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 (c) 2018, 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(void)
245 {}
246
247 static int
248 dtrace_enable_nullop(void)
249 {
250 return (0);
251 }
252
253 static dtrace_pops_t dtrace_provider_ops = {
254 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
255 (void (*)(void *, struct modctl *))dtrace_nullop,
256 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop,
257 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
258 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
259 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
260 NULL,
261 NULL,
262 NULL,
263 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
264 };
265
266 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
267 static dtrace_id_t dtrace_probeid_end; /* special END probe */
268 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
269
270 /*
271 * DTrace Helper Tracing Variables
272 *
273 * These variables should be set dynamically to enable helper tracing. The
274 * only variables that should be set are dtrace_helptrace_enable (which should
275 * be set to a non-zero value to allocate helper tracing buffers on the next
276 * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
277 * non-zero value to deallocate helper tracing buffers on the next close of
278 * /dev/dtrace). When (and only when) helper tracing is disabled, the
279 * buffer size may also be set via dtrace_helptrace_bufsize.
280 */
281 int dtrace_helptrace_enable = 0;
282 int dtrace_helptrace_disable = 0;
283 int dtrace_helptrace_bufsize = 16 * 1024 * 1024;
284 uint32_t dtrace_helptrace_nlocals;
285 static dtrace_helptrace_t *dtrace_helptrace_buffer;
286 static uint32_t dtrace_helptrace_next = 0;
287 static int dtrace_helptrace_wrapped = 0;
288
289 /*
290 * DTrace Error Hashing
291 *
292 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
293 * table. This is very useful for checking coverage of tests that are
294 * expected to induce DIF or DOF processing errors, and may be useful for
295 * debugging problems in the DIF code generator or in DOF generation . The
296 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
297 */
298 #ifdef DEBUG
299 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
300 static const char *dtrace_errlast;
301 static kthread_t *dtrace_errthread;
302 static kmutex_t dtrace_errlock;
303 #endif
304
305 /*
306 * DTrace Macros and Constants
307 *
308 * These are various macros that are useful in various spots in the
309 * implementation, along with a few random constants that have no meaning
310 * outside of the implementation. There is no real structure to this cpp
311 * mishmash -- but is there ever?
312 */
313 #define DTRACE_HASHSTR(hash, probe) \
314 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
315
316 #define DTRACE_HASHNEXT(hash, probe) \
317 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
318
319 #define DTRACE_HASHPREV(hash, probe) \
320 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
321
322 #define DTRACE_HASHEQ(hash, lhs, rhs) \
323 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
324 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
325
326 #define DTRACE_AGGHASHSIZE_SLEW 17
327
328 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
329
330 /*
331 * The key for a thread-local variable consists of the lower 61 bits of the
332 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
333 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
334 * equal to a variable identifier. This is necessary (but not sufficient) to
335 * assure that global associative arrays never collide with thread-local
336 * variables. To guarantee that they cannot collide, we must also define the
337 * order for keying dynamic variables. That order is:
338 *
339 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
340 *
341 * Because the variable-key and the tls-key are in orthogonal spaces, there is
342 * no way for a global variable key signature to match a thread-local key
343 * signature.
344 */
345 #define DTRACE_TLS_THRKEY(where) { \
346 uint_t intr = 0; \
347 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
348 for (; actv; actv >>= 1) \
349 intr++; \
350 ASSERT(intr < (1 << 3)); \
351 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
352 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
353 }
354
355 #define DT_BSWAP_8(x) ((x) & 0xff)
356 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
357 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
358 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
359
360 #define DT_MASK_LO 0x00000000FFFFFFFFULL
361
362 #define DTRACE_STORE(type, tomax, offset, what) \
363 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
364
365 #ifndef __x86
366 #define DTRACE_ALIGNCHECK(addr, size, flags) \
367 if (addr & (size - 1)) { \
368 *flags |= CPU_DTRACE_BADALIGN; \
369 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
370 return (0); \
371 }
372 #else
373 #define DTRACE_ALIGNCHECK(addr, size, flags)
374 #endif
375
376 /*
377 * Test whether a range of memory starting at testaddr of size testsz falls
378 * within the range of memory described by addr, sz. We take care to avoid
379 * problems with overflow and underflow of the unsigned quantities, and
380 * disallow all negative sizes. Ranges of size 0 are allowed.
381 */
382 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
383 ((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
384 (testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
385 (testaddr) + (testsz) >= (testaddr))
386
387 #define DTRACE_RANGE_REMAIN(remp, addr, baseaddr, basesz) \
388 do { \
389 if ((remp) != NULL) { \
390 *(remp) = (uintptr_t)(baseaddr) + (basesz) - (addr); \
391 } \
392 _NOTE(CONSTCOND) } while (0)
393
394
395 /*
396 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
397 * alloc_sz on the righthand side of the comparison in order to avoid overflow
398 * or underflow in the comparison with it. This is simpler than the INRANGE
399 * check above, because we know that the dtms_scratch_ptr is valid in the
400 * range. Allocations of size zero are allowed.
401 */
402 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
403 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
404 (mstate)->dtms_scratch_ptr >= (alloc_sz))
405
406 #define DTRACE_LOADFUNC(bits) \
407 /*CSTYLED*/ \
408 uint##bits##_t \
409 dtrace_load##bits(uintptr_t addr) \
410 { \
411 size_t size = bits / NBBY; \
412 /*CSTYLED*/ \
413 uint##bits##_t rval; \
414 int i; \
415 volatile uint16_t *flags = (volatile uint16_t *) \
416 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; \
417 \
418 DTRACE_ALIGNCHECK(addr, size, flags); \
419 \
420 for (i = 0; i < dtrace_toxranges; i++) { \
421 if (addr >= dtrace_toxrange[i].dtt_limit) \
422 continue; \
423 \
424 if (addr + size <= dtrace_toxrange[i].dtt_base) \
425 continue; \
426 \
427 /* \
428 * This address falls within a toxic region; return 0. \
429 */ \
430 *flags |= CPU_DTRACE_BADADDR; \
431 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
432 return (0); \
433 } \
434 \
435 *flags |= CPU_DTRACE_NOFAULT; \
436 /*CSTYLED*/ \
437 rval = *((volatile uint##bits##_t *)addr); \
438 *flags &= ~CPU_DTRACE_NOFAULT; \
439 \
440 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
441 }
442
443 #ifdef _LP64
444 #define dtrace_loadptr dtrace_load64
445 #else
446 #define dtrace_loadptr dtrace_load32
447 #endif
448
449 #define DTRACE_DYNHASH_FREE 0
450 #define DTRACE_DYNHASH_SINK 1
451 #define DTRACE_DYNHASH_VALID 2
452
453 #define DTRACE_MATCH_FAIL -1
454 #define DTRACE_MATCH_NEXT 0
455 #define DTRACE_MATCH_DONE 1
456 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
457 #define DTRACE_STATE_ALIGN 64
458
459 #define DTRACE_FLAGS2FLT(flags) \
460 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
461 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
462 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
463 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
464 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
465 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
466 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
467 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
468 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
469 DTRACEFLT_UNKNOWN)
470
471 #define DTRACEACT_ISSTRING(act) \
472 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
473 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
474
475 static size_t dtrace_strlen(const char *, size_t);
476 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
477 static void dtrace_enabling_provide(dtrace_provider_t *);
478 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
479 static void dtrace_enabling_matchall(void);
480 static void dtrace_enabling_reap(void);
481 static dtrace_state_t *dtrace_anon_grab(void);
482 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
483 dtrace_state_t *, uint64_t, uint64_t);
484 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
485 static void dtrace_buffer_drop(dtrace_buffer_t *);
486 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
487 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
488 dtrace_state_t *, dtrace_mstate_t *);
489 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
490 dtrace_optval_t);
491 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
492 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
493 static int dtrace_priv_proc(dtrace_state_t *, dtrace_mstate_t *);
494 static void dtrace_getf_barrier(void);
495 static int dtrace_canload_remains(uint64_t, size_t, size_t *,
496 dtrace_mstate_t *, dtrace_vstate_t *);
497 static int dtrace_canstore_remains(uint64_t, size_t, size_t *,
498 dtrace_mstate_t *, dtrace_vstate_t *);
499
500 /*
501 * DTrace Probe Context Functions
502 *
503 * These functions are called from probe context. Because probe context is
504 * any context in which C may be called, arbitrarily locks may be held,
505 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
506 * As a result, functions called from probe context may only call other DTrace
507 * support functions -- they may not interact at all with the system at large.
508 * (Note that the ASSERT macro is made probe-context safe by redefining it in
509 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
510 * loads are to be performed from probe context, they _must_ be in terms of
511 * the safe dtrace_load*() variants.
512 *
513 * Some functions in this block are not actually called from probe context;
514 * for these functions, there will be a comment above the function reading
515 * "Note: not called from probe context."
516 */
517 void
518 dtrace_panic(const char *format, ...)
519 {
520 va_list alist;
521
522 va_start(alist, format);
523 dtrace_vpanic(format, alist);
524 va_end(alist);
525 }
526
527 int
528 dtrace_assfail(const char *a, const char *f, int l)
529 {
530 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
531
532 /*
533 * We just need something here that even the most clever compiler
534 * cannot optimize away.
535 */
536 return (a[(uintptr_t)f]);
537 }
538
539 /*
540 * Atomically increment a specified error counter from probe context.
541 */
542 static void
543 dtrace_error(uint32_t *counter)
544 {
545 /*
546 * Most counters stored to in probe context are per-CPU counters.
547 * However, there are some error conditions that are sufficiently
548 * arcane that they don't merit per-CPU storage. If these counters
549 * are incremented concurrently on different CPUs, scalability will be
550 * adversely affected -- but we don't expect them to be white-hot in a
551 * correctly constructed enabling...
552 */
553 uint32_t oval, nval;
554
555 do {
556 oval = *counter;
557
558 if ((nval = oval + 1) == 0) {
559 /*
560 * If the counter would wrap, set it to 1 -- assuring
561 * that the counter is never zero when we have seen
562 * errors. (The counter must be 32-bits because we
563 * aren't guaranteed a 64-bit compare&swap operation.)
564 * To save this code both the infamy of being fingered
565 * by a priggish news story and the indignity of being
566 * the target of a neo-puritan witch trial, we're
567 * carefully avoiding any colorful description of the
568 * likelihood of this condition -- but suffice it to
569 * say that it is only slightly more likely than the
570 * overflow of predicate cache IDs, as discussed in
571 * dtrace_predicate_create().
572 */
573 nval = 1;
574 }
575 } while (dtrace_cas32(counter, oval, nval) != oval);
576 }
577
578 /*
579 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
580 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
581 */
582 /* BEGIN CSTYLED */
583 DTRACE_LOADFUNC(8)
584 DTRACE_LOADFUNC(16)
585 DTRACE_LOADFUNC(32)
586 DTRACE_LOADFUNC(64)
587 /* END CSTYLED */
588
589 static int
590 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
591 {
592 if (dest < mstate->dtms_scratch_base)
593 return (0);
594
595 if (dest + size < dest)
596 return (0);
597
598 if (dest + size > mstate->dtms_scratch_ptr)
599 return (0);
600
601 return (1);
602 }
603
604 static int
605 dtrace_canstore_statvar(uint64_t addr, size_t sz, size_t *remain,
606 dtrace_statvar_t **svars, int nsvars)
607 {
608 int i;
609 size_t maxglobalsize, maxlocalsize;
610
611 if (nsvars == 0)
612 return (0);
613
614 maxglobalsize = dtrace_statvar_maxsize + sizeof (uint64_t);
615 maxlocalsize = maxglobalsize * NCPU;
616
617 for (i = 0; i < nsvars; i++) {
618 dtrace_statvar_t *svar = svars[i];
619 uint8_t scope;
620 size_t size;
621
622 if (svar == NULL || (size = svar->dtsv_size) == 0)
623 continue;
624
625 scope = svar->dtsv_var.dtdv_scope;
626
627 /*
628 * We verify that our size is valid in the spirit of providing
629 * defense in depth: we want to prevent attackers from using
630 * DTrace to escalate an orthogonal kernel heap corruption bug
631 * into the ability to store to arbitrary locations in memory.
632 */
633 VERIFY((scope == DIFV_SCOPE_GLOBAL && size <= maxglobalsize) ||
634 (scope == DIFV_SCOPE_LOCAL && size <= maxlocalsize));
635
636 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data,
637 svar->dtsv_size)) {
638 DTRACE_RANGE_REMAIN(remain, addr, svar->dtsv_data,
639 svar->dtsv_size);
640 return (1);
641 }
642 }
643
644 return (0);
645 }
646
647 /*
648 * Check to see if the address is within a memory region to which a store may
649 * be issued. This includes the DTrace scratch areas, and any DTrace variable
650 * region. The caller of dtrace_canstore() is responsible for performing any
651 * alignment checks that are needed before stores are actually executed.
652 */
653 static int
654 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
655 dtrace_vstate_t *vstate)
656 {
657 return (dtrace_canstore_remains(addr, sz, NULL, mstate, vstate));
658 }
659
660 /*
661 * Implementation of dtrace_canstore which communicates the upper bound of the
662 * allowed memory region.
663 */
664 static int
665 dtrace_canstore_remains(uint64_t addr, size_t sz, size_t *remain,
666 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
667 {
668 /*
669 * First, check to see if the address is in scratch space...
670 */
671 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
672 mstate->dtms_scratch_size)) {
673 DTRACE_RANGE_REMAIN(remain, addr, mstate->dtms_scratch_base,
674 mstate->dtms_scratch_size);
675 return (1);
676 }
677
678 /*
679 * Now check to see if it's a dynamic variable. This check will pick
680 * up both thread-local variables and any global dynamically-allocated
681 * variables.
682 */
683 if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
684 vstate->dtvs_dynvars.dtds_size)) {
685 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
686 uintptr_t base = (uintptr_t)dstate->dtds_base +
687 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
688 uintptr_t chunkoffs;
689 dtrace_dynvar_t *dvar;
690
691 /*
692 * Before we assume that we can store here, we need to make
693 * sure that it isn't in our metadata -- storing to our
694 * dynamic variable metadata would corrupt our state. For
695 * the range to not include any dynamic variable metadata,
696 * it must:
697 *
698 * (1) Start above the hash table that is at the base of
699 * the dynamic variable space
700 *
701 * (2) Have a starting chunk offset that is beyond the
702 * dtrace_dynvar_t that is at the base of every chunk
703 *
704 * (3) Not span a chunk boundary
705 *
706 * (4) Not be in the tuple space of a dynamic variable
707 *
708 */
709 if (addr < base)
710 return (0);
711
712 chunkoffs = (addr - base) % dstate->dtds_chunksize;
713
714 if (chunkoffs < sizeof (dtrace_dynvar_t))
715 return (0);
716
717 if (chunkoffs + sz > dstate->dtds_chunksize)
718 return (0);
719
720 dvar = (dtrace_dynvar_t *)((uintptr_t)addr - chunkoffs);
721
722 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE)
723 return (0);
724
725 if (chunkoffs < sizeof (dtrace_dynvar_t) +
726 ((dvar->dtdv_tuple.dtt_nkeys - 1) * sizeof (dtrace_key_t)))
727 return (0);
728
729 DTRACE_RANGE_REMAIN(remain, addr, dvar, dstate->dtds_chunksize);
730 return (1);
731 }
732
733 /*
734 * Finally, check the static local and global variables. These checks
735 * take the longest, so we perform them last.
736 */
737 if (dtrace_canstore_statvar(addr, sz, remain,
738 vstate->dtvs_locals, vstate->dtvs_nlocals))
739 return (1);
740
741 if (dtrace_canstore_statvar(addr, sz, remain,
742 vstate->dtvs_globals, vstate->dtvs_nglobals))
743 return (1);
744
745 return (0);
746 }
747
748
749 /*
750 * Convenience routine to check to see if the address is within a memory
751 * region in which a load may be issued given the user's privilege level;
752 * if not, it sets the appropriate error flags and loads 'addr' into the
753 * illegal value slot.
754 *
755 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
756 * appropriate memory access protection.
757 */
758 static int
759 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
760 dtrace_vstate_t *vstate)
761 {
762 return (dtrace_canload_remains(addr, sz, NULL, mstate, vstate));
763 }
764
765 /*
766 * Implementation of dtrace_canload which communicates the upper bound of the
767 * allowed memory region.
768 */
769 static int
770 dtrace_canload_remains(uint64_t addr, size_t sz, size_t *remain,
771 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
772 {
773 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
774 file_t *fp;
775
776 /*
777 * If we hold the privilege to read from kernel memory, then
778 * everything is readable.
779 */
780 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
781 DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
782 return (1);
783 }
784
785 /*
786 * You can obviously read that which you can store.
787 */
788 if (dtrace_canstore_remains(addr, sz, remain, mstate, vstate))
789 return (1);
790
791 /*
792 * We're allowed to read from our own string table.
793 */
794 if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
795 mstate->dtms_difo->dtdo_strlen)) {
796 DTRACE_RANGE_REMAIN(remain, addr,
797 mstate->dtms_difo->dtdo_strtab,
798 mstate->dtms_difo->dtdo_strlen);
799 return (1);
800 }
801
802 if (vstate->dtvs_state != NULL &&
803 dtrace_priv_proc(vstate->dtvs_state, mstate)) {
804 proc_t *p;
805
806 /*
807 * When we have privileges to the current process, there are
808 * several context-related kernel structures that are safe to
809 * read, even absent the privilege to read from kernel memory.
810 * These reads are safe because these structures contain only
811 * state that (1) we're permitted to read, (2) is harmless or
812 * (3) contains pointers to additional kernel state that we're
813 * not permitted to read (and as such, do not present an
814 * opportunity for privilege escalation). Finally (and
815 * critically), because of the nature of their relation with
816 * the current thread context, the memory associated with these
817 * structures cannot change over the duration of probe context,
818 * and it is therefore impossible for this memory to be
819 * deallocated and reallocated as something else while it's
820 * being operated upon.
821 */
822 if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t))) {
823 DTRACE_RANGE_REMAIN(remain, addr, curthread,
824 sizeof (kthread_t));
825 return (1);
826 }
827
828 if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
829 sz, curthread->t_procp, sizeof (proc_t))) {
830 DTRACE_RANGE_REMAIN(remain, addr, curthread->t_procp,
831 sizeof (proc_t));
832 return (1);
833 }
834
835 if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
836 curthread->t_cred, sizeof (cred_t))) {
837 DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cred,
838 sizeof (cred_t));
839 return (1);
840 }
841
842 if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
843 &(p->p_pidp->pid_id), sizeof (pid_t))) {
844 DTRACE_RANGE_REMAIN(remain, addr, &(p->p_pidp->pid_id),
845 sizeof (pid_t));
846 return (1);
847 }
848
849 if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
850 curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
851 DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cpu,
852 offsetof(cpu_t, cpu_pause_thread));
853 return (1);
854 }
855 }
856
857 if ((fp = mstate->dtms_getf) != NULL) {
858 uintptr_t psz = sizeof (void *);
859 vnode_t *vp;
860 vnodeops_t *op;
861
862 /*
863 * When getf() returns a file_t, the enabling is implicitly
864 * granted the (transient) right to read the returned file_t
865 * as well as the v_path and v_op->vnop_name of the underlying
866 * vnode. These accesses are allowed after a successful
867 * getf() because the members that they refer to cannot change
868 * once set -- and the barrier logic in the kernel's closef()
869 * path assures that the file_t and its referenced vode_t
870 * cannot themselves be stale (that is, it impossible for
871 * either dtms_getf itself or its f_vnode member to reference
872 * freed memory).
873 */
874 if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t))) {
875 DTRACE_RANGE_REMAIN(remain, addr, fp, sizeof (file_t));
876 return (1);
877 }
878
879 if ((vp = fp->f_vnode) != NULL) {
880 size_t slen;
881
882 if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz)) {
883 DTRACE_RANGE_REMAIN(remain, addr, &vp->v_path,
884 psz);
885 return (1);
886 }
887
888 slen = strlen(vp->v_path) + 1;
889 if (DTRACE_INRANGE(addr, sz, vp->v_path, slen)) {
890 DTRACE_RANGE_REMAIN(remain, addr, vp->v_path,
891 slen);
892 return (1);
893 }
894
895 if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz)) {
896 DTRACE_RANGE_REMAIN(remain, addr, &vp->v_op,
897 psz);
898 return (1);
899 }
900
901 if ((op = vp->v_op) != NULL &&
902 DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
903 DTRACE_RANGE_REMAIN(remain, addr,
904 &op->vnop_name, psz);
905 return (1);
906 }
907
908 if (op != NULL && op->vnop_name != NULL &&
909 DTRACE_INRANGE(addr, sz, op->vnop_name,
910 (slen = strlen(op->vnop_name) + 1))) {
911 DTRACE_RANGE_REMAIN(remain, addr,
912 op->vnop_name, slen);
913 return (1);
914 }
915 }
916 }
917
918 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
919 *illval = addr;
920 return (0);
921 }
922
923 /*
924 * Convenience routine to check to see if a given string is within a memory
925 * region in which a load may be issued given the user's privilege level;
926 * this exists so that we don't need to issue unnecessary dtrace_strlen()
927 * calls in the event that the user has all privileges.
928 */
929 static int
930 dtrace_strcanload(uint64_t addr, size_t sz, size_t *remain,
931 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
932 {
933 size_t rsize;
934
935 /*
936 * If we hold the privilege to read from kernel memory, then
937 * everything is readable.
938 */
939 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
940 DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
941 return (1);
942 }
943
944 /*
945 * Even if the caller is uninterested in querying the remaining valid
946 * range, it is required to ensure that the access is allowed.
947 */
948 if (remain == NULL) {
949 remain = &rsize;
950 }
951 if (dtrace_canload_remains(addr, 0, remain, mstate, vstate)) {
952 size_t strsz;
953 /*
954 * Perform the strlen after determining the length of the
955 * memory region which is accessible. This prevents timing
956 * information from being used to find NULs in memory which is
957 * not accessible to the caller.
958 */
959 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr,
960 MIN(sz, *remain));
961 if (strsz <= *remain) {
962 return (1);
963 }
964 }
965
966 return (0);
967 }
968
969 /*
970 * Convenience routine to check to see if a given variable is within a memory
971 * region in which a load may be issued given the user's privilege level.
972 */
973 static int
974 dtrace_vcanload(void *src, dtrace_diftype_t *type, size_t *remain,
975 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
976 {
977 size_t sz;
978 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
979
980 /*
981 * Calculate the max size before performing any checks since even
982 * DTRACE_ACCESS_KERNEL-credentialed callers expect that this function
983 * return the max length via 'remain'.
984 */
985 if (type->dtdt_kind == DIF_TYPE_STRING) {
986 dtrace_state_t *state = vstate->dtvs_state;
987
988 if (state != NULL) {
989 sz = state->dts_options[DTRACEOPT_STRSIZE];
990 } else {
991 /*
992 * In helper context, we have a NULL state; fall back
993 * to using the system-wide default for the string size
994 * in this case.
995 */
996 sz = dtrace_strsize_default;
997 }
998 } else {
999 sz = type->dtdt_size;
1000 }
1001
1002 /*
1003 * If we hold the privilege to read from kernel memory, then
1004 * everything is readable.
1005 */
1006 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
1007 DTRACE_RANGE_REMAIN(remain, (uintptr_t)src, src, sz);
1008 return (1);
1009 }
1010
1011 if (type->dtdt_kind == DIF_TYPE_STRING) {
1012 return (dtrace_strcanload((uintptr_t)src, sz, remain, mstate,
1013 vstate));
1014 }
1015 return (dtrace_canload_remains((uintptr_t)src, sz, remain, mstate,
1016 vstate));
1017 }
1018
1019 /*
1020 * Convert a string to a signed integer using safe loads.
1021 *
1022 * NOTE: This function uses various macros from strtolctype.h to manipulate
1023 * digit values, etc -- these have all been checked to ensure they make
1024 * no additional function calls.
1025 */
1026 static int64_t
1027 dtrace_strtoll(char *input, int base, size_t limit)
1028 {
1029 uintptr_t pos = (uintptr_t)input;
1030 int64_t val = 0;
1031 int x;
1032 boolean_t neg = B_FALSE;
1033 char c, cc, ccc;
1034 uintptr_t end = pos + limit;
1035
1036 /*
1037 * Consume any whitespace preceding digits.
1038 */
1039 while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
1040 pos++;
1041
1042 /*
1043 * Handle an explicit sign if one is present.
1044 */
1045 if (c == '-' || c == '+') {
1046 if (c == '-')
1047 neg = B_TRUE;
1048 c = dtrace_load8(++pos);
1049 }
1050
1051 /*
1052 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
1053 * if present.
1054 */
1055 if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
1056 cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
1057 pos += 2;
1058 c = ccc;
1059 }
1060
1061 /*
1062 * Read in contiguous digits until the first non-digit character.
1063 */
1064 for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
1065 c = dtrace_load8(++pos))
1066 val = val * base + x;
1067
1068 return (neg ? -val : val);
1069 }
1070
1071 /*
1072 * Compare two strings using safe loads.
1073 */
1074 static int
1075 dtrace_strncmp(char *s1, char *s2, size_t limit)
1076 {
1077 uint8_t c1, c2;
1078 volatile uint16_t *flags;
1079
1080 if (s1 == s2 || limit == 0)
1081 return (0);
1082
1083 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1084
1085 do {
1086 if (s1 == NULL) {
1087 c1 = '\0';
1088 } else {
1089 c1 = dtrace_load8((uintptr_t)s1++);
1090 }
1091
1092 if (s2 == NULL) {
1093 c2 = '\0';
1094 } else {
1095 c2 = dtrace_load8((uintptr_t)s2++);
1096 }
1097
1098 if (c1 != c2)
1099 return (c1 - c2);
1100 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
1101
1102 return (0);
1103 }
1104
1105 /*
1106 * Compute strlen(s) for a string using safe memory accesses. The additional
1107 * len parameter is used to specify a maximum length to ensure completion.
1108 */
1109 static size_t
1110 dtrace_strlen(const char *s, size_t lim)
1111 {
1112 uint_t len;
1113
1114 for (len = 0; len != lim; len++) {
1115 if (dtrace_load8((uintptr_t)s++) == '\0')
1116 break;
1117 }
1118
1119 return (len);
1120 }
1121
1122 /*
1123 * Check if an address falls within a toxic region.
1124 */
1125 static int
1126 dtrace_istoxic(uintptr_t kaddr, size_t size)
1127 {
1128 uintptr_t taddr, tsize;
1129 int i;
1130
1131 for (i = 0; i < dtrace_toxranges; i++) {
1132 taddr = dtrace_toxrange[i].dtt_base;
1133 tsize = dtrace_toxrange[i].dtt_limit - taddr;
1134
1135 if (kaddr - taddr < tsize) {
1136 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1137 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
1138 return (1);
1139 }
1140
1141 if (taddr - kaddr < size) {
1142 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1143 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
1144 return (1);
1145 }
1146 }
1147
1148 return (0);
1149 }
1150
1151 /*
1152 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
1153 * memory specified by the DIF program. The dst is assumed to be safe memory
1154 * that we can store to directly because it is managed by DTrace. As with
1155 * standard bcopy, overlapping copies are handled properly.
1156 */
1157 static void
1158 dtrace_bcopy(const void *src, void *dst, size_t len)
1159 {
1160 if (len != 0) {
1161 uint8_t *s1 = dst;
1162 const uint8_t *s2 = src;
1163
1164 if (s1 <= s2) {
1165 do {
1166 *s1++ = dtrace_load8((uintptr_t)s2++);
1167 } while (--len != 0);
1168 } else {
1169 s2 += len;
1170 s1 += len;
1171
1172 do {
1173 *--s1 = dtrace_load8((uintptr_t)--s2);
1174 } while (--len != 0);
1175 }
1176 }
1177 }
1178
1179 /*
1180 * Copy src to dst using safe memory accesses, up to either the specified
1181 * length, or the point that a nul byte is encountered. The src is assumed to
1182 * be unsafe memory specified by the DIF program. The dst is assumed to be
1183 * safe memory that we can store to directly because it is managed by DTrace.
1184 * Unlike dtrace_bcopy(), overlapping regions are not handled.
1185 */
1186 static void
1187 dtrace_strcpy(const void *src, void *dst, size_t len)
1188 {
1189 if (len != 0) {
1190 uint8_t *s1 = dst, c;
1191 const uint8_t *s2 = src;
1192
1193 do {
1194 *s1++ = c = dtrace_load8((uintptr_t)s2++);
1195 } while (--len != 0 && c != '\0');
1196 }
1197 }
1198
1199 /*
1200 * Copy src to dst, deriving the size and type from the specified (BYREF)
1201 * variable type. The src is assumed to be unsafe memory specified by the DIF
1202 * program. The dst is assumed to be DTrace variable memory that is of the
1203 * specified type; we assume that we can store to directly.
1204 */
1205 static void
1206 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type, size_t limit)
1207 {
1208 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1209
1210 if (type->dtdt_kind == DIF_TYPE_STRING) {
1211 dtrace_strcpy(src, dst, MIN(type->dtdt_size, limit));
1212 } else {
1213 dtrace_bcopy(src, dst, MIN(type->dtdt_size, limit));
1214 }
1215 }
1216
1217 /*
1218 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1219 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1220 * safe memory that we can access directly because it is managed by DTrace.
1221 */
1222 static int
1223 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1224 {
1225 volatile uint16_t *flags;
1226
1227 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1228
1229 if (s1 == s2)
1230 return (0);
1231
1232 if (s1 == NULL || s2 == NULL)
1233 return (1);
1234
1235 if (s1 != s2 && len != 0) {
1236 const uint8_t *ps1 = s1;
1237 const uint8_t *ps2 = s2;
1238
1239 do {
1240 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1241 return (1);
1242 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1243 }
1244 return (0);
1245 }
1246
1247 /*
1248 * Zero the specified region using a simple byte-by-byte loop. Note that this
1249 * is for safe DTrace-managed memory only.
1250 */
1251 static void
1252 dtrace_bzero(void *dst, size_t len)
1253 {
1254 uchar_t *cp;
1255
1256 for (cp = dst; len != 0; len--)
1257 *cp++ = 0;
1258 }
1259
1260 static void
1261 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1262 {
1263 uint64_t result[2];
1264
1265 result[0] = addend1[0] + addend2[0];
1266 result[1] = addend1[1] + addend2[1] +
1267 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1268
1269 sum[0] = result[0];
1270 sum[1] = result[1];
1271 }
1272
1273 /*
1274 * Shift the 128-bit value in a by b. If b is positive, shift left.
