrace_dof_error(dof, "DOF has invalid data model");
13332 return (-1);
13333 }
13334
13335 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
13336 dtrace_dof_error(dof, "DOF encoding mismatch");
13337 return (-1);
13338 }
13339
13340 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
13341 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
13342 dtrace_dof_error(dof, "DOF version mismatch");
13343 return (-1);
13344 }
13345
13346 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
13347 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
13348 return (-1);
13349 }
13350
13351 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
13352 dtrace_dof_error(dof, "DOF uses too many integer registers");
13353 return (-1);
13354 }
13355
13356 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
13357 dtrace_dof_error(dof, "DOF uses too many tuple registers");
13358 return (-1);
13359 }
13360
13361 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
13362 if (dof->dofh_ident[i] != 0) {
13363 dtrace_dof_error(dof, "DOF has invalid ident byte set");
13364 return (-1);
13365 }
13366 }
13367
13368 if (dof->dofh_flags & ~DOF_FL_VALID) {
13369 dtrace_dof_error(dof, "DOF has invalid flag bits set");
13370 return (-1);
13371 }
13372
13373 if (dof->dofh_secsize == 0) {
13374 dtrace_dof_error(dof, "zero section header size");
13375 return (-1);
13376 }
13377
13378 /*
13379 * Check that the section headers don't exceed the amount of DOF
13380 * data. Note that we cast the section size and number of sections
13381 * to uint64_t's to prevent possible overflow in the multiplication.
13382 */
13383 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
13384
13385 if (dof->dofh_secoff > len || seclen > len ||
13386 dof->dofh_secoff + seclen > len) {
13387 dtrace_dof_error(dof, "truncated section headers");
13388 return (-1);
13389 }
13390
13391 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
13392 dtrace_dof_error(dof, "misaligned section headers");
13393 return (-1);
13394 }
13395
13396 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
13397 dtrace_dof_error(dof, "misaligned section size");
13398 return (-1);
13399 }
13400
13401 /*
13402 * Take an initial pass through the section headers to be sure that
13403 * the headers don't have stray offsets. If the 'noprobes' flag is
13404 * set, do not permit sections relating to providers, probes, or args.
13405 */
13406 for (i = 0; i < dof->dofh_secnum; i++) {
13407 dof_sec_t *sec = (dof_sec_t *)(daddr +
13408 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13409
13410 if (noprobes) {
13411 switch (sec->dofs_type) {
13412 case DOF_SECT_PROVIDER:
13413 case DOF_SECT_PROBES:
13414 case DOF_SECT_PRARGS:
13415 case DOF_SECT_PROFFS:
13416 dtrace_dof_error(dof, "illegal sections "
13417 "for enabling");
13418 return (-1);
13419 }
13420 }
13421
13422 if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
13423 !(sec->dofs_flags & DOF_SECF_LOAD)) {
13424 dtrace_dof_error(dof, "loadable section with load "
13425 "flag unset");
13426 return (-1);
13427 }
13428
13429 if (!(sec->dofs_flags & DOF_SECF_LOAD))
13430 continue; /* just ignore non-loadable sections */
13431
13432 if (!ISP2(sec->dofs_align)) {
13433 dtrace_dof_error(dof, "bad section alignment");
13434 return (-1);
13435 }
13436
13437 if (sec->dofs_offset & (sec->dofs_align - 1)) {
13438 dtrace_dof_error(dof, "misaligned section");
13439 return (-1);
13440 }
13441
13442 if (sec->dofs_offset > len || sec->dofs_size > len ||
13443 sec->dofs_offset + sec->dofs_size > len) {
13444 dtrace_dof_error(dof, "corrupt section header");
13445 return (-1);
13446 }
13447
13448 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
13449 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
13450 dtrace_dof_error(dof, "non-terminating string table");
13451 return (-1);
13452 }
13453 }
13454
13455 /*
13456 * Take a second pass through the sections and locate and perform any
13457 * relocations that are present. We do this after the first pass to
13458 * be sure that all sections have had their headers validated.
13459 */
13460 for (i = 0; i < dof->dofh_secnum; i++) {
13461 dof_sec_t *sec = (dof_sec_t *)(daddr +
13462 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13463
13464 if (!(sec->dofs_flags & DOF_SECF_LOAD))
13465 continue; /* skip sections that are not loadable */
13466
13467 switch (sec->dofs_type) {
13468 case DOF_SECT_URELHDR:
13469 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
13470 return (-1);
13471 break;
13472 }
13473 }
13474
13475 if ((enab = *enabp) == NULL)
13476 enab = *enabp = dtrace_enabling_create(vstate);
13477
13478 for (i = 0; i < dof->dofh_secnum; i++) {
13479 dof_sec_t *sec = (dof_sec_t *)(daddr +
13480 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13481
13482 if (sec->dofs_type != DOF_SECT_ECBDESC)
13483 continue;
13484
13485 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
13486 dtrace_enabling_destroy(enab);
13487 *enabp = NULL;
13488 return (-1);
13489 }
13490
13491 dtrace_enabling_add(enab, ep);
13492 }
13493
13494 return (0);
13495 }
13496
13497 /*
13498 * Process DOF for any options. This routine assumes that the DOF has been
13499 * at least processed by dtrace_dof_slurp().
13500 */
13501 static int
13502 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
13503 {
13504 int i, rval;
13505 uint32_t entsize;
13506 size_t offs;
13507 dof_optdesc_t *desc;
13508
13509 for (i = 0; i < dof->dofh_secnum; i++) {
13510 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
13511 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13512
13513 if (sec->dofs_type != DOF_SECT_OPTDESC)
13514 continue;
13515
13516 if (sec->dofs_align != sizeof (uint64_t)) {
13517 dtrace_dof_error(dof, "bad alignment in "
13518 "option description");
13519 return (EINVAL);
13520 }
13521
13522 if ((entsize = sec->dofs_entsize) == 0) {
13523 dtrace_dof_error(dof, "zeroed option entry size");
13524 return (EINVAL);
13525 }
13526
13527 if (entsize < sizeof (dof_optdesc_t)) {
13528 dtrace_dof_error(dof, "bad option entry size");
13529 return (EINVAL);
13530 }
13531
13532 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
13533 desc = (dof_optdesc_t *)((uintptr_t)dof +
13534 (uintptr_t)sec->dofs_offset + offs);
13535
13536 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
13537 dtrace_dof_error(dof, "non-zero option string");
13538 return (EINVAL);
13539 }
13540
13541 if (desc->dofo_value == DTRACEOPT_UNSET) {
13542 dtrace_dof_error(dof, "unset option");
13543 return (EINVAL);
13544 }
13545
13546 if ((rval = dtrace_state_option(state,
13547 desc->dofo_option, desc->dofo_value)) != 0) {
13548 dtrace_dof_error(dof, "rejected option");
13549 return (rval);
13550 }
13551 }
13552 }
13553
13554 return (0);
13555 }
13556
13557 /*
13558 * DTrace Consumer State Functions
13559 */
13560 int
13561 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
13562 {
13563 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
13564 void *base;
13565 uintptr_t limit;
13566 dtrace_dynvar_t *dvar, *next, *start;
13567 int i;
13568
13569 ASSERT(MUTEX_HELD(&dtrace_lock));
13570 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
13571
13572 bzero(dstate, sizeof (dtrace_dstate_t));
13573
13574 if ((dstate->dtds_chunksize = chunksize) == 0)
13575 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
13576
13577 VERIFY(dstate->dtds_chunksize < LONG_MAX);
13578
13579 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
13580 size = min;
13581
13582 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
13583 return (ENOMEM);
13584
13585 dstate->dtds_size = size;
13586 dstate->dtds_base = base;
13587 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
13588 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
13589
13590 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
13591
13592 if (hashsize != 1 && (hashsize & 1))
13593 hashsize--;
13594
13595 dstate->dtds_hashsize = hashsize;
13596 dstate->dtds_hash = dstate->dtds_base;
13597
13598 /*
13599 * Set all of our hash buckets to point to the single sink, and (if
13600 * it hasn't already been set), set the sink's hash value to be the
13601 * sink sentinel value. The sink is needed for dynamic variable
13602 * lookups to know that they have iterated over an entire, valid hash
13603 * chain.
13604 */
13605 for (i = 0; i < hashsize; i++)
13606 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
13607
13608 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
13609 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
13610
13611 /*
13612 * Determine number of active CPUs. Divide free list evenly among
13613 * active CPUs.
13614 */
13615 start = (dtrace_dynvar_t *)
13616 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
13617 limit = (uintptr_t)base + size;
13618
13619 VERIFY((uintptr_t)start < limit);
13620 VERIFY((uintptr_t)start >= (uintptr_t)base);
13621
13622 maxper = (limit - (uintptr_t)start) / NCPU;
13623 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13624
13625 for (i = 0; i < NCPU; i++) {
13626 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13627
13628 /*
13629 * If we don't even have enough chunks to make it once through
13630 * NCPUs, we're just going to allocate everything to the first
13631 * CPU. And if we're on the last CPU, we're going to allocate
13632 * whatever is left over. In either case, we set the limit to
13633 * be the limit of the dynamic variable space.
13634 */
13635 if (maxper == 0 || i == NCPU - 1) {
13636 limit = (uintptr_t)base + size;
13637 start = NULL;
13638 } else {
13639 limit = (uintptr_t)start + maxper;
13640 start = (dtrace_dynvar_t *)limit;
13641 }
13642
13643 VERIFY(limit <= (uintptr_t)base + size);
13644
13645 for (;;) {
13646 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13647 dstate->dtds_chunksize);
13648
13649 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13650 break;
13651
13652 VERIFY((uintptr_t)dvar >= (uintptr_t)base &&
13653 (uintptr_t)dvar <= (uintptr_t)base + size);
13654 dvar->dtdv_next = next;
13655 dvar = next;
13656 }
13657
13658 if (maxper == 0)
13659 break;
13660 }
13661
13662 return (0);
13663 }
13664
13665 void
13666 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13667 {
13668 ASSERT(MUTEX_HELD(&cpu_lock));
13669
13670 if (dstate->dtds_base == NULL)
13671 return;
13672
13673 kmem_free(dstate->dtds_base, dstate->dtds_size);
13674 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13675 }
13676
13677 static void
13678 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13679 {
13680 /*
13681 * Logical XOR, where are you?
13682 */
13683 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13684
13685 if (vstate->dtvs_nglobals > 0) {
13686 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13687 sizeof (dtrace_statvar_t *));
13688 }
13689
13690 if (vstate->dtvs_ntlocals > 0) {
13691 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13692 sizeof (dtrace_difv_t));
13693 }
13694
13695 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13696
13697 if (vstate->dtvs_nlocals > 0) {
13698 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13699 sizeof (dtrace_statvar_t *));
13700 }
13701 }
13702
13703 static void
13704 dtrace_state_clean(dtrace_state_t *state)
13705 {
13706 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13707 return;
13708
13709 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13710 dtrace_speculation_clean(state);
13711 }
13712
13713 static void
13714 dtrace_state_deadman(dtrace_state_t *state)
13715 {
13716 hrtime_t now;
13717
13718 dtrace_sync();
13719
13720 now = dtrace_gethrtime();
13721
13722 if (state != dtrace_anon.dta_state &&
13723 now - state->dts_laststatus >= dtrace_deadman_user)
13724 return;
13725
13726 /*
13727 * We must be sure that dts_alive never appears to be less than the
13728 * value upon entry to dtrace_state_deadman(), and because we lack a
13729 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13730 * store INT64_MAX to it, followed by a memory barrier, followed by
13731 * the new value. This assures that dts_alive never appears to be
13732 * less than its true value, regardless of the order in which the
13733 * stores to the underlying storage are issued.
13734 */
13735 state->dts_alive = INT64_MAX;
13736 dtrace_membar_producer();
13737 state->dts_alive = now;
13738 }
13739
13740 dtrace_state_t *
13741 dtrace_state_create(dev_t *devp, cred_t *cr)
13742 {
13743 minor_t minor;
13744 major_t major;
13745 char c[30];
13746 dtrace_state_t *state;
13747 dtrace_optval_t *opt;
13748 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13749
13750 ASSERT(MUTEX_HELD(&dtrace_lock));
13751 ASSERT(MUTEX_HELD(&cpu_lock));
13752
13753 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13754 VM_BESTFIT | VM_SLEEP);
13755
13756 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13757 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13758 return (NULL);
13759 }
13760
13761 state = ddi_get_soft_state(dtrace_softstate, minor);
13762 state->dts_epid = DTRACE_EPIDNONE + 1;
13763
13764 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
13765 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13766 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13767
13768 if (devp != NULL) {
13769 major = getemajor(*devp);
13770 } else {
13771 major = ddi_driver_major(dtrace_devi);
13772 }
13773
13774 state->dts_dev = makedevice(major, minor);
13775
13776 if (devp != NULL)
13777 *devp = state->dts_dev;
13778
13779 /*
13780 * We allocate NCPU buffers. On the one hand, this can be quite
13781 * a bit of memory per instance (nearly 36K on a Starcat). On the
13782 * other hand, it saves an additional memory reference in the probe
13783 * path.
13784 */
13785 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13786 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13787 state->dts_cleaner = CYCLIC_NONE;
13788 state->dts_deadman = CYCLIC_NONE;
13789 state->dts_vstate.dtvs_state = state;
13790
13791 for (i = 0; i < DTRACEOPT_MAX; i++)
13792 state->dts_options[i] = DTRACEOPT_UNSET;
13793
13794 /*
13795 * Set the default options.
13796 */
13797 opt = state->dts_options;
13798 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13799 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13800 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13801 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13802 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13803 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13804 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13805 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13806 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13807 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13808 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13809 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13810 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13811 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13812
13813 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13814
13815 /*
13816 * Depending on the user credentials, we set flag bits which alter probe
13817 * visibility or the amount of destructiveness allowed. In the case of
13818 * actual anonymous tracing, or the possession of all privileges, all of
13819 * the normal checks are bypassed.
13820 */
13821 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13822 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13823 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13824 } else {
13825 /*
13826 * Set up the credentials for this instantiation. We take a
13827 * hold on the credential to prevent it from disappearing on
13828 * us; this in turn prevents the zone_t referenced by this
13829 * credential from disappearing. This means that we can
13830 * examine the credential and the zone from probe context.
13831 */
13832 crhold(cr);
13833 state->dts_cred.dcr_cred = cr;
13834
13835 /*
13836 * CRA_PROC means "we have *some* privilege for dtrace" and
13837 * unlocks the use of variables like pid, zonename, etc.
13838 */
13839 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13840 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13841 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13842 }
13843
13844 /*
13845 * dtrace_user allows use of syscall and profile providers.
13846 * If the user also has proc_owner and/or proc_zone, we
13847 * extend the scope to include additional visibility and
13848 * destructive power.