1275 * If b is negative, shift right.
1276 */
1277 static void
1278 dtrace_shift_128(uint64_t *a, int b)
1279 {
1280 uint64_t mask;
1281
1282 if (b == 0)
1283 return;
1284
1285 if (b < 0) {
1286 b = -b;
1287 if (b >= 64) {
1288 a[0] = a[1] >> (b - 64);
1289 a[1] = 0;
1290 } else {
1291 a[0] >>= b;
1292 mask = 1LL << (64 - b);
1293 mask -= 1;
1294 a[0] |= ((a[1] & mask) << (64 - b));
1295 a[1] >>= b;
1296 }
1297 } else {
1298 if (b >= 64) {
1299 a[1] = a[0] << (b - 64);
1300 a[0] = 0;
1301 } else {
1302 a[1] <<= b;
1303 mask = a[0] >> (64 - b);
1304 a[1] |= mask;
1305 a[0] <<= b;
1306 }
1307 }
1308 }
1309
1310 /*
1311 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1312 * use native multiplication on those, and then re-combine into the
1313 * resulting 128-bit value.
1314 *
1315 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1316 * hi1 * hi2 << 64 +
1317 * hi1 * lo2 << 32 +
1318 * hi2 * lo1 << 32 +
1319 * lo1 * lo2
1320 */
1321 static void
1322 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1323 {
1324 uint64_t hi1, hi2, lo1, lo2;
1325 uint64_t tmp[2];
1326
1327 hi1 = factor1 >> 32;
1328 hi2 = factor2 >> 32;
1329
1330 lo1 = factor1 & DT_MASK_LO;
1331 lo2 = factor2 & DT_MASK_LO;
1332
1333 product[0] = lo1 * lo2;
1334 product[1] = hi1 * hi2;
1335
1336 tmp[0] = hi1 * lo2;
1337 tmp[1] = 0;
1338 dtrace_shift_128(tmp, 32);
1339 dtrace_add_128(product, tmp, product);
1340
1341 tmp[0] = hi2 * lo1;
1342 tmp[1] = 0;
1343 dtrace_shift_128(tmp, 32);
1344 dtrace_add_128(product, tmp, product);
1345 }
1346
1347 /*
1348 * This privilege check should be used by actions and subroutines to
1349 * verify that the user credentials of the process that enabled the
1350 * invoking ECB match the target credentials
1351 */
1352 static int
1353 dtrace_priv_proc_common_user(dtrace_state_t *state)
1354 {
1355 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1356
1357 /*
1358 * We should always have a non-NULL state cred here, since if cred
1359 * is null (anonymous tracing), we fast-path bypass this routine.
1360 */
1361 ASSERT(s_cr != NULL);
1362
1363 if ((cr = CRED()) != NULL &&
1364 s_cr->cr_uid == cr->cr_uid &&
1365 s_cr->cr_uid == cr->cr_ruid &&
1366 s_cr->cr_uid == cr->cr_suid &&
1367 s_cr->cr_gid == cr->cr_gid &&
1368 s_cr->cr_gid == cr->cr_rgid &&
1369 s_cr->cr_gid == cr->cr_sgid)
1370 return (1);
1371
1372 return (0);
1373 }
1374
1375 /*
1376 * This privilege check should be used by actions and subroutines to
1377 * verify that the zone of the process that enabled the invoking ECB
1378 * matches the target credentials
1379 */
1380 static int
1381 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1382 {
1383 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1384
1385 /*
1386 * We should always have a non-NULL state cred here, since if cred
1387 * is null (anonymous tracing), we fast-path bypass this routine.
1388 */
1389 ASSERT(s_cr != NULL);
1390
1391 if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1392 return (1);
1393
1394 return (0);
1395 }
1396
1397 /*
1398 * This privilege check should be used by actions and subroutines to
1399 * verify that the process has not setuid or changed credentials.
1400 */
1401 static int
1402 dtrace_priv_proc_common_nocd()
1403 {
1404 proc_t *proc;
1405
1406 if ((proc = ttoproc(curthread)) != NULL &&
1407 !(proc->p_flag & SNOCD))
1408 return (1);
1409
1410 return (0);
1411 }
1412
1413 static int
1414 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1415 {
1416 int action = state->dts_cred.dcr_action;
1417
1418 if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1419 goto bad;
1420
1421 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1422 dtrace_priv_proc_common_zone(state) == 0)
1423 goto bad;
1424
1425 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1426 dtrace_priv_proc_common_user(state) == 0)
1427 goto bad;
1428
1429 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1430 dtrace_priv_proc_common_nocd() == 0)
1431 goto bad;
1432
1433 return (1);
1434
1435 bad:
1436 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1437
1438 return (0);
1439 }
1440
1441 static int
1442 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1443 {
1444 if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1445 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1446 return (1);
1447
1448 if (dtrace_priv_proc_common_zone(state) &&
1449 dtrace_priv_proc_common_user(state) &&
1450 dtrace_priv_proc_common_nocd())
1451 return (1);
1452 }
1453
1454 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1455
1456 return (0);
1457 }
1458
1459 static int
1460 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1461 {
1462 if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1463 (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1464 return (1);
1465
1466 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1467
1468 return (0);
1469 }
1470
1471 static int
1472 dtrace_priv_kernel(dtrace_state_t *state)
1473 {
1474 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1475 return (1);
1476
1477 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1478
1479 return (0);
1480 }
1481
1482 static int
1483 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1484 {
1485 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1486 return (1);
1487
1488 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1489
1490 return (0);
1491 }
1492
1493 /*
1494 * Determine if the dte_cond of the specified ECB allows for processing of
1495 * the current probe to continue. Note that this routine may allow continued
1496 * processing, but with access(es) stripped from the mstate's dtms_access
1497 * field.
1498 */
1499 static int
1500 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1501 dtrace_ecb_t *ecb)
1502 {
1503 dtrace_probe_t *probe = ecb->dte_probe;
1504 dtrace_provider_t *prov = probe->dtpr_provider;
1505 dtrace_pops_t *pops = &prov->dtpv_pops;
1506 int mode = DTRACE_MODE_NOPRIV_DROP;
1507
1508 ASSERT(ecb->dte_cond);
1509
1510 if (pops->dtps_mode != NULL) {
1511 mode = pops->dtps_mode(prov->dtpv_arg,
1512 probe->dtpr_id, probe->dtpr_arg);
1513
1514 ASSERT(mode & (DTRACE_MODE_USER | DTRACE_MODE_KERNEL));
1515 ASSERT(mode & (DTRACE_MODE_NOPRIV_RESTRICT |
1516 DTRACE_MODE_NOPRIV_DROP));
1517 }
1518
1519 /*
1520 * If the dte_cond bits indicate that this consumer is only allowed to
1521 * see user-mode firings of this probe, check that the probe was fired
1522 * while in a user context. If that's not the case, use the policy
1523 * specified by the provider to determine if we drop the probe or
1524 * merely restrict operation.
1525 */
1526 if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1527 ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1528
1529 if (!(mode & DTRACE_MODE_USER)) {
1530 if (mode & DTRACE_MODE_NOPRIV_DROP)
1531 return (0);
1532
1533 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1534 }
1535 }
1536
1537 /*
1538 * This is more subtle than it looks. We have to be absolutely certain
1539 * that CRED() isn't going to change out from under us so it's only
1540 * legit to examine that structure if we're in constrained situations.
1541 * Currently, the only times we'll this check is if a non-super-user
1542 * has enabled the profile or syscall providers -- providers that
1543 * allow visibility of all processes. For the profile case, the check
1544 * above will ensure that we're examining a user context.
1545 */
1546 if (ecb->dte_cond & DTRACE_COND_OWNER) {
1547 cred_t *cr;
1548 cred_t *s_cr = state->dts_cred.dcr_cred;
1549 proc_t *proc;
1550
1551 ASSERT(s_cr != NULL);
1552
1553 if ((cr = CRED()) == NULL ||
1554 s_cr->cr_uid != cr->cr_uid ||
1555 s_cr->cr_uid != cr->cr_ruid ||
1556 s_cr->cr_uid != cr->cr_suid ||
1557 s_cr->cr_gid != cr->cr_gid ||
1558 s_cr->cr_gid != cr->cr_rgid ||
1559 s_cr->cr_gid != cr->cr_sgid ||
1560 (proc = ttoproc(curthread)) == NULL ||
1561 (proc->p_flag & SNOCD)) {
1562 if (mode & DTRACE_MODE_NOPRIV_DROP)
1563 return (0);
1564
1565 mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1566 }
1567 }
1568
1569 /*
1570 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1571 * in our zone, check to see if our mode policy is to restrict rather
1572 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1573 * and DTRACE_ACCESS_ARGS
1574 */
1575 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1576 cred_t *cr;
1577 cred_t *s_cr = state->dts_cred.dcr_cred;
1578
1579 ASSERT(s_cr != NULL);
1580
1581 if ((cr = CRED()) == NULL ||
1582 s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1583 if (mode & DTRACE_MODE_NOPRIV_DROP)
1584 return (0);
1585
1586 mstate->dtms_access &=
1587 ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1588 }
1589 }
1590
1591 /*
1592 * By merits of being in this code path at all, we have limited
1593 * privileges. If the provider has indicated that limited privileges
1594 * are to denote restricted operation, strip off the ability to access
1595 * arguments.
1596 */
1597 if (mode & DTRACE_MODE_LIMITEDPRIV_RESTRICT)
1598 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1599
1600 return (1);
1601 }
1602
1603 /*
1604 * Note: not called from probe context. This function is called
1605 * asynchronously (and at a regular interval) from outside of probe context to
1606 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1607 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1608 */
1609 void
1610 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1611 {
1612 dtrace_dynvar_t *dirty;
1613 dtrace_dstate_percpu_t *dcpu;
1614 dtrace_dynvar_t **rinsep;
1615 int i, j, work = 0;
1616
1617 for (i = 0; i < NCPU; i++) {
1618 dcpu = &dstate->dtds_percpu[i];
1619 rinsep = &dcpu->dtdsc_rinsing;
1620
1621 /*
1622 * If the dirty list is NULL, there is no dirty work to do.
1623 */
1624 if (dcpu->dtdsc_dirty == NULL)
1625 continue;
1626
1627 if (dcpu->dtdsc_rinsing != NULL) {
1628 /*
1629 * If the rinsing list is non-NULL, then it is because
1630 * this CPU was selected to accept another CPU's
1631 * dirty list -- and since that time, dirty buffers
1632 * have accumulated. This is a highly unlikely
1633 * condition, but we choose to ignore the dirty
1634 * buffers -- they'll be picked up a future cleanse.
1635 */
1636 continue;
1637 }
1638
1639 if (dcpu->dtdsc_clean != NULL) {
1640 /*
1641 * If the clean list is non-NULL, then we're in a
1642 * situation where a CPU has done deallocations (we
1643 * have a non-NULL dirty list) but no allocations (we
1644 * also have a non-NULL clean list). We can't simply
1645 * move the dirty list into the clean list on this
1646 * CPU, yet we also don't want to allow this condition
1647 * to persist, lest a short clean list prevent a
1648 * massive dirty list from being cleaned (which in
1649 * turn could lead to otherwise avoidable dynamic
1650 * drops). To deal with this, we look for some CPU
1651 * with a NULL clean list, NULL dirty list, and NULL
1652 * rinsing list -- and then we borrow this CPU to
1653 * rinse our dirty list.
1654 */
1655 for (j = 0; j < NCPU; j++) {
1656 dtrace_dstate_percpu_t *rinser;
1657
1658 rinser = &dstate->dtds_percpu[j];
1659
1660 if (rinser->dtdsc_rinsing != NULL)
1661 continue;
1662
1663 if (rinser->dtdsc_dirty != NULL)
1664 continue;
1665
1666 if (rinser->dtdsc_clean != NULL)
1667 continue;
1668
1669 rinsep = &rinser->dtdsc_rinsing;
1670 break;
1671 }
1672
1673 if (j == NCPU) {
1674 /*
1675 * We were unable to find another CPU that
1676 * could accept this dirty list -- we are
1677 * therefore unable to clean it now.
1678 */
1679 dtrace_dynvar_failclean++;
1680 continue;
1681 }
1682 }
1683
1684 work = 1;
1685
1686 /*
1687 * Atomically move the dirty list aside.
1688 */
1689 do {
1690 dirty = dcpu->dtdsc_dirty;
1691
1692 /*
1693 * Before we zap the dirty list, set the rinsing list.
1694 * (This allows for a potential assertion in
1695 * dtrace_dynvar(): if a free dynamic variable appears
1696 * on a hash chain, either the dirty list or the
1697 * rinsing list for some CPU must be non-NULL.)
1698 */
1699 *rinsep = dirty;
1700 dtrace_membar_producer();
1701 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1702 dirty, NULL) != dirty);
1703 }
1704
1705 if (!work) {
1706 /*
1707 * We have no work to do; we can simply return.
1708 */
1709 return;
1710 }
1711
1712 dtrace_sync();
1713
1714 for (i = 0; i < NCPU; i++) {
1715 dcpu = &dstate->dtds_percpu[i];
1716
1717 if (dcpu->dtdsc_rinsing == NULL)
1718 continue;
1719
1720 /*
1721 * We are now guaranteed that no hash chain contains a pointer
1722 * into this dirty list; we can make it clean.
1723 */
1724 ASSERT(dcpu->dtdsc_clean == NULL);
1725 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1726 dcpu->dtdsc_rinsing = NULL;
1727 }
1728
1729 /*
1730 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1731 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1732 * This prevents a race whereby a CPU incorrectly decides that
1733 * the state should be something other than DTRACE_DSTATE_CLEAN
1734 * after dtrace_dynvar_clean() has completed.
1735 */
1736 dtrace_sync();
1737
1738 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1739 }
1740
1741 /*
1742 * Depending on the value of the op parameter, this function looks-up,
1743 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1744 * allocation is requested, this function will return a pointer to a
1745 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1746 * variable can be allocated. If NULL is returned, the appropriate counter
1747 * will be incremented.
1748 */
1749 dtrace_dynvar_t *
1750 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1751 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1752 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1753 {
1754 uint64_t hashval = DTRACE_DYNHASH_VALID;
1755 dtrace_dynhash_t *hash = dstate->dtds_hash;
1756 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1757 processorid_t me = CPU->cpu_id, cpu = me;
1758 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1759 size_t bucket, ksize;
1760 size_t chunksize = dstate->dtds_chunksize;
1761 uintptr_t kdata, lock, nstate;
1762 uint_t i;
1763
1764 ASSERT(nkeys != 0);
1765
1766 /*
1767 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1768 * algorithm. For the by-value portions, we perform the algorithm in
1769 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1770 * bit, and seems to have only a minute effect on distribution. For
1771 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1772 * over each referenced byte. It's painful to do this, but it's much
1773 * better than pathological hash distribution. The efficacy of the
1774 * hashing algorithm (and a comparison with other algorithms) may be
1775 * found by running the ::dtrace_dynstat MDB dcmd.
1776 */
1777 for (i = 0; i < nkeys; i++) {
1778 if (key[i].dttk_size == 0) {
1779 uint64_t val = key[i].dttk_value;
1780
1781 hashval += (val >> 48) & 0xffff;
1782 hashval += (hashval << 10);
1783 hashval ^= (hashval >> 6);
1784
1785 hashval += (val >> 32) & 0xffff;
1786 hashval += (hashval << 10);
1787 hashval ^= (hashval >> 6);
1788
1789 hashval += (val >> 16) & 0xffff;
1790 hashval += (hashval << 10);
1791 hashval ^= (hashval >> 6);
1792
1793 hashval += val & 0xffff;
1794 hashval += (hashval << 10);
1795 hashval ^= (hashval >> 6);
1796 } else {
1797 /*
1798 * This is incredibly painful, but it beats the hell
1799 * out of the alternative.
1800 */
1801 uint64_t j, size = key[i].dttk_size;
1802 uintptr_t base = (uintptr_t)key[i].dttk_value;
1803
1804 if (!dtrace_canload(base, size, mstate, vstate))
1805 break;
1806
1807 for (j = 0; j < size; j++) {
1808 hashval += dtrace_load8(base + j);
1809 hashval += (hashval << 10);
1810 hashval ^= (hashval >> 6);
1811 }
1812 }
1813 }
1814
1815 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1816 return (NULL);
1817
1818 hashval += (hashval << 3);
1819 hashval ^= (hashval >> 11);
1820 hashval += (hashval << 15);
1821
1822 /*
1823 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1824 * comes out to be one of our two sentinel hash values. If this
1825 * actually happens, we set the hashval to be a value known to be a
1826 * non-sentinel value.
1827 */
1828 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1829 hashval = DTRACE_DYNHASH_VALID;
1830
1831 /*
1832 * Yes, it's painful to do a divide here. If the cycle count becomes
1833 * important here, tricks can be pulled to reduce it. (However, it's
1834 * critical that hash collisions be kept to an absolute minimum;
1835 * they're much more painful than a divide.) It's better to have a
1836 * solution that generates few collisions and still keeps things
1837 * relatively simple.
1838 */
1839 bucket = hashval % dstate->dtds_hashsize;
1840
1841 if (op == DTRACE_DYNVAR_DEALLOC) {
1842 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1843
1844 for (;;) {
1845 while ((lock = *lockp) & 1)
1846 continue;
1847
1848 if (dtrace_casptr((void *)lockp,
1849 (void *)lock, (void *)(lock + 1)) == (void *)lock)
1850 break;
1851 }
1852
1853 dtrace_membar_producer();
1854 }
1855
1856 top:
1857 prev = NULL;
1858 lock = hash[bucket].dtdh_lock;
1859
1860 dtrace_membar_consumer();
1861
1862 start = hash[bucket].dtdh_chain;
1863 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1864 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1865 op != DTRACE_DYNVAR_DEALLOC));
1866
1867 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1868 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1869 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1870
1871 if (dvar->dtdv_hashval != hashval) {
1872 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1873 /*
1874 * We've reached the sink, and therefore the
1875 * end of the hash chain; we can kick out of
1876 * the loop knowing that we have seen a valid
1877 * snapshot of state.
1878 */
1879 ASSERT(dvar->dtdv_next == NULL);
1880 ASSERT(dvar == &dtrace_dynhash_sink);
1881 break;
1882 }
1883
1884 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1885 /*
1886 * We've gone off the rails: somewhere along
1887 * the line, one of the members of this hash
1888 * chain was deleted. Note that we could also
1889 * detect this by simply letting this loop run
1890 * to completion, as we would eventually hit
1891 * the end of the dirty list. However, we
1892 * want to avoid running the length of the
1893 * dirty list unnecessarily (it might be quite
1894 * long), so we catch this as early as
1895 * possible by detecting the hash marker. In
1896 * this case, we simply set dvar to NULL and
1897 * break; the conditional after the loop will
1898 * send us back to top.
1899 */
1900 dvar = NULL;
1901 break;
1902 }
1903
1904 goto next;
1905 }
1906
1907 if (dtuple->dtt_nkeys != nkeys)
1908 goto next;
1909
1910 for (i = 0; i < nkeys; i++, dkey++) {
1911 if (dkey->dttk_size != key[i].dttk_size)
1912 goto next; /* size or type mismatch */
1913
1914 if (dkey->dttk_size != 0) {
1915 if (dtrace_bcmp(
1916 (void *)(uintptr_t)key[i].dttk_value,
1917 (void *)(uintptr_t)dkey->dttk_value,
1918 dkey->dttk_size))
1919 goto next;
1920 } else {
1921 if (dkey->dttk_value != key[i].dttk_value)
1922 goto next;
1923 }
1924 }
1925
1926 if (op != DTRACE_DYNVAR_DEALLOC)
1927 return (dvar);
1928
1929 ASSERT(dvar->dtdv_next == NULL ||
1930 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1931
1932 if (prev != NULL) {
1933 ASSERT(hash[bucket].dtdh_chain != dvar);
1934 ASSERT(start != dvar);
1935 ASSERT(prev->dtdv_next == dvar);
1936 prev->dtdv_next = dvar->dtdv_next;
1937 } else {
1938 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1939 start, dvar->dtdv_next) != start) {
1940 /*
1941 * We have failed to atomically swing the
1942 * hash table head pointer, presumably because
1943 * of a conflicting allocation on another CPU.
1944 * We need to reread the hash chain and try
1945 * again.
1946 */
1947 goto top;
1948 }
1949 }
1950
1951 dtrace_membar_producer();
1952
1953 /*
1954 * Now set the hash value to indicate that it's free.
1955 */
1956 ASSERT(hash[bucket].dtdh_chain != dvar);
1957 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1958
1959 dtrace_membar_producer();
1960
1961 /*
1962 * Set the next pointer to point at the dirty list, and
1963 * atomically swing the dirty pointer to the newly freed dvar.
1964 */
1965 do {
1966 next = dcpu->dtdsc_dirty;
1967 dvar->dtdv_next = next;
1968 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1969
1970 /*
1971 * Finally, unlock this hash bucket.
1972 */
1973 ASSERT(hash[bucket].dtdh_lock == lock);
1974 ASSERT(lock & 1);
1975 hash[bucket].dtdh_lock++;
1976
1977 return (NULL);
1978 next:
1979 prev = dvar;
1980 continue;
1981 }
1982
1983 if (dvar == NULL) {
1984 /*
1985 * If dvar is NULL, it is because we went off the rails:
1986 * one of the elements that we traversed in the hash chain
1987 * was deleted while we were traversing it. In this case,
1988 * we assert that we aren't doing a dealloc (deallocs lock
1989 * the hash bucket to prevent themselves from racing with
1990 * one another), and retry the hash chain traversal.
1991 */
1992 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1993 goto top;
1994 }
1995
1996 if (op != DTRACE_DYNVAR_ALLOC) {
1997 /*
1998 * If we are not to allocate a new variable, we want to
1999 * return NULL now. Before we return, check that the value
2000 * of the lock word hasn't changed. If it has, we may have
2001 * seen an inconsistent snapshot.
2002 */
2003 if (op == DTRACE_DYNVAR_NOALLOC) {
2004 if (hash[bucket].dtdh_lock != lock)
2005 goto top;
2006 } else {
2007 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
2008 ASSERT(hash[bucket].dtdh_lock == lock);
2009 ASSERT(lock & 1);
2010 hash[bucket].dtdh_lock++;
2011 }
2012
2013 return (NULL);
2014 }
2015
2016 /*
2017 * We need to allocate a new dynamic variable. The size we need is the
2018 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
2019 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
2020 * the size of any referred-to data (dsize). We then round the final
2021 * size up to the chunksize for allocation.
2022 */
2023 for (ksize = 0, i = 0; i < nkeys; i++)
2024 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
2025
2026 /*
2027 * This should be pretty much impossible, but could happen if, say,
2028 * strange DIF specified the tuple. Ideally, this should be an
2029 * assertion and not an error condition -- but that requires that the
2030 * chunksize calculation in dtrace_difo_chunksize() be absolutely
2031 * bullet-proof. (That is, it must not be able to be fooled by
2032 * malicious DIF.) Given the lack of backwards branches in DIF,
2033 * solving this would presumably not amount to solving the Halting
2034 * Problem -- but it still seems awfully hard.
2035 */
2036 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
2037 ksize + dsize > chunksize) {
2038 dcpu->dtdsc_drops++;
2039 return (NULL);
2040 }
2041
2042 nstate = DTRACE_DSTATE_EMPTY;
2043
2044 do {
2045 retry:
2046 free = dcpu->dtdsc_free;
2047
2048 if (free == NULL) {
2049 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
2050 void *rval;
2051
2052 if (clean == NULL) {
2053 /*
2054 * We're out of dynamic variable space on
2055 * this CPU. Unless we have tried all CPUs,
2056 * we'll try to allocate from a different
2057 * CPU.
2058 */
2059 switch (dstate->dtds_state) {
2060 case DTRACE_DSTATE_CLEAN: {
2061 void *sp = &dstate->dtds_state;
2062
2063 if (++cpu >= NCPU)
2064 cpu = 0;
2065
2066 if (dcpu->dtdsc_dirty != NULL &&
2067 nstate == DTRACE_DSTATE_EMPTY)
2068 nstate = DTRACE_DSTATE_DIRTY;
2069
2070 if (dcpu->dtdsc_rinsing != NULL)
2071 nstate = DTRACE_DSTATE_RINSING;
2072
2073 dcpu = &dstate->dtds_percpu[cpu];
2074
2075 if (cpu != me)
2076 goto retry;
2077
2078 (void) dtrace_cas32(sp,
2079 DTRACE_DSTATE_CLEAN, nstate);
2080
2081 /*
2082 * To increment the correct bean
2083 * counter, take another lap.
2084 */
2085 goto retry;
2086 }
2087
2088 case DTRACE_DSTATE_DIRTY:
2089 dcpu->dtdsc_dirty_drops++;
2090 break;
2091
2092 case DTRACE_DSTATE_RINSING:
2093 dcpu->dtdsc_rinsing_drops++;
2094 break;
2095
2096 case DTRACE_DSTATE_EMPTY:
2097 dcpu->dtdsc_drops++;
2098 break;
2099 }
2100
2101 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
2102 return (NULL);
2103 }
2104
2105 /*
2106 * The clean list appears to be non-empty. We want to
2107 * move the clean list to the free list; we start by
2108 * moving the clean pointer aside.
2109 */
2110 if (dtrace_casptr(&dcpu->dtdsc_clean,
2111 clean, NULL) != clean) {
2112 /*
2113 * We are in one of two situations:
2114 *
2115 * (a) The clean list was switched to the
2116 * free list by another CPU.
2117 *
2118 * (b) The clean list was added to by the
2119 * cleansing cyclic.
2120 *
2121 * In either of these situations, we can
2122 * just reattempt the free list allocation.
2123 */
2124 goto retry;
2125 }
2126
2127 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
2128
2129 /*
2130 * Now we'll move the clean list to our free list.
2131 * It's impossible for this to fail: the only way
2132 * the free list can be updated is through this
2133 * code path, and only one CPU can own the clean list.
2134 * Thus, it would only be possible for this to fail if
2135 * this code were racing with dtrace_dynvar_clean().
2136 * (That is, if dtrace_dynvar_clean() updated the clean
2137 * list, and we ended up racing to update the free
2138 * list.) This race is prevented by the dtrace_sync()
2139 * in dtrace_dynvar_clean() -- which flushes the
2140 * owners of the clean lists out before resetting
2141 * the clean lists.
2142 */
2143 dcpu = &dstate->dtds_percpu[me];
2144 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
2145 ASSERT(rval == NULL);
2146 goto retry;
2147 }
2148
2149 dvar = free;
2150 new_free = dvar->dtdv_next;
2151 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2152
2153 /*
2154 * We have now allocated a new chunk. We copy the tuple keys into the
2155 * tuple array and copy any referenced key data into the data space
2156 * following the tuple array. As we do this, we relocate dttk_value
2157 * in the final tuple to point to the key data address in the chunk.
2158 */
2159 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2160 dvar->dtdv_data = (void *)(kdata + ksize);
2161 dvar->dtdv_tuple.dtt_nkeys = nkeys;
2162
2163 for (i = 0; i < nkeys; i++) {
2164 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2165 size_t kesize = key[i].dttk_size;
2166
2167 if (kesize != 0) {
2168 dtrace_bcopy(
2169 (const void *)(uintptr_t)key[i].dttk_value,
2170 (void *)kdata, kesize);
2171 dkey->dttk_value = kdata;
2172 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2173 } else {
2174 dkey->dttk_value = key[i].dttk_value;
2175 }
2176
2177 dkey->dttk_size = kesize;
2178 }
2179
2180 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2181 dvar->dtdv_hashval = hashval;
2182 dvar->dtdv_next = start;
2183
2184 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2185 return (dvar);
2186
2187 /*
2188 * The cas has failed. Either another CPU is adding an element to
2189 * this hash chain, or another CPU is deleting an element from this
2190 * hash chain. The simplest way to deal with both of these cases
2191 * (though not necessarily the most efficient) is to free our
2192 * allocated block and re-attempt it all. Note that the free is
2193 * to the dirty list and _not_ to the free list. This is to prevent
2194 * races with allocators, above.
2195 */
2196 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2197
2198 dtrace_membar_producer();
2199
2200 do {
2201 free = dcpu->dtdsc_dirty;
2202 dvar->dtdv_next = free;
2203 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2204
2205 goto top;
2206 }
2207
2208 /*ARGSUSED*/
2209 static void
2210 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2211 {
2212 if ((int64_t)nval < (int64_t)*oval)
2213 *oval = nval;
2214 }
2215
2216 /*ARGSUSED*/
2217 static void
2218 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2219 {
2220 if ((int64_t)nval > (int64_t)*oval)
2221 *oval = nval;
2222 }
2223
2224 static void
2225 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2226 {
2227 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2228 int64_t val = (int64_t)nval;
2229
2230 if (val < 0) {
2231 for (i = 0; i < zero; i++) {
2232 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2233 quanta[i] += incr;
2234 return;
2235 }
2236 }
2237 } else {
2238 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2239 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2240 quanta[i - 1] += incr;
2241 return;
2242 }
2243 }
2244
2245 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2246 return;
2247 }
2248
2249 ASSERT(0);
2250 }
2251
2252 static void
2253 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2254 {
2255 uint64_t arg = *lquanta++;
2256 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2257 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2258 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2259 int32_t val = (int32_t)nval, level;
2260
2261 ASSERT(step != 0);
2262 ASSERT(levels != 0);
2263
2264 if (val < base) {
2265 /*
2266 * This is an underflow.
2267 */
2268 lquanta[0] += incr;
2269 return;
2270 }
2271
2272 level = (val - base) / step;
2273
2274 if (level < levels) {
2275 lquanta[level + 1] += incr;
2276 return;
2277 }
2278
2279 /*
2280 * This is an overflow.
2281 */
2282 lquanta[levels + 1] += incr;
2283 }
2284
2285 static int
2286 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2287 uint16_t high, uint16_t nsteps, int64_t value)
2288 {
2289 int64_t this = 1, last, next;
2290 int base = 1, order;
2291
2292 ASSERT(factor <= nsteps);
2293 ASSERT(nsteps % factor == 0);
2294
2295 for (order = 0; order < low; order++)
2296 this *= factor;
2297
2298 /*
2299 * If our value is less than our factor taken to the power of the
2300 * low order of magnitude, it goes into the zeroth bucket.
2301 */
2302 if (value < (last = this))
2303 return (0);
2304
2305 for (this *= factor; order <= high; order++) {
2306 int nbuckets = this > nsteps ? nsteps : this;
2307
2308 if ((next = this * factor) < this) {
2309 /*
2310 * We should not generally get log/linear quantizations
2311 * with a high magnitude that allows 64-bits to
2312 * overflow, but we nonetheless protect against this
2313 * by explicitly checking for overflow, and clamping
2314 * our value accordingly.