13849 */
13850 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13851 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13852 state->dts_cred.dcr_visible |=
13853 DTRACE_CRV_ALLPROC;
13854
13855 state->dts_cred.dcr_action |=
13856 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13857 }
13858
13859 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13860 state->dts_cred.dcr_visible |=
13861 DTRACE_CRV_ALLZONE;
13862
13863 state->dts_cred.dcr_action |=
13864 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13865 }
13866
13867 /*
13868 * If we have all privs in whatever zone this is,
13869 * we can do destructive things to processes which
13870 * have altered credentials.
13871 */
13872 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13873 cr->cr_zone->zone_privset)) {
13874 state->dts_cred.dcr_action |=
13875 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13876 }
13877 }
13878
13879 /*
13880 * Holding the dtrace_kernel privilege also implies that
13881 * the user has the dtrace_user privilege from a visibility
13882 * perspective. But without further privileges, some
13883 * destructive actions are not available.
13884 */
13885 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13886 /*
13887 * Make all probes in all zones visible. However,
13888 * this doesn't mean that all actions become available
13889 * to all zones.
13890 */
13891 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13892 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13893
13894 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13895 DTRACE_CRA_PROC;
13896 /*
13897 * Holding proc_owner means that destructive actions
13898 * for *this* zone are allowed.
13899 */
13900 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13901 state->dts_cred.dcr_action |=
13902 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13903
13904 /*
13905 * Holding proc_zone means that destructive actions
13906 * for this user/group ID in all zones is allowed.
13907 */
13908 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13909 state->dts_cred.dcr_action |=
13910 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13911
13912 /*
13913 * If we have all privs in whatever zone this is,
13914 * we can do destructive things to processes which
13915 * have altered credentials.
13916 */
13917 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13918 cr->cr_zone->zone_privset)) {
13919 state->dts_cred.dcr_action |=
13920 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13921 }
13922 }
13923
13924 /*
13925 * Holding the dtrace_proc privilege gives control over fasttrap
13926 * and pid providers. We need to grant wider destructive
13927 * privileges in the event that the user has proc_owner and/or
13928 * proc_zone.
13929 */
13930 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13931 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13932 state->dts_cred.dcr_action |=
13933 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13934
13935 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13936 state->dts_cred.dcr_action |=
13937 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13938 }
13939 }
13940
13941 return (state);
13942 }
13943
13944 static int
13945 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13946 {
13947 dtrace_optval_t *opt = state->dts_options, size;
13948 processorid_t cpu;
13949 int flags = 0, rval, factor, divisor = 1;
13950
13951 ASSERT(MUTEX_HELD(&dtrace_lock));
13952 ASSERT(MUTEX_HELD(&cpu_lock));
13953 ASSERT(which < DTRACEOPT_MAX);
13954 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13955 (state == dtrace_anon.dta_state &&
13956 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13957
13958 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13959 return (0);
13960
13961 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13962 cpu = opt[DTRACEOPT_CPU];
13963
13964 if (which == DTRACEOPT_SPECSIZE)
13965 flags |= DTRACEBUF_NOSWITCH;
13966
13967 if (which == DTRACEOPT_BUFSIZE) {
13968 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13969 flags |= DTRACEBUF_RING;
13970
13971 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13972 flags |= DTRACEBUF_FILL;
13973
13974 if (state != dtrace_anon.dta_state ||
13975 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13976 flags |= DTRACEBUF_INACTIVE;
13977 }
13978
13979 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
13980 /*
13981 * The size must be 8-byte aligned. If the size is not 8-byte
13982 * aligned, drop it down by the difference.
13983 */
13984 if (size & (sizeof (uint64_t) - 1))
13985 size -= size & (sizeof (uint64_t) - 1);
13986
13987 if (size < state->dts_reserve) {
13988 /*
13989 * Buffers always must be large enough to accommodate
13990 * their prereserved space. We return E2BIG instead
13991 * of ENOMEM in this case to allow for user-level
13992 * software to differentiate the cases.
13993 */
13994 return (E2BIG);
13995 }
13996
13997 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
13998
13999 if (rval != ENOMEM) {
14000 opt[which] = size;
14001 return (rval);
14002 }
14003
14004 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14005 return (rval);
14006
14007 for (divisor = 2; divisor < factor; divisor <<= 1)
14008 continue;
14009 }
14010
14011 return (ENOMEM);
14012 }
14013
14014 static int
14015 dtrace_state_buffers(dtrace_state_t *state)
14016 {
14017 dtrace_speculation_t *spec = state->dts_speculations;
14018 int rval, i;
14019
14020 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
14021 DTRACEOPT_BUFSIZE)) != 0)
14022 return (rval);
14023
14024 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
14025 DTRACEOPT_AGGSIZE)) != 0)
14026 return (rval);
14027
14028 for (i = 0; i < state->dts_nspeculations; i++) {
14029 if ((rval = dtrace_state_buffer(state,
14030 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
14031 return (rval);
14032 }
14033
14034 return (0);
14035 }
14036
14037 static void
14038 dtrace_state_prereserve(dtrace_state_t *state)
14039 {
14040 dtrace_ecb_t *ecb;
14041 dtrace_probe_t *probe;
14042
14043 state->dts_reserve = 0;
14044
14045 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
14046 return;
14047
14048 /*
14049 * If our buffer policy is a "fill" buffer policy, we need to set the
14050 * prereserved space to be the space required by the END probes.
14051 */
14052 probe = dtrace_probes[dtrace_probeid_end - 1];
14053 ASSERT(probe != NULL);
14054
14055 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
14056 if (ecb->dte_state != state)
14057 continue;
14058
14059 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
14060 }
14061 }
14062
14063 static int
14064 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
14065 {
14066 dtrace_optval_t *opt = state->dts_options, sz, nspec;
14067 dtrace_speculation_t *spec;
14068 dtrace_buffer_t *buf;
14069 cyc_handler_t hdlr;
14070 cyc_time_t when;
14071 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14072 dtrace_icookie_t cookie;
14073
14074 mutex_enter(&cpu_lock);
14075 mutex_enter(&dtrace_lock);
14076
14077 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
14078 rval = EBUSY;
14079 goto out;
14080 }
14081
14082 /*
14083 * Before we can perform any checks, we must prime all of the
14084 * retained enablings that correspond to this state.
14085 */
14086 dtrace_enabling_prime(state);
14087
14088 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
14089 rval = EACCES;
14090 goto out;
14091 }
14092
14093 dtrace_state_prereserve(state);
14094
14095 /*
14096 * Now we want to do is try to allocate our speculations.
14097 * We do not automatically resize the number of speculations; if
14098 * this fails, we will fail the operation.
14099 */
14100 nspec = opt[DTRACEOPT_NSPEC];
14101 ASSERT(nspec != DTRACEOPT_UNSET);
14102
14103 if (nspec > INT_MAX) {
14104 rval = ENOMEM;
14105 goto out;
14106 }
14107
14108 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
14109 KM_NOSLEEP | KM_NORMALPRI);
14110
14111 if (spec == NULL) {
14112 rval = ENOMEM;
14113 goto out;
14114 }
14115
14116 state->dts_speculations = spec;
14117 state->dts_nspeculations = (int)nspec;
14118
14119 for (i = 0; i < nspec; i++) {
14120 if ((buf = kmem_zalloc(bufsize,
14121 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
14122 rval = ENOMEM;
14123 goto err;
14124 }
14125
14126 spec[i].dtsp_buffer = buf;
14127 }
14128
14129 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
14130 if (dtrace_anon.dta_state == NULL) {
14131 rval = ENOENT;
14132 goto out;
14133 }
14134
14135 if (state->dts_necbs != 0) {
14136 rval = EALREADY;
14137 goto out;
14138 }
14139
14140 state->dts_anon = dtrace_anon_grab();
14141 ASSERT(state->dts_anon != NULL);
14142 state = state->dts_anon;
14143
14144 /*
14145 * We want "grabanon" to be set in the grabbed state, so we'll
14146 * copy that option value from the grabbing state into the
14147 * grabbed state.
14148 */
14149 state->dts_options[DTRACEOPT_GRABANON] =
14150 opt[DTRACEOPT_GRABANON];
14151
14152 *cpu = dtrace_anon.dta_beganon;
14153
14154 /*
14155 * If the anonymous state is active (as it almost certainly
14156 * is if the anonymous enabling ultimately matched anything),
14157 * we don't allow any further option processing -- but we
14158 * don't return failure.
14159 */
14160 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14161 goto out;
14162 }
14163
14164 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
14165 opt[DTRACEOPT_AGGSIZE] != 0) {
14166 if (state->dts_aggregations == NULL) {
14167 /*
14168 * We're not going to create an aggregation buffer
14169 * because we don't have any ECBs that contain
14170 * aggregations -- set this option to 0.
14171 */
14172 opt[DTRACEOPT_AGGSIZE] = 0;
14173 } else {
14174 /*
14175 * If we have an aggregation buffer, we must also have
14176 * a buffer to use as scratch.
14177 */
14178 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
14179 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
14180 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
14181 }
14182 }
14183 }
14184
14185 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
14186 opt[DTRACEOPT_SPECSIZE] != 0) {
14187 if (!state->dts_speculates) {
14188 /*
14189 * We're not going to create speculation buffers
14190 * because we don't have any ECBs that actually
14191 * speculate -- set the speculation size to 0.
14192 */
14193 opt[DTRACEOPT_SPECSIZE] = 0;
14194 }
14195 }
14196
14197 /*
14198 * The bare minimum size for any buffer that we're actually going to
14199 * do anything to is sizeof (uint64_t).
14200 */
14201 sz = sizeof (uint64_t);
14202
14203 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
14204 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
14205 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
14206 /*
14207 * A buffer size has been explicitly set to 0 (or to a size
14208 * that will be adjusted to 0) and we need the space -- we
14209 * need to return failure. We return ENOSPC to differentiate
14210 * it from failing to allocate a buffer due to failure to meet
14211 * the reserve (for which we return E2BIG).
14212 */
14213 rval = ENOSPC;
14214 goto out;
14215 }
14216
14217 if ((rval = dtrace_state_buffers(state)) != 0)
14218 goto err;
14219
14220 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
14221 sz = dtrace_dstate_defsize;
14222
14223 do {
14224 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
14225
14226 if (rval == 0)
14227 break;
14228
14229 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14230 goto err;
14231 } while (sz >>= 1);
14232
14233 opt[DTRACEOPT_DYNVARSIZE] = sz;
14234
14235 if (rval != 0)
14236 goto err;
14237
14238 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
14239 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
14240
14241 if (opt[DTRACEOPT_CLEANRATE] == 0)
14242 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14243
14244 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
14245 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
14246
14247 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
14248 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14249
14250 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
14251 hdlr.cyh_arg = state;
14252 hdlr.cyh_level = CY_LOW_LEVEL;
14253
14254 when.cyt_when = 0;
14255 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
14256
14257 state->dts_cleaner = cyclic_add(&hdlr, &when);
14258
14259 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
14260 hdlr.cyh_arg = state;
14261 hdlr.cyh_level = CY_LOW_LEVEL;
14262
14263 when.cyt_when = 0;
14264 when.cyt_interval = dtrace_deadman_interval;
14265
14266 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
14267 state->dts_deadman = cyclic_add(&hdlr, &when);
14268
14269 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
14270
14271 if (state->dts_getf != 0 &&
14272 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14273 /*
14274 * We don't have kernel privs but we have at least one call
14275 * to getf(); we need to bump our zone's count, and (if
14276 * this is the first enabling to have an unprivileged call
14277 * to getf()) we need to hook into closef().
14278 */
14279 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
14280
14281 if (dtrace_getf++ == 0) {
14282 ASSERT(dtrace_closef == NULL);
14283 dtrace_closef = dtrace_getf_barrier;
14284 }
14285 }
14286
14287 /*
14288 * Now it's time to actually fire the BEGIN probe. We need to disable
14289 * interrupts here both to record the CPU on which we fired the BEGIN
14290 * probe (the data from this CPU will be processed first at user
14291 * level) and to manually activate the buffer for this CPU.
14292 */
14293 cookie = dtrace_interrupt_disable();
14294 *cpu = CPU->cpu_id;
14295 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
14296 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
14297
14298 dtrace_probe(dtrace_probeid_begin,
14299 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
14300 dtrace_interrupt_enable(cookie);
14301 /*
14302 * We may have had an exit action from a BEGIN probe; only change our
14303 * state to ACTIVE if we're still in WARMUP.
14304 */
14305 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
14306 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
14307
14308 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
14309 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
14310
14311 /*
14312 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
14313 * want each CPU to transition its principal buffer out of the
14314 * INACTIVE state. Doing this assures that no CPU will suddenly begin
14315 * processing an ECB halfway down a probe's ECB chain; all CPUs will
14316 * atomically transition from processing none of a state's ECBs to
14317 * processing all of them.
14318 */
14319 dtrace_xcall(DTRACE_CPUALL,
14320 (dtrace_xcall_t)dtrace_buffer_activate, state);
14321 goto out;
14322
14323 err:
14324 dtrace_buffer_free(state->dts_buffer);
14325 dtrace_buffer_free(state->dts_aggbuffer);
14326
14327 if ((nspec = state->dts_nspeculations) == 0) {
14328 ASSERT(state->dts_speculations == NULL);
14329 goto out;
14330 }
14331
14332 spec = state->dts_speculations;
14333 ASSERT(spec != NULL);
14334
14335 for (i = 0; i < state->dts_nspeculations; i++) {
14336 if ((buf = spec[i].dtsp_buffer) == NULL)
14337 break;
14338
14339 dtrace_buffer_free(buf);
14340 kmem_free(buf, bufsize);
14341 }
14342
14343 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14344 state->dts_nspeculations = 0;
14345 state->dts_speculations = NULL;
14346
14347 out:
14348 mutex_exit(&dtrace_lock);
14349 mutex_exit(&cpu_lock);
14350
14351 return (rval);
14352 }
14353
14354 static int
14355 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
14356 {
14357 dtrace_icookie_t cookie;
14358
14359 ASSERT(MUTEX_HELD(&dtrace_lock));
14360
14361 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
14362 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
14363 return (EINVAL);
14364
14365 /*
14366 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
14367 * to be sure that every CPU has seen it. See below for the details
14368 * on why this is done.
14369 */
14370 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
14371 dtrace_sync();
14372
14373 /*
14374 * By this point, it is impossible for any CPU to be still processing
14375 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
14376 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
14377 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
14378 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
14379 * iff we're in the END probe.
14380 */
14381 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
14382 dtrace_sync();
14383 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
14384
14385 /*
14386 * Finally, we can release the reserve and call the END probe. We
14387 * disable interrupts across calling the END probe to allow us to
14388 * return the CPU on which we actually called the END probe. This
14389 * allows user-land to be sure that this CPU's principal buffer is
14390 * processed last.
14391 */
14392 state->dts_reserve = 0;
14393
14394 cookie = dtrace_interrupt_disable();
14395 *cpu = CPU->cpu_id;
14396 dtrace_probe(dtrace_probeid_end,
14397 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
14398 dtrace_interrupt_enable(cookie);
14399
14400 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
14401 dtrace_sync();
14402
14403 if (state->dts_getf != 0 &&
14404 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14405 /*
14406 * We don't have kernel privs but we have at least one call
14407 * to getf(); we need to lower our zone's count, and (if
14408 * this is the last enabling to have an unprivileged call
14409 * to getf()) we need to clear the closef() hook.
14410 */
14411 ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
14412 ASSERT(dtrace_closef == dtrace_getf_barrier);
14413 ASSERT(dtrace_getf > 0);
14414
14415 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
14416
14417 if (--dtrace_getf == 0)
14418 dtrace_closef = NULL;
14419 }
14420
14421 return (0);
14422 }
14423
14424 static int
14425 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
14426 dtrace_optval_t val)
14427 {
14428 ASSERT(MUTEX_HELD(&dtrace_lock));
14429
14430 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14431 return (EBUSY);
14432
14433 if (option >= DTRACEOPT_MAX)
14434 return (EINVAL);
14435
14436 if (option != DTRACEOPT_CPU && val < 0)
14437 return (EINVAL);
14438
14439 switch (option) {
14440 case DTRACEOPT_DESTRUCTIVE:
14441 if (dtrace_destructive_disallow)
14442 return (EACCES);
14443
14444 state->dts_cred.dcr_destructive = 1;
14445 break;
14446
14447 case DTRACEOPT_BUFSIZE:
14448 case DTRACEOPT_DYNVARSIZE:
14449 case DTRACEOPT_AGGSIZE:
14450 case DTRACEOPT_SPECSIZE:
14451 case DTRACEOPT_STRSIZE:
14452 if (val < 0)
14453 return (EINVAL);
14454
14455 if (val >= LONG_MAX) {
14456 /*
14457 * If this is an otherwise negative value, set it to
14458 * the highest multiple of 128m less than LONG_MAX.
14459 * Technically, we're adjusting the size without
14460 * regard to the buffer resizing policy, but in fact,
14461 * this has no effect -- if we set the buffer size to
14462 * ~LONG_MAX and the buffer policy is ultimately set to
14463 * be "manual", the buffer allocation is guaranteed to
14464 * fail, if only because the allocation requires two
14465 * buffers. (We set the the size to the highest
14466 * multiple of 128m because it ensures that the size
14467 * will remain a multiple of a megabyte when
14468 * repeatedly halved -- all the way down to 15m.)
14469 */
14470 val = LONG_MAX - (1 << 27) + 1;
14471 }
14472 }
14473
14474 state->dts_options[option] = val;
14475
14476 return (0);
14477 }
14478
14479 static void
14480 dtrace_state_destroy(dtrace_state_t *state)
14481 {
14482 dtrace_ecb_t *ecb;
14483 dtrace_vstate_t *vstate = &state->dts_vstate;
14484 minor_t minor = getminor(state->dts_dev);
14485 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14486 dtrace_speculation_t *spec = state->dts_speculations;
14487 int nspec = state->dts_nspeculations;
14488 uint32_t match;
14489
14490 ASSERT(MUTEX_HELD(&dtrace_lock));
14491 ASSERT(MUTEX_HELD(&cpu_lock));
14492
14493 /*
14494 * First, retract any retained enablings for this state.
14495 */
14496 dtrace_enabling_retract(state);
14497 ASSERT(state->dts_nretained == 0);
14498
14499 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
14500 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
14501 /*
14502 * We have managed to come into dtrace_state_destroy() on a
14503 * hot enabling -- almost certainly because of a disorderly
14504 * shutdown of a consumer. (That is, a consumer that is
14505 * exiting without having called dtrace_stop().) In this case,
14506 * we're going to set our activity to be KILLED, and then
14507 * issue a sync to be sure that everyone is out of probe
14508 * context before we start blowing away ECBs.
14509 */
14510 state->dts_activity = DTRACE_ACTIVITY_KILLED;
14511 dtrace_sync();
14512 }
14513
14514 /*
14515 * Release the credential hold we took in dtrace_state_create().
14516 */
14517 if (state->dts_cred.dcr_cred != NULL)
14518 crfree(state->dts_cred.dcr_cred);
14519
14520 /*
14521 * Now we can safely disable and destroy any enabled probes. Because
14522 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
14523 * (especially if they're all enabled), we take two passes through the
14524 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
14525 * in the second we disable whatever is left over.
14526 */
14527 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
14528 for (i = 0; i < state->dts_necbs; i++) {
14529 if ((ecb = state->dts_ecbs[i]) == NULL)
14530 continue;
14531
14532 if (match && ecb->dte_probe != NULL) {
14533 dtrace_probe_t *probe = ecb->dte_probe;
14534 dtrace_provider_t *prov = probe->dtpr_provider;
14535
14536 if (!(prov->dtpv_priv.dtpp_flags & match))
14537 continue;
14538 }
14539
14540 dtrace_ecb_disable(ecb);
14541 dtrace_ecb_destroy(ecb);
14542 }
14543
14544 if (!match)
14545 break;
14546 }
14547
14548 /*
14549 * Before we free the buffers, perform one more sync to assure that
14550 * every CPU is out of probe context.
14551 */
14552 dtrace_sync();
14553
14554 dtrace_buffer_free(state->dts_buffer);
14555 dtrace_buffer_free(state->dts_aggbuffer);
14556
14557 for (i = 0; i < nspec; i++)
14558 dtrace_buffer_free(spec[i].dtsp_buffer);
14559
14560 if (state->dts_cleaner != CYCLIC_NONE)
14561 cyclic_remove(state->dts_cleaner);
14562
14563 if (state->dts_deadman != CYCLIC_NONE)
14564 cyclic_remove(state->dts_deadman);
14565
14566 dtrace_dstate_fini(&vstate->dtvs_dynvars);
14567 dtrace_vstate_fini(vstate);
14568 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14569
14570 if (state->dts_aggregations != NULL) {
14571 #ifdef DEBUG
14572 for (i = 0; i < state->dts_naggregations; i++)
14573 ASSERT(state->dts_aggregations[i] == NULL);
14574 #endif
14575 ASSERT(state->dts_naggregations > 0);
14576 kmem_free(state->dts_aggregations,
14577 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14578 }
14579
14580 kmem_free(state->dts_buffer, bufsize);
14581 kmem_free(state->dts_aggbuffer, bufsize);
14582
14583 for (i = 0; i < nspec; i++)
14584 kmem_free(spec[i].dtsp_buffer, bufsize);
14585
14586 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14587
14588 dtrace_format_destroy(state);
14589
14590 vmem_destroy(state->dts_aggid_arena);
14591 ddi_soft_state_free(dtrace_softstate, minor);
14592 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14593 }
14594
14595 /*
14596 * DTrace Anonymous Enabling Functions
14597 */
14598 static dtrace_state_t *
14599 dtrace_anon_grab(void)
14600 {
14601 dtrace_state_t *state;
14602
14603 ASSERT(MUTEX_HELD(&dtrace_lock));
14604
14605 if ((state = dtrace_anon.dta_state) == NULL) {
14606 ASSERT(dtrace_anon.dta_enabling == NULL);
14607 return (NULL);
14608 }
14609
14610 ASSERT(dtrace_anon.dta_enabling != NULL);
14611 ASSERT(dtrace_retained != NULL);
14612
14613 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14614 dtrace_anon.dta_enabling = NULL;
14615 dtrace_anon.dta_state = NULL;
14616
14617 return (state);
14618 }
14619
14620 static void
14621 dtrace_anon_property(void)
14622 {
14623 int i, rv;
14624 dtrace_state_t *state;
14625 dof_hdr_t *dof;
14626 char c[32]; /* enough for "dof-data-" + digits */
14627
14628 ASSERT(MUTEX_HELD(&dtrace_lock));
14629 ASSERT(MUTEX_HELD(&cpu_lock));
14630
14631 for (i = 0; ; i++) {
14632 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14633
14634 dtrace_err_verbose = 1;
14635
14636 if ((dof = dtrace_dof_property(c)) == NULL) {
14637 dtrace_err_verbose = 0;
14638 break;
14639 }
14640
14641 /*
14642 * We want to create anonymous state, so we need to transition
14643 * the kernel debugger to indicate that DTrace is active. If
14644 * this fails (e.g. because the debugger has modified text in
14645 * some way), we won't continue with the processing.
14646 */
14647 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14648 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14649 "enabling ignored.");
14650 dtrace_dof_destroy(dof);
14651 break;
14652 }
14653
14654 /*
14655 * If we haven't allocated an anonymous state, we'll do so now.
14656 */
14657 if ((state = dtrace_anon.dta_state) == NULL) {
14658 state = dtrace_state_create(NULL, NULL);
14659 dtrace_anon.dta_state = state;
14660
14661 if (state == NULL) {
14662 /*
14663 * This basically shouldn't happen: the only
14664 * failure mode from dtrace_state_create() is a
14665 * failure of ddi_soft_state_zalloc() that
14666 * itself should never happen. Still, the
14667 * interface allows for a failure mode, and
14668 * we want to fail as gracefully as possible:
14669 * we'll emit an error message and cease
14670 * processing anonymous state in this case.
14671 */
14672 cmn_err(CE_WARN, "failed to create "
14673 "anonymous state");
14674 dtrace_dof_destroy(dof);
14675 break;
14676 }
14677 }
14678
14679 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14680 &dtrace_anon.dta_enabling, 0, B_TRUE);
14681
14682 if (rv == 0)
14683 rv = dtrace_dof_options(dof, state);
14684
14685 dtrace_err_verbose = 0;
14686 dtrace_dof_destroy(dof);
14687
14688 if (rv != 0) {
14689 /*
14690 * This is malformed DOF; chuck any anonymous state
14691 * that we created.
14692 */
14693 ASSERT(dtrace_anon.dta_enabling == NULL);
14694 dtrace_state_destroy(state);
14695 dtrace_anon.dta_state = NULL;
14696 break;
14697 }
14698
14699 ASSERT(dtrace_anon.dta_enabling != NULL);
14700 }
14701
14702 if (dtrace_anon.dta_enabling != NULL) {
14703 int rval;
14704
14705 /*
14706 * dtrace_enabling_retain() can only fail because we are
14707 * trying to retain more enablings than are allowed -- but
14708 * we only have one anonymous enabling, and we are guaranteed
14709 * to be allowed at least one retained enabling; we assert
14710 * that dtrace_enabling_retain() returns success.
14711 */
14712 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14713 ASSERT(rval == 0);
14714
14715 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14716 }
14717 }
14718
14719 /*
14720 * DTrace Helper Functions
14721 */
14722 static void
14723 dtrace_helper_trace(dtrace_helper_action_t *helper,
14724 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14725 {
14726 uint32_t size, next, nnext, i;
14727 dtrace_helptrace_t *ent, *buffer;
14728 uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14729
14730 if ((buffer = dtrace_helptrace_buffer) == NULL)
14731 return;
14732
14733 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14734
14735 /*
14736 * What would a tracing framework be without its own tracing
14737 * framework? (Well, a hell of a lot simpler, for starters...)
14738 */
14739 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14740 sizeof (uint64_t) - sizeof (uint64_t);
14741
14742 /*
14743 * Iterate until we can allocate a slot in the trace buffer.
14744 */
14745 do {
14746 next = dtrace_helptrace_next;
14747
14748 if (next + size < dtrace_helptrace_bufsize) {
14749 nnext = next + size;
14750 } else {
14751 nnext = size;
14752 }
14753 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14754
14755 /*
14756 * We have our slot; fill it in.
14757 */
14758 if (nnext == size) {
14759 dtrace_helptrace_wrapped++;
14760 next = 0;
14761 }
14762
14763 ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
14764 ent->dtht_helper = helper;
14765 ent->dtht_where = where;
14766 ent->dtht_nlocals = vstate->dtvs_nlocals;
14767
14768 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14769 mstate->dtms_fltoffs : -1;
14770 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14771 ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
14772
14773 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14774 dtrace_statvar_t *svar;
14775
14776 if ((svar = vstate->dtvs_locals[i]) == NULL)
14777 continue;
14778
14779 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14780 ent->dtht_locals[i] =
14781 ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
14782 }
14783 }
14784
14785 static uint64_t
14786 dtrace_helper(int which, dtrace_mstate_t *mstate,
14787 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14788 {
14789 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14790 uint64_t sarg0 = mstate->dtms_arg[0];
14791 uint64_t sarg1 = mstate->dtms_arg[1];
14792 uint64_t rval;
14793 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14794 dtrace_helper_action_t *helper;
14795 dtrace_vstate_t *vstate;
14796 dtrace_difo_t *pred;
14797 int i, trace = dtrace_helptrace_buffer != NULL;
14798
14799 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14800
14801 if (helpers == NULL)
14802 return (0);
14803
14804 if ((helper = helpers->dthps_actions[which]) == NULL)
14805 return (0);
14806
14807 vstate = &helpers->dthps_vstate;
14808 mstate->dtms_arg[0] = arg0;
14809 mstate->dtms_arg[1] = arg1;
14810
14811 /*
14812 * Now iterate over each helper. If its predicate evaluates to 'true',
14813 * we'll call the corresponding actions. Note that the below calls
14814 * to dtrace_dif_emulate() may set faults in machine state. This is
14815 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14816 * the stored DIF offset with its own (which is the desired behavior).
14817 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14818 * from machine state; this is okay, too.
14819 */
14820 for (; helper != NULL; helper = helper->dtha_next) {
14821 if ((pred = helper->dtha_predicate) != NULL) {
14822 if (trace)
14823 dtrace_helper_trace(helper, mstate, vstate, 0);
14824
14825 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14826 goto next;
14827
14828 if (*flags & CPU_DTRACE_FAULT)
14829 goto err;
14830 }
14831
14832 for (i = 0; i < helper->dtha_nactions; i++) {
14833 if (trace)
14834 dtrace_helper_trace(helper,
14835 mstate, vstate, i + 1);
14836
14837 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14838 mstate, vstate, state);
14839
14840 if (*flags & CPU_DTRACE_FAULT)
14841 goto err;
14842 }
14843
14844 next:
14845 if (trace)
14846 dtrace_helper_trace(helper, mstate, vstate,
14847 DTRACE_HELPTRACE_NEXT);
14848 }
14849
14850 if (trace)
14851 dtrace_helper_trace(helper, mstate, vstate,
14852 DTRACE_HELPTRACE_DONE);
14853
14854 /*
14855 * Restore the arg0 that we saved upon entry.