2315 */
2316 value = this - 1;
2317 }
2318
2319 if (value < this) {
2320 /*
2321 * If our value lies within this order of magnitude,
2322 * determine its position by taking the offset within
2323 * the order of magnitude, dividing by the bucket
2324 * width, and adding to our (accumulated) base.
2325 */
2326 return (base + (value - last) / (this / nbuckets));
2327 }
2328
2329 base += nbuckets - (nbuckets / factor);
2330 last = this;
2331 this = next;
2332 }
2333
2334 /*
2335 * Our value is greater than or equal to our factor taken to the
2336 * power of one plus the high magnitude -- return the top bucket.
2337 */
2338 return (base);
2339 }
2340
2341 static void
2342 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2343 {
2344 uint64_t arg = *llquanta++;
2345 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2346 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2347 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2348 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2349
2350 llquanta[dtrace_aggregate_llquantize_bucket(factor,
2351 low, high, nsteps, nval)] += incr;
2352 }
2353
2354 /*ARGSUSED*/
2355 static void
2356 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2357 {
2358 data[0]++;
2359 data[1] += nval;
2360 }
2361
2362 /*ARGSUSED*/
2363 static void
2364 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2365 {
2366 int64_t snval = (int64_t)nval;
2367 uint64_t tmp[2];
2368
2369 data[0]++;
2370 data[1] += nval;
2371
2372 /*
2373 * What we want to say here is:
2374 *
2375 * data[2] += nval * nval;
2376 *
2377 * But given that nval is 64-bit, we could easily overflow, so
2378 * we do this as 128-bit arithmetic.
2379 */
2380 if (snval < 0)
2381 snval = -snval;
2382
2383 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2384 dtrace_add_128(data + 2, tmp, data + 2);
2385 }
2386
2387 /*ARGSUSED*/
2388 static void
2389 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2390 {
2391 *oval = *oval + 1;
2392 }
2393
2394 /*ARGSUSED*/
2395 static void
2396 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2397 {
2398 *oval += nval;
2399 }
2400
2401 /*
2402 * Aggregate given the tuple in the principal data buffer, and the aggregating
2403 * action denoted by the specified dtrace_aggregation_t. The aggregation
2404 * buffer is specified as the buf parameter. This routine does not return
2405 * failure; if there is no space in the aggregation buffer, the data will be
2406 * dropped, and a corresponding counter incremented.
2407 */
2408 static void
2409 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2410 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2411 {
2412 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2413 uint32_t i, ndx, size, fsize;
2414 uint32_t align = sizeof (uint64_t) - 1;
2415 dtrace_aggbuffer_t *agb;
2416 dtrace_aggkey_t *key;
2417 uint32_t hashval = 0, limit, isstr;
2418 caddr_t tomax, data, kdata;
2419 dtrace_actkind_t action;
2420 dtrace_action_t *act;
2421 uintptr_t offs;
2422
2423 if (buf == NULL)
2424 return;
2425
2426 if (!agg->dtag_hasarg) {
2427 /*
2428 * Currently, only quantize() and lquantize() take additional
2429 * arguments, and they have the same semantics: an increment
2430 * value that defaults to 1 when not present. If additional
2431 * aggregating actions take arguments, the setting of the
2432 * default argument value will presumably have to become more
2433 * sophisticated...
2434 */
2435 arg = 1;
2436 }
2437
2438 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2439 size = rec->dtrd_offset - agg->dtag_base;
2440 fsize = size + rec->dtrd_size;
2441
2442 ASSERT(dbuf->dtb_tomax != NULL);
2443 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2444
2445 if ((tomax = buf->dtb_tomax) == NULL) {
2446 dtrace_buffer_drop(buf);
2447 return;
2448 }
2449
2450 /*
2451 * The metastructure is always at the bottom of the buffer.
2452 */
2453 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2454 sizeof (dtrace_aggbuffer_t));
2455
2456 if (buf->dtb_offset == 0) {
2457 /*
2458 * We just kludge up approximately 1/8th of the size to be
2459 * buckets. If this guess ends up being routinely
2460 * off-the-mark, we may need to dynamically readjust this
2461 * based on past performance.
2462 */
2463 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2464
2465 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2466 (uintptr_t)tomax || hashsize == 0) {
2467 /*
2468 * We've been given a ludicrously small buffer;
2469 * increment our drop count and leave.
2470 */
2471 dtrace_buffer_drop(buf);
2472 return;
2473 }
2474
2475 /*
2476 * And now, a pathetic attempt to try to get a an odd (or
2477 * perchance, a prime) hash size for better hash distribution.
2478 */
2479 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2480 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2481
2482 agb->dtagb_hashsize = hashsize;
2483 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2484 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2485 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2486
2487 for (i = 0; i < agb->dtagb_hashsize; i++)
2488 agb->dtagb_hash[i] = NULL;
2489 }
2490
2491 ASSERT(agg->dtag_first != NULL);
2492 ASSERT(agg->dtag_first->dta_intuple);
2493
2494 /*
2495 * Calculate the hash value based on the key. Note that we _don't_
2496 * include the aggid in the hashing (but we will store it as part of
2497 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2498 * algorithm: a simple, quick algorithm that has no known funnels, and
2499 * gets good distribution in practice. The efficacy of the hashing
2500 * algorithm (and a comparison with other algorithms) may be found by
2501 * running the ::dtrace_aggstat MDB dcmd.
2502 */
2503 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2504 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2505 limit = i + act->dta_rec.dtrd_size;
2506 ASSERT(limit <= size);
2507 isstr = DTRACEACT_ISSTRING(act);
2508
2509 for (; i < limit; i++) {
2510 hashval += data[i];
2511 hashval += (hashval << 10);
2512 hashval ^= (hashval >> 6);
2513
2514 if (isstr && data[i] == '\0')
2515 break;
2516 }
2517 }
2518
2519 hashval += (hashval << 3);
2520 hashval ^= (hashval >> 11);
2521 hashval += (hashval << 15);
2522
2523 /*
2524 * Yes, the divide here is expensive -- but it's generally the least
2525 * of the performance issues given the amount of data that we iterate
2526 * over to compute hash values, compare data, etc.
2527 */
2528 ndx = hashval % agb->dtagb_hashsize;
2529
2530 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2531 ASSERT((caddr_t)key >= tomax);
2532 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2533
2534 if (hashval != key->dtak_hashval || key->dtak_size != size)
2535 continue;
2536
2537 kdata = key->dtak_data;
2538 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2539
2540 for (act = agg->dtag_first; act->dta_intuple;
2541 act = act->dta_next) {
2542 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2543 limit = i + act->dta_rec.dtrd_size;
2544 ASSERT(limit <= size);
2545 isstr = DTRACEACT_ISSTRING(act);
2546
2547 for (; i < limit; i++) {
2548 if (kdata[i] != data[i])
2549 goto next;
2550
2551 if (isstr && data[i] == '\0')
2552 break;
2553 }
2554 }
2555
2556 if (action != key->dtak_action) {
2557 /*
2558 * We are aggregating on the same value in the same
2559 * aggregation with two different aggregating actions.
2560 * (This should have been picked up in the compiler,
2561 * so we may be dealing with errant or devious DIF.)
2562 * This is an error condition; we indicate as much,
2563 * and return.
2564 */
2565 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2566 return;
2567 }
2568
2569 /*
2570 * This is a hit: we need to apply the aggregator to
2571 * the value at this key.
2572 */
2573 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2574 return;
2575 next:
2576 continue;
2577 }
2578
2579 /*
2580 * We didn't find it. We need to allocate some zero-filled space,
2581 * link it into the hash table appropriately, and apply the aggregator
2582 * to the (zero-filled) value.
2583 */
2584 offs = buf->dtb_offset;
2585 while (offs & (align - 1))
2586 offs += sizeof (uint32_t);
2587
2588 /*
2589 * If we don't have enough room to both allocate a new key _and_
2590 * its associated data, increment the drop count and return.
2591 */
2592 if ((uintptr_t)tomax + offs + fsize >
2593 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2594 dtrace_buffer_drop(buf);
2595 return;
2596 }
2597
2598 /*CONSTCOND*/
2599 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2600 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2601 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2602
2603 key->dtak_data = kdata = tomax + offs;
2604 buf->dtb_offset = offs + fsize;
2605
2606 /*
2607 * Now copy the data across.
2608 */
2609 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2610
2611 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2612 kdata[i] = data[i];
2613
2614 /*
2615 * Because strings are not zeroed out by default, we need to iterate
2616 * looking for actions that store strings, and we need to explicitly
2617 * pad these strings out with zeroes.
2618 */
2619 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2620 int nul;
2621
2622 if (!DTRACEACT_ISSTRING(act))
2623 continue;
2624
2625 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2626 limit = i + act->dta_rec.dtrd_size;
2627 ASSERT(limit <= size);
2628
2629 for (nul = 0; i < limit; i++) {
2630 if (nul) {
2631 kdata[i] = '\0';
2632 continue;
2633 }
2634
2635 if (data[i] != '\0')
2636 continue;
2637
2638 nul = 1;
2639 }
2640 }
2641
2642 for (i = size; i < fsize; i++)
2643 kdata[i] = 0;
2644
2645 key->dtak_hashval = hashval;
2646 key->dtak_size = size;
2647 key->dtak_action = action;
2648 key->dtak_next = agb->dtagb_hash[ndx];
2649 agb->dtagb_hash[ndx] = key;
2650
2651 /*
2652 * Finally, apply the aggregator.
2653 */
2654 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2655 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2656 }
2657
2658 /*
2659 * Given consumer state, this routine finds a speculation in the INACTIVE
2660 * state and transitions it into the ACTIVE state. If there is no speculation
2661 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2662 * incremented -- it is up to the caller to take appropriate action.
2663 */
2664 static int
2665 dtrace_speculation(dtrace_state_t *state)
2666 {
2667 int i = 0;
2668 dtrace_speculation_state_t current;
2669 uint32_t *stat = &state->dts_speculations_unavail, count;
2670
2671 while (i < state->dts_nspeculations) {
2672 dtrace_speculation_t *spec = &state->dts_speculations[i];
2673
2674 current = spec->dtsp_state;
2675
2676 if (current != DTRACESPEC_INACTIVE) {
2677 if (current == DTRACESPEC_COMMITTINGMANY ||
2678 current == DTRACESPEC_COMMITTING ||
2679 current == DTRACESPEC_DISCARDING)
2680 stat = &state->dts_speculations_busy;
2681 i++;
2682 continue;
2683 }
2684
2685 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2686 current, DTRACESPEC_ACTIVE) == current)
2687 return (i + 1);
2688 }
2689
2690 /*
2691 * We couldn't find a speculation. If we found as much as a single
2692 * busy speculation buffer, we'll attribute this failure as "busy"
2693 * instead of "unavail".
2694 */
2695 do {
2696 count = *stat;
2697 } while (dtrace_cas32(stat, count, count + 1) != count);
2698
2699 return (0);
2700 }
2701
2702 /*
2703 * This routine commits an active speculation. If the specified speculation
2704 * is not in a valid state to perform a commit(), this routine will silently do
2705 * nothing. The state of the specified speculation is transitioned according
2706 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2707 */
2708 static void
2709 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2710 dtrace_specid_t which)
2711 {
2712 dtrace_speculation_t *spec;
2713 dtrace_buffer_t *src, *dest;
2714 uintptr_t daddr, saddr, dlimit, slimit;
2715 dtrace_speculation_state_t current, new;
2716 intptr_t offs;
2717 uint64_t timestamp;
2718
2719 if (which == 0)
2720 return;
2721
2722 if (which > state->dts_nspeculations) {
2723 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2724 return;
2725 }
2726
2727 spec = &state->dts_speculations[which - 1];
2728 src = &spec->dtsp_buffer[cpu];
2729 dest = &state->dts_buffer[cpu];
2730
2731 do {
2732 current = spec->dtsp_state;
2733
2734 if (current == DTRACESPEC_COMMITTINGMANY)
2735 break;
2736
2737 switch (current) {
2738 case DTRACESPEC_INACTIVE:
2739 case DTRACESPEC_DISCARDING:
2740 return;
2741
2742 case DTRACESPEC_COMMITTING:
2743 /*
2744 * This is only possible if we are (a) commit()'ing
2745 * without having done a prior speculate() on this CPU
2746 * and (b) racing with another commit() on a different
2747 * CPU. There's nothing to do -- we just assert that
2748 * our offset is 0.
2749 */
2750 ASSERT(src->dtb_offset == 0);
2751 return;
2752
2753 case DTRACESPEC_ACTIVE:
2754 new = DTRACESPEC_COMMITTING;
2755 break;
2756
2757 case DTRACESPEC_ACTIVEONE:
2758 /*
2759 * This speculation is active on one CPU. If our
2760 * buffer offset is non-zero, we know that the one CPU
2761 * must be us. Otherwise, we are committing on a
2762 * different CPU from the speculate(), and we must
2763 * rely on being asynchronously cleaned.
2764 */
2765 if (src->dtb_offset != 0) {
2766 new = DTRACESPEC_COMMITTING;
2767 break;
2768 }
2769 /*FALLTHROUGH*/
2770
2771 case DTRACESPEC_ACTIVEMANY:
2772 new = DTRACESPEC_COMMITTINGMANY;
2773 break;
2774
2775 default:
2776 ASSERT(0);
2777 }
2778 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2779 current, new) != current);
2780
2781 /*
2782 * We have set the state to indicate that we are committing this
2783 * speculation. Now reserve the necessary space in the destination
2784 * buffer.
2785 */
2786 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2787 sizeof (uint64_t), state, NULL)) < 0) {
2788 dtrace_buffer_drop(dest);
2789 goto out;
2790 }
2791
2792 /*
2793 * We have sufficient space to copy the speculative buffer into the
2794 * primary buffer. First, modify the speculative buffer, filling
2795 * in the timestamp of all entries with the current time. The data
2796 * must have the commit() time rather than the time it was traced,
2797 * so that all entries in the primary buffer are in timestamp order.
2798 */
2799 timestamp = dtrace_gethrtime();
2800 saddr = (uintptr_t)src->dtb_tomax;
2801 slimit = saddr + src->dtb_offset;
2802 while (saddr < slimit) {
2803 size_t size;
2804 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2805
2806 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2807 saddr += sizeof (dtrace_epid_t);
2808 continue;
2809 }
2810 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2811 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2812
2813 ASSERT3U(saddr + size, <=, slimit);
2814 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2815 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2816
2817 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2818
2819 saddr += size;
2820 }
2821
2822 /*
2823 * Copy the buffer across. (Note that this is a
2824 * highly subobtimal bcopy(); in the unlikely event that this becomes
2825 * a serious performance issue, a high-performance DTrace-specific
2826 * bcopy() should obviously be invented.)
2827 */
2828 daddr = (uintptr_t)dest->dtb_tomax + offs;
2829 dlimit = daddr + src->dtb_offset;
2830 saddr = (uintptr_t)src->dtb_tomax;
2831
2832 /*
2833 * First, the aligned portion.
2834 */
2835 while (dlimit - daddr >= sizeof (uint64_t)) {
2836 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2837
2838 daddr += sizeof (uint64_t);
2839 saddr += sizeof (uint64_t);
2840 }
2841
2842 /*
2843 * Now any left-over bit...
2844 */
2845 while (dlimit - daddr)
2846 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2847
2848 /*
2849 * Finally, commit the reserved space in the destination buffer.
2850 */
2851 dest->dtb_offset = offs + src->dtb_offset;
2852
2853 out:
2854 /*
2855 * If we're lucky enough to be the only active CPU on this speculation
2856 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2857 */
2858 if (current == DTRACESPEC_ACTIVE ||
2859 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2860 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2861 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2862
2863 ASSERT(rval == DTRACESPEC_COMMITTING);
2864 }
2865
2866 src->dtb_offset = 0;
2867 src->dtb_xamot_drops += src->dtb_drops;
2868 src->dtb_drops = 0;
2869 }
2870
2871 /*
2872 * This routine discards an active speculation. If the specified speculation
2873 * is not in a valid state to perform a discard(), this routine will silently
2874 * do nothing. The state of the specified speculation is transitioned
2875 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2876 */
2877 static void
2878 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2879 dtrace_specid_t which)
2880 {
2881 dtrace_speculation_t *spec;
2882 dtrace_speculation_state_t current, new;
2883 dtrace_buffer_t *buf;
2884
2885 if (which == 0)
2886 return;
2887
2888 if (which > state->dts_nspeculations) {
2889 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2890 return;
2891 }
2892
2893 spec = &state->dts_speculations[which - 1];
2894 buf = &spec->dtsp_buffer[cpu];
2895
2896 do {
2897 current = spec->dtsp_state;
2898
2899 switch (current) {
2900 case DTRACESPEC_INACTIVE:
2901 case DTRACESPEC_COMMITTINGMANY:
2902 case DTRACESPEC_COMMITTING:
2903 case DTRACESPEC_DISCARDING:
2904 return;
2905
2906 case DTRACESPEC_ACTIVE:
2907 case DTRACESPEC_ACTIVEMANY:
2908 new = DTRACESPEC_DISCARDING;
2909 break;
2910
2911 case DTRACESPEC_ACTIVEONE:
2912 if (buf->dtb_offset != 0) {
2913 new = DTRACESPEC_INACTIVE;
2914 } else {
2915 new = DTRACESPEC_DISCARDING;
2916 }
2917 break;
2918
2919 default:
2920 ASSERT(0);
2921 }
2922 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2923 current, new) != current);
2924
2925 buf->dtb_offset = 0;
2926 buf->dtb_drops = 0;
2927 }
2928
2929 /*
2930 * Note: not called from probe context. This function is called
2931 * asynchronously from cross call context to clean any speculations that are
2932 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2933 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2934 * speculation.
2935 */
2936 static void
2937 dtrace_speculation_clean_here(dtrace_state_t *state)
2938 {
2939 dtrace_icookie_t cookie;
2940 processorid_t cpu = CPU->cpu_id;
2941 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2942 dtrace_specid_t i;
2943
2944 cookie = dtrace_interrupt_disable();
2945
2946 if (dest->dtb_tomax == NULL) {
2947 dtrace_interrupt_enable(cookie);
2948 return;
2949 }
2950
2951 for (i = 0; i < state->dts_nspeculations; i++) {
2952 dtrace_speculation_t *spec = &state->dts_speculations[i];
2953 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2954
2955 if (src->dtb_tomax == NULL)
2956 continue;
2957
2958 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2959 src->dtb_offset = 0;
2960 continue;
2961 }
2962
2963 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2964 continue;
2965
2966 if (src->dtb_offset == 0)
2967 continue;
2968
2969 dtrace_speculation_commit(state, cpu, i + 1);
2970 }
2971
2972 dtrace_interrupt_enable(cookie);
2973 }
2974
2975 /*
2976 * Note: not called from probe context. This function is called
2977 * asynchronously (and at a regular interval) to clean any speculations that
2978 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2979 * is work to be done, it cross calls all CPUs to perform that work;
2980 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2981 * INACTIVE state until they have been cleaned by all CPUs.
2982 */
2983 static void
2984 dtrace_speculation_clean(dtrace_state_t *state)
2985 {
2986 int work = 0, rv;
2987 dtrace_specid_t i;
2988
2989 for (i = 0; i < state->dts_nspeculations; i++) {
2990 dtrace_speculation_t *spec = &state->dts_speculations[i];
2991
2992 ASSERT(!spec->dtsp_cleaning);
2993
2994 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2995 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2996 continue;
2997
2998 work++;
2999 spec->dtsp_cleaning = 1;
3000 }
3001
3002 if (!work)
3003 return;
3004
3005 dtrace_xcall(DTRACE_CPUALL,
3006 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
3007
3008 /*
3009 * We now know that all CPUs have committed or discarded their
3010 * speculation buffers, as appropriate. We can now set the state
3011 * to inactive.
3012 */
3013 for (i = 0; i < state->dts_nspeculations; i++) {
3014 dtrace_speculation_t *spec = &state->dts_speculations[i];
3015 dtrace_speculation_state_t current, new;
3016
3017 if (!spec->dtsp_cleaning)
3018 continue;
3019
3020 current = spec->dtsp_state;
3021 ASSERT(current == DTRACESPEC_DISCARDING ||
3022 current == DTRACESPEC_COMMITTINGMANY);
3023
3024 new = DTRACESPEC_INACTIVE;
3025
3026 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
3027 ASSERT(rv == current);
3028 spec->dtsp_cleaning = 0;
3029 }
3030 }
3031
3032 /*
3033 * Called as part of a speculate() to get the speculative buffer associated
3034 * with a given speculation. Returns NULL if the specified speculation is not
3035 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
3036 * the active CPU is not the specified CPU -- the speculation will be
3037 * atomically transitioned into the ACTIVEMANY state.
3038 */
3039 static dtrace_buffer_t *
3040 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
3041 dtrace_specid_t which)
3042 {
3043 dtrace_speculation_t *spec;
3044 dtrace_speculation_state_t current, new;
3045 dtrace_buffer_t *buf;
3046
3047 if (which == 0)
3048 return (NULL);
3049
3050 if (which > state->dts_nspeculations) {
3051 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3052 return (NULL);
3053 }
3054
3055 spec = &state->dts_speculations[which - 1];
3056 buf = &spec->dtsp_buffer[cpuid];
3057
3058 do {
3059 current = spec->dtsp_state;
3060
3061 switch (current) {
3062 case DTRACESPEC_INACTIVE:
3063 case DTRACESPEC_COMMITTINGMANY:
3064 case DTRACESPEC_DISCARDING:
3065 return (NULL);
3066
3067 case DTRACESPEC_COMMITTING:
3068 ASSERT(buf->dtb_offset == 0);
3069 return (NULL);
3070
3071 case DTRACESPEC_ACTIVEONE:
3072 /*
3073 * This speculation is currently active on one CPU.
3074 * Check the offset in the buffer; if it's non-zero,
3075 * that CPU must be us (and we leave the state alone).
3076 * If it's zero, assume that we're starting on a new
3077 * CPU -- and change the state to indicate that the
3078 * speculation is active on more than one CPU.
3079 */
3080 if (buf->dtb_offset != 0)
3081 return (buf);
3082
3083 new = DTRACESPEC_ACTIVEMANY;
3084 break;
3085
3086 case DTRACESPEC_ACTIVEMANY:
3087 return (buf);
3088
3089 case DTRACESPEC_ACTIVE:
3090 new = DTRACESPEC_ACTIVEONE;
3091 break;
3092
3093 default:
3094 ASSERT(0);
3095 }
3096 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3097 current, new) != current);
3098
3099 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
3100 return (buf);
3101 }
3102
3103 /*
3104 * Return a string. In the event that the user lacks the privilege to access
3105 * arbitrary kernel memory, we copy the string out to scratch memory so that we
3106 * don't fail access checking.
3107 *
3108 * dtrace_dif_variable() uses this routine as a helper for various
3109 * builtin values such as 'execname' and 'probefunc.'
3110 */
3111 uintptr_t
3112 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
3113 dtrace_mstate_t *mstate)
3114 {
3115 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3116 uintptr_t ret;
3117 size_t strsz;
3118
3119 /*
3120 * The easy case: this probe is allowed to read all of memory, so
3121 * we can just return this as a vanilla pointer.
3122 */
3123 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
3124 return (addr);
3125
3126 /*
3127 * This is the tougher case: we copy the string in question from
3128 * kernel memory into scratch memory and return it that way: this
3129 * ensures that we won't trip up when access checking tests the
3130 * BYREF return value.
3131 */
3132 strsz = dtrace_strlen((char *)addr, size) + 1;
3133
3134 if (mstate->dtms_scratch_ptr + strsz >
3135 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3136 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3137 return (NULL);
3138 }
3139
3140 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3141 strsz);
3142 ret = mstate->dtms_scratch_ptr;
3143 mstate->dtms_scratch_ptr += strsz;
3144 return (ret);
3145 }
3146
3147 /*
3148 * This function implements the DIF emulator's variable lookups. The emulator
3149 * passes a reserved variable identifier and optional built-in array index.
3150 */
3151 static uint64_t
3152 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3153 uint64_t ndx)
3154 {
3155 /*
3156 * If we're accessing one of the uncached arguments, we'll turn this
3157 * into a reference in the args array.
3158 */
3159 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3160 ndx = v - DIF_VAR_ARG0;
3161 v = DIF_VAR_ARGS;
3162 }
3163
3164 switch (v) {
3165 case DIF_VAR_ARGS:
3166 if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
3167 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
3168 CPU_DTRACE_KPRIV;
3169 return (0);
3170 }
3171
3172 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3173 if (ndx >= sizeof (mstate->dtms_arg) /
3174 sizeof (mstate->dtms_arg[0])) {
3175 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3176 dtrace_provider_t *pv;
3177 uint64_t val;
3178
3179 pv = mstate->dtms_probe->dtpr_provider;
3180 if (pv->dtpv_pops.dtps_getargval != NULL)
3181 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3182 mstate->dtms_probe->dtpr_id,
3183 mstate->dtms_probe->dtpr_arg, ndx, aframes);
3184 else
3185 val = dtrace_getarg(ndx, aframes);
3186
3187 /*
3188 * This is regrettably required to keep the compiler
3189 * from tail-optimizing the call to dtrace_getarg().
3190 * The condition always evaluates to true, but the
3191 * compiler has no way of figuring that out a priori.
3192 * (None of this would be necessary if the compiler
3193 * could be relied upon to _always_ tail-optimize
3194 * the call to dtrace_getarg() -- but it can't.)
3195 */
3196 if (mstate->dtms_probe != NULL)
3197 return (val);
3198
3199 ASSERT(0);
3200 }
3201
3202 return (mstate->dtms_arg[ndx]);
3203
3204 case DIF_VAR_UREGS: {
3205 klwp_t *lwp;
3206
3207 if (!dtrace_priv_proc(state, mstate))
3208 return (0);
3209
3210 if ((lwp = curthread->t_lwp) == NULL) {
3211 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3212 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
3213 return (0);
3214 }
3215
3216 return (dtrace_getreg(lwp->lwp_regs, ndx));
3217 }
3218
3219 case DIF_VAR_VMREGS: {
3220 uint64_t rval;
3221
3222 if (!dtrace_priv_kernel(state))
3223 return (0);
3224
3225 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3226
3227 rval = dtrace_getvmreg(ndx,
3228 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
3229
3230 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3231
3232 return (rval);
3233 }
3234
3235 case DIF_VAR_CURTHREAD:
3236 if (!dtrace_priv_proc(state, mstate))
3237 return (0);
3238 return ((uint64_t)(uintptr_t)curthread);
3239
3240 case DIF_VAR_TIMESTAMP:
3241 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3242 mstate->dtms_timestamp = dtrace_gethrtime();
3243 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3244 }
3245 return (mstate->dtms_timestamp);
3246
3247 case DIF_VAR_VTIMESTAMP:
3248 ASSERT(dtrace_vtime_references != 0);
3249 return (curthread->t_dtrace_vtime);
3250
3251 case DIF_VAR_WALLTIMESTAMP:
3252 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3253 mstate->dtms_walltimestamp = dtrace_gethrestime();
3254 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3255 }
3256 return (mstate->dtms_walltimestamp);
3257
3258 case DIF_VAR_IPL:
3259 if (!dtrace_priv_kernel(state))
3260 return (0);
3261 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3262 mstate->dtms_ipl = dtrace_getipl();
3263 mstate->dtms_present |= DTRACE_MSTATE_IPL;
3264 }
3265 return (mstate->dtms_ipl);
3266
3267 case DIF_VAR_EPID:
3268 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3269 return (mstate->dtms_epid);
3270
3271 case DIF_VAR_ID:
3272 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3273 return (mstate->dtms_probe->dtpr_id);
3274
3275 case DIF_VAR_STACKDEPTH:
3276 if (!dtrace_priv_kernel(state))
3277 return (0);
3278 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3279 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3280
3281 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3282 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3283 }
3284 return (mstate->dtms_stackdepth);
3285
3286 case DIF_VAR_USTACKDEPTH:
3287 if (!dtrace_priv_proc(state, mstate))
3288 return (0);
3289 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3290 /*
3291 * See comment in DIF_VAR_PID.
3292 */
3293 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3294 CPU_ON_INTR(CPU)) {
3295 mstate->dtms_ustackdepth = 0;
3296 } else {
3297 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3298 mstate->dtms_ustackdepth =
3299 dtrace_getustackdepth();
3300 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3301 }
3302 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3303 }
3304 return (mstate->dtms_ustackdepth);
3305
3306 case DIF_VAR_CALLER:
3307 if (!dtrace_priv_kernel(state))
3308 return (0);
3309 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3310 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3311
3312 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3313 /*
3314 * If this is an unanchored probe, we are
3315 * required to go through the slow path:
3316 * dtrace_caller() only guarantees correct
3317 * results for anchored probes.
3318 */
3319 pc_t caller[2];
3320
3321 dtrace_getpcstack(caller, 2, aframes,
3322 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3323 mstate->dtms_caller = caller[1];
3324 } else if ((mstate->dtms_caller =
3325 dtrace_caller(aframes)) == -1) {
3326 /*
3327 * We have failed to do this the quick way;
3328 * we must resort to the slower approach of
3329 * calling dtrace_getpcstack().
3330 */
3331 pc_t caller;
3332
3333 dtrace_getpcstack(&caller, 1, aframes, NULL);
3334 mstate->dtms_caller = caller;
3335 }
3336
3337 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3338 }
3339 return (mstate->dtms_caller);
3340
3341 case DIF_VAR_UCALLER:
3342 if (!dtrace_priv_proc(state, mstate))
3343 return (0);
3344
3345 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3346 uint64_t ustack[3];
3347
3348 /*
3349 * dtrace_getupcstack() fills in the first uint64_t
3350 * with the current PID. The second uint64_t will
3351 * be the program counter at user-level. The third
3352 * uint64_t will contain the caller, which is what
3353 * we're after.