14856 */
14857 mstate->dtms_arg[0] = sarg0;
14858 mstate->dtms_arg[1] = sarg1;
14859
14860 return (rval);
14861
14862 err:
14863 if (trace)
14864 dtrace_helper_trace(helper, mstate, vstate,
14865 DTRACE_HELPTRACE_ERR);
14866
14867 /*
14868 * Restore the arg0 that we saved upon entry.
14869 */
14870 mstate->dtms_arg[0] = sarg0;
14871 mstate->dtms_arg[1] = sarg1;
14872
14873 return (0);
14874 }
14875
14876 static void
14877 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14878 dtrace_vstate_t *vstate)
14879 {
14880 int i;
14881
14882 if (helper->dtha_predicate != NULL)
14883 dtrace_difo_release(helper->dtha_predicate, vstate);
14884
14885 for (i = 0; i < helper->dtha_nactions; i++) {
14886 ASSERT(helper->dtha_actions[i] != NULL);
14887 dtrace_difo_release(helper->dtha_actions[i], vstate);
14888 }
14889
14890 kmem_free(helper->dtha_actions,
14891 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14892 kmem_free(helper, sizeof (dtrace_helper_action_t));
14893 }
14894
14895 static int
14896 dtrace_helper_destroygen(int gen)
14897 {
14898 proc_t *p = curproc;
14899 dtrace_helpers_t *help = p->p_dtrace_helpers;
14900 dtrace_vstate_t *vstate;
14901 int i;
14902
14903 ASSERT(MUTEX_HELD(&dtrace_lock));
14904
14905 if (help == NULL || gen > help->dthps_generation)
14906 return (EINVAL);
14907
14908 vstate = &help->dthps_vstate;
14909
14910 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14911 dtrace_helper_action_t *last = NULL, *h, *next;
14912
14913 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14914 next = h->dtha_next;
14915
14916 if (h->dtha_generation == gen) {
14917 if (last != NULL) {
14918 last->dtha_next = next;
14919 } else {
14920 help->dthps_actions[i] = next;
14921 }
14922
14923 dtrace_helper_action_destroy(h, vstate);
14924 } else {
14925 last = h;
14926 }
14927 }
14928 }
14929
14930 /*
14931 * Interate until we've cleared out all helper providers with the
14932 * given generation number.
14933 */
14934 for (;;) {
14935 dtrace_helper_provider_t *prov;
14936
14937 /*
14938 * Look for a helper provider with the right generation. We
14939 * have to start back at the beginning of the list each time
14940 * because we drop dtrace_lock. It's unlikely that we'll make
14941 * more than two passes.
14942 */
14943 for (i = 0; i < help->dthps_nprovs; i++) {
14944 prov = help->dthps_provs[i];
14945
14946 if (prov->dthp_generation == gen)
14947 break;
14948 }
14949
14950 /*
14951 * If there were no matches, we're done.
14952 */
14953 if (i == help->dthps_nprovs)
14954 break;
14955
14956 /*
14957 * Move the last helper provider into this slot.
14958 */
14959 help->dthps_nprovs--;
14960 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14961 help->dthps_provs[help->dthps_nprovs] = NULL;
14962
14963 mutex_exit(&dtrace_lock);
14964
14965 /*
14966 * If we have a meta provider, remove this helper provider.
14967 */
14968 mutex_enter(&dtrace_meta_lock);
14969 if (dtrace_meta_pid != NULL) {
14970 ASSERT(dtrace_deferred_pid == NULL);
14971 dtrace_helper_provider_remove(&prov->dthp_prov,
14972 p->p_pid);
14973 }
14974 mutex_exit(&dtrace_meta_lock);
14975
14976 dtrace_helper_provider_destroy(prov);
14977
14978 mutex_enter(&dtrace_lock);
14979 }
14980
14981 return (0);
14982 }
14983
14984 static int
14985 dtrace_helper_validate(dtrace_helper_action_t *helper)
14986 {
14987 int err = 0, i;
14988 dtrace_difo_t *dp;
14989
14990 if ((dp = helper->dtha_predicate) != NULL)
14991 err += dtrace_difo_validate_helper(dp);
14992
14993 for (i = 0; i < helper->dtha_nactions; i++)
14994 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14995
14996 return (err == 0);
14997 }
14998
14999 static int
15000 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
15001 {
15002 dtrace_helpers_t *help;
15003 dtrace_helper_action_t *helper, *last;
15004 dtrace_actdesc_t *act;
15005 dtrace_vstate_t *vstate;
15006 dtrace_predicate_t *pred;
15007 int count = 0, nactions = 0, i;
15008
15009 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
15010 return (EINVAL);
15011
15012 help = curproc->p_dtrace_helpers;
15013 last = help->dthps_actions[which];
15014 vstate = &help->dthps_vstate;
15015
15016 for (count = 0; last != NULL; last = last->dtha_next) {
15017 count++;
15018 if (last->dtha_next == NULL)
15019 break;
15020 }
15021
15022 /*
15023 * If we already have dtrace_helper_actions_max helper actions for this
15024 * helper action type, we'll refuse to add a new one.
15025 */
15026 if (count >= dtrace_helper_actions_max)
15027 return (ENOSPC);
15028
15029 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
15030 helper->dtha_generation = help->dthps_generation;
15031
15032 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
15033 ASSERT(pred->dtp_difo != NULL);
15034 dtrace_difo_hold(pred->dtp_difo);
15035 helper->dtha_predicate = pred->dtp_difo;
15036 }
15037
15038 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
15039 if (act->dtad_kind != DTRACEACT_DIFEXPR)
15040 goto err;
15041
15042 if (act->dtad_difo == NULL)
15043 goto err;
15044
15045 nactions++;
15046 }
15047
15048 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
15049 (helper->dtha_nactions = nactions), KM_SLEEP);
15050
15051 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
15052 dtrace_difo_hold(act->dtad_difo);
15053 helper->dtha_actions[i++] = act->dtad_difo;
15054 }
15055
15056 if (!dtrace_helper_validate(helper))
15057 goto err;
15058
15059 if (last == NULL) {
15060 help->dthps_actions[which] = helper;
15061 } else {
15062 last->dtha_next = helper;
15063 }
15064
15065 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
15066 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
15067 dtrace_helptrace_next = 0;
15068 }
15069
15070 return (0);
15071 err:
15072 dtrace_helper_action_destroy(helper, vstate);
15073 return (EINVAL);
15074 }
15075
15076 static void
15077 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
15078 dof_helper_t *dofhp)
15079 {
15080 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
15081
15082 mutex_enter(&dtrace_meta_lock);
15083 mutex_enter(&dtrace_lock);
15084
15085 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
15086 /*
15087 * If the dtrace module is loaded but not attached, or if
15088 * there aren't isn't a meta provider registered to deal with
15089 * these provider descriptions, we need to postpone creating
15090 * the actual providers until later.
15091 */
15092
15093 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
15094 dtrace_deferred_pid != help) {
15095 help->dthps_deferred = 1;
15096 help->dthps_pid = p->p_pid;
15097 help->dthps_next = dtrace_deferred_pid;
15098 help->dthps_prev = NULL;
15099 if (dtrace_deferred_pid != NULL)
15100 dtrace_deferred_pid->dthps_prev = help;
15101 dtrace_deferred_pid = help;
15102 }
15103
15104 mutex_exit(&dtrace_lock);
15105
15106 } else if (dofhp != NULL) {
15107 /*
15108 * If the dtrace module is loaded and we have a particular
15109 * helper provider description, pass that off to the
15110 * meta provider.
15111 */
15112
15113 mutex_exit(&dtrace_lock);
15114
15115 dtrace_helper_provide(dofhp, p->p_pid);
15116
15117 } else {
15118 /*
15119 * Otherwise, just pass all the helper provider descriptions
15120 * off to the meta provider.
15121 */
15122
15123 int i;
15124 mutex_exit(&dtrace_lock);
15125
15126 for (i = 0; i < help->dthps_nprovs; i++) {
15127 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
15128 p->p_pid);
15129 }
15130 }
15131
15132 mutex_exit(&dtrace_meta_lock);
15133 }
15134
15135 static int
15136 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
15137 {
15138 dtrace_helpers_t *help;
15139 dtrace_helper_provider_t *hprov, **tmp_provs;
15140 uint_t tmp_maxprovs, i;
15141
15142 ASSERT(MUTEX_HELD(&dtrace_lock));
15143
15144 help = curproc->p_dtrace_helpers;
15145 ASSERT(help != NULL);
15146
15147 /*
15148 * If we already have dtrace_helper_providers_max helper providers,
15149 * we're refuse to add a new one.
15150 */
15151 if (help->dthps_nprovs >= dtrace_helper_providers_max)
15152 return (ENOSPC);
15153
15154 /*
15155 * Check to make sure this isn't a duplicate.
15156 */
15157 for (i = 0; i < help->dthps_nprovs; i++) {
15158 if (dofhp->dofhp_addr ==
15159 help->dthps_provs[i]->dthp_prov.dofhp_addr)
15160 return (EALREADY);
15161 }
15162
15163 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
15164 hprov->dthp_prov = *dofhp;
15165 hprov->dthp_ref = 1;
15166 hprov->dthp_generation = gen;
15167
15168 /*
15169 * Allocate a bigger table for helper providers if it's already full.
15170 */
15171 if (help->dthps_maxprovs == help->dthps_nprovs) {
15172 tmp_maxprovs = help->dthps_maxprovs;
15173 tmp_provs = help->dthps_provs;
15174
15175 if (help->dthps_maxprovs == 0)
15176 help->dthps_maxprovs = 2;
15177 else
15178 help->dthps_maxprovs *= 2;
15179 if (help->dthps_maxprovs > dtrace_helper_providers_max)
15180 help->dthps_maxprovs = dtrace_helper_providers_max;
15181
15182 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
15183
15184 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
15185 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15186
15187 if (tmp_provs != NULL) {
15188 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
15189 sizeof (dtrace_helper_provider_t *));
15190 kmem_free(tmp_provs, tmp_maxprovs *
15191 sizeof (dtrace_helper_provider_t *));
15192 }
15193 }
15194
15195 help->dthps_provs[help->dthps_nprovs] = hprov;
15196 help->dthps_nprovs++;
15197
15198 return (0);
15199 }
15200
15201 static void
15202 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
15203 {
15204 mutex_enter(&dtrace_lock);
15205
15206 if (--hprov->dthp_ref == 0) {
15207 dof_hdr_t *dof;
15208 mutex_exit(&dtrace_lock);
15209 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
15210 dtrace_dof_destroy(dof);
15211 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
15212 } else {
15213 mutex_exit(&dtrace_lock);
15214 }
15215 }
15216
15217 static int
15218 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
15219 {
15220 uintptr_t daddr = (uintptr_t)dof;
15221 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
15222 dof_provider_t *provider;
15223 dof_probe_t *probe;
15224 uint8_t *arg;
15225 char *strtab, *typestr;
15226 dof_stridx_t typeidx;
15227 size_t typesz;
15228 uint_t nprobes, j, k;
15229
15230 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
15231
15232 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
15233 dtrace_dof_error(dof, "misaligned section offset");
15234 return (-1);
15235 }
15236
15237 /*
15238 * The section needs to be large enough to contain the DOF provider
15239 * structure appropriate for the given version.
15240 */
15241 if (sec->dofs_size <
15242 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
15243 offsetof(dof_provider_t, dofpv_prenoffs) :
15244 sizeof (dof_provider_t))) {
15245 dtrace_dof_error(dof, "provider section too small");
15246 return (-1);
15247 }
15248
15249 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
15250 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
15251 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
15252 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
15253 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
15254
15255 if (str_sec == NULL || prb_sec == NULL ||
15256 arg_sec == NULL || off_sec == NULL)
15257 return (-1);
15258
15259 enoff_sec = NULL;
15260
15261 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
15262 provider->dofpv_prenoffs != DOF_SECT_NONE &&
15263 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
15264 provider->dofpv_prenoffs)) == NULL)
15265 return (-1);
15266
15267 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
15268
15269 if (provider->dofpv_name >= str_sec->dofs_size ||
15270 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
15271 dtrace_dof_error(dof, "invalid provider name");
15272 return (-1);
15273 }
15274
15275 if (prb_sec->dofs_entsize == 0 ||
15276 prb_sec->dofs_entsize > prb_sec->dofs_size) {
15277 dtrace_dof_error(dof, "invalid entry size");
15278 return (-1);
15279 }
15280
15281 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
15282 dtrace_dof_error(dof, "misaligned entry size");
15283 return (-1);
15284 }
15285
15286 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
15287 dtrace_dof_error(dof, "invalid entry size");
15288 return (-1);
15289 }
15290
15291 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
15292 dtrace_dof_error(dof, "misaligned section offset");
15293 return (-1);
15294 }
15295
15296 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
15297 dtrace_dof_error(dof, "invalid entry size");
15298 return (-1);
15299 }
15300
15301 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
15302
15303 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
15304
15305 /*
15306 * Take a pass through the probes to check for errors.
15307 */
15308 for (j = 0; j < nprobes; j++) {
15309 probe = (dof_probe_t *)(uintptr_t)(daddr +
15310 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
15311
15312 if (probe->dofpr_func >= str_sec->dofs_size) {
15313 dtrace_dof_error(dof, "invalid function name");
15314 return (-1);
15315 }
15316
15317 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
15318 dtrace_dof_error(dof, "function name too long");
15319 return (-1);
15320 }
15321
15322 if (probe->dofpr_name >= str_sec->dofs_size ||
15323 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
15324 dtrace_dof_error(dof, "invalid probe name");
15325 return (-1);
15326 }
15327
15328 /*
15329 * The offset count must not wrap the index, and the offsets
15330 * must also not overflow the section's data.
15331 */
15332 if (probe->dofpr_offidx + probe->dofpr_noffs <
15333 probe->dofpr_offidx ||
15334 (probe->dofpr_offidx + probe->dofpr_noffs) *
15335 off_sec->dofs_entsize > off_sec->dofs_size) {
15336 dtrace_dof_error(dof, "invalid probe offset");
15337 return (-1);
15338 }
15339
15340 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
15341 /*
15342 * If there's no is-enabled offset section, make sure
15343 * there aren't any is-enabled offsets. Otherwise
15344 * perform the same checks as for probe offsets
15345 * (immediately above).