3354 */
3355 ustack[2] = NULL;
3356 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3357 dtrace_getupcstack(ustack, 3);
3358 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3359 mstate->dtms_ucaller = ustack[2];
3360 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3361 }
3362
3363 return (mstate->dtms_ucaller);
3364
3365 case DIF_VAR_PROBEPROV:
3366 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3367 return (dtrace_dif_varstr(
3368 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3369 state, mstate));
3370
3371 case DIF_VAR_PROBEMOD:
3372 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3373 return (dtrace_dif_varstr(
3374 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3375 state, mstate));
3376
3377 case DIF_VAR_PROBEFUNC:
3378 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3379 return (dtrace_dif_varstr(
3380 (uintptr_t)mstate->dtms_probe->dtpr_func,
3381 state, mstate));
3382
3383 case DIF_VAR_PROBENAME:
3384 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3385 return (dtrace_dif_varstr(
3386 (uintptr_t)mstate->dtms_probe->dtpr_name,
3387 state, mstate));
3388
3389 case DIF_VAR_PID:
3390 if (!dtrace_priv_proc(state, mstate))
3391 return (0);
3392
3393 /*
3394 * Note that we are assuming that an unanchored probe is
3395 * always due to a high-level interrupt. (And we're assuming
3396 * that there is only a single high level interrupt.)
3397 */
3398 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3399 return (pid0.pid_id);
3400
3401 /*
3402 * It is always safe to dereference one's own t_procp pointer:
3403 * it always points to a valid, allocated proc structure.
3404 * Further, it is always safe to dereference the p_pidp member
3405 * of one's own proc structure. (These are truisms becuase
3406 * threads and processes don't clean up their own state --
3407 * they leave that task to whomever reaps them.)
3408 */
3409 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3410
3411 case DIF_VAR_PPID:
3412 if (!dtrace_priv_proc(state, mstate))
3413 return (0);
3414
3415 /*
3416 * See comment in DIF_VAR_PID.
3417 */
3418 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3419 return (pid0.pid_id);
3420
3421 /*
3422 * It is always safe to dereference one's own t_procp pointer:
3423 * it always points to a valid, allocated proc structure.
3424 * (This is true because threads don't clean up their own
3425 * state -- they leave that task to whomever reaps them.)
3426 */
3427 return ((uint64_t)curthread->t_procp->p_ppid);
3428
3429 case DIF_VAR_TID:
3430 /*
3431 * See comment in DIF_VAR_PID.
3432 */
3433 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3434 return (0);
3435
3436 return ((uint64_t)curthread->t_tid);
3437
3438 case DIF_VAR_EXECNAME:
3439 if (!dtrace_priv_proc(state, mstate))
3440 return (0);
3441
3442 /*
3443 * See comment in DIF_VAR_PID.
3444 */
3445 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3446 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3447
3448 /*
3449 * It is always safe to dereference one's own t_procp pointer:
3450 * it always points to a valid, allocated proc structure.
3451 * (This is true because threads don't clean up their own
3452 * state -- they leave that task to whomever reaps them.)
3453 */
3454 return (dtrace_dif_varstr(
3455 (uintptr_t)curthread->t_procp->p_user.u_comm,
3456 state, mstate));
3457
3458 case DIF_VAR_ZONENAME:
3459 if (!dtrace_priv_proc(state, mstate))
3460 return (0);
3461
3462 /*
3463 * See comment in DIF_VAR_PID.
3464 */
3465 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3466 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3467
3468 /*
3469 * It is always safe to dereference one's own t_procp pointer:
3470 * it always points to a valid, allocated proc structure.
3471 * (This is true because threads don't clean up their own
3472 * state -- they leave that task to whomever reaps them.)
3473 */
3474 return (dtrace_dif_varstr(
3475 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3476 state, mstate));
3477
3478 case DIF_VAR_UID:
3479 if (!dtrace_priv_proc(state, mstate))
3480 return (0);
3481
3482 /*
3483 * See comment in DIF_VAR_PID.
3484 */
3485 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3486 return ((uint64_t)p0.p_cred->cr_uid);
3487
3488 /*
3489 * It is always safe to dereference one's own t_procp pointer:
3490 * it always points to a valid, allocated proc structure.
3491 * (This is true because threads don't clean up their own
3492 * state -- they leave that task to whomever reaps them.)
3493 *
3494 * Additionally, it is safe to dereference one's own process
3495 * credential, since this is never NULL after process birth.
3496 */
3497 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3498
3499 case DIF_VAR_GID:
3500 if (!dtrace_priv_proc(state, mstate))
3501 return (0);
3502
3503 /*
3504 * See comment in DIF_VAR_PID.
3505 */
3506 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3507 return ((uint64_t)p0.p_cred->cr_gid);
3508
3509 /*
3510 * It is always safe to dereference one's own t_procp pointer:
3511 * it always points to a valid, allocated proc structure.
3512 * (This is true because threads don't clean up their own
3513 * state -- they leave that task to whomever reaps them.)
3514 *
3515 * Additionally, it is safe to dereference one's own process
3516 * credential, since this is never NULL after process birth.
3517 */
3518 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3519
3520 case DIF_VAR_ERRNO: {
3521 klwp_t *lwp;
3522 if (!dtrace_priv_proc(state, mstate))
3523 return (0);
3524
3525 /*
3526 * See comment in DIF_VAR_PID.
3527 */
3528 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3529 return (0);
3530
3531 /*
3532 * It is always safe to dereference one's own t_lwp pointer in
3533 * the event that this pointer is non-NULL. (This is true
3534 * because threads and lwps don't clean up their own state --
3535 * they leave that task to whomever reaps them.)
3536 */
3537 if ((lwp = curthread->t_lwp) == NULL)
3538 return (0);
3539
3540 return ((uint64_t)lwp->lwp_errno);
3541 }
3542
3543 case DIF_VAR_THREADNAME:
3544 /*
3545 * See comment in DIF_VAR_PID.
3546 */
3547 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3548 return (0);
3549
3550 if (curthread->t_name == NULL)
3551 return (0);
3552
3553 /*
3554 * Once set, ->t_name itself is never changed: any updates are
3555 * made to the same buffer that we are pointing out. So we are
3556 * safe to dereference it here.
3557 */
3558 return (dtrace_dif_varstr((uintptr_t)curthread->t_name,
3559 state, mstate));
3560
3561 default:
3562 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3563 return (0);
3564 }
3565 }
3566
3567 static void
3568 dtrace_dif_variable_write(dtrace_mstate_t *mstate, dtrace_state_t *state,
3569 uint64_t v, uint64_t ndx, uint64_t data)
3570 {
3571 switch (v) {
3572 case DIF_VAR_UREGS: {
3573 klwp_t *lwp;
3574
3575 if (dtrace_destructive_disallow ||
3576 !dtrace_priv_proc_control(state, mstate)) {
3577 return;
3578 }
3579
3580 if ((lwp = curthread->t_lwp) == NULL) {
3581 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3582 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
3583 return;
3584 }
3585
3586 dtrace_setreg(lwp->lwp_regs, ndx, data);
3587 return;
3588 }
3589
3590 default:
3591 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3592 return;
3593 }
3594 }
3595
3596 typedef enum dtrace_json_state {
3597 DTRACE_JSON_REST = 1,
3598 DTRACE_JSON_OBJECT,
3599 DTRACE_JSON_STRING,
3600 DTRACE_JSON_STRING_ESCAPE,
3601 DTRACE_JSON_STRING_ESCAPE_UNICODE,
3602 DTRACE_JSON_COLON,
3603 DTRACE_JSON_COMMA,
3604 DTRACE_JSON_VALUE,
3605 DTRACE_JSON_IDENTIFIER,
3606 DTRACE_JSON_NUMBER,
3607 DTRACE_JSON_NUMBER_FRAC,
3608 DTRACE_JSON_NUMBER_EXP,
3609 DTRACE_JSON_COLLECT_OBJECT
3610 } dtrace_json_state_t;
3611
3612 /*
3613 * This function possesses just enough knowledge about JSON to extract a single
3614 * value from a JSON string and store it in the scratch buffer. It is able
3615 * to extract nested object values, and members of arrays by index.
3616 *
3617 * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3618 * be looked up as we descend into the object tree. e.g.
3619 *
3620 * foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3621 * with nelems = 5.
3622 *
3623 * The run time of this function must be bounded above by strsize to limit the
3624 * amount of work done in probe context. As such, it is implemented as a
3625 * simple state machine, reading one character at a time using safe loads
3626 * until we find the requested element, hit a parsing error or run off the
3627 * end of the object or string.
3628 *
3629 * As there is no way for a subroutine to return an error without interrupting
3630 * clause execution, we simply return NULL in the event of a missing key or any
3631 * other error condition. Each NULL return in this function is commented with
3632 * the error condition it represents -- parsing or otherwise.
3633 *
3634 * The set of states for the state machine closely matches the JSON
3635 * specification (http://json.org/). Briefly:
3636 *
3637 * DTRACE_JSON_REST:
3638 * Skip whitespace until we find either a top-level Object, moving
3639 * to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3640 *
3641 * DTRACE_JSON_OBJECT:
3642 * Locate the next key String in an Object. Sets a flag to denote
3643 * the next String as a key string and moves to DTRACE_JSON_STRING.
3644 *
3645 * DTRACE_JSON_COLON:
3646 * Skip whitespace until we find the colon that separates key Strings
3647 * from their values. Once found, move to DTRACE_JSON_VALUE.
3648 *
3649 * DTRACE_JSON_VALUE:
3650 * Detects the type of the next value (String, Number, Identifier, Object
3651 * or Array) and routes to the states that process that type. Here we also
3652 * deal with the element selector list if we are requested to traverse down
3653 * into the object tree.
3654 *
3655 * DTRACE_JSON_COMMA:
3656 * Skip whitespace until we find the comma that separates key-value pairs
3657 * in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3658 * (similarly DTRACE_JSON_VALUE). All following literal value processing
3659 * states return to this state at the end of their value, unless otherwise
3660 * noted.
3661 *
3662 * DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3663 * Processes a Number literal from the JSON, including any exponent
3664 * component that may be present. Numbers are returned as strings, which
3665 * may be passed to strtoll() if an integer is required.
3666 *
3667 * DTRACE_JSON_IDENTIFIER:
3668 * Processes a "true", "false" or "null" literal in the JSON.
3669 *
3670 * DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3671 * DTRACE_JSON_STRING_ESCAPE_UNICODE:
3672 * Processes a String literal from the JSON, whether the String denotes
3673 * a key, a value or part of a larger Object. Handles all escape sequences
3674 * present in the specification, including four-digit unicode characters,
3675 * but merely includes the escape sequence without converting it to the
3676 * actual escaped character. If the String is flagged as a key, we
3677 * move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3678 *
3679 * DTRACE_JSON_COLLECT_OBJECT:
3680 * This state collects an entire Object (or Array), correctly handling
3681 * embedded strings. If the full element selector list matches this nested
3682 * object, we return the Object in full as a string. If not, we use this
3683 * state to skip to the next value at this level and continue processing.
3684 *
3685 * NOTE: This function uses various macros from strtolctype.h to manipulate
3686 * digit values, etc -- these have all been checked to ensure they make
3687 * no additional function calls.
3688 */
3689 static char *
3690 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3691 char *dest)
3692 {
3693 dtrace_json_state_t state = DTRACE_JSON_REST;
3694 int64_t array_elem = INT64_MIN;
3695 int64_t array_pos = 0;
3696 uint8_t escape_unicount = 0;
3697 boolean_t string_is_key = B_FALSE;
3698 boolean_t collect_object = B_FALSE;
3699 boolean_t found_key = B_FALSE;
3700 boolean_t in_array = B_FALSE;
3701 uint32_t braces = 0, brackets = 0;
3702 char *elem = elemlist;
3703 char *dd = dest;
3704 uintptr_t cur;
3705
3706 for (cur = json; cur < json + size; cur++) {
3707 char cc = dtrace_load8(cur);
3708 if (cc == '\0')
3709 return (NULL);
3710
3711 switch (state) {
3712 case DTRACE_JSON_REST:
3713 if (isspace(cc))
3714 break;
3715
3716 if (cc == '{') {
3717 state = DTRACE_JSON_OBJECT;
3718 break;
3719 }
3720
3721 if (cc == '[') {
3722 in_array = B_TRUE;
3723 array_pos = 0;
3724 array_elem = dtrace_strtoll(elem, 10, size);
3725 found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3726 state = DTRACE_JSON_VALUE;
3727 break;
3728 }
3729
3730 /*
3731 * ERROR: expected to find a top-level object or array.
3732 */
3733 return (NULL);
3734 case DTRACE_JSON_OBJECT:
3735 if (isspace(cc))
3736 break;
3737
3738 if (cc == '"') {
3739 state = DTRACE_JSON_STRING;
3740 string_is_key = B_TRUE;
3741 break;
3742 }
3743
3744 /*
3745 * ERROR: either the object did not start with a key
3746 * string, or we've run off the end of the object
3747 * without finding the requested key.
3748 */
3749 return (NULL);
3750 case DTRACE_JSON_STRING:
3751 if (cc == '\\') {
3752 *dd++ = '\\';
3753 state = DTRACE_JSON_STRING_ESCAPE;
3754 break;
3755 }
3756
3757 if (cc == '"') {
3758 if (collect_object) {
3759 /*
3760 * We don't reset the dest here, as
3761 * the string is part of a larger
3762 * object being collected.
3763 */
3764 *dd++ = cc;
3765 collect_object = B_FALSE;
3766 state = DTRACE_JSON_COLLECT_OBJECT;
3767 break;
3768 }
3769 *dd = '\0';
3770 dd = dest; /* reset string buffer */
3771 if (string_is_key) {
3772 if (dtrace_strncmp(dest, elem,
3773 size) == 0)
3774 found_key = B_TRUE;
3775 } else if (found_key) {
3776 if (nelems > 1) {
3777 /*
3778 * We expected an object, not
3779 * this string.
3780 */
3781 return (NULL);
3782 }
3783 return (dest);
3784 }
3785 state = string_is_key ? DTRACE_JSON_COLON :
3786 DTRACE_JSON_COMMA;
3787 string_is_key = B_FALSE;
3788 break;
3789 }
3790
3791 *dd++ = cc;
3792 break;
3793 case DTRACE_JSON_STRING_ESCAPE:
3794 *dd++ = cc;
3795 if (cc == 'u') {
3796 escape_unicount = 0;
3797 state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3798 } else {
3799 state = DTRACE_JSON_STRING;
3800 }
3801 break;
3802 case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3803 if (!isxdigit(cc)) {
3804 /*
3805 * ERROR: invalid unicode escape, expected
3806 * four valid hexidecimal digits.
3807 */
3808 return (NULL);
3809 }
3810
3811 *dd++ = cc;
3812 if (++escape_unicount == 4)
3813 state = DTRACE_JSON_STRING;
3814 break;
3815 case DTRACE_JSON_COLON:
3816 if (isspace(cc))
3817 break;
3818
3819 if (cc == ':') {
3820 state = DTRACE_JSON_VALUE;
3821 break;
3822 }
3823
3824 /*
3825 * ERROR: expected a colon.
3826 */
3827 return (NULL);
3828 case DTRACE_JSON_COMMA:
3829 if (isspace(cc))
3830 break;
3831
3832 if (cc == ',') {
3833 if (in_array) {
3834 state = DTRACE_JSON_VALUE;
3835 if (++array_pos == array_elem)
3836 found_key = B_TRUE;
3837 } else {
3838 state = DTRACE_JSON_OBJECT;
3839 }
3840 break;
3841 }
3842
3843 /*
3844 * ERROR: either we hit an unexpected character, or
3845 * we reached the end of the object or array without
3846 * finding the requested key.
3847 */
3848 return (NULL);
3849 case DTRACE_JSON_IDENTIFIER:
3850 if (islower(cc)) {
3851 *dd++ = cc;
3852 break;
3853 }
3854
3855 *dd = '\0';
3856 dd = dest; /* reset string buffer */
3857
3858 if (dtrace_strncmp(dest, "true", 5) == 0 ||
3859 dtrace_strncmp(dest, "false", 6) == 0 ||
3860 dtrace_strncmp(dest, "null", 5) == 0) {
3861 if (found_key) {
3862 if (nelems > 1) {
3863 /*
3864 * ERROR: We expected an object,
3865 * not this identifier.
3866 */
3867 return (NULL);
3868 }
3869 return (dest);
3870 } else {
3871 cur--;
3872 state = DTRACE_JSON_COMMA;
3873 break;
3874 }
3875 }
3876
3877 /*
3878 * ERROR: we did not recognise the identifier as one
3879 * of those in the JSON specification.
3880 */
3881 return (NULL);
3882 case DTRACE_JSON_NUMBER:
3883 if (cc == '.') {
3884 *dd++ = cc;
3885 state = DTRACE_JSON_NUMBER_FRAC;
3886 break;
3887 }
3888
3889 if (cc == 'x' || cc == 'X') {
3890 /*
3891 * ERROR: specification explicitly excludes
3892 * hexidecimal or octal numbers.
3893 */
3894 return (NULL);
3895 }
3896
3897 /* FALLTHRU */
3898 case DTRACE_JSON_NUMBER_FRAC:
3899 if (cc == 'e' || cc == 'E') {
3900 *dd++ = cc;
3901 state = DTRACE_JSON_NUMBER_EXP;
3902 break;
3903 }
3904
3905 if (cc == '+' || cc == '-') {
3906 /*
3907 * ERROR: expect sign as part of exponent only.
3908 */
3909 return (NULL);
3910 }
3911 /* FALLTHRU */
3912 case DTRACE_JSON_NUMBER_EXP:
3913 if (isdigit(cc) || cc == '+' || cc == '-') {
3914 *dd++ = cc;
3915 break;
3916 }
3917
3918 *dd = '\0';
3919 dd = dest; /* reset string buffer */
3920 if (found_key) {
3921 if (nelems > 1) {
3922 /*
3923 * ERROR: We expected an object, not
3924 * this number.
3925 */
3926 return (NULL);
3927 }
3928 return (dest);
3929 }
3930
3931 cur--;
3932 state = DTRACE_JSON_COMMA;
3933 break;
3934 case DTRACE_JSON_VALUE:
3935 if (isspace(cc))
3936 break;
3937
3938 if (cc == '{' || cc == '[') {
3939 if (nelems > 1 && found_key) {
3940 in_array = cc == '[' ? B_TRUE : B_FALSE;
3941 /*
3942 * If our element selector directs us
3943 * to descend into this nested object,
3944 * then move to the next selector
3945 * element in the list and restart the
3946 * state machine.
3947 */
3948 while (*elem != '\0')
3949 elem++;
3950 elem++; /* skip the inter-element NUL */
3951 nelems--;
3952 dd = dest;
3953 if (in_array) {
3954 state = DTRACE_JSON_VALUE;
3955 array_pos = 0;
3956 array_elem = dtrace_strtoll(
3957 elem, 10, size);
3958 found_key = array_elem == 0 ?
3959 B_TRUE : B_FALSE;
3960 } else {
3961 found_key = B_FALSE;
3962 state = DTRACE_JSON_OBJECT;
3963 }
3964 break;
3965 }
3966
3967 /*
3968 * Otherwise, we wish to either skip this
3969 * nested object or return it in full.
3970 */
3971 if (cc == '[')
3972 brackets = 1;
3973 else
3974 braces = 1;
3975 *dd++ = cc;
3976 state = DTRACE_JSON_COLLECT_OBJECT;
3977 break;
3978 }
3979
3980 if (cc == '"') {
3981 state = DTRACE_JSON_STRING;
3982 break;
3983 }
3984
3985 if (islower(cc)) {
3986 /*
3987 * Here we deal with true, false and null.
3988 */
3989 *dd++ = cc;
3990 state = DTRACE_JSON_IDENTIFIER;
3991 break;
3992 }
3993
3994 if (cc == '-' || isdigit(cc)) {
3995 *dd++ = cc;
3996 state = DTRACE_JSON_NUMBER;
3997 break;
3998 }
3999
4000 /*
4001 * ERROR: unexpected character at start of value.
4002 */
4003 return (NULL);
4004 case DTRACE_JSON_COLLECT_OBJECT:
4005 if (cc == '\0')
4006 /*
4007 * ERROR: unexpected end of input.
4008 */
4009 return (NULL);
4010
4011 *dd++ = cc;
4012 if (cc == '"') {
4013 collect_object = B_TRUE;
4014 state = DTRACE_JSON_STRING;
4015 break;
4016 }
4017
4018 if (cc == ']') {
4019 if (brackets-- == 0) {
4020 /*
4021 * ERROR: unbalanced brackets.
4022 */
4023 return (NULL);
4024 }
4025 } else if (cc == '}') {
4026 if (braces-- == 0) {
4027 /*
4028 * ERROR: unbalanced braces.
4029 */
4030 return (NULL);
4031 }
4032 } else if (cc == '{') {
4033 braces++;
4034 } else if (cc == '[') {
4035 brackets++;
4036 }
4037
4038 if (brackets == 0 && braces == 0) {
4039 if (found_key) {
4040 *dd = '\0';
4041 return (dest);
4042 }
4043 dd = dest; /* reset string buffer */
4044 state = DTRACE_JSON_COMMA;
4045 }
4046 break;
4047 }
4048 }
4049 return (NULL);
4050 }
4051
4052 /*
4053 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
4054 * Notice that we don't bother validating the proper number of arguments or
4055 * their types in the tuple stack. This isn't needed because all argument
4056 * interpretation is safe because of our load safety -- the worst that can
4057 * happen is that a bogus program can obtain bogus results.
4058 */
4059 static void
4060 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
4061 dtrace_key_t *tupregs, int nargs,
4062 dtrace_mstate_t *mstate, dtrace_state_t *state)
4063 {
4064 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4065 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
4066 dtrace_vstate_t *vstate = &state->dts_vstate;
4067
4068 union {
4069 mutex_impl_t mi;
4070 uint64_t mx;
4071 } m;
4072
4073 union {
4074 krwlock_t ri;
4075 uintptr_t rw;
4076 } r;
4077
4078 switch (subr) {
4079 case DIF_SUBR_RAND:
4080 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
4081 break;
4082
4083 case DIF_SUBR_MUTEX_OWNED:
4084 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4085 mstate, vstate)) {
4086 regs[rd] = NULL;
4087 break;
4088 }
4089
4090 m.mx = dtrace_load64(tupregs[0].dttk_value);
4091 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
4092 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
4093 else
4094 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
4095 break;
4096
4097 case DIF_SUBR_MUTEX_OWNER:
4098 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4099 mstate, vstate)) {
4100 regs[rd] = NULL;
4101 break;
4102 }
4103
4104 m.mx = dtrace_load64(tupregs[0].dttk_value);
4105 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
4106 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
4107 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
4108 else
4109 regs[rd] = 0;
4110 break;
4111
4112 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4113 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4114 mstate, vstate)) {
4115 regs[rd] = NULL;
4116 break;
4117 }
4118
4119 m.mx = dtrace_load64(tupregs[0].dttk_value);
4120 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
4121 break;
4122
4123 case DIF_SUBR_MUTEX_TYPE_SPIN:
4124 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4125 mstate, vstate)) {
4126 regs[rd] = NULL;
4127 break;
4128 }
4129
4130 m.mx = dtrace_load64(tupregs[0].dttk_value);
4131 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
4132 break;
4133
4134 case DIF_SUBR_RW_READ_HELD: {
4135 uintptr_t tmp;
4136
4137 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4138 mstate, vstate)) {
4139 regs[rd] = NULL;
4140 break;
4141 }
4142
4143 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4144 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
4145 break;
4146 }
4147
4148 case DIF_SUBR_RW_WRITE_HELD:
4149 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4150 mstate, vstate)) {
4151 regs[rd] = NULL;
4152 break;
4153 }
4154
4155 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4156 regs[rd] = _RW_WRITE_HELD(&r.ri);
4157 break;
4158
4159 case DIF_SUBR_RW_ISWRITER:
4160 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4161 mstate, vstate)) {
4162 regs[rd] = NULL;
4163 break;
4164 }
4165
4166 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4167 regs[rd] = _RW_ISWRITER(&r.ri);
4168 break;
4169
4170 case DIF_SUBR_BCOPY: {
4171 /*
4172 * We need to be sure that the destination is in the scratch
4173 * region -- no other region is allowed.
4174 */
4175 uintptr_t src = tupregs[0].dttk_value;
4176 uintptr_t dest = tupregs[1].dttk_value;
4177 size_t size = tupregs[2].dttk_value;
4178
4179 if (!dtrace_inscratch(dest, size, mstate)) {
4180 *flags |= CPU_DTRACE_BADADDR;
4181 *illval = regs[rd];
4182 break;
4183 }
4184
4185 if (!dtrace_canload(src, size, mstate, vstate)) {
4186 regs[rd] = NULL;
4187 break;
4188 }
4189
4190 dtrace_bcopy((void *)src, (void *)dest, size);
4191 break;
4192 }
4193
4194 case DIF_SUBR_ALLOCA:
4195 case DIF_SUBR_COPYIN: {
4196 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4197 uint64_t size =
4198 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
4199 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
4200
4201 /*
4202 * This action doesn't require any credential checks since
4203 * probes will not activate in user contexts to which the
4204 * enabling user does not have permissions.
4205 */
4206
4207 /*
4208 * Rounding up the user allocation size could have overflowed
4209 * a large, bogus allocation (like -1ULL) to 0.
4210 */
4211 if (scratch_size < size ||
4212 !DTRACE_INSCRATCH(mstate, scratch_size)) {
4213 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4214 regs[rd] = NULL;
4215 break;
4216 }
4217
4218 if (subr == DIF_SUBR_COPYIN) {
4219 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4220 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4221 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4222 }
4223
4224 mstate->dtms_scratch_ptr += scratch_size;
4225 regs[rd] = dest;
4226 break;
4227 }
4228
4229 case DIF_SUBR_COPYINTO: {
4230 uint64_t size = tupregs[1].dttk_value;
4231 uintptr_t dest = tupregs[2].dttk_value;
4232
4233 /*
4234 * This action doesn't require any credential checks since
4235 * probes will not activate in user contexts to which the
4236 * enabling user does not have permissions.
4237 */
4238 if (!dtrace_inscratch(dest, size, mstate)) {
4239 *flags |= CPU_DTRACE_BADADDR;
4240 *illval = regs[rd];
4241 break;
4242 }
4243
4244 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4245 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4246 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4247 break;
4248 }
4249
4250 case DIF_SUBR_COPYINSTR: {
4251 uintptr_t dest = mstate->dtms_scratch_ptr;
4252 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4253
4254 if (nargs > 1 && tupregs[1].dttk_value < size)
4255 size = tupregs[1].dttk_value + 1;
4256
4257 /*
4258 * This action doesn't require any credential checks since
4259 * probes will not activate in user contexts to which the
4260 * enabling user does not have permissions.
4261 */
4262 if (!DTRACE_INSCRATCH(mstate, size)) {
4263 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4264 regs[rd] = NULL;
4265 break;
4266 }
4267
4268 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4269 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4270 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4271
4272 ((char *)dest)[size - 1] = '\0';
4273 mstate->dtms_scratch_ptr += size;
4274 regs[rd] = dest;
4275 break;
4276 }
4277
4278 case DIF_SUBR_MSGSIZE:
4279 case DIF_SUBR_MSGDSIZE: {
4280 uintptr_t baddr = tupregs[0].dttk_value, daddr;
4281 uintptr_t wptr, rptr;
4282 size_t count = 0;
4283 int cont = 0;
4284
4285 while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4286
4287 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4288 vstate)) {
4289 regs[rd] = NULL;
4290 break;
4291 }
4292
4293 wptr = dtrace_loadptr(baddr +
4294 offsetof(mblk_t, b_wptr));
4295
4296 rptr = dtrace_loadptr(baddr +
4297 offsetof(mblk_t, b_rptr));
4298
4299 if (wptr < rptr) {
4300 *flags |= CPU_DTRACE_BADADDR;
4301 *illval = tupregs[0].dttk_value;
4302 break;
4303 }
4304
4305 daddr = dtrace_loadptr(baddr +
4306 offsetof(mblk_t, b_datap));
4307
4308 baddr = dtrace_loadptr(baddr +
4309 offsetof(mblk_t, b_cont));
4310
4311 /*
4312 * We want to prevent against denial-of-service here,
4313 * so we're only going to search the list for
4314 * dtrace_msgdsize_max mblks.
4315 */
4316 if (cont++ > dtrace_msgdsize_max) {
4317 *flags |= CPU_DTRACE_ILLOP;
4318 break;
4319 }
4320
4321 if (subr == DIF_SUBR_MSGDSIZE) {
4322 if (dtrace_load8(daddr +
4323 offsetof(dblk_t, db_type)) != M_DATA)
4324 continue;
4325 }
4326
4327 count += wptr - rptr;
4328 }
4329
4330 if (!(*flags & CPU_DTRACE_FAULT))
4331 regs[rd] = count;
4332
4333 break;
4334 }
4335
4336 case DIF_SUBR_PROGENYOF: {
4337 pid_t pid = tupregs[0].dttk_value;
4338 proc_t *p;
4339 int rval = 0;
4340
4341 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4342
4343 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4344 if (p->p_pidp->pid_id == pid) {
4345 rval = 1;
4346 break;
4347 }
4348 }
4349
4350 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4351
4352 regs[rd] = rval;
4353 break;
4354 }
4355
4356 case DIF_SUBR_SPECULATION:
4357 regs[rd] = dtrace_speculation(state);
4358 break;
4359
4360 case DIF_SUBR_COPYOUT: {
4361 uintptr_t kaddr = tupregs[0].dttk_value;
4362 uintptr_t uaddr = tupregs[1].dttk_value;
4363 uint64_t size = tupregs[2].dttk_value;
4364
4365 if (!dtrace_destructive_disallow &&
4366 dtrace_priv_proc_control(state, mstate) &&
4367 !dtrace_istoxic(kaddr, size) &&
4368 dtrace_canload(kaddr, size, mstate, vstate)) {
4369 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4370 dtrace_copyout(kaddr, uaddr, size, flags);
4371 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4372 }
4373 break;
4374 }
4375
4376 case DIF_SUBR_COPYOUTSTR: {
4377 uintptr_t kaddr = tupregs[0].dttk_value;
4378 uintptr_t uaddr = tupregs[1].dttk_value;
4379 uint64_t size = tupregs[2].dttk_value;
4380 size_t lim;
4381
4382 if (!dtrace_destructive_disallow &&
4383 dtrace_priv_proc_control(state, mstate) &&
4384 !dtrace_istoxic(kaddr, size) &&
4385 dtrace_strcanload(kaddr, size, &lim, mstate, vstate)) {
4386 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4387 dtrace_copyoutstr(kaddr, uaddr, lim, flags);
4388 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4389 }
4390 break;
4391 }
4392
4393 case DIF_SUBR_STRLEN: {
4394 size_t size = state->dts_options[DTRACEOPT_STRSIZE];
4395 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4396 size_t lim;
4397
4398 if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4399 regs[rd] = NULL;
4400 break;
4401 }
4402 regs[rd] = dtrace_strlen((char *)addr, lim);
4403
4404 break;
4405 }
4406
4407 case DIF_SUBR_STRCHR:
4408 case DIF_SUBR_STRRCHR: {
4409 /*
4410 * We're going to iterate over the string looking for the
4411 * specified character. We will iterate until we have reached
4412 * the string length or we have found the character. If this
4413 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4414 * of the specified character instead of the first.