15346 */
15347 if (enoff_sec == NULL) {
15348 if (probe->dofpr_enoffidx != 0 ||
15349 probe->dofpr_nenoffs != 0) {
15350 dtrace_dof_error(dof, "is-enabled "
15351 "offsets with null section");
15352 return (-1);
15353 }
15354 } else if (probe->dofpr_enoffidx +
15355 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
15356 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
15357 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
15358 dtrace_dof_error(dof, "invalid is-enabled "
15359 "offset");
15360 return (-1);
15361 }
15362
15363 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
15364 dtrace_dof_error(dof, "zero probe and "
15365 "is-enabled offsets");
15366 return (-1);
15367 }
15368 } else if (probe->dofpr_noffs == 0) {
15369 dtrace_dof_error(dof, "zero probe offsets");
15370 return (-1);
15371 }
15372
15373 if (probe->dofpr_argidx + probe->dofpr_xargc <
15374 probe->dofpr_argidx ||
15375 (probe->dofpr_argidx + probe->dofpr_xargc) *
15376 arg_sec->dofs_entsize > arg_sec->dofs_size) {
15377 dtrace_dof_error(dof, "invalid args");
15378 return (-1);
15379 }
15380
15381 typeidx = probe->dofpr_nargv;
15382 typestr = strtab + probe->dofpr_nargv;
15383 for (k = 0; k < probe->dofpr_nargc; k++) {
15384 if (typeidx >= str_sec->dofs_size) {
15385 dtrace_dof_error(dof, "bad "
15386 "native argument type");
15387 return (-1);
15388 }
15389
15390 typesz = strlen(typestr) + 1;
15391 if (typesz > DTRACE_ARGTYPELEN) {
15392 dtrace_dof_error(dof, "native "
15393 "argument type too long");
15394 return (-1);
15395 }
15396 typeidx += typesz;
15397 typestr += typesz;
15398 }
15399
15400 typeidx = probe->dofpr_xargv;
15401 typestr = strtab + probe->dofpr_xargv;
15402 for (k = 0; k < probe->dofpr_xargc; k++) {
15403 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
15404 dtrace_dof_error(dof, "bad "
15405 "native argument index");
15406 return (-1);
15407 }
15408
15409 if (typeidx >= str_sec->dofs_size) {
15410 dtrace_dof_error(dof, "bad "
15411 "translated argument type");
15412 return (-1);
15413 }
15414
15415 typesz = strlen(typestr) + 1;
15416 if (typesz > DTRACE_ARGTYPELEN) {
15417 dtrace_dof_error(dof, "translated argument "
15418 "type too long");
15419 return (-1);
15420 }
15421
15422 typeidx += typesz;
15423 typestr += typesz;
15424 }
15425 }
15426
15427 return (0);
15428 }
15429
15430 static int
15431 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
15432 {
15433 dtrace_helpers_t *help;
15434 dtrace_vstate_t *vstate;
15435 dtrace_enabling_t *enab = NULL;
15436 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
15437 uintptr_t daddr = (uintptr_t)dof;
15438
15439 ASSERT(MUTEX_HELD(&dtrace_lock));
15440
15441 if ((help = curproc->p_dtrace_helpers) == NULL)
15442 help = dtrace_helpers_create(curproc);
15443
15444 vstate = &help->dthps_vstate;
15445
15446 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
15447 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
15448 dtrace_dof_destroy(dof);
15449 return (rv);
15450 }
15451
15452 /*
15453 * Look for helper providers and validate their descriptions.
15454 */
15455 if (dhp != NULL) {
15456 for (i = 0; i < dof->dofh_secnum; i++) {
15457 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
15458 dof->dofh_secoff + i * dof->dofh_secsize);
15459
15460 if (sec->dofs_type != DOF_SECT_PROVIDER)
15461 continue;
15462
15463 if (dtrace_helper_provider_validate(dof, sec) != 0) {
15464 dtrace_enabling_destroy(enab);
15465 dtrace_dof_destroy(dof);
15466 return (-1);
15467 }
15468
15469 nprovs++;
15470 }
15471 }
15472
15473 /*
15474 * Now we need to walk through the ECB descriptions in the enabling.
15475 */
15476 for (i = 0; i < enab->dten_ndesc; i++) {
15477 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
15478 dtrace_probedesc_t *desc = &ep->dted_probe;
15479
15480 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
15481 continue;
15482
15483 if (strcmp(desc->dtpd_mod, "helper") != 0)
15484 continue;
15485
15486 if (strcmp(desc->dtpd_func, "ustack") != 0)
15487 continue;
15488
15489 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
15490 ep)) != 0) {
15491 /*
15492 * Adding this helper action failed -- we are now going
15493 * to rip out the entire generation and return failure.
15494 */
15495 (void) dtrace_helper_destroygen(help->dthps_generation);
15496 dtrace_enabling_destroy(enab);
15497 dtrace_dof_destroy(dof);
15498 return (-1);
15499 }
15500
15501 nhelpers++;
15502 }
15503
15504 if (nhelpers < enab->dten_ndesc)
15505 dtrace_dof_error(dof, "unmatched helpers");
15506
15507 gen = help->dthps_generation++;
15508 dtrace_enabling_destroy(enab);
15509
15510 if (dhp != NULL && nprovs > 0) {
15511 /*
15512 * Now that this is in-kernel, we change the sense of the
15513 * members: dofhp_dof denotes the in-kernel copy of the DOF
15514 * and dofhp_addr denotes the address at user-level.
15515 */
15516 dhp->dofhp_addr = dhp->dofhp_dof;
15517 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
15518
15519 if (dtrace_helper_provider_add(dhp, gen) == 0) {
15520 mutex_exit(&dtrace_lock);
15521 dtrace_helper_provider_register(curproc, help, dhp);
15522 mutex_enter(&dtrace_lock);
15523
15524 destroy = 0;
15525 }
15526 }
15527
15528 if (destroy)
15529 dtrace_dof_destroy(dof);
15530
15531 return (gen);
15532 }
15533
15534 static dtrace_helpers_t *
15535 dtrace_helpers_create(proc_t *p)
15536 {
15537 dtrace_helpers_t *help;
15538
15539 ASSERT(MUTEX_HELD(&dtrace_lock));
15540 ASSERT(p->p_dtrace_helpers == NULL);
15541
15542 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
15543 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
15544 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
15545
15546 p->p_dtrace_helpers = help;
15547 dtrace_helpers++;
15548
15549 return (help);
15550 }
15551
15552 static void
15553 dtrace_helpers_destroy(proc_t *p)
15554 {
15555 dtrace_helpers_t *help;
15556 dtrace_vstate_t *vstate;
15557 int i;
15558
15559 mutex_enter(&dtrace_lock);
15560
15561 ASSERT(p->p_dtrace_helpers != NULL);
15562 ASSERT(dtrace_helpers > 0);
15563
15564 help = p->p_dtrace_helpers;
15565 vstate = &help->dthps_vstate;
15566
15567 /*
15568 * We're now going to lose the help from this process.
15569 */
15570 p->p_dtrace_helpers = NULL;
15571 if (p == curproc) {
15572 dtrace_sync();
15573 } else {
15574 /*
15575 * It is sometimes necessary to clean up dtrace helpers from a
15576 * an incomplete child process as part of a failed fork
15577 * operation. In such situations, a dtrace_sync() call should
15578 * be unnecessary as the process should be devoid of threads,
15579 * much less any in probe context.
15580 */
15581 VERIFY(p->p_stat == SIDL);
15582 }
15583
15584 /*
15585 * Destroy the helper actions.
15586 */
15587 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15588 dtrace_helper_action_t *h, *next;
15589
15590 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15591 next = h->dtha_next;
15592 dtrace_helper_action_destroy(h, vstate);
15593 h = next;
15594 }
15595 }
15596
15597 mutex_exit(&dtrace_lock);
15598
15599 /*
15600 * Destroy the helper providers.
15601 */
15602 if (help->dthps_maxprovs > 0) {
15603 mutex_enter(&dtrace_meta_lock);
15604 if (dtrace_meta_pid != NULL) {
15605 ASSERT(dtrace_deferred_pid == NULL);
15606
15607 for (i = 0; i < help->dthps_nprovs; i++) {
15608 dtrace_helper_provider_remove(
15609 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15610 }
15611 } else {
15612 mutex_enter(&dtrace_lock);
15613 ASSERT(help->dthps_deferred == 0 ||
15614 help->dthps_next != NULL ||
15615 help->dthps_prev != NULL ||
15616 help == dtrace_deferred_pid);
15617
15618 /*
15619 * Remove the helper from the deferred list.
15620 */
15621 if (help->dthps_next != NULL)
15622 help->dthps_next->dthps_prev = help->dthps_prev;
15623 if (help->dthps_prev != NULL)
15624 help->dthps_prev->dthps_next = help->dthps_next;
15625 if (dtrace_deferred_pid == help) {
15626 dtrace_deferred_pid = help->dthps_next;
15627 ASSERT(help->dthps_prev == NULL);
15628 }
15629
15630 mutex_exit(&dtrace_lock);
15631 }
15632
15633 mutex_exit(&dtrace_meta_lock);
15634
15635 for (i = 0; i < help->dthps_nprovs; i++) {
15636 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15637 }
15638
15639 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15640 sizeof (dtrace_helper_provider_t *));
15641 }
15642
15643 mutex_enter(&dtrace_lock);
15644
15645 dtrace_vstate_fini(&help->dthps_vstate);
15646 kmem_free(help->dthps_actions,
15647 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15648 kmem_free(help, sizeof (dtrace_helpers_t));
15649
15650 --dtrace_helpers;
15651 mutex_exit(&dtrace_lock);
15652 }
15653
15654 static void
15655 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15656 {
15657 dtrace_helpers_t *help, *newhelp;
15658 dtrace_helper_action_t *helper, *new, *last;
15659 dtrace_difo_t *dp;
15660 dtrace_vstate_t *vstate;
15661 int i, j, sz, hasprovs = 0;
15662
15663 mutex_enter(&dtrace_lock);
15664 ASSERT(from->p_dtrace_helpers != NULL);
15665 ASSERT(dtrace_helpers > 0);
15666
15667 help = from->p_dtrace_helpers;
15668 newhelp = dtrace_helpers_create(to);
15669 ASSERT(to->p_dtrace_helpers != NULL);
15670
15671 newhelp->dthps_generation = help->dthps_generation;
15672 vstate = &newhelp->dthps_vstate;
15673
15674 /*
15675 * Duplicate the helper actions.
15676 */
15677 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15678 if ((helper = help->dthps_actions[i]) == NULL)
15679 continue;
15680
15681 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15682 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15683 KM_SLEEP);
15684 new->dtha_generation = helper->dtha_generation;
15685
15686 if ((dp = helper->dtha_predicate) != NULL) {
15687 dp = dtrace_difo_duplicate(dp, vstate);
15688 new->dtha_predicate = dp;
15689 }
15690
15691 new->dtha_nactions = helper->dtha_nactions;
15692 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15693 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15694
15695 for (j = 0; j < new->dtha_nactions; j++) {
15696 dtrace_difo_t *dp = helper->dtha_actions[j];
15697
15698 ASSERT(dp != NULL);
15699 dp = dtrace_difo_duplicate(dp, vstate);
15700 new->dtha_actions[j] = dp;
15701 }
15702
15703 if (last != NULL) {
15704 last->dtha_next = new;
15705 } else {
15706 newhelp->dthps_actions[i] = new;
15707 }
15708
15709 last = new;
15710 }
15711 }
15712
15713 /*
15714 * Duplicate the helper providers and register them with the
15715 * DTrace framework.
15716 */
15717 if (help->dthps_nprovs > 0) {
15718 newhelp->dthps_nprovs = help->dthps_nprovs;
15719 newhelp->dthps_maxprovs = help->dthps_nprovs;
15720 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15721 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15722 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15723 newhelp->dthps_provs[i] = help->dthps_provs[i];
15724 newhelp->dthps_provs[i]->dthp_ref++;
15725 }
15726
15727 hasprovs = 1;
15728 }
15729
15730 mutex_exit(&dtrace_lock);
15731
15732 if (hasprovs)
15733 dtrace_helper_provider_register(to, newhelp, NULL);
15734 }
15735
15736 /*
15737 * DTrace Hook Functions
15738 */
15739 static void
15740 dtrace_module_loaded(struct modctl *ctl)
15741 {
15742 dtrace_provider_t *prv;
15743
15744 mutex_enter(&dtrace_provider_lock);
15745 mutex_enter(&mod_lock);
15746
15747 ASSERT(ctl->mod_busy);
15748
15749 /*
15750 * We're going to call each providers per-module provide operation
15751 * specifying only this module.
15752 */
15753 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15754 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15755
15756 mutex_exit(&mod_lock);
15757 mutex_exit(&dtrace_provider_lock);
15758
15759 /*
15760 * If we have any retained enablings, we need to match against them.
15761 * Enabling probes requires that cpu_lock be held, and we cannot hold
15762 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15763 * module. (In particular, this happens when loading scheduling
15764 * classes.) So if we have any retained enablings, we need to dispatch
15765 * our task queue to do the match for us.
15766 */
15767 mutex_enter(&dtrace_lock);
15768
15769 if (dtrace_retained == NULL) {
15770 mutex_exit(&dtrace_lock);
15771 return;
15772 }
15773
15774 (void) taskq_dispatch(dtrace_taskq,
15775 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15776
15777 mutex_exit(&dtrace_lock);
15778
15779 /*
15780 * And now, for a little heuristic sleaze: in general, we want to
15781 * match modules as soon as they load. However, we cannot guarantee
15782 * this, because it would lead us to the lock ordering violation
15783 * outlined above. The common case, of course, is that cpu_lock is
15784 * _not_ held -- so we delay here for a clock tick, hoping that that's
15785 * long enough for the task queue to do its work. If it's not, it's
15786 * not a serious problem -- it just means that the module that we
15787 * just loaded may not be immediately instrumentable.
15788 */
15789 delay(1);
15790 }
15791
15792 static void
15793 dtrace_module_unloaded(struct modctl *ctl)
15794 {
15795 dtrace_probe_t template, *probe, *first, *next;
15796 dtrace_provider_t *prov;
15797
15798 template.dtpr_mod = ctl->mod_modname;
15799
15800 mutex_enter(&dtrace_provider_lock);
15801 mutex_enter(&mod_lock);
15802 mutex_enter(&dtrace_lock);
15803
15804 if (dtrace_bymod == NULL) {
15805 /*
15806 * The DTrace module is loaded (obviously) but not attached;
15807 * we don't have any work to do.
15808 */
15809 mutex_exit(&dtrace_provider_lock);
15810 mutex_exit(&mod_lock);
15811 mutex_exit(&dtrace_lock);
15812 return;
15813 }
15814
15815 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15816 probe != NULL; probe = probe->dtpr_nextmod) {
15817 if (probe->dtpr_ecb != NULL) {
15818 mutex_exit(&dtrace_provider_lock);
15819 mutex_exit(&mod_lock);
15820 mutex_exit(&dtrace_lock);
15821
15822 /*
15823 * This shouldn't _actually_ be possible -- we're
15824 * unloading a module that has an enabled probe in it.
15825 * (It's normally up to the provider to make sure that
15826 * this can't happen.) However, because dtps_enable()
15827 * doesn't have a failure mode, there can be an
15828 * enable/unload race. Upshot: we don't want to
15829 * assert, but we're not going to disable the
15830 * probe, either.