4415 */
4416 uintptr_t addr = tupregs[0].dttk_value;
4417 uintptr_t addr_limit;
4418 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4419 size_t lim;
4420 char c, target = (char)tupregs[1].dttk_value;
4421
4422 if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4423 regs[rd] = NULL;
4424 break;
4425 }
4426 addr_limit = addr + lim;
4427
4428 for (regs[rd] = NULL; addr < addr_limit; addr++) {
4429 if ((c = dtrace_load8(addr)) == target) {
4430 regs[rd] = addr;
4431
4432 if (subr == DIF_SUBR_STRCHR)
4433 break;
4434 }
4435 if (c == '\0')
4436 break;
4437 }
4438
4439 break;
4440 }
4441
4442 case DIF_SUBR_STRSTR:
4443 case DIF_SUBR_INDEX:
4444 case DIF_SUBR_RINDEX: {
4445 /*
4446 * We're going to iterate over the string looking for the
4447 * specified string. We will iterate until we have reached
4448 * the string length or we have found the string. (Yes, this
4449 * is done in the most naive way possible -- but considering
4450 * that the string we're searching for is likely to be
4451 * relatively short, the complexity of Rabin-Karp or similar
4452 * hardly seems merited.)
4453 */
4454 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4455 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4456 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4457 size_t len = dtrace_strlen(addr, size);
4458 size_t sublen = dtrace_strlen(substr, size);
4459 char *limit = addr + len, *orig = addr;
4460 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4461 int inc = 1;
4462
4463 regs[rd] = notfound;
4464
4465 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4466 regs[rd] = NULL;
4467 break;
4468 }
4469
4470 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4471 vstate)) {
4472 regs[rd] = NULL;
4473 break;
4474 }
4475
4476 /*
4477 * strstr() and index()/rindex() have similar semantics if
4478 * both strings are the empty string: strstr() returns a
4479 * pointer to the (empty) string, and index() and rindex()
4480 * both return index 0 (regardless of any position argument).
4481 */
4482 if (sublen == 0 && len == 0) {
4483 if (subr == DIF_SUBR_STRSTR)
4484 regs[rd] = (uintptr_t)addr;
4485 else
4486 regs[rd] = 0;
4487 break;
4488 }
4489
4490 if (subr != DIF_SUBR_STRSTR) {
4491 if (subr == DIF_SUBR_RINDEX) {
4492 limit = orig - 1;
4493 addr += len;
4494 inc = -1;
4495 }
4496
4497 /*
4498 * Both index() and rindex() take an optional position
4499 * argument that denotes the starting position.
4500 */
4501 if (nargs == 3) {
4502 int64_t pos = (int64_t)tupregs[2].dttk_value;
4503
4504 /*
4505 * If the position argument to index() is
4506 * negative, Perl implicitly clamps it at
4507 * zero. This semantic is a little surprising
4508 * given the special meaning of negative
4509 * positions to similar Perl functions like
4510 * substr(), but it appears to reflect a
4511 * notion that index() can start from a
4512 * negative index and increment its way up to
4513 * the string. Given this notion, Perl's
4514 * rindex() is at least self-consistent in
4515 * that it implicitly clamps positions greater
4516 * than the string length to be the string
4517 * length. Where Perl completely loses
4518 * coherence, however, is when the specified
4519 * substring is the empty string (""). In
4520 * this case, even if the position is
4521 * negative, rindex() returns 0 -- and even if
4522 * the position is greater than the length,
4523 * index() returns the string length. These
4524 * semantics violate the notion that index()
4525 * should never return a value less than the
4526 * specified position and that rindex() should
4527 * never return a value greater than the
4528 * specified position. (One assumes that
4529 * these semantics are artifacts of Perl's
4530 * implementation and not the results of
4531 * deliberate design -- it beggars belief that
4532 * even Larry Wall could desire such oddness.)
4533 * While in the abstract one would wish for
4534 * consistent position semantics across
4535 * substr(), index() and rindex() -- or at the
4536 * very least self-consistent position
4537 * semantics for index() and rindex() -- we
4538 * instead opt to keep with the extant Perl
4539 * semantics, in all their broken glory. (Do
4540 * we have more desire to maintain Perl's
4541 * semantics than Perl does? Probably.)
4542 */
4543 if (subr == DIF_SUBR_RINDEX) {
4544 if (pos < 0) {
4545 if (sublen == 0)
4546 regs[rd] = 0;
4547 break;
4548 }
4549
4550 if (pos > len)
4551 pos = len;
4552 } else {
4553 if (pos < 0)
4554 pos = 0;
4555
4556 if (pos >= len) {
4557 if (sublen == 0)
4558 regs[rd] = len;
4559 break;
4560 }
4561 }
4562
4563 addr = orig + pos;
4564 }
4565 }
4566
4567 for (regs[rd] = notfound; addr != limit; addr += inc) {
4568 if (dtrace_strncmp(addr, substr, sublen) == 0) {
4569 if (subr != DIF_SUBR_STRSTR) {
4570 /*
4571 * As D index() and rindex() are
4572 * modeled on Perl (and not on awk),
4573 * we return a zero-based (and not a
4574 * one-based) index. (For you Perl
4575 * weenies: no, we're not going to add
4576 * $[ -- and shouldn't you be at a con
4577 * or something?)
4578 */
4579 regs[rd] = (uintptr_t)(addr - orig);
4580 break;
4581 }
4582
4583 ASSERT(subr == DIF_SUBR_STRSTR);
4584 regs[rd] = (uintptr_t)addr;
4585 break;
4586 }
4587 }
4588
4589 break;
4590 }
4591
4592 case DIF_SUBR_STRTOK: {
4593 uintptr_t addr = tupregs[0].dttk_value;
4594 uintptr_t tokaddr = tupregs[1].dttk_value;
4595 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4596 uintptr_t limit, toklimit;
4597 size_t clim;
4598 uint8_t c, tokmap[32]; /* 256 / 8 */
4599 char *dest = (char *)mstate->dtms_scratch_ptr;
4600 int i;
4601
4602 /*
4603 * Check both the token buffer and (later) the input buffer,
4604 * since both could be non-scratch addresses.
4605 */
4606 if (!dtrace_strcanload(tokaddr, size, &clim, mstate, vstate)) {
4607 regs[rd] = NULL;
4608 break;
4609 }
4610 toklimit = tokaddr + clim;
4611
4612 if (!DTRACE_INSCRATCH(mstate, size)) {
4613 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4614 regs[rd] = NULL;
4615 break;
4616 }
4617
4618 if (addr == NULL) {
4619 /*
4620 * If the address specified is NULL, we use our saved
4621 * strtok pointer from the mstate. Note that this
4622 * means that the saved strtok pointer is _only_
4623 * valid within multiple enablings of the same probe --
4624 * it behaves like an implicit clause-local variable.
4625 */
4626 addr = mstate->dtms_strtok;
4627 limit = mstate->dtms_strtok_limit;
4628 } else {
4629 /*
4630 * If the user-specified address is non-NULL we must
4631 * access check it. This is the only time we have
4632 * a chance to do so, since this address may reside
4633 * in the string table of this clause-- future calls
4634 * (when we fetch addr from mstate->dtms_strtok)
4635 * would fail this access check.
4636 */
4637 if (!dtrace_strcanload(addr, size, &clim, mstate,
4638 vstate)) {
4639 regs[rd] = NULL;
4640 break;
4641 }
4642 limit = addr + clim;
4643 }
4644
4645 /*
4646 * First, zero the token map, and then process the token
4647 * string -- setting a bit in the map for every character
4648 * found in the token string.
4649 */
4650 for (i = 0; i < sizeof (tokmap); i++)
4651 tokmap[i] = 0;
4652
4653 for (; tokaddr < toklimit; tokaddr++) {
4654 if ((c = dtrace_load8(tokaddr)) == '\0')
4655 break;
4656
4657 ASSERT((c >> 3) < sizeof (tokmap));
4658 tokmap[c >> 3] |= (1 << (c & 0x7));
4659 }
4660
4661 for (; addr < limit; addr++) {
4662 /*
4663 * We're looking for a character that is _not_
4664 * contained in the token string.
4665 */
4666 if ((c = dtrace_load8(addr)) == '\0')
4667 break;
4668
4669 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4670 break;
4671 }
4672
4673 if (c == '\0') {
4674 /*
4675 * We reached the end of the string without finding
4676 * any character that was not in the token string.
4677 * We return NULL in this case, and we set the saved
4678 * address to NULL as well.
4679 */
4680 regs[rd] = NULL;
4681 mstate->dtms_strtok = NULL;
4682 mstate->dtms_strtok_limit = NULL;
4683 break;
4684 }
4685
4686 /*
4687 * From here on, we're copying into the destination string.
4688 */
4689 for (i = 0; addr < limit && i < size - 1; addr++) {
4690 if ((c = dtrace_load8(addr)) == '\0')
4691 break;
4692
4693 if (tokmap[c >> 3] & (1 << (c & 0x7)))
4694 break;
4695
4696 ASSERT(i < size);
4697 dest[i++] = c;
4698 }
4699
4700 ASSERT(i < size);
4701 dest[i] = '\0';
4702 regs[rd] = (uintptr_t)dest;
4703 mstate->dtms_scratch_ptr += size;
4704 mstate->dtms_strtok = addr;
4705 mstate->dtms_strtok_limit = limit;
4706 break;
4707 }
4708
4709 case DIF_SUBR_SUBSTR: {
4710 uintptr_t s = tupregs[0].dttk_value;
4711 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4712 char *d = (char *)mstate->dtms_scratch_ptr;
4713 int64_t index = (int64_t)tupregs[1].dttk_value;
4714 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4715 size_t len = dtrace_strlen((char *)s, size);
4716 int64_t i;
4717
4718 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4719 regs[rd] = NULL;
4720 break;
4721 }
4722
4723 if (!DTRACE_INSCRATCH(mstate, size)) {
4724 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4725 regs[rd] = NULL;
4726 break;
4727 }
4728
4729 if (nargs <= 2)
4730 remaining = (int64_t)size;
4731
4732 if (index < 0) {
4733 index += len;
4734
4735 if (index < 0 && index + remaining > 0) {
4736 remaining += index;
4737 index = 0;
4738 }
4739 }
4740
4741 if (index >= len || index < 0) {
4742 remaining = 0;
4743 } else if (remaining < 0) {
4744 remaining += len - index;
4745 } else if (index + remaining > size) {
4746 remaining = size - index;
4747 }
4748
4749 for (i = 0; i < remaining; i++) {
4750 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4751 break;
4752 }
4753
4754 d[i] = '\0';
4755
4756 mstate->dtms_scratch_ptr += size;
4757 regs[rd] = (uintptr_t)d;
4758 break;
4759 }
4760
4761 case DIF_SUBR_JSON: {
4762 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4763 uintptr_t json = tupregs[0].dttk_value;
4764 size_t jsonlen = dtrace_strlen((char *)json, size);
4765 uintptr_t elem = tupregs[1].dttk_value;
4766 size_t elemlen = dtrace_strlen((char *)elem, size);
4767
4768 char *dest = (char *)mstate->dtms_scratch_ptr;
4769 char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
4770 char *ee = elemlist;
4771 int nelems = 1;
4772 uintptr_t cur;
4773
4774 if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
4775 !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
4776 regs[rd] = NULL;
4777 break;
4778 }
4779
4780 if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
4781 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4782 regs[rd] = NULL;
4783 break;
4784 }
4785
4786 /*
4787 * Read the element selector and split it up into a packed list
4788 * of strings.
4789 */
4790 for (cur = elem; cur < elem + elemlen; cur++) {
4791 char cc = dtrace_load8(cur);
4792
4793 if (cur == elem && cc == '[') {
4794 /*
4795 * If the first element selector key is
4796 * actually an array index then ignore the
4797 * bracket.
4798 */
4799 continue;
4800 }
4801
4802 if (cc == ']')
4803 continue;
4804
4805 if (cc == '.' || cc == '[') {
4806 nelems++;
4807 cc = '\0';
4808 }
4809
4810 *ee++ = cc;
4811 }
4812 *ee++ = '\0';
4813
4814 if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
4815 nelems, dest)) != NULL)
4816 mstate->dtms_scratch_ptr += jsonlen + 1;
4817 break;
4818 }
4819
4820 case DIF_SUBR_TOUPPER:
4821 case DIF_SUBR_TOLOWER: {
4822 uintptr_t s = tupregs[0].dttk_value;
4823 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4824 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4825 size_t len = dtrace_strlen((char *)s, size);
4826 char lower, upper, convert;
4827 int64_t i;
4828
4829 if (subr == DIF_SUBR_TOUPPER) {
4830 lower = 'a';
4831 upper = 'z';
4832 convert = 'A';
4833 } else {
4834 lower = 'A';
4835 upper = 'Z';
4836 convert = 'a';
4837 }
4838
4839 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4840 regs[rd] = NULL;
4841 break;
4842 }
4843
4844 if (!DTRACE_INSCRATCH(mstate, size)) {
4845 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4846 regs[rd] = NULL;
4847 break;
4848 }
4849
4850 for (i = 0; i < size - 1; i++) {
4851 if ((c = dtrace_load8(s + i)) == '\0')
4852 break;
4853
4854 if (c >= lower && c <= upper)
4855 c = convert + (c - lower);
4856
4857 dest[i] = c;
4858 }
4859
4860 ASSERT(i < size);
4861 dest[i] = '\0';
4862 regs[rd] = (uintptr_t)dest;
4863 mstate->dtms_scratch_ptr += size;
4864 break;
4865 }
4866
4867 case DIF_SUBR_GETMAJOR:
4868 #ifdef _LP64
4869 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4870 #else
4871 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4872 #endif
4873 break;
4874
4875 case DIF_SUBR_GETMINOR:
4876 #ifdef _LP64
4877 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4878 #else
4879 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4880 #endif
4881 break;
4882
4883 case DIF_SUBR_DDI_PATHNAME: {
4884 /*
4885 * This one is a galactic mess. We are going to roughly
4886 * emulate ddi_pathname(), but it's made more complicated
4887 * by the fact that we (a) want to include the minor name and
4888 * (b) must proceed iteratively instead of recursively.
4889 */
4890 uintptr_t dest = mstate->dtms_scratch_ptr;
4891 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4892 char *start = (char *)dest, *end = start + size - 1;
4893 uintptr_t daddr = tupregs[0].dttk_value;
4894 int64_t minor = (int64_t)tupregs[1].dttk_value;
4895 char *s;
4896 int i, len, depth = 0;
4897
4898 /*
4899 * Due to all the pointer jumping we do and context we must
4900 * rely upon, we just mandate that the user must have kernel
4901 * read privileges to use this routine.
4902 */
4903 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4904 *flags |= CPU_DTRACE_KPRIV;
4905 *illval = daddr;
4906 regs[rd] = NULL;
4907 }
4908
4909 if (!DTRACE_INSCRATCH(mstate, size)) {
4910 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4911 regs[rd] = NULL;
4912 break;
4913 }
4914
4915 *end = '\0';
4916
4917 /*
4918 * We want to have a name for the minor. In order to do this,
4919 * we need to walk the minor list from the devinfo. We want
4920 * to be sure that we don't infinitely walk a circular list,
4921 * so we check for circularity by sending a scout pointer
4922 * ahead two elements for every element that we iterate over;
4923 * if the list is circular, these will ultimately point to the
4924 * same element. You may recognize this little trick as the
4925 * answer to a stupid interview question -- one that always
4926 * seems to be asked by those who had to have it laboriously
4927 * explained to them, and who can't even concisely describe
4928 * the conditions under which one would be forced to resort to
4929 * this technique. Needless to say, those conditions are
4930 * found here -- and probably only here. Is this the only use
4931 * of this infamous trick in shipping, production code? If it
4932 * isn't, it probably should be...
4933 */
4934 if (minor != -1) {
4935 uintptr_t maddr = dtrace_loadptr(daddr +
4936 offsetof(struct dev_info, devi_minor));
4937
4938 uintptr_t next = offsetof(struct ddi_minor_data, next);
4939 uintptr_t name = offsetof(struct ddi_minor_data,
4940 d_minor) + offsetof(struct ddi_minor, name);
4941 uintptr_t dev = offsetof(struct ddi_minor_data,
4942 d_minor) + offsetof(struct ddi_minor, dev);
4943 uintptr_t scout;
4944
4945 if (maddr != NULL)
4946 scout = dtrace_loadptr(maddr + next);
4947
4948 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4949 uint64_t m;
4950 #ifdef _LP64
4951 m = dtrace_load64(maddr + dev) & MAXMIN64;
4952 #else
4953 m = dtrace_load32(maddr + dev) & MAXMIN;
4954 #endif
4955 if (m != minor) {
4956 maddr = dtrace_loadptr(maddr + next);
4957
4958 if (scout == NULL)
4959 continue;
4960
4961 scout = dtrace_loadptr(scout + next);
4962
4963 if (scout == NULL)
4964 continue;
4965
4966 scout = dtrace_loadptr(scout + next);
4967
4968 if (scout == NULL)
4969 continue;
4970
4971 if (scout == maddr) {
4972 *flags |= CPU_DTRACE_ILLOP;
4973 break;
4974 }
4975
4976 continue;
4977 }
4978
4979 /*
4980 * We have the minor data. Now we need to
4981 * copy the minor's name into the end of the
4982 * pathname.
4983 */
4984 s = (char *)dtrace_loadptr(maddr + name);
4985 len = dtrace_strlen(s, size);
4986
4987 if (*flags & CPU_DTRACE_FAULT)
4988 break;
4989
4990 if (len != 0) {
4991 if ((end -= (len + 1)) < start)
4992 break;
4993
4994 *end = ':';
4995 }
4996
4997 for (i = 1; i <= len; i++)
4998 end[i] = dtrace_load8((uintptr_t)s++);
4999 break;
5000 }
5001 }
5002
5003 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5004 ddi_node_state_t devi_state;
5005
5006 devi_state = dtrace_load32(daddr +
5007 offsetof(struct dev_info, devi_node_state));
5008
5009 if (*flags & CPU_DTRACE_FAULT)
5010 break;
5011
5012 if (devi_state >= DS_INITIALIZED) {
5013 s = (char *)dtrace_loadptr(daddr +
5014 offsetof(struct dev_info, devi_addr));
5015 len = dtrace_strlen(s, size);
5016
5017 if (*flags & CPU_DTRACE_FAULT)
5018 break;
5019
5020 if (len != 0) {
5021 if ((end -= (len + 1)) < start)
5022 break;
5023
5024 *end = '@';
5025 }
5026
5027 for (i = 1; i <= len; i++)
5028 end[i] = dtrace_load8((uintptr_t)s++);
5029 }
5030
5031 /*
5032 * Now for the node name...
5033 */
5034 s = (char *)dtrace_loadptr(daddr +
5035 offsetof(struct dev_info, devi_node_name));
5036
5037 daddr = dtrace_loadptr(daddr +
5038 offsetof(struct dev_info, devi_parent));
5039
5040 /*
5041 * If our parent is NULL (that is, if we're the root
5042 * node), we're going to use the special path
5043 * "devices".
5044 */
5045 if (daddr == NULL)
5046 s = "devices";
5047
5048 len = dtrace_strlen(s, size);
5049 if (*flags & CPU_DTRACE_FAULT)
5050 break;
5051
5052 if ((end -= (len + 1)) < start)
5053 break;
5054
5055 for (i = 1; i <= len; i++)
5056 end[i] = dtrace_load8((uintptr_t)s++);
5057 *end = '/';
5058
5059 if (depth++ > dtrace_devdepth_max) {
5060 *flags |= CPU_DTRACE_ILLOP;
5061 break;
5062 }
5063 }
5064
5065 if (end < start)
5066 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5067
5068 if (daddr == NULL) {
5069 regs[rd] = (uintptr_t)end;
5070 mstate->dtms_scratch_ptr += size;
5071 }
5072
5073 break;
5074 }
5075
5076 case DIF_SUBR_STRJOIN: {
5077 char *d = (char *)mstate->dtms_scratch_ptr;
5078 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5079 uintptr_t s1 = tupregs[0].dttk_value;
5080 uintptr_t s2 = tupregs[1].dttk_value;
5081 int i = 0, j = 0;
5082 size_t lim1, lim2;
5083 char c;
5084
5085 if (!dtrace_strcanload(s1, size, &lim1, mstate, vstate) ||
5086 !dtrace_strcanload(s2, size, &lim2, mstate, vstate)) {
5087 regs[rd] = NULL;
5088 break;
5089 }
5090
5091 if (!DTRACE_INSCRATCH(mstate, size)) {
5092 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5093 regs[rd] = NULL;
5094 break;
5095 }
5096
5097 for (;;) {
5098 if (i >= size) {
5099 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5100 regs[rd] = NULL;
5101 break;
5102 }
5103 c = (i >= lim1) ? '\0' : dtrace_load8(s1++);
5104 if ((d[i++] = c) == '\0') {
5105 i--;
5106 break;
5107 }
5108 }
5109
5110 for (;;) {
5111 if (i >= size) {
5112 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5113 regs[rd] = NULL;
5114 break;
5115 }
5116
5117 c = (j++ >= lim2) ? '\0' : dtrace_load8(s2++);
5118 if ((d[i++] = c) == '\0')
5119 break;
5120 }
5121
5122 if (i < size) {
5123 mstate->dtms_scratch_ptr += i;
5124 regs[rd] = (uintptr_t)d;
5125 }
5126
5127 break;
5128 }
5129
5130 case DIF_SUBR_STRTOLL: {
5131 uintptr_t s = tupregs[0].dttk_value;
5132 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5133 size_t lim;
5134 int base = 10;
5135
5136 if (nargs > 1) {
5137 if ((base = tupregs[1].dttk_value) <= 1 ||
5138 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5139 *flags |= CPU_DTRACE_ILLOP;
5140 break;
5141 }
5142 }
5143
5144 if (!dtrace_strcanload(s, size, &lim, mstate, vstate)) {
5145 regs[rd] = INT64_MIN;
5146 break;
5147 }
5148
5149 regs[rd] = dtrace_strtoll((char *)s, base, lim);
5150 break;
5151 }
5152
5153 case DIF_SUBR_LLTOSTR: {
5154 int64_t i = (int64_t)tupregs[0].dttk_value;
5155 uint64_t val, digit;
5156 uint64_t size = 65; /* enough room for 2^64 in binary */
5157 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
5158 int base = 10;
5159
5160 if (nargs > 1) {
5161 if ((base = tupregs[1].dttk_value) <= 1 ||
5162 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5163 *flags |= CPU_DTRACE_ILLOP;
5164 break;
5165 }
5166 }
5167
5168 val = (base == 10 && i < 0) ? i * -1 : i;
5169
5170 if (!DTRACE_INSCRATCH(mstate, size)) {
5171 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5172 regs[rd] = NULL;
5173 break;
5174 }
5175
5176 for (*end-- = '\0'; val; val /= base) {
5177 if ((digit = val % base) <= '9' - '0') {
5178 *end-- = '0' + digit;
5179 } else {
5180 *end-- = 'a' + (digit - ('9' - '0') - 1);
5181 }
5182 }
5183
5184 if (i == 0 && base == 16)
5185 *end-- = '0';
5186
5187 if (base == 16)
5188 *end-- = 'x';
5189
5190 if (i == 0 || base == 8 || base == 16)
5191 *end-- = '0';
5192
5193 if (i < 0 && base == 10)
5194 *end-- = '-';
5195
5196 regs[rd] = (uintptr_t)end + 1;
5197 mstate->dtms_scratch_ptr += size;
5198 break;
5199 }
5200
5201 case DIF_SUBR_HTONS:
5202 case DIF_SUBR_NTOHS:
5203 #ifdef _BIG_ENDIAN
5204 regs[rd] = (uint16_t)tupregs[0].dttk_value;
5205 #else
5206 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
5207 #endif
5208 break;
5209
5210
5211 case DIF_SUBR_HTONL:
5212 case DIF_SUBR_NTOHL:
5213 #ifdef _BIG_ENDIAN
5214 regs[rd] = (uint32_t)tupregs[0].dttk_value;
5215 #else
5216 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5217 #endif
5218 break;
5219
5220
5221 case DIF_SUBR_HTONLL:
5222 case DIF_SUBR_NTOHLL:
5223 #ifdef _BIG_ENDIAN
5224 regs[rd] = (uint64_t)tupregs[0].dttk_value;
5225 #else
5226 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5227 #endif
5228 break;
5229
5230
5231 case DIF_SUBR_DIRNAME:
5232 case DIF_SUBR_BASENAME: {
5233 char *dest = (char *)mstate->dtms_scratch_ptr;
5234 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5235 uintptr_t src = tupregs[0].dttk_value;
5236 int i, j, len = dtrace_strlen((char *)src, size);
5237 int lastbase = -1, firstbase = -1, lastdir = -1;
5238 int start, end;
5239
5240 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5241 regs[rd] = NULL;
5242 break;
5243 }
5244
5245 if (!DTRACE_INSCRATCH(mstate, size)) {
5246 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5247 regs[rd] = NULL;
5248 break;
5249 }
5250
5251 /*
5252 * The basename and dirname for a zero-length string is
5253 * defined to be "."
5254 */
5255 if (len == 0) {
5256 len = 1;
5257 src = (uintptr_t)".";
5258 }
5259
5260 /*
5261 * Start from the back of the string, moving back toward the
5262 * front until we see a character that isn't a slash. That
5263 * character is the last character in the basename.
5264 */
5265 for (i = len - 1; i >= 0; i--) {
5266 if (dtrace_load8(src + i) != '/')
5267 break;
5268 }
5269
5270 if (i >= 0)
5271 lastbase = i;
5272
5273 /*
5274 * Starting from the last character in the basename, move
5275 * towards the front until we find a slash. The character
5276 * that we processed immediately before that is the first
5277 * character in the basename.
5278 */
5279 for (; i >= 0; i--) {
5280 if (dtrace_load8(src + i) == '/')
5281 break;
5282 }
5283
5284 if (i >= 0)
5285 firstbase = i + 1;
5286
5287 /*
5288 * Now keep going until we find a non-slash character. That
5289 * character is the last character in the dirname.
5290 */
5291 for (; i >= 0; i--) {
5292 if (dtrace_load8(src + i) != '/')
5293 break;
5294 }
5295
5296 if (i >= 0)
5297 lastdir = i;
5298
5299 ASSERT(!(lastbase == -1 && firstbase != -1));
5300 ASSERT(!(firstbase == -1 && lastdir != -1));
5301
5302 if (lastbase == -1) {
5303 /*
5304 * We didn't find a non-slash character. We know that
5305 * the length is non-zero, so the whole string must be
5306 * slashes. In either the dirname or the basename
5307 * case, we return '/'.
5308 */
5309 ASSERT(firstbase == -1);
5310 firstbase = lastbase = lastdir = 0;
5311 }
5312
5313 if (firstbase == -1) {
5314 /*
5315 * The entire string consists only of a basename
5316 * component. If we're looking for dirname, we need
5317 * to change our string to be just "."; if we're
5318 * looking for a basename, we'll just set the first
5319 * character of the basename to be 0.
5320 */
5321 if (subr == DIF_SUBR_DIRNAME) {
5322 ASSERT(lastdir == -1);
5323 src = (uintptr_t)".";
5324 lastdir = 0;
5325 } else {
5326 firstbase = 0;
5327 }
5328 }
5329
5330 if (subr == DIF_SUBR_DIRNAME) {
5331 if (lastdir == -1) {
5332 /*
5333 * We know that we have a slash in the name --
5334 * or lastdir would be set to 0, above. And
5335 * because lastdir is -1, we know that this
5336 * slash must be the first character. (That
5337 * is, the full string must be of the form
5338 * "/basename".) In this case, the last
5339 * character of the directory name is 0.
5340 */
5341 lastdir = 0;
5342 }
5343
5344 start = 0;
5345 end = lastdir;
5346 } else {
5347 ASSERT(subr == DIF_SUBR_BASENAME);
5348 ASSERT(firstbase != -1 && lastbase != -1);
5349 start = firstbase;
5350 end = lastbase;
5351 }
5352
5353 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5354 dest[j] = dtrace_load8(src + i);
5355
5356 dest[j] = '\0';
5357 regs[rd] = (uintptr_t)dest;
5358 mstate->dtms_scratch_ptr += size;
5359 break;
5360 }
5361
5362 case DIF_SUBR_GETF: {
5363 uintptr_t fd = tupregs[0].dttk_value;
5364 uf_info_t *finfo = &curthread->t_procp->p_user.u_finfo;
5365 file_t *fp;
5366
5367 if (!dtrace_priv_proc(state, mstate)) {
5368 regs[rd] = NULL;
5369 break;
5370 }
5371
5372 /*
5373 * This is safe because fi_nfiles only increases, and the
5374 * fi_list array is not freed when the array size doubles.
5375 * (See the comment in flist_grow() for details on the
5376 * management of the u_finfo structure.)
5377 */
5378 fp = fd < finfo->fi_nfiles ? finfo->fi_list[fd].uf_file : NULL;
5379
5380 mstate->dtms_getf = fp;
5381 regs[rd] = (uintptr_t)fp;
5382 break;
5383 }
5384
5385 case DIF_SUBR_CLEANPATH: {
5386 char *dest = (char *)mstate->dtms_scratch_ptr, c;
5387 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5388 uintptr_t src = tupregs[0].dttk_value;
5389 size_t lim;
5390 int i = 0, j = 0;
5391 zone_t *z;
5392
5393 if (!dtrace_strcanload(src, size, &lim, mstate, vstate)) {
5394 regs[rd] = NULL;
5395 break;
5396 }
5397
5398 if (!DTRACE_INSCRATCH(mstate, size)) {
5399 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5400 regs[rd] = NULL;
5401 break;
5402 }
5403
5404 /*
5405 * Move forward, loading each character.
5406 */
5407 do {
5408 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5409 next:
5410 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
5411 break;
5412
5413 if (c != '/') {
5414 dest[j++] = c;
5415 continue;
5416 }
5417
5418 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5419
5420 if (c == '/') {
5421 /*
5422 * We have two slashes -- we can just advance
5423 * to the next character.
5424 */
5425 goto next;
5426 }
5427
5428 if (c != '.') {
5429 /*
5430 * This is not "." and it's not ".." -- we can
5431 * just store the "/" and this character and
5432 * drive on.
5433 */
5434 dest[j++] = '/';
5435 dest[j++] = c;
5436 continue;
5437 }
5438
5439 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5440
5441 if (c == '/') {
5442 /*
5443 * This is a "/./" component. We're not going
5444 * to store anything in the destination buffer;
5445 * we're just going to go to the next component.
5446 */
5447 goto next;
5448 }
5449
5450 if (c != '.') {
5451 /*
5452 * This is not ".." -- we can just store the
5453 * "/." and this character and continue
5454 * processing.
5455 */
5456 dest[j++] = '/';
5457 dest[j++] = '.';
5458 dest[j++] = c;
5459 continue;
5460 }
5461
5462 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5463
5464 if (c != '/' && c != '\0') {
5465 /*
5466 * This is not ".." -- it's "..[mumble]".
5467 * We'll store the "/.." and this character
5468 * and continue processing.
5469 */
5470 dest[j++] = '/';
5471 dest[j++] = '.';
5472 dest[j++] = '.';
5473 dest[j++] = c;
5474 continue;
5475 }
5476
5477 /*
5478 * This is "/../" or "/..\0". We need to back up
5479 * our destination pointer until we find a "/".
5480 */
5481 i--;
5482 while (j != 0 && dest[--j] != '/')
5483 continue;
5484
5485 if (c == '\0')
5486 dest[++j] = '/';
5487 } while (c != '\0');
5488
5489 dest[j] = '\0';
5490
5491 if (mstate->dtms_getf != NULL &&
5492 !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5493 (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5494 /*
5495 * If we've done a getf() as a part of this ECB and we
5496 * don't have kernel access (and we're not in the global
5497 * zone), check if the path we cleaned up begins with
5498 * the zone's root path, and trim it off if so. Note
5499 * that this is an output cleanliness issue, not a
5500 * security issue: knowing one's zone root path does
5501 * not enable privilege escalation.
5502 */
5503 if (strstr(dest, z->zone_rootpath) == dest)
5504 dest += strlen(z->zone_rootpath) - 1;
5505 }
5506
5507 regs[rd] = (uintptr_t)dest;
5508 mstate->dtms_scratch_ptr += size;
5509 break;
5510 }
5511
5512 case DIF_SUBR_INET_NTOA:
5513 case DIF_SUBR_INET_NTOA6:
5514 case DIF_SUBR_INET_NTOP: {
5515 size_t size;
5516 int af, argi, i;
5517 char *base, *end;
5518
5519 if (subr == DIF_SUBR_INET_NTOP) {
5520 af = (int)tupregs[0].dttk_value;
5521 argi = 1;
5522 } else {
5523 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5524 argi = 0;
5525 }
5526
5527 if (af == AF_INET) {
5528 ipaddr_t ip4;
5529 uint8_t *ptr8, val;
5530
5531 if (!dtrace_canload(tupregs[argi].dttk_value,
5532 sizeof (ipaddr_t), mstate, vstate)) {
5533 regs[rd] = NULL;
5534 break;
5535 }
5536
5537 /*
5538 * Safely load the IPv4 address.
5539 */
5540 ip4 = dtrace_load32(tupregs[argi].dttk_value);
5541
5542 /*
5543 * Check an IPv4 string will fit in scratch.
5544 */
5545 size = INET_ADDRSTRLEN;
5546 if (!DTRACE_INSCRATCH(mstate, size)) {
5547 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5548 regs[rd] = NULL;
5549 break;
5550 }
5551 base = (char *)mstate->dtms_scratch_ptr;
5552 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5553
5554 /*
5555 * Stringify as a dotted decimal quad.
5556 */
5557 *end-- = '\0';
5558 ptr8 = (uint8_t *)&ip4;
5559 for (i = 3; i >= 0; i--) {
5560 val = ptr8[i];
5561
5562 if (val == 0) {
5563 *end-- = '0';
5564 } else {
5565 for (; val; val /= 10) {
5566 *end-- = '0' + (val % 10);
5567 }
5568 }
5569
5570 if (i > 0)
5571 *end-- = '.';
5572 }
5573 ASSERT(end + 1 >= base);
5574
5575 } else if (af == AF_INET6) {
5576 struct in6_addr ip6;
5577 int firstzero, tryzero, numzero, v6end;
5578 uint16_t val;
5579 const char digits[] = "0123456789abcdef";
5580
5581 /*
5582 * Stringify using RFC 1884 convention 2 - 16 bit
5583 * hexadecimal values with a zero-run compression.
5584 * Lower case hexadecimal digits are used.
5585 * eg, fe80::214:4fff:fe0b:76c8.
5586 * The IPv4 embedded form is returned for inet_ntop,
5587 * just the IPv4 string is returned for inet_ntoa6.
5588 */
5589
5590 if (!dtrace_canload(tupregs[argi].dttk_value,
5591 sizeof (struct in6_addr), mstate, vstate)) {
5592 regs[rd] = NULL;
5593 break;
5594 }
5595
5596 /*
5597 * Safely load the IPv6 address.
5598 */
5599 dtrace_bcopy(
5600 (void *)(uintptr_t)tupregs[argi].dttk_value,
5601 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5602
5603 /*
5604 * Check an IPv6 string will fit in scratch.
5605 */
5606 size = INET6_ADDRSTRLEN;
5607 if (!DTRACE_INSCRATCH(mstate, size)) {
5608 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5609 regs[rd] = NULL;
5610 break;
5611 }
5612 base = (char *)mstate->dtms_scratch_ptr;
5613 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5614 *end-- = '\0';
5615
5616 /*
5617 * Find the longest run of 16 bit zero values
5618 * for the single allowed zero compression - "::".
5619 */
5620 firstzero = -1;
5621 tryzero = -1;
5622 numzero = 1;
5623 for (i = 0; i < sizeof (struct in6_addr); i++) {
5624 if (ip6._S6_un._S6_u8[i] == 0 &&
5625 tryzero == -1 && i % 2 == 0) {
5626 tryzero = i;
5627 continue;
5628 }
5629
5630 if (tryzero != -1 &&
5631 (ip6._S6_un._S6_u8[i] != 0 ||
5632 i == sizeof (struct in6_addr) - 1)) {
5633
5634 if (i - tryzero <= numzero) {
5635 tryzero = -1;
5636 continue;
5637 }
5638
5639 firstzero = tryzero;
5640 numzero = i - i % 2 - tryzero;
5641 tryzero = -1;
5642
5643 if (ip6._S6_un._S6_u8[i] == 0 &&
5644 i == sizeof (struct in6_addr) - 1)
5645 numzero += 2;
5646 }
5647 }
5648 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5649
5650 /*
5651 * Check for an IPv4 embedded address.
5652 */
5653 v6end = sizeof (struct in6_addr) - 2;
5654 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5655 IN6_IS_ADDR_V4COMPAT(&ip6)) {
5656 for (i = sizeof (struct in6_addr) - 1;
5657 i >= DTRACE_V4MAPPED_OFFSET; i--) {
5658 ASSERT(end >= base);
5659
5660 val = ip6._S6_un._S6_u8[i];
5661
5662 if (val == 0) {
5663 *end-- = '0';
5664 } else {
5665 for (; val; val /= 10) {
5666 *end-- = '0' + val % 10;
5667 }
5668 }
5669
5670 if (i > DTRACE_V4MAPPED_OFFSET)
5671 *end-- = '.';
5672 }
5673
5674 if (subr == DIF_SUBR_INET_NTOA6)
5675 goto inetout;
5676
5677 /*
5678 * Set v6end to skip the IPv4 address that
5679 * we have already stringified.
5680 */
5681 v6end = 10;
5682 }
5683
5684 /*
5685 * Build the IPv6 string by working through the
5686 * address in reverse.
5687 */
5688 for (i = v6end; i >= 0; i -= 2) {
5689 ASSERT(end >= base);
5690
5691 if (i == firstzero + numzero - 2) {
5692 *end-- = ':';
5693 *end-- = ':';
5694 i -= numzero - 2;
5695 continue;
5696 }
5697
5698 if (i < 14 && i != firstzero - 2)
5699 *end-- = ':';
5700
5701 val = (ip6._S6_un._S6_u8[i] << 8) +
5702 ip6._S6_un._S6_u8[i + 1];
5703
5704 if (val == 0) {
5705 *end-- = '0';
5706 } else {
5707 for (; val; val /= 16) {
5708 *end-- = digits[val % 16];
5709 }
5710 }
5711 }
5712 ASSERT(end + 1 >= base);
5713
5714 } else {
5715 /*
5716 * The user didn't use AH_INET or AH_INET6.
5717 */
5718 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5719 regs[rd] = NULL;
5720 break;
5721 }
5722
5723 inetout: regs[rd] = (uintptr_t)end + 1;
5724 mstate->dtms_scratch_ptr += size;
5725 break;
5726 }
5727
5728 }
5729 }
5730
5731 /*
5732 * Emulate the execution of DTrace IR instructions specified by the given
5733 * DIF object. This function is deliberately void of assertions as all of
5734 * the necessary checks are handled by a call to dtrace_difo_validate().
5735 */
5736 static uint64_t
5737 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
5738 dtrace_vstate_t *vstate, dtrace_state_t *state)
5739 {
5740 const dif_instr_t *text = difo->dtdo_buf;
5741 const uint_t textlen = difo->dtdo_len;
5742 const char *strtab = difo->dtdo_strtab;
5743 const uint64_t *inttab = difo->dtdo_inttab;
5744
5745 uint64_t rval = 0;
5746 dtrace_statvar_t *svar;
5747 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5748 dtrace_difv_t *v;
5749 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5750 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
5751
5752 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5753 uint64_t regs[DIF_DIR_NREGS];
5754 uint64_t *tmp;
5755
5756 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5757 int64_t cc_r;
5758 uint_t pc = 0, id, opc;
5759 uint8_t ttop = 0;
5760 dif_instr_t instr;
5761 uint_t r1, r2, rd;
5762
5763 /*
5764 * We stash the current DIF object into the machine state: we need it
5765 * for subsequent access checking.
5766 */
5767 mstate->dtms_difo = difo;
5768
5769 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
5770
5771 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5772 opc = pc;
5773
5774 instr = text[pc++];
5775 r1 = DIF_INSTR_R1(instr);
5776 r2 = DIF_INSTR_R2(instr);
5777 rd = DIF_INSTR_RD(instr);
5778
5779 switch (DIF_INSTR_OP(instr)) {
5780 case DIF_OP_OR:
5781 regs[rd] = regs[r1] | regs[r2];
5782 break;
5783 case DIF_OP_XOR:
5784 regs[rd] = regs[r1] ^ regs[r2];
5785 break;
5786 case DIF_OP_AND:
5787 regs[rd] = regs[r1] & regs[r2];
5788 break;
5789 case DIF_OP_SLL:
5790 regs[rd] = regs[r1] << regs[r2];
5791 break;
5792 case DIF_OP_SRL:
5793 regs[rd] = regs[r1] >> regs[r2];
5794 break;
5795 case DIF_OP_SUB:
5796 regs[rd] = regs[r1] - regs[r2];
5797 break;
5798 case DIF_OP_ADD:
5799 regs[rd] = regs[r1] + regs[r2];
5800 break;
5801 case DIF_OP_MUL:
5802 regs[rd] = regs[r1] * regs[r2];
5803 break;
5804 case DIF_OP_SDIV:
5805 if (regs[r2] == 0) {
5806 regs[rd] = 0;
5807 *flags |= CPU_DTRACE_DIVZERO;
5808 } else {
5809 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5810 regs[rd] = (int64_t)regs[r1] /
5811 (int64_t)regs[r2];
5812 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5813 }
5814 break;
5815
5816 case DIF_OP_UDIV:
5817 if (regs[r2] == 0) {
5818 regs[rd] = 0;
5819 *flags |= CPU_DTRACE_DIVZERO;
5820 } else {
5821 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5822 regs[rd] = regs[r1] / regs[r2];
5823 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5824 }
5825 break;
5826
5827 case DIF_OP_SREM:
5828 if (regs[r2] == 0) {
5829 regs[rd] = 0;
5830 *flags |= CPU_DTRACE_DIVZERO;
5831 } else {
5832 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5833 regs[rd] = (int64_t)regs[r1] %
5834 (int64_t)regs[r2];
5835 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5836 }
5837 break;
5838
5839 case DIF_OP_UREM:
5840 if (regs[r2] == 0) {
5841 regs[rd] = 0;
5842 *flags |= CPU_DTRACE_DIVZERO;
5843 } else {
5844 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5845 regs[rd] = regs[r1] % regs[r2];
5846 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5847 }
5848 break;
5849
5850 case DIF_OP_NOT:
5851 regs[rd] = ~regs[r1];
5852 break;
5853 case DIF_OP_MOV:
5854 regs[rd] = regs[r1];
5855 break;
5856 case DIF_OP_CMP:
5857 cc_r = regs[r1] - regs[r2];
5858 cc_n = cc_r < 0;
5859 cc_z = cc_r == 0;
5860 cc_v = 0;
5861 cc_c = regs[r1] < regs[r2];
5862 break;
5863 case DIF_OP_TST:
5864 cc_n = cc_v = cc_c = 0;
5865 cc_z = regs[r1] == 0;
5866 break;
5867 case DIF_OP_BA:
5868 pc = DIF_INSTR_LABEL(instr);
5869 break;
5870 case DIF_OP_BE:
5871 if (cc_z)
5872 pc = DIF_INSTR_LABEL(instr);
5873 break;
5874 case DIF_OP_BNE:
5875 if (cc_z == 0)
5876 pc = DIF_INSTR_LABEL(instr);
5877 break;
5878 case DIF_OP_BG:
5879 if ((cc_z | (cc_n ^ cc_v)) == 0)
5880 pc = DIF_INSTR_LABEL(instr);
5881 break;
5882 case DIF_OP_BGU:
5883 if ((cc_c | cc_z) == 0)
5884 pc = DIF_INSTR_LABEL(instr);
5885 break;
5886 case DIF_OP_BGE:
5887 if ((cc_n ^ cc_v) == 0)
5888 pc = DIF_INSTR_LABEL(instr);
5889 break;
5890 case DIF_OP_BGEU:
5891 if (cc_c == 0)
5892 pc = DIF_INSTR_LABEL(instr);
5893 break;
5894 case DIF_OP_BL:
5895 if (cc_n ^ cc_v)
5896 pc = DIF_INSTR_LABEL(instr);
5897 break;
5898 case DIF_OP_BLU:
5899 if (cc_c)
5900 pc = DIF_INSTR_LABEL(instr);
5901 break;
5902 case DIF_OP_BLE:
5903 if (cc_z | (cc_n ^ cc_v))
5904 pc = DIF_INSTR_LABEL(instr);
5905 break;
5906 case DIF_OP_BLEU:
5907 if (cc_c | cc_z)
5908 pc = DIF_INSTR_LABEL(instr);
5909 break;
5910 case DIF_OP_RLDSB:
5911 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5912 break;
5913 /*FALLTHROUGH*/
5914 case DIF_OP_LDSB:
5915 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5916 break;
5917 case DIF_OP_RLDSH:
5918 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5919 break;
5920 /*FALLTHROUGH*/
5921 case DIF_OP_LDSH:
5922 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5923 break;
5924 case DIF_OP_RLDSW:
5925 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5926 break;
5927 /*FALLTHROUGH*/
5928 case DIF_OP_LDSW:
5929 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5930 break;
5931 case DIF_OP_RLDUB:
5932 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5933 break;
5934 /*FALLTHROUGH*/
5935 case DIF_OP_LDUB:
5936 regs[rd] = dtrace_load8(regs[r1]);
5937 break;
5938 case DIF_OP_RLDUH:
5939 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5940 break;
5941 /*FALLTHROUGH*/
5942 case DIF_OP_LDUH:
5943 regs[rd] = dtrace_load16(regs[r1]);
5944 break;
5945 case DIF_OP_RLDUW:
5946 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5947 break;
5948 /*FALLTHROUGH*/
5949 case DIF_OP_LDUW:
5950 regs[rd] = dtrace_load32(regs[r1]);
5951 break;
5952 case DIF_OP_RLDX:
5953 if (!dtrace_canload(regs[r1], 8, mstate, vstate))
5954 break;
5955 /*FALLTHROUGH*/
5956 case DIF_OP_LDX:
5957 regs[rd] = dtrace_load64(regs[r1]);
5958 break;
5959 case DIF_OP_ULDSB:
5960 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5961 regs[rd] = (int8_t)
5962 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5963 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5964 break;
5965 case DIF_OP_ULDSH:
5966 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5967 regs[rd] = (int16_t)
5968 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5969 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5970 break;
5971 case DIF_OP_ULDSW:
5972 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5973 regs[rd] = (int32_t)
5974 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5975 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5976 break;
5977 case DIF_OP_ULDUB:
5978 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5979 regs[rd] =
5980 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5981 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5982 break;
5983 case DIF_OP_ULDUH:
5984 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5985 regs[rd] =
5986 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5987 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5988 break;
5989 case DIF_OP_ULDUW:
5990 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5991 regs[rd] =
5992 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5993 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5994 break;
5995 case DIF_OP_ULDX:
5996 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5997 regs[rd] =
5998 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5999 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6000 break;
6001 case DIF_OP_RET:
6002 rval = regs[rd];
6003 pc = textlen;
6004 break;
6005 case DIF_OP_NOP:
6006 break;
6007 case DIF_OP_SETX:
6008 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
6009 break;
6010 case DIF_OP_SETS:
6011 regs[rd] = (uint64_t)(uintptr_t)
6012 (strtab + DIF_INSTR_STRING(instr));
6013 break;
6014 case DIF_OP_SCMP: {
6015 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
6016 uintptr_t s1 = regs[r1];
6017 uintptr_t s2 = regs[r2];
6018 size_t lim1, lim2;
6019
6020 if (s1 != NULL &&
6021 !dtrace_strcanload(s1, sz, &lim1, mstate, vstate))
6022 break;
6023 if (s2 != NULL &&
6024 !dtrace_strcanload(s2, sz, &lim2, mstate, vstate))
6025 break;
6026
6027 cc_r = dtrace_strncmp((char *)s1, (char *)s2,
6028 MIN(lim1, lim2));
6029
6030 cc_n = cc_r < 0;
6031 cc_z = cc_r == 0;
6032 cc_v = cc_c = 0;
6033 break;
6034 }
6035 case DIF_OP_LDGA:
6036 regs[rd] = dtrace_dif_variable(mstate, state,
6037 r1, regs[r2]);
6038 break;
6039 case DIF_OP_LDGS:
6040 id = DIF_INSTR_VAR(instr);
6041
6042 if (id >= DIF_VAR_OTHER_UBASE) {
6043 uintptr_t a;
6044
6045 id -= DIF_VAR_OTHER_UBASE;
6046 svar = vstate->dtvs_globals[id];
6047 ASSERT(svar != NULL);
6048 v = &svar->dtsv_var;
6049
6050 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
6051 regs[rd] = svar->dtsv_data;
6052 break;
6053 }
6054
6055 a = (uintptr_t)svar->dtsv_data;
6056
6057 if (*(uint8_t *)a == UINT8_MAX) {
6058 /*
6059 * If the 0th byte is set to UINT8_MAX
6060 * then this is to be treated as a
6061 * reference to a NULL variable.
6062 */
6063 regs[rd] = NULL;
6064 } else {
6065 regs[rd] = a + sizeof (uint64_t);
6066 }
6067
6068 break;
6069 }
6070
6071 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
6072 break;
6073
6074 case DIF_OP_STGA:
6075 dtrace_dif_variable_write(mstate, state, r1, regs[r2],
6076 regs[rd]);
6077 break;
6078
6079 case DIF_OP_STGS:
6080 id = DIF_INSTR_VAR(instr);
6081
6082 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6083 id -= DIF_VAR_OTHER_UBASE;
6084
6085 VERIFY(id < vstate->dtvs_nglobals);
6086 svar = vstate->dtvs_globals[id];
6087 ASSERT(svar != NULL);
6088 v = &svar->dtsv_var;
6089
6090 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6091 uintptr_t a = (uintptr_t)svar->dtsv_data;
6092 size_t lim;
6093
6094 ASSERT(a != NULL);
6095 ASSERT(svar->dtsv_size != 0);
6096
6097 if (regs[rd] == NULL) {
6098 *(uint8_t *)a = UINT8_MAX;
6099 break;
6100 } else {
6101 *(uint8_t *)a = 0;
6102 a += sizeof (uint64_t);
6103 }
6104 if (!dtrace_vcanload(
6105 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6106 &lim, mstate, vstate))
6107 break;
6108
6109 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6110 (void *)a, &v->dtdv_type, lim);
6111 break;
6112 }
6113
6114 svar->dtsv_data = regs[rd];
6115 break;
6116
6117 case DIF_OP_LDTA:
6118 /*
6119 * There are no DTrace built-in thread-local arrays at
6120 * present. This opcode is saved for future work.
6121 */
6122 *flags |= CPU_DTRACE_ILLOP;
6123 regs[rd] = 0;
6124 break;
6125
6126 case DIF_OP_LDLS:
6127 id = DIF_INSTR_VAR(instr);
6128
6129 if (id < DIF_VAR_OTHER_UBASE) {
6130 /*
6131 * For now, this has no meaning.
6132 */
6133 regs[rd] = 0;
6134 break;
6135 }
6136
6137 id -= DIF_VAR_OTHER_UBASE;
6138
6139 ASSERT(id < vstate->dtvs_nlocals);
6140 ASSERT(vstate->dtvs_locals != NULL);
6141
6142 svar = vstate->dtvs_locals[id];
6143 ASSERT(svar != NULL);
6144 v = &svar->dtsv_var;
6145
6146 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6147 uintptr_t a = (uintptr_t)svar->dtsv_data;
6148 size_t sz = v->dtdv_type.dtdt_size;
6149
6150 sz += sizeof (uint64_t);
6151 ASSERT(svar->dtsv_size == NCPU * sz);
6152 a += CPU->cpu_id * sz;
6153
6154 if (*(uint8_t *)a == UINT8_MAX) {
6155 /*
6156 * If the 0th byte is set to UINT8_MAX
6157 * then this is to be treated as a
6158 * reference to a NULL variable.
6159 */
6160 regs[rd] = NULL;
6161 } else {
6162 regs[rd] = a + sizeof (uint64_t);
6163 }
6164
6165 break;
6166 }
6167
6168 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6169 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6170 regs[rd] = tmp[CPU->cpu_id];
6171 break;
6172
6173 case DIF_OP_STLS:
6174 id = DIF_INSTR_VAR(instr);
6175
6176 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6177 id -= DIF_VAR_OTHER_UBASE;
6178 VERIFY(id < vstate->dtvs_nlocals);
6179
6180 ASSERT(vstate->dtvs_locals != NULL);
6181 svar = vstate->dtvs_locals[id];
6182 ASSERT(svar != NULL);
6183 v = &svar->dtsv_var;
6184
6185 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6186 uintptr_t a = (uintptr_t)svar->dtsv_data;
6187 size_t sz = v->dtdv_type.dtdt_size;
6188 size_t lim;
6189
6190 sz += sizeof (uint64_t);
6191 ASSERT(svar->dtsv_size == NCPU * sz);
6192 a += CPU->cpu_id * sz;
6193
6194 if (regs[rd] == NULL) {
6195 *(uint8_t *)a = UINT8_MAX;
6196 break;
6197 } else {
6198 *(uint8_t *)a = 0;
6199 a += sizeof (uint64_t);
6200 }
6201
6202 if (!dtrace_vcanload(
6203 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6204 &lim, mstate, vstate))
6205 break;
6206
6207 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6208 (void *)a, &v->dtdv_type, lim);
6209 break;
6210 }
6211
6212 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6213 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6214 tmp[CPU->cpu_id] = regs[rd];
6215 break;
6216
6217 case DIF_OP_LDTS: {
6218 dtrace_dynvar_t *dvar;
6219 dtrace_key_t *key;
6220
6221 id = DIF_INSTR_VAR(instr);
6222 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6223 id -= DIF_VAR_OTHER_UBASE;
6224 v = &vstate->dtvs_tlocals[id];
6225
6226 key = &tupregs[DIF_DTR_NREGS];
6227 key[0].dttk_value = (uint64_t)id;
6228 key[0].dttk_size = 0;
6229 DTRACE_TLS_THRKEY(key[1].dttk_value);
6230 key[1].dttk_size = 0;
6231
6232 dvar = dtrace_dynvar(dstate, 2, key,
6233 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
6234 mstate, vstate);
6235
6236 if (dvar == NULL) {
6237 regs[rd] = 0;
6238 break;
6239 }
6240
6241 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6242 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6243 } else {
6244 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6245 }
6246
6247 break;
6248 }
6249
6250 case DIF_OP_STTS: {
6251 dtrace_dynvar_t *dvar;
6252 dtrace_key_t *key;
6253
6254 id = DIF_INSTR_VAR(instr);
6255 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6256 id -= DIF_VAR_OTHER_UBASE;
6257 VERIFY(id < vstate->dtvs_ntlocals);
6258
6259 key = &tupregs[DIF_DTR_NREGS];
6260 key[0].dttk_value = (uint64_t)id;
6261 key[0].dttk_size = 0;
6262 DTRACE_TLS_THRKEY(key[1].dttk_value);
6263 key[1].dttk_size = 0;
6264 v = &vstate->dtvs_tlocals[id];
6265
6266 dvar = dtrace_dynvar(dstate, 2, key,
6267 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6268 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6269 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6270 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6271
6272 /*
6273 * Given that we're storing to thread-local data,
6274 * we need to flush our predicate cache.
6275 */
6276 curthread->t_predcache = NULL;
6277
6278 if (dvar == NULL)
6279 break;
6280
6281 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6282 size_t lim;
6283
6284 if (!dtrace_vcanload(
6285 (void *)(uintptr_t)regs[rd],
6286 &v->dtdv_type, &lim, mstate, vstate))
6287 break;
6288
6289 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6290 dvar->dtdv_data, &v->dtdv_type, lim);
6291 } else {
6292 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6293 }
6294
6295 break;
6296 }
6297
6298 case DIF_OP_SRA:
6299 regs[rd] = (int64_t)regs[r1] >> regs[r2];
6300 break;
6301
6302 case DIF_OP_CALL:
6303 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6304 regs, tupregs, ttop, mstate, state);
6305 break;
6306
6307 case DIF_OP_PUSHTR:
6308 if (ttop == DIF_DTR_NREGS) {
6309 *flags |= CPU_DTRACE_TUPOFLOW;
6310 break;
6311 }
6312
6313 if (r1 == DIF_TYPE_STRING) {
6314 /*
6315 * If this is a string type and the size is 0,
6316 * we'll use the system-wide default string
6317 * size. Note that we are _not_ looking at
6318 * the value of the DTRACEOPT_STRSIZE option;
6319 * had this been set, we would expect to have
6320 * a non-zero size value in the "pushtr".
6321 */
6322 tupregs[ttop].dttk_size =
6323 dtrace_strlen((char *)(uintptr_t)regs[rd],
6324 regs[r2] ? regs[r2] :
6325 dtrace_strsize_default) + 1;
6326 } else {
6327 if (regs[r2] > LONG_MAX) {
6328 *flags |= CPU_DTRACE_ILLOP;
6329 break;
6330 }
6331
6332 tupregs[ttop].dttk_size = regs[r2];
6333 }
6334
6335 tupregs[ttop++].dttk_value = regs[rd];
6336 break;
6337
6338 case DIF_OP_PUSHTV:
6339 if (ttop == DIF_DTR_NREGS) {
6340 *flags |= CPU_DTRACE_TUPOFLOW;
6341 break;
6342 }
6343
6344 tupregs[ttop].dttk_value = regs[rd];
6345 tupregs[ttop++].dttk_size = 0;
6346 break;
6347
6348 case DIF_OP_POPTS:
6349 if (ttop != 0)
6350 ttop--;
6351 break;
6352
6353 case DIF_OP_FLUSHTS:
6354 ttop = 0;
6355 break;
6356
6357 case DIF_OP_LDGAA:
6358 case DIF_OP_LDTAA: {
6359 dtrace_dynvar_t *dvar;
6360 dtrace_key_t *key = tupregs;
6361 uint_t nkeys = ttop;
6362
6363 id = DIF_INSTR_VAR(instr);
6364 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6365 id -= DIF_VAR_OTHER_UBASE;
6366
6367 key[nkeys].dttk_value = (uint64_t)id;
6368 key[nkeys++].dttk_size = 0;
6369
6370 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6371 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6372 key[nkeys++].dttk_size = 0;
6373 VERIFY(id < vstate->dtvs_ntlocals);
6374 v = &vstate->dtvs_tlocals[id];
6375 } else {
6376 VERIFY(id < vstate->dtvs_nglobals);
6377 v = &vstate->dtvs_globals[id]->dtsv_var;
6378 }
6379
6380 dvar = dtrace_dynvar(dstate, nkeys, key,
6381 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6382 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6383 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6384
6385 if (dvar == NULL) {
6386 regs[rd] = 0;
6387 break;
6388 }
6389
6390 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6391 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6392 } else {
6393 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6394 }
6395
6396 break;
6397 }
6398
6399 case DIF_OP_STGAA:
6400 case DIF_OP_STTAA: {
6401 dtrace_dynvar_t *dvar;
6402 dtrace_key_t *key = tupregs;
6403 uint_t nkeys = ttop;
6404
6405 id = DIF_INSTR_VAR(instr);
6406 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6407 id -= DIF_VAR_OTHER_UBASE;
6408
6409 key[nkeys].dttk_value = (uint64_t)id;
6410 key[nkeys++].dttk_size = 0;
6411
6412 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6413 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6414 key[nkeys++].dttk_size = 0;
6415 VERIFY(id < vstate->dtvs_ntlocals);
6416 v = &vstate->dtvs_tlocals[id];
6417 } else {
6418 VERIFY(id < vstate->dtvs_nglobals);
6419 v = &vstate->dtvs_globals[id]->dtsv_var;
6420 }
6421
6422 dvar = dtrace_dynvar(dstate, nkeys, key,
6423 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6424 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6425 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6426 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6427
6428 if (dvar == NULL)
6429 break;
6430
6431 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6432 size_t lim;
6433
6434 if (!dtrace_vcanload(
6435 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6436 &lim, mstate, vstate))
6437 break;
6438
6439 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6440 dvar->dtdv_data, &v->dtdv_type, lim);
6441 } else {
6442 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6443 }
6444
6445 break;
6446 }
6447
6448 case DIF_OP_ALLOCS: {
6449 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6450 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6451
6452 /*
6453 * Rounding up the user allocation size could have
6454 * overflowed large, bogus allocations (like -1ULL) to
6455 * 0.