15831 */
15832 if (dtrace_err_verbose) {
15833 cmn_err(CE_WARN, "unloaded module '%s' had "
15834 "enabled probes", ctl->mod_modname);
15835 }
15836
15837 return;
15838 }
15839 }
15840
15841 probe = first;
15842
15843 for (first = NULL; probe != NULL; probe = next) {
15844 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15845
15846 dtrace_probes[probe->dtpr_id - 1] = NULL;
15847
15848 next = probe->dtpr_nextmod;
15849 dtrace_hash_remove(dtrace_bymod, probe);
15850 dtrace_hash_remove(dtrace_byfunc, probe);
15851 dtrace_hash_remove(dtrace_byname, probe);
15852
15853 if (first == NULL) {
15854 first = probe;
15855 probe->dtpr_nextmod = NULL;
15856 } else {
15857 probe->dtpr_nextmod = first;
15858 first = probe;
15859 }
15860 }
15861
15862 /*
15863 * We've removed all of the module's probes from the hash chains and
15864 * from the probe array. Now issue a dtrace_sync() to be sure that
15865 * everyone has cleared out from any probe array processing.
15866 */
15867 dtrace_sync();
15868
15869 for (probe = first; probe != NULL; probe = first) {
15870 first = probe->dtpr_nextmod;
15871 prov = probe->dtpr_provider;
15872 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15873 probe->dtpr_arg);
15874 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15875 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15876 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15877 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15878 kmem_free(probe, sizeof (dtrace_probe_t));
15879 }
15880
15881 mutex_exit(&dtrace_lock);
15882 mutex_exit(&mod_lock);
15883 mutex_exit(&dtrace_provider_lock);
15884 }
15885
15886 void
15887 dtrace_suspend(void)
15888 {
15889 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15890 }
15891
15892 void
15893 dtrace_resume(void)
15894 {
15895 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15896 }
15897
15898 static int
15899 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu, void *ptr __unused)
15900 {
15901 ASSERT(MUTEX_HELD(&cpu_lock));
15902 mutex_enter(&dtrace_lock);
15903
15904 switch (what) {
15905 case CPU_CONFIG: {
15906 dtrace_state_t *state;
15907 dtrace_optval_t *opt, rs, c;
15908
15909 /*
15910 * For now, we only allocate a new buffer for anonymous state.
15911 */
15912 if ((state = dtrace_anon.dta_state) == NULL)
15913 break;
15914
15915 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15916 break;
15917
15918 opt = state->dts_options;
15919 c = opt[DTRACEOPT_CPU];
15920
15921 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15922 break;
15923
15924 /*
15925 * Regardless of what the actual policy is, we're going to
15926 * temporarily set our resize policy to be manual. We're
15927 * also going to temporarily set our CPU option to denote
15928 * the newly configured CPU.
15929 */
15930 rs = opt[DTRACEOPT_BUFRESIZE];
15931 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15932 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15933
15934 (void) dtrace_state_buffers(state);
15935
15936 opt[DTRACEOPT_BUFRESIZE] = rs;
15937 opt[DTRACEOPT_CPU] = c;
15938
15939 break;
15940 }
15941
15942 case CPU_UNCONFIG:
15943 /*
15944 * We don't free the buffer in the CPU_UNCONFIG case. (The
15945 * buffer will be freed when the consumer exits.)
15946 */
15947 break;
15948
15949 default:
15950 break;
15951 }
15952
15953 mutex_exit(&dtrace_lock);
15954 return (0);
15955 }
15956
15957 static void
15958 dtrace_cpu_setup_initial(processorid_t cpu)
15959 {
15960 (void) dtrace_cpu_setup(CPU_CONFIG, cpu, NULL);
15961 }
15962
15963 static void
15964 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15965 {
15966 if (dtrace_toxranges >= dtrace_toxranges_max) {
15967 int osize, nsize;
15968 dtrace_toxrange_t *range;
15969
15970 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15971
15972 if (osize == 0) {
15973 ASSERT(dtrace_toxrange == NULL);
15974 ASSERT(dtrace_toxranges_max == 0);
15975 dtrace_toxranges_max = 1;
15976 } else {
15977 dtrace_toxranges_max <<= 1;
15978 }
15979
15980 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15981 range = kmem_zalloc(nsize, KM_SLEEP);
15982
15983 if (dtrace_toxrange != NULL) {
15984 ASSERT(osize != 0);
15985 bcopy(dtrace_toxrange, range, osize);
15986 kmem_free(dtrace_toxrange, osize);
15987 }
15988
15989 dtrace_toxrange = range;
15990 }
15991
15992 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == (uintptr_t)NULL);
15993 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == (uintptr_t)NULL);
15994
15995 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15996 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15997 dtrace_toxranges++;
15998 }
15999
16000 static void
16001 dtrace_getf_barrier()
16002 {
16003 /*
16004 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
16005 * that contain calls to getf(), this routine will be called on every
16006 * closef() before either the underlying vnode is released or the
16007 * file_t itself is freed. By the time we are here, it is essential
16008 * that the file_t can no longer be accessed from a call to getf()
16009 * in probe context -- that assures that a dtrace_sync() can be used
16010 * to clear out any enablings referring to the old structures.
16011 */
16012 if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
16013 kcred->cr_zone->zone_dtrace_getf != 0)
16014 dtrace_sync();
16015 }
16016
16017 /*
16018 * DTrace Driver Cookbook Functions
16019 */
16020 /*ARGSUSED*/
16021 static int
16022 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
16023 {
16024 dtrace_provider_id_t id;
16025 dtrace_state_t *state = NULL;
16026 dtrace_enabling_t *enab;
16027
16028 mutex_enter(&cpu_lock);
16029 mutex_enter(&dtrace_provider_lock);
16030 mutex_enter(&dtrace_lock);
16031
16032 if (ddi_soft_state_init(&dtrace_softstate,
16033 sizeof (dtrace_state_t), 0) != 0) {
16034 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
16035 mutex_exit(&cpu_lock);
16036 mutex_exit(&dtrace_provider_lock);
16037 mutex_exit(&dtrace_lock);
16038 return (DDI_FAILURE);
16039 }
16040
16041 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
16042 DTRACEMNRN_DTRACE, DDI_PSEUDO, 0) == DDI_FAILURE ||
16043 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
16044 DTRACEMNRN_HELPER, DDI_PSEUDO, 0) == DDI_FAILURE) {
16045 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
16046 ddi_remove_minor_node(devi, NULL);
16047 ddi_soft_state_fini(&dtrace_softstate);
16048 mutex_exit(&cpu_lock);
16049 mutex_exit(&dtrace_provider_lock);
16050 mutex_exit(&dtrace_lock);
16051 return (DDI_FAILURE);
16052 }
16053
16054 ddi_report_dev(devi);
16055 dtrace_devi = devi;
16056
16057 dtrace_modload = dtrace_module_loaded;
16058 dtrace_modunload = dtrace_module_unloaded;
16059 dtrace_cpu_init = dtrace_cpu_setup_initial;
16060 dtrace_helpers_cleanup = dtrace_helpers_destroy;
16061 dtrace_helpers_fork = dtrace_helpers_duplicate;
16062 dtrace_cpustart_init = dtrace_suspend;
16063 dtrace_cpustart_fini = dtrace_resume;
16064 dtrace_debugger_init = dtrace_suspend;
16065 dtrace_debugger_fini = dtrace_resume;
16066
16067 register_cpu_setup_func(dtrace_cpu_setup, NULL);
16068
16069 ASSERT(MUTEX_HELD(&cpu_lock));
16070
16071 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
16072 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
16073 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
16074 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
16075 VM_SLEEP | VMC_IDENTIFIER);
16076 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
16077 1, INT_MAX, 0);
16078
16079 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
16080 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
16081 NULL, NULL, NULL, NULL, NULL, 0);
16082
16083 ASSERT(MUTEX_HELD(&cpu_lock));
16084 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
16085 offsetof(dtrace_probe_t, dtpr_nextmod),
16086 offsetof(dtrace_probe_t, dtpr_prevmod));
16087
16088 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
16089 offsetof(dtrace_probe_t, dtpr_nextfunc),
16090 offsetof(dtrace_probe_t, dtpr_prevfunc));
16091
16092 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
16093 offsetof(dtrace_probe_t, dtpr_nextname),
16094 offsetof(dtrace_probe_t, dtpr_prevname));
16095
16096 if (dtrace_retain_max < 1) {
16097 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
16098 "setting to 1", dtrace_retain_max);
16099 dtrace_retain_max = 1;
16100 }
16101
16102 /*
16103 * Now discover our toxic ranges.
16104 */
16105 dtrace_toxic_ranges(dtrace_toxrange_add);
16106
16107 /*
16108 * Before we register ourselves as a provider to our own framework,
16109 * we would like to assert that dtrace_provider is NULL -- but that's
16110 * not true if we were loaded as a dependency of a DTrace provider.
16111 * Once we've registered, we can assert that dtrace_provider is our
16112 * pseudo provider.
16113 */
16114 (void) dtrace_register("dtrace", &dtrace_provider_attr,
16115 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
16116
16117 ASSERT(dtrace_provider != NULL);
16118 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
16119
16120 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
16121 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
16122 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
16123 dtrace_provider, NULL, NULL, "END", 0, NULL);
16124 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
16125 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
16126
16127 dtrace_anon_property();
16128 mutex_exit(&cpu_lock);
16129
16130 /*
16131 * If there are already providers, we must ask them to provide their
16132 * probes, and then match any anonymous enabling against them. Note
16133 * that there should be no other retained enablings at this time:
16134 * the only retained enablings at this time should be the anonymous
16135 * enabling.
16136 */
16137 if (dtrace_anon.dta_enabling != NULL) {
16138 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
16139
16140 dtrace_enabling_provide(NULL);
16141 state = dtrace_anon.dta_state;
16142
16143 /*
16144 * We couldn't hold cpu_lock across the above call to
16145 * dtrace_enabling_provide(), but we must hold it to actually
16146 * enable the probes. We have to drop all of our locks, pick
16147 * up cpu_lock, and regain our locks before matching the
16148 * retained anonymous enabling.
16149 */
16150 mutex_exit(&dtrace_lock);
16151 mutex_exit(&dtrace_provider_lock);
16152
16153 mutex_enter(&cpu_lock);
16154 mutex_enter(&dtrace_provider_lock);
16155 mutex_enter(&dtrace_lock);
16156
16157 if ((enab = dtrace_anon.dta_enabling) != NULL)
16158 (void) dtrace_enabling_match(enab, NULL);
16159
16160 mutex_exit(&cpu_lock);
16161 }
16162
16163 mutex_exit(&dtrace_lock);
16164 mutex_exit(&dtrace_provider_lock);
16165
16166 if (state != NULL) {
16167 /*
16168 * If we created any anonymous state, set it going now.
16169 */
16170 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
16171 }
16172
16173 return (DDI_SUCCESS);
16174 }
16175
16176 /*ARGSUSED*/
16177 static int
16178 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
16179 {
16180 dtrace_state_t *state;
16181 uint32_t priv;
16182 uid_t uid;
16183 zoneid_t zoneid;
16184
16185 if (getminor(*devp) == DTRACEMNRN_HELPER)
16186 return (0);
16187
16188 /*
16189 * If this wasn't an open with the "helper" minor, then it must be
16190 * the "dtrace" minor.
16191 */
16192 if (getminor(*devp) != DTRACEMNRN_DTRACE)
16193 return (ENXIO);
16194
16195 /*
16196 * If no DTRACE_PRIV_* bits are set in the credential, then the
16197 * caller lacks sufficient permission to do anything with DTrace.
16198 */
16199 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
16200 if (priv == DTRACE_PRIV_NONE)
16201 return (EACCES);
16202
16203 /*
16204 * Ask all providers to provide all their probes.
16205 */
16206 mutex_enter(&dtrace_provider_lock);
16207 dtrace_probe_provide(NULL, NULL);
16208 mutex_exit(&dtrace_provider_lock);
16209
16210 mutex_enter(&cpu_lock);
16211 mutex_enter(&dtrace_lock);
16212 dtrace_opens++;
16213 dtrace_membar_producer();
16214
16215 /*
16216 * If the kernel debugger is active (that is, if the kernel debugger
16217 * modified text in some way), we won't allow the open.
16218 */
16219 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
16220 dtrace_opens--;
16221 mutex_exit(&cpu_lock);
16222 mutex_exit(&dtrace_lock);
16223 return (EBUSY);
16224 }
16225
16226 if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
16227 /*
16228 * If DTrace helper tracing is enabled, we need to allocate the
16229 * trace buffer and initialize the values.
16230 */
16231 dtrace_helptrace_buffer =
16232 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
16233 dtrace_helptrace_next = 0;
16234 dtrace_helptrace_wrapped = 0;
16235 dtrace_helptrace_enable = 0;
16236 }
16237
16238 state = dtrace_state_create(devp, cred_p);
16239 mutex_exit(&cpu_lock);
16240
16241 if (state == NULL) {
16242 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
16243 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16244 mutex_exit(&dtrace_lock);
16245 return (EAGAIN);
16246 }
16247
16248 mutex_exit(&dtrace_lock);
16249
16250 return (0);
16251 }
16252
16253 /*ARGSUSED*/
16254 static int
16255 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
16256 {
16257 minor_t minor = getminor(dev);
16258 dtrace_state_t *state;
16259 dtrace_helptrace_t *buf = NULL;
16260
16261 if (minor == DTRACEMNRN_HELPER)
16262 return (0);
16263
16264 state = ddi_get_soft_state(dtrace_softstate, minor);
16265
16266 mutex_enter(&cpu_lock);
16267 mutex_enter(&dtrace_lock);
16268
16269 if (state->dts_anon) {
16270 /*
16271 * There is anonymous state. Destroy that first.
16272 */
16273 ASSERT(dtrace_anon.dta_state == NULL);
16274 dtrace_state_destroy(state->dts_anon);
16275 }
16276
16277 if (dtrace_helptrace_disable) {
16278 /*
16279 * If we have been told to disable helper tracing, set the
16280 * buffer to NULL before calling into dtrace_state_destroy();
16281 * we take advantage of its dtrace_sync() to know that no
16282 * CPU is in probe context with enabled helper tracing
16283 * after it returns.
16284 */
16285 buf = dtrace_helptrace_buffer;
16286 dtrace_helptrace_buffer = NULL;
16287 }
16288
16289 dtrace_state_destroy(state);
16290 ASSERT(dtrace_opens > 0);
16291
16292 /*
16293 * Only relinquish control of the kernel debugger interface when there
16294 * are no consumers and no anonymous enablings.