6456 */
6457 if (size < regs[r1] ||
6458 !DTRACE_INSCRATCH(mstate, size)) {
6459 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6460 regs[rd] = NULL;
6461 break;
6462 }
6463
6464 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6465 mstate->dtms_scratch_ptr += size;
6466 regs[rd] = ptr;
6467 break;
6468 }
6469
6470 case DIF_OP_COPYS:
6471 if (!dtrace_canstore(regs[rd], regs[r2],
6472 mstate, vstate)) {
6473 *flags |= CPU_DTRACE_BADADDR;
6474 *illval = regs[rd];
6475 break;
6476 }
6477
6478 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6479 break;
6480
6481 dtrace_bcopy((void *)(uintptr_t)regs[r1],
6482 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6483 break;
6484
6485 case DIF_OP_STB:
6486 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6487 *flags |= CPU_DTRACE_BADADDR;
6488 *illval = regs[rd];
6489 break;
6490 }
6491 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6492 break;
6493
6494 case DIF_OP_STH:
6495 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6496 *flags |= CPU_DTRACE_BADADDR;
6497 *illval = regs[rd];
6498 break;
6499 }
6500 if (regs[rd] & 1) {
6501 *flags |= CPU_DTRACE_BADALIGN;
6502 *illval = regs[rd];
6503 break;
6504 }
6505 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6506 break;
6507
6508 case DIF_OP_STW:
6509 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6510 *flags |= CPU_DTRACE_BADADDR;
6511 *illval = regs[rd];
6512 break;
6513 }
6514 if (regs[rd] & 3) {
6515 *flags |= CPU_DTRACE_BADALIGN;
6516 *illval = regs[rd];
6517 break;
6518 }
6519 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6520 break;
6521
6522 case DIF_OP_STX:
6523 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6524 *flags |= CPU_DTRACE_BADADDR;
6525 *illval = regs[rd];
6526 break;
6527 }
6528 if (regs[rd] & 7) {
6529 *flags |= CPU_DTRACE_BADALIGN;
6530 *illval = regs[rd];
6531 break;
6532 }
6533 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6534 break;
6535 }
6536 }
6537
6538 if (!(*flags & CPU_DTRACE_FAULT))
6539 return (rval);
6540
6541 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6542 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6543
6544 return (0);
6545 }
6546
6547 static void
6548 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6549 {
6550 dtrace_probe_t *probe = ecb->dte_probe;
6551 dtrace_provider_t *prov = probe->dtpr_provider;
6552 char c[DTRACE_FULLNAMELEN + 80], *str;
6553 char *msg = "dtrace: breakpoint action at probe ";
6554 char *ecbmsg = " (ecb ";
6555 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6556 uintptr_t val = (uintptr_t)ecb;
6557 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6558
6559 if (dtrace_destructive_disallow)
6560 return;
6561
6562 /*
6563 * It's impossible to be taking action on the NULL probe.
6564 */
6565 ASSERT(probe != NULL);
6566
6567 /*
6568 * This is a poor man's (destitute man's?) sprintf(): we want to
6569 * print the provider name, module name, function name and name of
6570 * the probe, along with the hex address of the ECB with the breakpoint
6571 * action -- all of which we must place in the character buffer by
6572 * hand.
6573 */
6574 while (*msg != '\0')
6575 c[i++] = *msg++;
6576
6577 for (str = prov->dtpv_name; *str != '\0'; str++)
6578 c[i++] = *str;
6579 c[i++] = ':';
6580
6581 for (str = probe->dtpr_mod; *str != '\0'; str++)
6582 c[i++] = *str;
6583 c[i++] = ':';
6584
6585 for (str = probe->dtpr_func; *str != '\0'; str++)
6586 c[i++] = *str;
6587 c[i++] = ':';
6588
6589 for (str = probe->dtpr_name; *str != '\0'; str++)
6590 c[i++] = *str;
6591
6592 while (*ecbmsg != '\0')
6593 c[i++] = *ecbmsg++;
6594
6595 while (shift >= 0) {
6596 mask = (uintptr_t)0xf << shift;
6597
6598 if (val >= ((uintptr_t)1 << shift))
6599 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6600 shift -= 4;
6601 }
6602
6603 c[i++] = ')';
6604 c[i] = '\0';
6605
6606 debug_enter(c);
6607 }
6608
6609 static void
6610 dtrace_action_panic(dtrace_ecb_t *ecb)
6611 {
6612 dtrace_probe_t *probe = ecb->dte_probe;
6613
6614 /*
6615 * It's impossible to be taking action on the NULL probe.
6616 */
6617 ASSERT(probe != NULL);
6618
6619 if (dtrace_destructive_disallow)
6620 return;
6621
6622 if (dtrace_panicked != NULL)
6623 return;
6624
6625 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6626 return;
6627
6628 /*
6629 * We won the right to panic. (We want to be sure that only one
6630 * thread calls panic() from dtrace_probe(), and that panic() is
6631 * called exactly once.)
6632 */
6633 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6634 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6635 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6636 }
6637
6638 static void
6639 dtrace_action_raise(uint64_t sig)
6640 {
6641 if (dtrace_destructive_disallow)
6642 return;
6643
6644 if (sig >= NSIG) {
6645 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6646 return;
6647 }
6648
6649 /*
6650 * raise() has a queue depth of 1 -- we ignore all subsequent
6651 * invocations of the raise() action.
6652 */
6653 if (curthread->t_dtrace_sig == 0)
6654 curthread->t_dtrace_sig = (uint8_t)sig;
6655
6656 curthread->t_sig_check = 1;
6657 aston(curthread);
6658 }
6659
6660 static void
6661 dtrace_action_stop(void)
6662 {
6663 if (dtrace_destructive_disallow)
6664 return;
6665
6666 if (!curthread->t_dtrace_stop) {
6667 curthread->t_dtrace_stop = 1;
6668 curthread->t_sig_check = 1;
6669 aston(curthread);
6670 }
6671 }
6672
6673 static void
6674 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
6675 {
6676 hrtime_t now;
6677 volatile uint16_t *flags;
6678 cpu_t *cpu = CPU;
6679
6680 if (dtrace_destructive_disallow)
6681 return;
6682
6683 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
6684
6685 now = dtrace_gethrtime();
6686
6687 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
6688 /*
6689 * We need to advance the mark to the current time.
6690 */
6691 cpu->cpu_dtrace_chillmark = now;
6692 cpu->cpu_dtrace_chilled = 0;
6693 }
6694
6695 /*
6696 * Now check to see if the requested chill time would take us over
6697 * the maximum amount of time allowed in the chill interval. (Or
6698 * worse, if the calculation itself induces overflow.)
6699 */
6700 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
6701 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
6702 *flags |= CPU_DTRACE_ILLOP;
6703 return;
6704 }
6705
6706 while (dtrace_gethrtime() - now < val)
6707 continue;
6708
6709 /*
6710 * Normally, we assure that the value of the variable "timestamp" does
6711 * not change within an ECB. The presence of chill() represents an
6712 * exception to this rule, however.
6713 */
6714 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
6715 cpu->cpu_dtrace_chilled += val;
6716 }
6717
6718 static void
6719 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
6720 uint64_t *buf, uint64_t arg)
6721 {
6722 int nframes = DTRACE_USTACK_NFRAMES(arg);
6723 int strsize = DTRACE_USTACK_STRSIZE(arg);
6724 uint64_t *pcs = &buf[1], *fps;
6725 char *str = (char *)&pcs[nframes];
6726 int size, offs = 0, i, j;
6727 size_t rem;
6728 uintptr_t old = mstate->dtms_scratch_ptr, saved;
6729 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6730 char *sym;
6731
6732 /*
6733 * Should be taking a faster path if string space has not been
6734 * allocated.
6735 */
6736 ASSERT(strsize != 0);
6737
6738 /*
6739 * We will first allocate some temporary space for the frame pointers.
6740 */
6741 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6742 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
6743 (nframes * sizeof (uint64_t));
6744
6745 if (!DTRACE_INSCRATCH(mstate, size)) {
6746 /*
6747 * Not enough room for our frame pointers -- need to indicate
6748 * that we ran out of scratch space.
6749 */
6750 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6751 return;
6752 }
6753
6754 mstate->dtms_scratch_ptr += size;
6755 saved = mstate->dtms_scratch_ptr;
6756
6757 /*
6758 * Now get a stack with both program counters and frame pointers.
6759 */
6760 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6761 dtrace_getufpstack(buf, fps, nframes + 1);
6762 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6763
6764 /*
6765 * If that faulted, we're cooked.
6766 */
6767 if (*flags & CPU_DTRACE_FAULT)
6768 goto out;
6769
6770 /*
6771 * Now we want to walk up the stack, calling the USTACK helper. For
6772 * each iteration, we restore the scratch pointer.
6773 */
6774 for (i = 0; i < nframes; i++) {
6775 mstate->dtms_scratch_ptr = saved;
6776
6777 if (offs >= strsize)
6778 break;
6779
6780 sym = (char *)(uintptr_t)dtrace_helper(
6781 DTRACE_HELPER_ACTION_USTACK,
6782 mstate, state, pcs[i], fps[i]);
6783
6784 /*
6785 * If we faulted while running the helper, we're going to
6786 * clear the fault and null out the corresponding string.
6787 */
6788 if (*flags & CPU_DTRACE_FAULT) {
6789 *flags &= ~CPU_DTRACE_FAULT;
6790 str[offs++] = '\0';
6791 continue;
6792 }
6793
6794 if (sym == NULL) {
6795 str[offs++] = '\0';
6796 continue;
6797 }
6798
6799 if (!dtrace_strcanload((uintptr_t)sym, strsize, &rem, mstate,
6800 &(state->dts_vstate))) {
6801 str[offs++] = '\0';
6802 continue;
6803 }
6804
6805 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6806
6807 /*
6808 * Now copy in the string that the helper returned to us.
6809 */
6810 for (j = 0; offs + j < strsize && j < rem; j++) {
6811 if ((str[offs + j] = sym[j]) == '\0')
6812 break;
6813 }
6814
6815 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6816
6817 offs += j + 1;
6818 }
6819
6820 if (offs >= strsize) {
6821 /*
6822 * If we didn't have room for all of the strings, we don't
6823 * abort processing -- this needn't be a fatal error -- but we
6824 * still want to increment a counter (dts_stkstroverflows) to
6825 * allow this condition to be warned about. (If this is from
6826 * a jstack() action, it is easily tuned via jstackstrsize.)
6827 */
6828 dtrace_error(&state->dts_stkstroverflows);
6829 }
6830
6831 while (offs < strsize)
6832 str[offs++] = '\0';
6833
6834 out:
6835 mstate->dtms_scratch_ptr = old;
6836 }
6837
6838 static void
6839 dtrace_store_by_ref(dtrace_difo_t *dp, caddr_t tomax, size_t size,
6840 size_t *valoffsp, uint64_t *valp, uint64_t end, int intuple, int dtkind)
6841 {
6842 volatile uint16_t *flags;
6843 uint64_t val = *valp;
6844 size_t valoffs = *valoffsp;
6845
6846 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6847 ASSERT(dtkind == DIF_TF_BYREF || dtkind == DIF_TF_BYUREF);
6848
6849 /*
6850 * If this is a string, we're going to only load until we find the zero
6851 * byte -- after which we'll store zero bytes.
6852 */
6853 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
6854 char c = '\0' + 1;
6855 size_t s;
6856
6857 for (s = 0; s < size; s++) {
6858 if (c != '\0' && dtkind == DIF_TF_BYREF) {
6859 c = dtrace_load8(val++);
6860 } else if (c != '\0' && dtkind == DIF_TF_BYUREF) {
6861 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6862 c = dtrace_fuword8((void *)(uintptr_t)val++);
6863 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6864 if (*flags & CPU_DTRACE_FAULT)
6865 break;
6866 }
6867
6868 DTRACE_STORE(uint8_t, tomax, valoffs++, c);
6869
6870 if (c == '\0' && intuple)
6871 break;
6872 }
6873 } else {
6874 uint8_t c;
6875 while (valoffs < end) {
6876 if (dtkind == DIF_TF_BYREF) {
6877 c = dtrace_load8(val++);
6878 } else if (dtkind == DIF_TF_BYUREF) {
6879 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6880 c = dtrace_fuword8((void *)(uintptr_t)val++);
6881 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6882 if (*flags & CPU_DTRACE_FAULT)
6883 break;
6884 }
6885
6886 DTRACE_STORE(uint8_t, tomax,
6887 valoffs++, c);
6888 }
6889 }
6890
6891 *valp = val;
6892 *valoffsp = valoffs;
6893 }
6894
6895 /*
6896 * If you're looking for the epicenter of DTrace, you just found it. This
6897 * is the function called by the provider to fire a probe -- from which all
6898 * subsequent probe-context DTrace activity emanates.
6899 */
6900 void
6901 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6902 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6903 {
6904 processorid_t cpuid;
6905 dtrace_icookie_t cookie;
6906 dtrace_probe_t *probe;
6907 dtrace_mstate_t mstate;
6908 dtrace_ecb_t *ecb;
6909 dtrace_action_t *act;
6910 intptr_t offs;
6911 size_t size;
6912 int vtime, onintr;
6913 volatile uint16_t *flags;
6914 hrtime_t now, end;
6915
6916 /*
6917 * Kick out immediately if this CPU is still being born (in which case
6918 * curthread will be set to -1) or the current thread can't allow
6919 * probes in its current context.
6920 */
6921 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6922 return;
6923
6924 cookie = dtrace_interrupt_disable();
6925 probe = dtrace_probes[id - 1];
6926 cpuid = CPU->cpu_id;
6927 onintr = CPU_ON_INTR(CPU);
6928
6929 CPU->cpu_dtrace_probes++;
6930
6931 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6932 probe->dtpr_predcache == curthread->t_predcache) {
6933 /*
6934 * We have hit in the predicate cache; we know that
6935 * this predicate would evaluate to be false.
6936 */
6937 dtrace_interrupt_enable(cookie);
6938 return;
6939 }
6940
6941 if (panic_quiesce) {
6942 /*
6943 * We don't trace anything if we're panicking.
6944 */
6945 dtrace_interrupt_enable(cookie);
6946 return;
6947 }
6948
6949 now = mstate.dtms_timestamp = dtrace_gethrtime();
6950 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6951 vtime = dtrace_vtime_references != 0;
6952
6953 if (vtime && curthread->t_dtrace_start)
6954 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6955
6956 mstate.dtms_difo = NULL;
6957 mstate.dtms_probe = probe;
6958 mstate.dtms_strtok = NULL;
6959 mstate.dtms_arg[0] = arg0;
6960 mstate.dtms_arg[1] = arg1;
6961 mstate.dtms_arg[2] = arg2;
6962 mstate.dtms_arg[3] = arg3;
6963 mstate.dtms_arg[4] = arg4;
6964
6965 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6966
6967 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6968 dtrace_predicate_t *pred = ecb->dte_predicate;
6969 dtrace_state_t *state = ecb->dte_state;
6970 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6971 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6972 dtrace_vstate_t *vstate = &state->dts_vstate;
6973 dtrace_provider_t *prov = probe->dtpr_provider;
6974 uint64_t tracememsize = 0;
6975 int committed = 0;
6976 caddr_t tomax;
6977
6978 /*
6979 * A little subtlety with the following (seemingly innocuous)
6980 * declaration of the automatic 'val': by looking at the
6981 * code, you might think that it could be declared in the
6982 * action processing loop, below. (That is, it's only used in
6983 * the action processing loop.) However, it must be declared
6984 * out of that scope because in the case of DIF expression
6985 * arguments to aggregating actions, one iteration of the
6986 * action loop will use the last iteration's value.
6987 */
6988 #ifdef lint
6989 uint64_t val = 0;
6990 #else
6991 uint64_t val;
6992 #endif
6993
6994 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6995 mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
6996 mstate.dtms_getf = NULL;
6997
6998 *flags &= ~CPU_DTRACE_ERROR;
6999
7000 if (prov == dtrace_provider) {
7001 /*
7002 * If dtrace itself is the provider of this probe,
7003 * we're only going to continue processing the ECB if
7004 * arg0 (the dtrace_state_t) is equal to the ECB's
7005 * creating state. (This prevents disjoint consumers
7006 * from seeing one another's metaprobes.)
7007 */
7008 if (arg0 != (uint64_t)(uintptr_t)state)
7009 continue;
7010 }
7011
7012 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
7013 /*
7014 * We're not currently active. If our provider isn't
7015 * the dtrace pseudo provider, we're not interested.
7016 */
7017 if (prov != dtrace_provider)
7018 continue;
7019
7020 /*
7021 * Now we must further check if we are in the BEGIN
7022 * probe. If we are, we will only continue processing
7023 * if we're still in WARMUP -- if one BEGIN enabling
7024 * has invoked the exit() action, we don't want to
7025 * evaluate subsequent BEGIN enablings.
7026 */
7027 if (probe->dtpr_id == dtrace_probeid_begin &&
7028 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
7029 ASSERT(state->dts_activity ==
7030 DTRACE_ACTIVITY_DRAINING);
7031 continue;
7032 }
7033 }
7034
7035 if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
7036 continue;
7037
7038 if (now - state->dts_alive > dtrace_deadman_timeout) {
7039 /*
7040 * We seem to be dead. Unless we (a) have kernel
7041 * destructive permissions (b) have explicitly enabled
7042 * destructive actions and (c) destructive actions have
7043 * not been disabled, we're going to transition into
7044 * the KILLED state, from which no further processing
7045 * on this state will be performed.
7046 */
7047 if (!dtrace_priv_kernel_destructive(state) ||
7048 !state->dts_cred.dcr_destructive ||
7049 dtrace_destructive_disallow) {
7050 void *activity = &state->dts_activity;
7051 dtrace_activity_t current;
7052
7053 do {
7054 current = state->dts_activity;
7055 } while (dtrace_cas32(activity, current,
7056 DTRACE_ACTIVITY_KILLED) != current);
7057
7058 continue;
7059 }
7060 }
7061
7062 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
7063 ecb->dte_alignment, state, &mstate)) < 0)
7064 continue;
7065
7066 tomax = buf->dtb_tomax;
7067 ASSERT(tomax != NULL);
7068
7069 if (ecb->dte_size != 0) {
7070 dtrace_rechdr_t dtrh;
7071 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
7072 mstate.dtms_timestamp = dtrace_gethrtime();
7073 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7074 }
7075 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
7076 dtrh.dtrh_epid = ecb->dte_epid;
7077 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
7078 mstate.dtms_timestamp);
7079 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
7080 }
7081
7082 mstate.dtms_epid = ecb->dte_epid;
7083 mstate.dtms_present |= DTRACE_MSTATE_EPID;
7084
7085 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
7086 mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
7087
7088 if (pred != NULL) {
7089 dtrace_difo_t *dp = pred->dtp_difo;
7090 int rval;
7091
7092 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
7093
7094 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
7095 dtrace_cacheid_t cid = probe->dtpr_predcache;
7096
7097 if (cid != DTRACE_CACHEIDNONE && !onintr) {
7098 /*
7099 * Update the predicate cache...
7100 */
7101 ASSERT(cid == pred->dtp_cacheid);
7102 curthread->t_predcache = cid;
7103 }
7104
7105 continue;
7106 }
7107 }
7108
7109 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
7110 act != NULL; act = act->dta_next) {
7111 size_t valoffs;
7112 dtrace_difo_t *dp;
7113 dtrace_recdesc_t *rec = &act->dta_rec;
7114
7115 size = rec->dtrd_size;
7116 valoffs = offs + rec->dtrd_offset;
7117
7118 if (DTRACEACT_ISAGG(act->dta_kind)) {
7119 uint64_t v = 0xbad;
7120 dtrace_aggregation_t *agg;
7121
7122 agg = (dtrace_aggregation_t *)act;
7123
7124 if ((dp = act->dta_difo) != NULL)
7125 v = dtrace_dif_emulate(dp,
7126 &mstate, vstate, state);
7127
7128 if (*flags & CPU_DTRACE_ERROR)
7129 continue;
7130
7131 /*
7132 * Note that we always pass the expression
7133 * value from the previous iteration of the
7134 * action loop. This value will only be used
7135 * if there is an expression argument to the
7136 * aggregating action, denoted by the
7137 * dtag_hasarg field.
7138 */
7139 dtrace_aggregate(agg, buf,
7140 offs, aggbuf, v, val);
7141 continue;
7142 }
7143
7144 switch (act->dta_kind) {
7145 case DTRACEACT_STOP:
7146 if (dtrace_priv_proc_destructive(state,
7147 &mstate))
7148 dtrace_action_stop();
7149 continue;
7150
7151 case DTRACEACT_BREAKPOINT:
7152 if (dtrace_priv_kernel_destructive(state))
7153 dtrace_action_breakpoint(ecb);
7154 continue;
7155
7156 case DTRACEACT_PANIC:
7157 if (dtrace_priv_kernel_destructive(state))
7158 dtrace_action_panic(ecb);
7159 continue;
7160
7161 case DTRACEACT_STACK:
7162 if (!dtrace_priv_kernel(state))
7163 continue;
7164
7165 dtrace_getpcstack((pc_t *)(tomax + valoffs),
7166 size / sizeof (pc_t), probe->dtpr_aframes,
7167 DTRACE_ANCHORED(probe) ? NULL :
7168 (uint32_t *)arg0);
7169
7170 continue;
7171
7172 case DTRACEACT_JSTACK:
7173 case DTRACEACT_USTACK:
7174 if (!dtrace_priv_proc(state, &mstate))
7175 continue;
7176
7177 /*
7178 * See comment in DIF_VAR_PID.
7179 */
7180 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
7181 CPU_ON_INTR(CPU)) {
7182 int depth = DTRACE_USTACK_NFRAMES(
7183 rec->dtrd_arg) + 1;
7184
7185 dtrace_bzero((void *)(tomax + valoffs),
7186 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
7187 + depth * sizeof (uint64_t));
7188
7189 continue;
7190 }
7191
7192 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
7193 curproc->p_dtrace_helpers != NULL) {
7194 /*
7195 * This is the slow path -- we have
7196 * allocated string space, and we're
7197 * getting the stack of a process that
7198 * has helpers. Call into a separate
7199 * routine to perform this processing.
7200 */
7201 dtrace_action_ustack(&mstate, state,
7202 (uint64_t *)(tomax + valoffs),
7203 rec->dtrd_arg);
7204 continue;
7205 }
7206
7207 /*
7208 * Clear the string space, since there's no
7209 * helper to do it for us.
7210 */
7211 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
7212 int depth = DTRACE_USTACK_NFRAMES(
7213 rec->dtrd_arg);
7214 size_t strsize = DTRACE_USTACK_STRSIZE(
7215 rec->dtrd_arg);
7216 uint64_t *buf = (uint64_t *)(tomax +
7217 valoffs);
7218 void *strspace = &buf[depth + 1];
7219
7220 dtrace_bzero(strspace,
7221 MIN(depth, strsize));
7222 }
7223
7224 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7225 dtrace_getupcstack((uint64_t *)
7226 (tomax + valoffs),
7227 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
7228 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7229 continue;
7230
7231 default:
7232 break;
7233 }
7234
7235 dp = act->dta_difo;
7236 ASSERT(dp != NULL);
7237
7238 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
7239
7240 if (*flags & CPU_DTRACE_ERROR)
7241 continue;
7242
7243 switch (act->dta_kind) {
7244 case DTRACEACT_SPECULATE: {
7245 dtrace_rechdr_t *dtrh;
7246
7247 ASSERT(buf == &state->dts_buffer[cpuid]);
7248 buf = dtrace_speculation_buffer(state,
7249 cpuid, val);
7250
7251 if (buf == NULL) {
7252 *flags |= CPU_DTRACE_DROP;
7253 continue;
7254 }
7255
7256 offs = dtrace_buffer_reserve(buf,
7257 ecb->dte_needed, ecb->dte_alignment,
7258 state, NULL);
7259
7260 if (offs < 0) {
7261 *flags |= CPU_DTRACE_DROP;
7262 continue;
7263 }
7264
7265 tomax = buf->dtb_tomax;
7266 ASSERT(tomax != NULL);
7267
7268 if (ecb->dte_size == 0)
7269 continue;
7270
7271 ASSERT3U(ecb->dte_size, >=,
7272 sizeof (dtrace_rechdr_t));
7273 dtrh = ((void *)(tomax + offs));
7274 dtrh->dtrh_epid = ecb->dte_epid;
7275 /*
7276 * When the speculation is committed, all of
7277 * the records in the speculative buffer will
7278 * have their timestamps set to the commit
7279 * time. Until then, it is set to a sentinel
7280 * value, for debugability.
7281 */
7282 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
7283 continue;
7284 }
7285
7286 case DTRACEACT_CHILL:
7287 if (dtrace_priv_kernel_destructive(state))
7288 dtrace_action_chill(&mstate, val);
7289 continue;
7290
7291 case DTRACEACT_RAISE:
7292 if (dtrace_priv_proc_destructive(state,
7293 &mstate))
7294 dtrace_action_raise(val);
7295 continue;
7296
7297 case DTRACEACT_COMMIT:
7298 ASSERT(!committed);
7299
7300 /*
7301 * We need to commit our buffer state.
7302 */
7303 if (ecb->dte_size)
7304 buf->dtb_offset = offs + ecb->dte_size;
7305 buf = &state->dts_buffer[cpuid];
7306 dtrace_speculation_commit(state, cpuid, val);
7307 committed = 1;
7308 continue;
7309
7310 case DTRACEACT_DISCARD:
7311 dtrace_speculation_discard(state, cpuid, val);
7312 continue;
7313
7314 case DTRACEACT_DIFEXPR:
7315 case DTRACEACT_LIBACT:
7316 case DTRACEACT_PRINTF:
7317 case DTRACEACT_PRINTA:
7318 case DTRACEACT_SYSTEM:
7319 case DTRACEACT_FREOPEN:
7320 case DTRACEACT_TRACEMEM:
7321 break;
7322
7323 case DTRACEACT_TRACEMEM_DYNSIZE:
7324 tracememsize = val;
7325 break;
7326
7327 case DTRACEACT_SYM:
7328 case DTRACEACT_MOD:
7329 if (!dtrace_priv_kernel(state))
7330 continue;
7331 break;
7332
7333 case DTRACEACT_USYM:
7334 case DTRACEACT_UMOD:
7335 case DTRACEACT_UADDR: {
7336 struct pid *pid = curthread->t_procp->p_pidp;
7337
7338 if (!dtrace_priv_proc(state, &mstate))
7339 continue;
7340
7341 DTRACE_STORE(uint64_t, tomax,
7342 valoffs, (uint64_t)pid->pid_id);
7343 DTRACE_STORE(uint64_t, tomax,
7344 valoffs + sizeof (uint64_t), val);
7345
7346 continue;
7347 }
7348
7349 case DTRACEACT_EXIT: {
7350 /*
7351 * For the exit action, we are going to attempt
7352 * to atomically set our activity to be
7353 * draining. If this fails (either because
7354 * another CPU has beat us to the exit action,
7355 * or because our current activity is something
7356 * other than ACTIVE or WARMUP), we will
7357 * continue. This assures that the exit action
7358 * can be successfully recorded at most once
7359 * when we're in the ACTIVE state. If we're
7360 * encountering the exit() action while in
7361 * COOLDOWN, however, we want to honor the new
7362 * status code. (We know that we're the only
7363 * thread in COOLDOWN, so there is no race.)
7364 */
7365 void *activity = &state->dts_activity;
7366 dtrace_activity_t current = state->dts_activity;
7367
7368 if (current == DTRACE_ACTIVITY_COOLDOWN)
7369 break;
7370
7371 if (current != DTRACE_ACTIVITY_WARMUP)
7372 current = DTRACE_ACTIVITY_ACTIVE;
7373
7374 if (dtrace_cas32(activity, current,
7375 DTRACE_ACTIVITY_DRAINING) != current) {
7376 *flags |= CPU_DTRACE_DROP;
7377 continue;
7378 }
7379
7380 break;
7381 }
7382
7383 default:
7384 ASSERT(0);
7385 }
7386
7387 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ||
7388 dp->dtdo_rtype.dtdt_flags & DIF_TF_BYUREF) {
7389 uintptr_t end = valoffs + size;
7390
7391 if (tracememsize != 0 &&
7392 valoffs + tracememsize < end) {
7393 end = valoffs + tracememsize;
7394 tracememsize = 0;
7395 }
7396
7397 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF &&
7398 !dtrace_vcanload((void *)(uintptr_t)val,
7399 &dp->dtdo_rtype, NULL, &mstate, vstate))
7400 continue;
7401
7402 dtrace_store_by_ref(dp, tomax, size, &valoffs,
7403 &val, end, act->dta_intuple,
7404 dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ?
7405 DIF_TF_BYREF: DIF_TF_BYUREF);
7406 continue;
7407 }
7408
7409 switch (size) {
7410 case 0:
7411 break;
7412
7413 case sizeof (uint8_t):
7414 DTRACE_STORE(uint8_t, tomax, valoffs, val);
7415 break;
7416 case sizeof (uint16_t):
7417 DTRACE_STORE(uint16_t, tomax, valoffs, val);
7418 break;
7419 case sizeof (uint32_t):
7420 DTRACE_STORE(uint32_t, tomax, valoffs, val);
7421 break;
7422 case sizeof (uint64_t):
7423 DTRACE_STORE(uint64_t, tomax, valoffs, val);
7424 break;
7425 default:
7426 /*
7427 * Any other size should have been returned by
7428 * reference, not by value.
7429 */
7430 ASSERT(0);
7431 break;
7432 }
7433 }
7434
7435 if (*flags & CPU_DTRACE_DROP)
7436 continue;
7437
7438 if (*flags & CPU_DTRACE_FAULT) {
7439 int ndx;
7440 dtrace_action_t *err;
7441
7442 buf->dtb_errors++;
7443
7444 if (probe->dtpr_id == dtrace_probeid_error) {
7445 /*
7446 * There's nothing we can do -- we had an
7447 * error on the error probe. We bump an
7448 * error counter to at least indicate that
7449 * this condition happened.
7450 */
7451 dtrace_error(&state->dts_dblerrors);
7452 continue;
7453 }
7454
7455 if (vtime) {
7456 /*
7457 * Before recursing on dtrace_probe(), we
7458 * need to explicitly clear out our start
7459 * time to prevent it from being accumulated
7460 * into t_dtrace_vtime.