16295 */
16296 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
16297 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16298
16299 if (buf != NULL) {
16300 kmem_free(buf, dtrace_helptrace_bufsize);
16301 dtrace_helptrace_disable = 0;
16302 }
16303
16304 mutex_exit(&dtrace_lock);
16305 mutex_exit(&cpu_lock);
16306
16307 return (0);
16308 }
16309
16310 /*ARGSUSED*/
16311 static int
16312 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
16313 {
16314 int rval;
16315 dof_helper_t help, *dhp = NULL;
16316
16317 switch (cmd) {
16318 case DTRACEHIOC_ADDDOF:
16319 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
16320 dtrace_dof_error(NULL, "failed to copyin DOF helper");
16321 return (EFAULT);
16322 }
16323
16324 dhp = &help;
16325 arg = (intptr_t)help.dofhp_dof;
16326 /*FALLTHROUGH*/
16327
16328 case DTRACEHIOC_ADD: {
16329 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
16330
16331 if (dof == NULL)
16332 return (rval);
16333
16334 mutex_enter(&dtrace_lock);
16335
16336 /*
16337 * dtrace_helper_slurp() takes responsibility for the dof --
16338 * it may free it now or it may save it and free it later.
16339 */
16340 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
16341 *rv = rval;
16342 rval = 0;
16343 } else {
16344 rval = EINVAL;
16345 }
16346
16347 mutex_exit(&dtrace_lock);
16348 return (rval);
16349 }
16350
16351 case DTRACEHIOC_REMOVE: {
16352 mutex_enter(&dtrace_lock);
16353 rval = dtrace_helper_destroygen(arg);
16354 mutex_exit(&dtrace_lock);
16355
16356 return (rval);
16357 }
16358
16359 default:
16360 break;
16361 }
16362
16363 return (ENOTTY);
16364 }
16365
16366 /*ARGSUSED*/
16367 static int
16368 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
16369 {
16370 minor_t minor = getminor(dev);
16371 dtrace_state_t *state;
16372 int rval;
16373
16374 if (minor == DTRACEMNRN_HELPER)
16375 return (dtrace_ioctl_helper(cmd, arg, rv));
16376
16377 state = ddi_get_soft_state(dtrace_softstate, minor);
16378
16379 if (state->dts_anon) {
16380 ASSERT(dtrace_anon.dta_state == NULL);
16381 state = state->dts_anon;
16382 }
16383
16384 switch (cmd) {
16385 case DTRACEIOC_PROVIDER: {
16386 dtrace_providerdesc_t pvd;
16387 dtrace_provider_t *pvp;
16388
16389 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
16390 return (EFAULT);
16391
16392 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
16393 mutex_enter(&dtrace_provider_lock);
16394
16395 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
16396 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
16397 break;
16398 }
16399
16400 mutex_exit(&dtrace_provider_lock);
16401
16402 if (pvp == NULL)
16403 return (ESRCH);
16404
16405 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
16406 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
16407 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
16408 return (EFAULT);
16409
16410 return (0);
16411 }
16412
16413 case DTRACEIOC_EPROBE: {
16414 dtrace_eprobedesc_t epdesc;
16415 dtrace_ecb_t *ecb;
16416 dtrace_action_t *act;
16417 void *buf;
16418 size_t size;
16419 uintptr_t dest;
16420 int nrecs;
16421
16422 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
16423 return (EFAULT);
16424
16425 mutex_enter(&dtrace_lock);
16426
16427 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
16428 mutex_exit(&dtrace_lock);
16429 return (EINVAL);
16430 }
16431
16432 if (ecb->dte_probe == NULL) {
16433 mutex_exit(&dtrace_lock);
16434 return (EINVAL);
16435 }
16436
16437 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
16438 epdesc.dtepd_uarg = ecb->dte_uarg;
16439 epdesc.dtepd_size = ecb->dte_size;
16440
16441 nrecs = epdesc.dtepd_nrecs;
16442 epdesc.dtepd_nrecs = 0;
16443 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16444 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16445 continue;
16446
16447 epdesc.dtepd_nrecs++;
16448 }
16449
16450 /*
16451 * Now that we have the size, we need to allocate a temporary
16452 * buffer in which to store the complete description. We need
16453 * the temporary buffer to be able to drop dtrace_lock()
16454 * across the copyout(), below.
16455 */
16456 size = sizeof (dtrace_eprobedesc_t) +
16457 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
16458
16459 buf = kmem_alloc(size, KM_SLEEP);
16460 dest = (uintptr_t)buf;
16461
16462 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
16463 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
16464
16465 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16466 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16467 continue;
16468
16469 if (nrecs-- == 0)
16470 break;
16471
16472 bcopy(&act->dta_rec, (void *)dest,
16473 sizeof (dtrace_recdesc_t));
16474 dest += sizeof (dtrace_recdesc_t);
16475 }
16476
16477 mutex_exit(&dtrace_lock);
16478
16479 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16480 kmem_free(buf, size);
16481 return (EFAULT);
16482 }
16483
16484 kmem_free(buf, size);
16485 return (0);
16486 }
16487
16488 case DTRACEIOC_AGGDESC: {
16489 dtrace_aggdesc_t aggdesc;
16490 dtrace_action_t *act;
16491 dtrace_aggregation_t *agg;
16492 int nrecs;
16493 uint32_t offs;
16494 dtrace_recdesc_t *lrec;
16495 void *buf;
16496 size_t size;
16497 uintptr_t dest;
16498
16499 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16500 return (EFAULT);
16501
16502 mutex_enter(&dtrace_lock);
16503
16504 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16505 mutex_exit(&dtrace_lock);
16506 return (EINVAL);
16507 }
16508
16509 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16510
16511 nrecs = aggdesc.dtagd_nrecs;
16512 aggdesc.dtagd_nrecs = 0;
16513
16514 offs = agg->dtag_base;
16515 lrec = &agg->dtag_action.dta_rec;
16516 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16517
16518 for (act = agg->dtag_first; ; act = act->dta_next) {
16519 ASSERT(act->dta_intuple ||
16520 DTRACEACT_ISAGG(act->dta_kind));
16521
16522 /*
16523 * If this action has a record size of zero, it
16524 * denotes an argument to the aggregating action.
16525 * Because the presence of this record doesn't (or
16526 * shouldn't) affect the way the data is interpreted,
16527 * we don't copy it out to save user-level the
16528 * confusion of dealing with a zero-length record.
16529 */
16530 if (act->dta_rec.dtrd_size == 0) {
16531 ASSERT(agg->dtag_hasarg);
16532 continue;
16533 }
16534
16535 aggdesc.dtagd_nrecs++;
16536
16537 if (act == &agg->dtag_action)
16538 break;
16539 }
16540
16541 /*
16542 * Now that we have the size, we need to allocate a temporary
16543 * buffer in which to store the complete description. We need
16544 * the temporary buffer to be able to drop dtrace_lock()
16545 * across the copyout(), below.
16546 */
16547 size = sizeof (dtrace_aggdesc_t) +
16548 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16549
16550 buf = kmem_alloc(size, KM_SLEEP);
16551 dest = (uintptr_t)buf;
16552
16553 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16554 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16555
16556 for (act = agg->dtag_first; ; act = act->dta_next) {
16557 dtrace_recdesc_t rec = act->dta_rec;
16558
16559 /*
16560 * See the comment in the above loop for why we pass
16561 * over zero-length records.
16562 */
16563 if (rec.dtrd_size == 0) {
16564 ASSERT(agg->dtag_hasarg);
16565 continue;
16566 }
16567
16568 if (nrecs-- == 0)
16569 break;
16570
16571 rec.dtrd_offset -= offs;
16572 bcopy(&rec, (void *)dest, sizeof (rec));
16573 dest += sizeof (dtrace_recdesc_t);
16574
16575 if (act == &agg->dtag_action)
16576 break;
16577 }
16578
16579 mutex_exit(&dtrace_lock);
16580
16581 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16582 kmem_free(buf, size);
16583 return (EFAULT);
16584 }
16585
16586 kmem_free(buf, size);
16587 return (0);
16588 }
16589
16590 case DTRACEIOC_ENABLE: {
16591 dof_hdr_t *dof;
16592 dtrace_enabling_t *enab = NULL;
16593 dtrace_vstate_t *vstate;
16594 int err = 0;
16595
16596 *rv = 0;
16597
16598 /*
16599 * If a NULL argument has been passed, we take this as our
16600 * cue to reevaluate our enablings.
16601 */
16602 if (arg == 0) {
16603 dtrace_enabling_matchall();
16604
16605 return (0);
16606 }
16607
16608 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16609 return (rval);
16610
16611 mutex_enter(&cpu_lock);
16612 mutex_enter(&dtrace_lock);
16613 vstate = &state->dts_vstate;
16614
16615 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16616 mutex_exit(&dtrace_lock);
16617 mutex_exit(&cpu_lock);
16618 dtrace_dof_destroy(dof);
16619 return (EBUSY);
16620 }
16621
16622 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16623 mutex_exit(&dtrace_lock);
16624 mutex_exit(&cpu_lock);
16625 dtrace_dof_destroy(dof);
16626 return (EINVAL);
16627 }
16628
16629 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16630 dtrace_enabling_destroy(enab);
16631 mutex_exit(&dtrace_lock);
16632 mutex_exit(&cpu_lock);
16633 dtrace_dof_destroy(dof);
16634 return (rval);
16635 }
16636
16637 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16638 err = dtrace_enabling_retain(enab);
16639 } else {
16640 dtrace_enabling_destroy(enab);
16641 }
16642
16643 mutex_exit(&cpu_lock);
16644 mutex_exit(&dtrace_lock);
16645 dtrace_dof_destroy(dof);
16646
16647 return (err);
16648 }
16649
16650 case DTRACEIOC_REPLICATE: {
16651 dtrace_repldesc_t desc;
16652 dtrace_probedesc_t *match = &desc.dtrpd_match;
16653 dtrace_probedesc_t *create = &desc.dtrpd_create;
16654 int err;
16655
16656 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16657 return (EFAULT);
16658
16659 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16660 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16661 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16662 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16663
16664 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16665 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16666 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16667 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16668
16669 mutex_enter(&dtrace_lock);
16670 err = dtrace_enabling_replicate(state, match, create);
16671 mutex_exit(&dtrace_lock);
16672
16673 return (err);
16674 }
16675
16676 case DTRACEIOC_PROBEMATCH:
16677 case DTRACEIOC_PROBES: {
16678 dtrace_probe_t *probe = NULL;
16679 dtrace_probedesc_t desc;
16680 dtrace_probekey_t pkey;
16681 dtrace_id_t i;
16682 int m = 0;
16683 uint32_t priv;
16684 uid_t uid;
16685 zoneid_t zoneid;
16686
16687 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16688 return (EFAULT);
16689
16690 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16691 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16692 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16693 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16694
16695 /*
16696 * Before we attempt to match this probe, we want to give
16697 * all providers the opportunity to provide it.
16698 */
16699 if (desc.dtpd_id == DTRACE_IDNONE) {
16700 mutex_enter(&dtrace_provider_lock);
16701 dtrace_probe_provide(&desc, NULL);
16702 mutex_exit(&dtrace_provider_lock);
16703 desc.dtpd_id++;
16704 }
16705
16706 if (cmd == DTRACEIOC_PROBEMATCH) {
16707 dtrace_probekey(&desc, &pkey);
16708 pkey.dtpk_id = DTRACE_IDNONE;
16709 }
16710
16711 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16712
16713 mutex_enter(&dtrace_lock);
16714
16715 if (cmd == DTRACEIOC_PROBEMATCH) {
16716 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16717 if ((probe = dtrace_probes[i - 1]) != NULL &&
16718 (m = dtrace_match_probe(probe, &pkey,
16719 priv, uid, zoneid)) != 0)
16720 break;
16721 }
16722
16723 if (m < 0) {
16724 mutex_exit(&dtrace_lock);
16725 return (EINVAL);
16726 }
16727
16728 } else {
16729 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16730 if ((probe = dtrace_probes[i - 1]) != NULL &&
16731 dtrace_match_priv(probe, priv, uid, zoneid))
16732 break;
16733 }
16734 }
16735
16736 if (probe == NULL) {
16737 mutex_exit(&dtrace_lock);
16738 return (ESRCH);
16739 }
16740
16741 dtrace_probe_description(probe, &desc);
16742 mutex_exit(&dtrace_lock);
16743
16744 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16745 return (EFAULT);
16746
16747 return (0);
16748 }
16749
16750 case DTRACEIOC_PROBEARG: {
16751 dtrace_argdesc_t desc;
16752 dtrace_probe_t *probe;
16753 dtrace_provider_t *prov;
16754
16755 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16756 return (EFAULT);
16757
16758 if (desc.dtargd_id == DTRACE_IDNONE)
16759 return (EINVAL);
16760
16761 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16762 return (EINVAL);
16763
16764 mutex_enter(&dtrace_provider_lock);
16765 mutex_enter(&mod_lock);
16766 mutex_enter(&dtrace_lock);
16767
16768 if (desc.dtargd_id > dtrace_nprobes) {
16769 mutex_exit(&dtrace_lock);
16770 mutex_exit(&mod_lock);
16771 mutex_exit(&dtrace_provider_lock);
16772 return (EINVAL);
16773 }
16774
16775 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16776 mutex_exit(&dtrace_lock);
16777 mutex_exit(&mod_lock);
16778 mutex_exit(&dtrace_provider_lock);
16779 return (EINVAL);
16780 }
16781
16782 mutex_exit(&dtrace_lock);
16783
16784 prov = probe->dtpr_provider;
16785
16786 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16787 /*
16788 * There isn't any typed information for this probe.
16789 * Set the argument number to DTRACE_ARGNONE.
16790 */
16791 desc.dtargd_ndx = DTRACE_ARGNONE;
16792 } else {
16793 desc.dtargd_native[0] = '\0';
16794 desc.dtargd_xlate[0] = '\0';
16795 desc.dtargd_mapping = desc.dtargd_ndx;
16796
16797 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16798 probe->dtpr_id, probe->dtpr_arg, &desc);
16799 }
16800
16801 mutex_exit(&mod_lock);
16802 mutex_exit(&dtrace_provider_lock);
16803
16804 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16805 return (EFAULT);
16806
16807 return (0);
16808 }
16809
16810 case DTRACEIOC_GO: {
16811 processorid_t cpuid;
16812 rval = dtrace_state_go(state, &cpuid);
16813
16814 if (rval != 0)
16815 return (rval);
16816
16817 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16818 return (EFAULT);
16819
16820 return (0);
16821 }
16822
16823 case DTRACEIOC_STOP: {
16824 processorid_t cpuid;
16825
16826 mutex_enter(&dtrace_lock);
16827 rval = dtrace_state_stop(state, &cpuid);
16828 mutex_exit(&dtrace_lock);
16829
16830 if (rval != 0)
16831 return (rval);
16832
16833 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16834 return (EFAULT);
16835
16836 return (0);
16837 }
16838
16839 case DTRACEIOC_DOFGET: {
16840 dof_hdr_t hdr, *dof;
16841 uint64_t len;
16842
16843 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16844 return (EFAULT);
16845
16846 mutex_enter(&dtrace_lock);
16847 dof = dtrace_dof_create(state);
16848 mutex_exit(&dtrace_lock);
16849
16850 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16851 rval = copyout(dof, (void *)arg, len);
16852 dtrace_dof_destroy(dof);
16853
16854 return (rval == 0 ? 0 : EFAULT);
16855 }
16856
16857 case DTRACEIOC_AGGSNAP:
16858 case DTRACEIOC_BUFSNAP: {
16859 dtrace_bufdesc_t desc;
16860 caddr_t cached;
16861 dtrace_buffer_t *buf;
16862
16863 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16864 return (EFAULT);
16865
16866 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16867 return (EINVAL);
16868
16869 mutex_enter(&dtrace_lock);
16870
16871 if (cmd == DTRACEIOC_BUFSNAP) {
16872 buf = &state->dts_buffer[desc.dtbd_cpu];
16873 } else {
16874 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16875 }
16876
16877 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16878 size_t sz = buf->dtb_offset;
16879
16880 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16881 mutex_exit(&dtrace_lock);
16882 return (EBUSY);
16883 }
16884
16885 /*
16886 * If this buffer has already been consumed, we're
16887 * going to indicate that there's nothing left here
16888 * to consume.
16889 */
16890 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16891 mutex_exit(&dtrace_lock);
16892
16893 desc.dtbd_size = 0;
16894 desc.dtbd_drops = 0;
16895 desc.dtbd_errors = 0;
16896 desc.dtbd_oldest = 0;
16897 sz = sizeof (desc);
16898
16899 if (copyout(&desc, (void *)arg, sz) != 0)
16900 return (EFAULT);
16901
16902 return (0);
16903 }
16904
16905 /*
16906 * If this is a ring buffer that has wrapped, we want
16907 * to copy the whole thing out.
16908 */
16909 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16910 dtrace_buffer_polish(buf);
16911 sz = buf->dtb_size;
16912 }
16913
16914 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16915 mutex_exit(&dtrace_lock);
16916 return (EFAULT);
16917 }
16918
16919 desc.dtbd_size = sz;
16920 desc.dtbd_drops = buf->dtb_drops;
16921 desc.dtbd_errors = buf->dtb_errors;
16922 desc.dtbd_oldest = buf->dtb_xamot_offset;
16923 desc.dtbd_timestamp = dtrace_gethrtime();
16924
16925 mutex_exit(&dtrace_lock);
16926
16927 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16928 return (EFAULT);
16929
16930 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16931
16932 return (0);
16933 }
16934
16935 if (buf->dtb_tomax == NULL) {
16936 ASSERT(buf->dtb_xamot == NULL);
16937 mutex_exit(&dtrace_lock);
16938 return (ENOENT);
16939 }
16940
16941 cached = buf->dtb_tomax;
16942 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16943
16944 dtrace_xcall(desc.dtbd_cpu,
16945 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16946
16947 state->dts_errors += buf->dtb_xamot_errors;
16948
16949 /*
16950 * If the buffers did not actually switch, then the cross call
16951 * did not take place -- presumably because the given CPU is
16952 * not in the ready set. If this is the case, we'll return
16953 * ENOENT.
16954 */
16955 if (buf->dtb_tomax == cached) {
16956 ASSERT(buf->dtb_xamot != cached);
16957 mutex_exit(&dtrace_lock);
16958 return (ENOENT);
16959 }
16960
16961 ASSERT(cached == buf->dtb_xamot);
16962
16963 /*
16964 * We have our snapshot; now copy it out.
16965 */
16966 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16967 buf->dtb_xamot_offset) != 0) {
16968 mutex_exit(&dtrace_lock);
16969 return (EFAULT);
16970 }
16971
16972 desc.dtbd_size = buf->dtb_xamot_offset;
16973 desc.dtbd_drops = buf->dtb_xamot_drops;
16974 desc.dtbd_errors = buf->dtb_xamot_errors;
16975 desc.dtbd_oldest = 0;
16976 desc.dtbd_timestamp = buf->dtb_switched;
16977
16978 mutex_exit(&dtrace_lock);
16979
16980 /*
16981 * Finally, copy out the buffer description.
16982 */
16983 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16984 return (EFAULT);
16985
16986 return (0);
16987 }
16988
16989 case DTRACEIOC_CONF: {
16990 dtrace_conf_t conf;
16991
16992 bzero(&conf, sizeof (conf));
16993 conf.dtc_difversion = DIF_VERSION;
16994 conf.dtc_difintregs = DIF_DIR_NREGS;
16995 conf.dtc_diftupregs = DIF_DTR_NREGS;
16996 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16997
16998 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16999 return (EFAULT);
17000
17001 return (0);
17002 }
17003
17004 case DTRACEIOC_STATUS: {
17005 dtrace_status_t stat;
17006 dtrace_dstate_t *dstate;
17007 int i, j;
17008 uint64_t nerrs;
17009
17010 /*
17011 * See the comment in dtrace_state_deadman() for the reason
17012 * for setting dts_laststatus to INT64_MAX before setting
17013 * it to the correct value.
17014 */
17015 state->dts_laststatus = INT64_MAX;
17016 dtrace_membar_producer();
17017 state->dts_laststatus = dtrace_gethrtime();
17018
17019 bzero(&stat, sizeof (stat));
17020
17021 mutex_enter(&dtrace_lock);
17022
17023 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
17024 mutex_exit(&dtrace_lock);
17025 return (ENOENT);
17026 }
17027
17028 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
17029 stat.dtst_exiting = 1;
17030
17031 nerrs = state->dts_errors;
17032 dstate = &state->dts_vstate.dtvs_dynvars;
17033
17034 for (i = 0; i < NCPU; i++) {
17035 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
17036
17037 stat.dtst_dyndrops += dcpu->dtdsc_drops;
17038 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
17039 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
17040
17041 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
17042 stat.dtst_filled++;
17043
17044 nerrs += state->dts_buffer[i].dtb_errors;
17045
17046 for (j = 0; j < state->dts_nspeculations; j++) {
17047 dtrace_speculation_t *spec;
17048 dtrace_buffer_t *buf;
17049
17050 spec = &state->dts_speculations[j];
17051 buf = &spec->dtsp_buffer[i];
17052 stat.dtst_specdrops += buf->dtb_xamot_drops;
17053 }
17054 }
17055
17056 stat.dtst_specdrops_busy = state->dts_speculations_busy;
17057 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
17058 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
17059 stat.dtst_dblerrors = state->dts_dblerrors;
17060 stat.dtst_killed =
17061 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
17062 stat.dtst_errors = nerrs;
17063
17064 mutex_exit(&dtrace_lock);
17065
17066 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
17067 return (EFAULT);
17068
17069 return (0);
17070 }
17071
17072 case DTRACEIOC_FORMAT: {
17073 dtrace_fmtdesc_t fmt;
17074 char *str;
17075 int len;
17076
17077 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
17078 return (EFAULT);
17079
17080 mutex_enter(&dtrace_lock);
17081
17082 if (fmt.dtfd_format == 0 ||
17083 fmt.dtfd_format > state->dts_nformats) {
17084 mutex_exit(&dtrace_lock);
17085 return (EINVAL);
17086 }
17087
17088 /*
17089 * Format strings are allocated contiguously and they are
17090 * never freed; if a format index is less than the number
17091 * of formats, we can assert that the format map is non-NULL
17092 * and that the format for the specified index is non-NULL.
17093 */
17094 ASSERT(state->dts_formats != NULL);
17095 str = state->dts_formats[fmt.dtfd_format - 1];
17096 ASSERT(str != NULL);
17097
17098 len = strlen(str) + 1;
17099
17100 if (len > fmt.dtfd_length) {
17101 fmt.dtfd_length = len;
17102
17103 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
17104 mutex_exit(&dtrace_lock);
17105 return (EINVAL);
17106 }
17107 } else {
17108 if (copyout(str, fmt.dtfd_string, len) != 0) {
17109 mutex_exit(&dtrace_lock);
17110 return (EINVAL);
17111 }
17112 }
17113
17114 mutex_exit(&dtrace_lock);
17115 return (0);
17116 }
17117
17118 default:
17119 break;
17120 }
17121
17122 return (ENOTTY);
17123 }
17124
17125 /*ARGSUSED*/
17126 static int
17127 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
17128 {
17129 dtrace_state_t *state;
17130
17131 switch (cmd) {
17132 case DDI_DETACH:
17133 break;
17134
17135 case DDI_SUSPEND:
17136 return (DDI_SUCCESS);
17137
17138 default:
17139 return (DDI_FAILURE);
17140 }
17141
17142 mutex_enter(&cpu_lock);
17143 mutex_enter(&dtrace_provider_lock);
17144 mutex_enter(&dtrace_lock);
17145
17146 ASSERT(dtrace_opens == 0);
17147
17148 if (dtrace_helpers > 0) {
17149 mutex_exit(&dtrace_provider_lock);
17150 mutex_exit(&dtrace_lock);
17151 mutex_exit(&cpu_lock);
17152 return (DDI_FAILURE);
17153 }
17154
17155 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
17156 mutex_exit(&dtrace_provider_lock);
17157 mutex_exit(&dtrace_lock);
17158 mutex_exit(&cpu_lock);
17159 return (DDI_FAILURE);
17160 }
17161
17162 dtrace_provider = NULL;
17163
17164 if ((state = dtrace_anon_grab()) != NULL) {
17165 /*
17166 * If there were ECBs on this state, the provider should
17167 * have not been allowed to detach; assert that there is
17168 * none.
17169 */
17170 ASSERT(state->dts_necbs == 0);
17171 dtrace_state_destroy(state);
17172
17173 /*
17174 * If we're being detached with anonymous state, we need to
17175 * indicate to the kernel debugger that DTrace is now inactive.
17176 */
17177 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17178 }
17179
17180 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
17181 unregister_cpu_setup_func(dtrace_cpu_setup, NULL);
17182 dtrace_cpu_init = NULL;
17183 dtrace_helpers_cleanup = NULL;
17184 dtrace_helpers_fork = NULL;
17185 dtrace_cpustart_init = NULL;
17186 dtrace_cpustart_fini = NULL;
17187 dtrace_debugger_init = NULL;
17188 dtrace_debugger_fini = NULL;
17189 dtrace_modload = NULL;
17190 dtrace_modunload = NULL;
17191
17192 ASSERT(dtrace_getf == 0);
17193 ASSERT(dtrace_closef == NULL);
17194
17195 mutex_exit(&cpu_lock);
17196
17197 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
17198 dtrace_probes = NULL;
17199 dtrace_nprobes = 0;
17200
17201 dtrace_hash_destroy(dtrace_bymod);
17202 dtrace_hash_destroy(dtrace_byfunc);
17203 dtrace_hash_destroy(dtrace_byname);
17204 dtrace_bymod = NULL;
17205 dtrace_byfunc = NULL;
17206 dtrace_byname = NULL;
17207
17208 kmem_cache_destroy(dtrace_state_cache);
17209 vmem_destroy(dtrace_minor);
17210 vmem_destroy(dtrace_arena);
17211
17212 if (dtrace_toxrange != NULL) {
17213 kmem_free(dtrace_toxrange,
17214 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
17215 dtrace_toxrange = NULL;
17216 dtrace_toxranges = 0;
17217 dtrace_toxranges_max = 0;
17218 }
17219
17220 ddi_remove_minor_node(dtrace_devi, NULL);
17221 dtrace_devi = NULL;
17222
17223 ddi_soft_state_fini(&dtrace_softstate);
17224
17225 ASSERT(dtrace_vtime_references == 0);
17226 ASSERT(dtrace_opens == 0);
17227 ASSERT(dtrace_retained == NULL);
17228
17229 mutex_exit(&dtrace_lock);
17230 mutex_exit(&dtrace_provider_lock);
17231
17232 /*
17233 * We don't destroy the task queue until after we have dropped our
17234 * locks (taskq_destroy() may block on running tasks). To prevent
17235 * attempting to do work after we have effectively detached but before
17236 * the task queue has been destroyed, all tasks dispatched via the
17237 * task queue must check that DTrace is still attached before
17238 * performing any operation.
17239 */
17240 taskq_destroy(dtrace_taskq);
17241 dtrace_taskq = NULL;
17242
17243 return (DDI_SUCCESS);
17244 }
17245
17246 /*ARGSUSED*/
17247 static int
17248 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
17249 {
17250 int error;
17251
17252 switch (infocmd) {
17253 case DDI_INFO_DEVT2DEVINFO:
17254 *result = (void *)dtrace_devi;
17255 error = DDI_SUCCESS;
17256 break;
17257 case DDI_INFO_DEVT2INSTANCE:
17258 *result = (void *)0;
17259 error = DDI_SUCCESS;
17260 break;
17261 default:
17262 error = DDI_FAILURE;
17263 }
17264 return (error);
17265 }
17266
17267 static struct cb_ops dtrace_cb_ops = {
17268 dtrace_open, /* open */
17269 dtrace_close, /* close */
17270 nulldev, /* strategy */
17271 nulldev, /* print */
17272 nodev, /* dump */
17273 nodev, /* read */
17274 nodev, /* write */
17275 dtrace_ioctl, /* ioctl */
17276 nodev, /* devmap */
17277 nodev, /* mmap */
17278 nodev, /* segmap */
17279 nochpoll, /* poll */
17280 ddi_prop_op, /* cb_prop_op */
17281 0, /* streamtab */
17282 D_NEW | D_MP /* Driver compatibility flag */
17283 };
17284
17285 static struct dev_ops dtrace_ops = {
17286 DEVO_REV, /* devo_rev */
17287 0, /* refcnt */
17288 dtrace_info, /* get_dev_info */
17289 nulldev, /* identify */
17290 nulldev, /* probe */
17291 dtrace_attach, /* attach */
17292 dtrace_detach, /* detach */
17293 nodev, /* reset */
17294 &dtrace_cb_ops, /* driver operations */
17295 NULL, /* bus operations */
17296 nodev, /* dev power */
17297 ddi_quiesce_not_needed, /* quiesce */
17298 };
17299
17300 static struct modldrv modldrv = {
17301 &mod_driverops, /* module type (this is a pseudo driver) */
17302 "Dynamic Tracing", /* name of module */
17303 &dtrace_ops, /* driver ops */
17304 };
17305
17306 static struct modlinkage modlinkage = {
17307 MODREV_1,
17308 (void *)&modldrv,
17309 NULL
17310 };
17311
17312 int
17313 _init(void)
17314 {
17315 return (mod_install(&modlinkage));
17316 }
17317
17318 int
17319 _info(struct modinfo *modinfop)
17320 {
17321 return (mod_info(&modlinkage, modinfop));
17322 }
17323
17324 int
17325 _fini(void)
17326 {
17327 return (mod_remove(&modlinkage));
17328 }