7461 */
7462 curthread->t_dtrace_start = 0;
7463 }
7464
7465 /*
7466 * Iterate over the actions to figure out which action
7467 * we were processing when we experienced the error.
7468 * Note that act points _past_ the faulting action; if
7469 * act is ecb->dte_action, the fault was in the
7470 * predicate, if it's ecb->dte_action->dta_next it's
7471 * in action #1, and so on.
7472 */
7473 for (err = ecb->dte_action, ndx = 0;
7474 err != act; err = err->dta_next, ndx++)
7475 continue;
7476
7477 dtrace_probe_error(state, ecb->dte_epid, ndx,
7478 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7479 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7480 cpu_core[cpuid].cpuc_dtrace_illval);
7481
7482 continue;
7483 }
7484
7485 if (!committed)
7486 buf->dtb_offset = offs + ecb->dte_size;
7487 }
7488
7489 end = dtrace_gethrtime();
7490 if (vtime)
7491 curthread->t_dtrace_start = end;
7492
7493 CPU->cpu_dtrace_nsec += end - now;
7494
7495 dtrace_interrupt_enable(cookie);
7496 }
7497
7498 /*
7499 * DTrace Probe Hashing Functions
7500 *
7501 * The functions in this section (and indeed, the functions in remaining
7502 * sections) are not _called_ from probe context. (Any exceptions to this are
7503 * marked with a "Note:".) Rather, they are called from elsewhere in the
7504 * DTrace framework to look-up probes in, add probes to and remove probes from
7505 * the DTrace probe hashes. (Each probe is hashed by each element of the
7506 * probe tuple -- allowing for fast lookups, regardless of what was
7507 * specified.)
7508 */
7509 static uint_t
7510 dtrace_hash_str(char *p)
7511 {
7512 unsigned int g;
7513 uint_t hval = 0;
7514
7515 while (*p) {
7516 hval = (hval << 4) + *p++;
7517 if ((g = (hval & 0xf0000000)) != 0)
7518 hval ^= g >> 24;
7519 hval &= ~g;
7520 }
7521 return (hval);
7522 }
7523
7524 static dtrace_hash_t *
7525 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
7526 {
7527 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
7528
7529 hash->dth_stroffs = stroffs;
7530 hash->dth_nextoffs = nextoffs;
7531 hash->dth_prevoffs = prevoffs;
7532
7533 hash->dth_size = 1;
7534 hash->dth_mask = hash->dth_size - 1;
7535
7536 hash->dth_tab = kmem_zalloc(hash->dth_size *
7537 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
7538
7539 return (hash);
7540 }
7541
7542 static void
7543 dtrace_hash_destroy(dtrace_hash_t *hash)
7544 {
7545 #ifdef DEBUG
7546 int i;
7547
7548 for (i = 0; i < hash->dth_size; i++)
7549 ASSERT(hash->dth_tab[i] == NULL);
7550 #endif
7551
7552 kmem_free(hash->dth_tab,
7553 hash->dth_size * sizeof (dtrace_hashbucket_t *));
7554 kmem_free(hash, sizeof (dtrace_hash_t));
7555 }
7556
7557 static void
7558 dtrace_hash_resize(dtrace_hash_t *hash)
7559 {
7560 int size = hash->dth_size, i, ndx;
7561 int new_size = hash->dth_size << 1;
7562 int new_mask = new_size - 1;
7563 dtrace_hashbucket_t **new_tab, *bucket, *next;
7564
7565 ASSERT((new_size & new_mask) == 0);
7566
7567 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
7568
7569 for (i = 0; i < size; i++) {
7570 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
7571 dtrace_probe_t *probe = bucket->dthb_chain;
7572
7573 ASSERT(probe != NULL);
7574 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
7575
7576 next = bucket->dthb_next;
7577 bucket->dthb_next = new_tab[ndx];
7578 new_tab[ndx] = bucket;
7579 }
7580 }
7581
7582 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
7583 hash->dth_tab = new_tab;
7584 hash->dth_size = new_size;
7585 hash->dth_mask = new_mask;
7586 }
7587
7588 static void
7589 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
7590 {
7591 int hashval = DTRACE_HASHSTR(hash, new);
7592 int ndx = hashval & hash->dth_mask;
7593 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7594 dtrace_probe_t **nextp, **prevp;
7595
7596 for (; bucket != NULL; bucket = bucket->dthb_next) {
7597 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
7598 goto add;
7599 }
7600
7601 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
7602 dtrace_hash_resize(hash);
7603 dtrace_hash_add(hash, new);
7604 return;
7605 }
7606
7607 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
7608 bucket->dthb_next = hash->dth_tab[ndx];
7609 hash->dth_tab[ndx] = bucket;
7610 hash->dth_nbuckets++;
7611
7612 add:
7613 nextp = DTRACE_HASHNEXT(hash, new);
7614 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
7615 *nextp = bucket->dthb_chain;
7616
7617 if (bucket->dthb_chain != NULL) {
7618 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
7619 ASSERT(*prevp == NULL);
7620 *prevp = new;
7621 }
7622
7623 bucket->dthb_chain = new;
7624 bucket->dthb_len++;
7625 }
7626
7627 static dtrace_probe_t *
7628 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
7629 {
7630 int hashval = DTRACE_HASHSTR(hash, template);
7631 int ndx = hashval & hash->dth_mask;
7632 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7633
7634 for (; bucket != NULL; bucket = bucket->dthb_next) {
7635 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7636 return (bucket->dthb_chain);
7637 }
7638
7639 return (NULL);
7640 }
7641
7642 static int
7643 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
7644 {
7645 int hashval = DTRACE_HASHSTR(hash, template);
7646 int ndx = hashval & hash->dth_mask;
7647 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7648
7649 for (; bucket != NULL; bucket = bucket->dthb_next) {
7650 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7651 return (bucket->dthb_len);
7652 }
7653
7654 return (NULL);
7655 }
7656
7657 static void
7658 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7659 {
7660 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7661 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7662
7663 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7664 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7665
7666 /*
7667 * Find the bucket that we're removing this probe from.
7668 */
7669 for (; bucket != NULL; bucket = bucket->dthb_next) {
7670 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7671 break;
7672 }
7673
7674 ASSERT(bucket != NULL);
7675
7676 if (*prevp == NULL) {
7677 if (*nextp == NULL) {
7678 /*
7679 * The removed probe was the only probe on this
7680 * bucket; we need to remove the bucket.
7681 */
7682 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7683
7684 ASSERT(bucket->dthb_chain == probe);
7685 ASSERT(b != NULL);
7686
7687 if (b == bucket) {
7688 hash->dth_tab[ndx] = bucket->dthb_next;
7689 } else {
7690 while (b->dthb_next != bucket)
7691 b = b->dthb_next;
7692 b->dthb_next = bucket->dthb_next;
7693 }
7694
7695 ASSERT(hash->dth_nbuckets > 0);
7696 hash->dth_nbuckets--;
7697 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7698 return;
7699 }
7700
7701 bucket->dthb_chain = *nextp;
7702 } else {
7703 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7704 }
7705
7706 if (*nextp != NULL)
7707 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7708 }
7709
7710 /*
7711 * DTrace Utility Functions
7712 *
7713 * These are random utility functions that are _not_ called from probe context.
7714 */
7715 static int
7716 dtrace_badattr(const dtrace_attribute_t *a)
7717 {
7718 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7719 a->dtat_data > DTRACE_STABILITY_MAX ||
7720 a->dtat_class > DTRACE_CLASS_MAX);
7721 }
7722
7723 /*
7724 * Return a duplicate copy of a string. If the specified string is NULL,
7725 * this function returns a zero-length string.
7726 */
7727 static char *
7728 dtrace_strdup(const char *str)
7729 {
7730 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7731
7732 if (str != NULL)
7733 (void) strcpy(new, str);
7734
7735 return (new);
7736 }
7737
7738 #define DTRACE_ISALPHA(c) \
7739 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7740
7741 static int
7742 dtrace_badname(const char *s)
7743 {
7744 char c;
7745
7746 if (s == NULL || (c = *s++) == '\0')
7747 return (0);
7748
7749 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7750 return (1);
7751
7752 while ((c = *s++) != '\0') {
7753 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7754 c != '-' && c != '_' && c != '.' && c != '`')
7755 return (1);
7756 }
7757
7758 return (0);
7759 }
7760
7761 static void
7762 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7763 {
7764 uint32_t priv;
7765
7766 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7767 /*
7768 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7769 */
7770 priv = DTRACE_PRIV_ALL;
7771 } else {
7772 *uidp = crgetuid(cr);
7773 *zoneidp = crgetzoneid(cr);
7774
7775 priv = 0;
7776 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7777 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7778 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7779 priv |= DTRACE_PRIV_USER;
7780 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7781 priv |= DTRACE_PRIV_PROC;
7782 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7783 priv |= DTRACE_PRIV_OWNER;
7784 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7785 priv |= DTRACE_PRIV_ZONEOWNER;
7786 }
7787
7788 *privp = priv;
7789 }
7790
7791 #ifdef DTRACE_ERRDEBUG
7792 static void
7793 dtrace_errdebug(const char *str)
7794 {
7795 int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
7796 int occupied = 0;
7797
7798 mutex_enter(&dtrace_errlock);
7799 dtrace_errlast = str;
7800 dtrace_errthread = curthread;
7801
7802 while (occupied++ < DTRACE_ERRHASHSZ) {
7803 if (dtrace_errhash[hval].dter_msg == str) {
7804 dtrace_errhash[hval].dter_count++;
7805 goto out;
7806 }
7807
7808 if (dtrace_errhash[hval].dter_msg != NULL) {
7809 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7810 continue;
7811 }
7812
7813 dtrace_errhash[hval].dter_msg = str;
7814 dtrace_errhash[hval].dter_count = 1;
7815 goto out;
7816 }
7817
7818 panic("dtrace: undersized error hash");
7819 out:
7820 mutex_exit(&dtrace_errlock);
7821 }
7822 #endif
7823
7824 /*
7825 * DTrace Matching Functions
7826 *
7827 * These functions are used to match groups of probes, given some elements of
7828 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7829 */
7830 static int
7831 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7832 zoneid_t zoneid)
7833 {
7834 if (priv != DTRACE_PRIV_ALL) {
7835 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7836 uint32_t match = priv & ppriv;
7837
7838 /*
7839 * No PRIV_DTRACE_* privileges...
7840 */
7841 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7842 DTRACE_PRIV_KERNEL)) == 0)
7843 return (0);
7844
7845 /*
7846 * No matching bits, but there were bits to match...
7847 */
7848 if (match == 0 && ppriv != 0)
7849 return (0);
7850
7851 /*
7852 * Need to have permissions to the process, but don't...
7853 */
7854 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7855 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7856 return (0);
7857 }
7858
7859 /*
7860 * Need to be in the same zone unless we possess the
7861 * privilege to examine all zones.
7862 */
7863 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7864 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7865 return (0);
7866 }
7867 }
7868
7869 return (1);
7870 }
7871
7872 /*
7873 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7874 * consists of input pattern strings and an ops-vector to evaluate them.
7875 * This function returns >0 for match, 0 for no match, and <0 for error.
7876 */
7877 static int
7878 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7879 uint32_t priv, uid_t uid, zoneid_t zoneid)
7880 {
7881 dtrace_provider_t *pvp = prp->dtpr_provider;
7882 int rv;
7883
7884 if (pvp->dtpv_defunct)
7885 return (0);
7886
7887 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7888 return (rv);
7889
7890 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7891 return (rv);
7892
7893 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7894 return (rv);
7895
7896 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7897 return (rv);
7898
7899 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7900 return (0);
7901
7902 return (rv);
7903 }
7904
7905 /*
7906 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7907 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7908 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7909 * In addition, all of the recursion cases except for '*' matching have been
7910 * unwound. For '*', we still implement recursive evaluation, but a depth
7911 * counter is maintained and matching is aborted if we recurse too deep.
7912 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7913 */
7914 static int
7915 dtrace_match_glob(const char *s, const char *p, int depth)
7916 {
7917 const char *olds;
7918 char s1, c;
7919 int gs;
7920
7921 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7922 return (-1);
7923
7924 if (s == NULL)
7925 s = ""; /* treat NULL as empty string */
7926
7927 top:
7928 olds = s;
7929 s1 = *s++;
7930
7931 if (p == NULL)
7932 return (0);
7933
7934 if ((c = *p++) == '\0')
7935 return (s1 == '\0');
7936
7937 switch (c) {
7938 case '[': {
7939 int ok = 0, notflag = 0;
7940 char lc = '\0';
7941
7942 if (s1 == '\0')
7943 return (0);
7944
7945 if (*p == '!') {
7946 notflag = 1;
7947 p++;
7948 }
7949
7950 if ((c = *p++) == '\0')
7951 return (0);
7952
7953 do {
7954 if (c == '-' && lc != '\0' && *p != ']') {
7955 if ((c = *p++) == '\0')
7956 return (0);
7957 if (c == '\\' && (c = *p++) == '\0')
7958 return (0);
7959
7960 if (notflag) {
7961 if (s1 < lc || s1 > c)
7962 ok++;
7963 else
7964 return (0);
7965 } else if (lc <= s1 && s1 <= c)
7966 ok++;
7967
7968 } else if (c == '\\' && (c = *p++) == '\0')
7969 return (0);
7970
7971 lc = c; /* save left-hand 'c' for next iteration */
7972
7973 if (notflag) {
7974 if (s1 != c)
7975 ok++;
7976 else
7977 return (0);
7978 } else if (s1 == c)
7979 ok++;
7980
7981 if ((c = *p++) == '\0')
7982 return (0);
7983
7984 } while (c != ']');
7985
7986 if (ok)
7987 goto top;
7988
7989 return (0);
7990 }
7991
7992 case '\\':
7993 if ((c = *p++) == '\0')
7994 return (0);
7995 /*FALLTHRU*/
7996
7997 default:
7998 if (c != s1)
7999 return (0);
8000 /*FALLTHRU*/
8001
8002 case '?':
8003 if (s1 != '\0')
8004 goto top;
8005 return (0);
8006
8007 case '*':
8008 while (*p == '*')
8009 p++; /* consecutive *'s are identical to a single one */
8010
8011 if (*p == '\0')
8012 return (1);
8013
8014 for (s = olds; *s != '\0'; s++) {
8015 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
8016 return (gs);
8017 }
8018
8019 return (0);
8020 }
8021 }
8022
8023 /*ARGSUSED*/
8024 static int
8025 dtrace_match_string(const char *s, const char *p, int depth)
8026 {
8027 return (s != NULL && strcmp(s, p) == 0);
8028 }
8029
8030 /*ARGSUSED*/
8031 static int
8032 dtrace_match_nul(const char *s, const char *p, int depth)
8033 {
8034 return (1); /* always match the empty pattern */
8035 }
8036
8037 /*ARGSUSED*/
8038 static int
8039 dtrace_match_nonzero(const char *s, const char *p, int depth)
8040 {
8041 return (s != NULL && s[0] != '\0');
8042 }
8043
8044 static int
8045 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
8046 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
8047 {
8048 dtrace_probe_t template, *probe;
8049 dtrace_hash_t *hash = NULL;
8050 int len, rc, best = INT_MAX, nmatched = 0;
8051 dtrace_id_t i;
8052
8053 ASSERT(MUTEX_HELD(&dtrace_lock));
8054
8055 /*
8056 * If the probe ID is specified in the key, just lookup by ID and
8057 * invoke the match callback once if a matching probe is found.
8058 */
8059 if (pkp->dtpk_id != DTRACE_IDNONE) {
8060 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
8061 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
8062 if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
8063 return (DTRACE_MATCH_FAIL);
8064 nmatched++;
8065 }
8066 return (nmatched);
8067 }
8068
8069 template.dtpr_mod = (char *)pkp->dtpk_mod;
8070 template.dtpr_func = (char *)pkp->dtpk_func;
8071 template.dtpr_name = (char *)pkp->dtpk_name;
8072
8073 /*
8074 * We want to find the most distinct of the module name, function
8075 * name, and name. So for each one that is not a glob pattern or
8076 * empty string, we perform a lookup in the corresponding hash and
8077 * use the hash table with the fewest collisions to do our search.
8078 */
8079 if (pkp->dtpk_mmatch == &dtrace_match_string &&
8080 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
8081 best = len;
8082 hash = dtrace_bymod;
8083 }
8084
8085 if (pkp->dtpk_fmatch == &dtrace_match_string &&
8086 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
8087 best = len;
8088 hash = dtrace_byfunc;
8089 }
8090
8091 if (pkp->dtpk_nmatch == &dtrace_match_string &&
8092 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
8093 best = len;
8094 hash = dtrace_byname;
8095 }
8096
8097 /*
8098 * If we did not select a hash table, iterate over every probe and
8099 * invoke our callback for each one that matches our input probe key.
8100 */
8101 if (hash == NULL) {
8102 for (i = 0; i < dtrace_nprobes; i++) {
8103 if ((probe = dtrace_probes[i]) == NULL ||
8104 dtrace_match_probe(probe, pkp, priv, uid,
8105 zoneid) <= 0)
8106 continue;
8107
8108 nmatched++;
8109
8110 if ((rc = (*matched)(probe, arg)) !=
8111 DTRACE_MATCH_NEXT) {
8112 if (rc == DTRACE_MATCH_FAIL)
8113 return (DTRACE_MATCH_FAIL);
8114 break;
8115 }
8116 }
8117
8118 return (nmatched);
8119 }
8120
8121 /*
8122 * If we selected a hash table, iterate over each probe of the same key
8123 * name and invoke the callback for every probe that matches the other
8124 * attributes of our input probe key.
8125 */
8126 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
8127 probe = *(DTRACE_HASHNEXT(hash, probe))) {
8128
8129 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
8130 continue;
8131
8132 nmatched++;
8133
8134 if ((rc = (*matched)(probe, arg)) != 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 * Return the function pointer dtrace_probecmp() should use to compare the
8146 * specified pattern with a string. For NULL or empty patterns, we select
8147 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
8148 * For non-empty non-glob strings, we use dtrace_match_string().
8149 */
8150 static dtrace_probekey_f *
8151 dtrace_probekey_func(const char *p)
8152 {
8153 char c;
8154
8155 if (p == NULL || *p == '\0')
8156 return (&dtrace_match_nul);
8157
8158 while ((c = *p++) != '\0') {
8159 if (c == '[' || c == '?' || c == '*' || c == '\\')
8160 return (&dtrace_match_glob);
8161 }
8162
8163 return (&dtrace_match_string);
8164 }
8165
8166 /*
8167 * Build a probe comparison key for use with dtrace_match_probe() from the
8168 * given probe description. By convention, a null key only matches anchored
8169 * probes: if each field is the empty string, reset dtpk_fmatch to
8170 * dtrace_match_nonzero().
8171 */
8172 static void
8173 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
8174 {
8175 pkp->dtpk_prov = pdp->dtpd_provider;
8176 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
8177
8178 pkp->dtpk_mod = pdp->dtpd_mod;
8179 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
8180
8181 pkp->dtpk_func = pdp->dtpd_func;
8182 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
8183
8184 pkp->dtpk_name = pdp->dtpd_name;
8185 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
8186
8187 pkp->dtpk_id = pdp->dtpd_id;
8188
8189 if (pkp->dtpk_id == DTRACE_IDNONE &&
8190 pkp->dtpk_pmatch == &dtrace_match_nul &&
8191 pkp->dtpk_mmatch == &dtrace_match_nul &&
8192 pkp->dtpk_fmatch == &dtrace_match_nul &&
8193 pkp->dtpk_nmatch == &dtrace_match_nul)
8194 pkp->dtpk_fmatch = &dtrace_match_nonzero;
8195 }
8196
8197 /*
8198 * DTrace Provider-to-Framework API Functions
8199 *
8200 * These functions implement much of the Provider-to-Framework API, as
8201 * described in <sys/dtrace.h>. The parts of the API not in this section are
8202 * the functions in the API for probe management (found below), and
8203 * dtrace_probe() itself (found above).
8204 */
8205
8206 /*
8207 * Register the calling provider with the DTrace framework. This should
8208 * generally be called by DTrace providers in their attach(9E) entry point.
8209 */
8210 int
8211 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
8212 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
8213 {
8214 dtrace_provider_t *provider;
8215
8216 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
8217 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8218 "arguments", name ? name : "<NULL>");
8219 return (EINVAL);
8220 }
8221
8222 if (name[0] == '\0' || dtrace_badname(name)) {
8223 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8224 "provider name", name);
8225 return (EINVAL);
8226 }
8227
8228 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
8229 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
8230 pops->dtps_destroy == NULL ||
8231 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
8232 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8233 "provider ops", name);
8234 return (EINVAL);
8235 }
8236
8237 if (dtrace_badattr(&pap->dtpa_provider) ||
8238 dtrace_badattr(&pap->dtpa_mod) ||
8239 dtrace_badattr(&pap->dtpa_func) ||
8240 dtrace_badattr(&pap->dtpa_name) ||
8241 dtrace_badattr(&pap->dtpa_args)) {
8242 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8243 "provider attributes", name);
8244 return (EINVAL);
8245 }
8246
8247 if (priv & ~DTRACE_PRIV_ALL) {
8248 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8249 "privilege attributes", name);
8250 return (EINVAL);
8251 }
8252
8253 if ((priv & DTRACE_PRIV_KERNEL) &&
8254 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
8255 pops->dtps_mode == NULL) {
8256 cmn_err(CE_WARN, "failed to register provider '%s': need "
8257 "dtps_mode() op for given privilege attributes", name);
8258 return (EINVAL);
8259 }
8260
8261 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
8262 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8263 (void) strcpy(provider->dtpv_name, name);
8264
8265 provider->dtpv_attr = *pap;
8266 provider->dtpv_priv.dtpp_flags = priv;
8267 if (cr != NULL) {
8268 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
8269 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
8270 }
8271 provider->dtpv_pops = *pops;
8272
8273 if (pops->dtps_provide == NULL) {
8274 ASSERT(pops->dtps_provide_module != NULL);
8275 provider->dtpv_pops.dtps_provide =
8276 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
8277 }
8278
8279 if (pops->dtps_provide_module == NULL) {
8280 ASSERT(pops->dtps_provide != NULL);
8281 provider->dtpv_pops.dtps_provide_module =
8282 (void (*)(void *, struct modctl *))dtrace_nullop;
8283 }
8284
8285 if (pops->dtps_suspend == NULL) {
8286 ASSERT(pops->dtps_resume == NULL);
8287 provider->dtpv_pops.dtps_suspend =
8288 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8289 provider->dtpv_pops.dtps_resume =
8290 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8291 }
8292
8293 provider->dtpv_arg = arg;
8294 *idp = (dtrace_provider_id_t)provider;
8295
8296 if (pops == &dtrace_provider_ops) {
8297 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8298 ASSERT(MUTEX_HELD(&dtrace_lock));
8299 ASSERT(dtrace_anon.dta_enabling == NULL);
8300
8301 /*
8302 * We make sure that the DTrace provider is at the head of
8303 * the provider chain.
8304 */
8305 provider->dtpv_next = dtrace_provider;
8306 dtrace_provider = provider;
8307 return (0);
8308 }
8309
8310 mutex_enter(&dtrace_provider_lock);
8311 mutex_enter(&dtrace_lock);
8312
8313 /*
8314 * If there is at least one provider registered, we'll add this
8315 * provider after the first provider.
8316 */
8317 if (dtrace_provider != NULL) {
8318 provider->dtpv_next = dtrace_provider->dtpv_next;
8319 dtrace_provider->dtpv_next = provider;
8320 } else {
8321 dtrace_provider = provider;
8322 }
8323
8324 if (dtrace_retained != NULL) {
8325 dtrace_enabling_provide(provider);
8326
8327 /*
8328 * Now we need to call dtrace_enabling_matchall() -- which
8329 * will acquire cpu_lock and dtrace_lock. We therefore need
8330 * to drop all of our locks before calling into it...
8331 */
8332 mutex_exit(&dtrace_lock);
8333 mutex_exit(&dtrace_provider_lock);
8334 dtrace_enabling_matchall();
8335
8336 return (0);
8337 }
8338
8339 mutex_exit(&dtrace_lock);
8340 mutex_exit(&dtrace_provider_lock);
8341
8342 return (0);
8343 }
8344
8345 /*
8346 * Unregister the specified provider from the DTrace framework. This should
8347 * generally be called by DTrace providers in their detach(9E) entry point.
8348 */
8349 int
8350 dtrace_unregister(dtrace_provider_id_t id)
8351 {
8352 dtrace_provider_t *old = (dtrace_provider_t *)id;
8353 dtrace_provider_t *prev = NULL;
8354 int i, self = 0, noreap = 0;
8355 dtrace_probe_t *probe, *first = NULL;
8356
8357 if (old->dtpv_pops.dtps_enable ==
8358 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
8359 /*
8360 * If DTrace itself is the provider, we're called with locks
8361 * already held.
8362 */
8363 ASSERT(old == dtrace_provider);
8364 ASSERT(dtrace_devi != NULL);
8365 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8366 ASSERT(MUTEX_HELD(&dtrace_lock));
8367 self = 1;
8368
8369 if (dtrace_provider->dtpv_next != NULL) {
8370 /*
8371 * There's another provider here; return failure.
8372 */
8373 return (EBUSY);
8374 }
8375 } else {
8376 mutex_enter(&dtrace_provider_lock);
8377 mutex_enter(&mod_lock);
8378 mutex_enter(&dtrace_lock);
8379 }
8380
8381 /*
8382 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8383 * probes, we refuse to let providers slither away, unless this
8384 * provider has already been explicitly invalidated.
8385 */
8386 if (!old->dtpv_defunct &&
8387 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8388 dtrace_anon.dta_state->dts_necbs > 0))) {
8389 if (!self) {
8390 mutex_exit(&dtrace_lock);
8391 mutex_exit(&mod_lock);
8392 mutex_exit(&dtrace_provider_lock);
8393 }
8394 return (EBUSY);
8395 }
8396
8397 /*
8398 * Attempt to destroy the probes associated with this provider.
8399 */
8400 for (i = 0; i < dtrace_nprobes; i++) {
8401 if ((probe = dtrace_probes[i]) == NULL)
8402 continue;
8403
8404 if (probe->dtpr_provider != old)
8405 continue;
8406
8407 if (probe->dtpr_ecb == NULL)
8408 continue;
8409
8410 /*
8411 * If we are trying to unregister a defunct provider, and the
8412 * provider was made defunct within the interval dictated by
8413 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8414 * attempt to reap our enablings. To denote that the provider
8415 * should reattempt to unregister itself at some point in the
8416 * future, we will return a differentiable error code (EAGAIN
8417 * instead of EBUSY) in this case.
8418 */
8419 if (dtrace_gethrtime() - old->dtpv_defunct >
8420 dtrace_unregister_defunct_reap)
8421 noreap = 1;
8422
8423 if (!self) {
8424 mutex_exit(&dtrace_lock);
8425 mutex_exit(&mod_lock);
8426 mutex_exit(&dtrace_provider_lock);
8427 }
8428
8429 if (noreap)
8430 return (EBUSY);
8431
8432 (void) taskq_dispatch(dtrace_taskq,
8433 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8434
8435 return (EAGAIN);
8436 }
8437
8438 /*
8439 * All of the probes for this provider are disabled; we can safely
8440 * remove all of them from their hash chains and from the probe array.
8441 */
8442 for (i = 0; i < dtrace_nprobes; i++) {
8443 if ((probe = dtrace_probes[i]) == NULL)
8444 continue;
8445
8446 if (probe->dtpr_provider != old)
8447 continue;
8448
8449 dtrace_probes[i] = NULL;
8450
8451 dtrace_hash_remove(dtrace_bymod, probe);
8452 dtrace_hash_remove(dtrace_byfunc, probe);
8453 dtrace_hash_remove(dtrace_byname, probe);
8454
8455 if (first == NULL) {
8456 first = probe;
8457 probe->dtpr_nextmod = NULL;
8458 } else {
8459 probe->dtpr_nextmod = first;
8460 first = probe;
8461 }
8462 }
8463
8464 /*
8465 * The provider's probes have been removed from the hash chains and
8466 * from the probe array. Now issue a dtrace_sync() to be sure that
8467 * everyone has cleared out from any probe array processing.
8468 */
8469 dtrace_sync();
8470
8471 for (probe = first; probe != NULL; probe = first) {
8472 first = probe->dtpr_nextmod;
8473
8474 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8475 probe->dtpr_arg);
8476 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8477 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8478 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8479 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8480 kmem_free(probe, sizeof (dtrace_probe_t));
8481 }
8482
8483 if ((prev = dtrace_provider) == old) {
8484 ASSERT(self || dtrace_devi == NULL);
8485 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8486 dtrace_provider = old->dtpv_next;
8487 } else {
8488 while (prev != NULL && prev->dtpv_next != old)
8489 prev = prev->dtpv_next;
8490
8491 if (prev == NULL) {
8492 panic("attempt to unregister non-existent "
8493 "dtrace provider %p\n", (void *)id);
8494 }
8495
8496 prev->dtpv_next = old->dtpv_next;
8497 }
8498
8499 if (!self) {
8500 mutex_exit(&dtrace_lock);
8501 mutex_exit(&mod_lock);
8502 mutex_exit(&dtrace_provider_lock);
8503 }
8504
8505 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
8506 kmem_free(old, sizeof (dtrace_provider_t));
8507
8508 return (0);
8509 }
8510
8511 /*
8512 * Invalidate the specified provider. All subsequent probe lookups for the
8513 * specified provider will fail, but its probes will not be removed.
8514 */
8515 void
8516 dtrace_invalidate(dtrace_provider_id_t id)
8517 {
8518 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
8519
8520 ASSERT(pvp->dtpv_pops.dtps_enable !=
8521 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8522
8523 mutex_enter(&dtrace_provider_lock);
8524 mutex_enter(&dtrace_lock);
8525
8526 pvp->dtpv_defunct = dtrace_gethrtime();
8527
8528 mutex_exit(&dtrace_lock);
8529 mutex_exit(&dtrace_provider_lock);
8530 }
8531
8532 /*
8533 * Indicate whether or not DTrace has attached.
8534 */
8535 int
8536 dtrace_attached(void)
8537 {
8538 /*
8539 * dtrace_provider will be non-NULL iff the DTrace driver has
8540 * attached. (It's non-NULL because DTrace is always itself a
8541 * provider.)
8542 */
8543 return (dtrace_provider != NULL);
8544 }
8545
8546 /*
8547 * Remove all the unenabled probes for the given provider. This function is
8548 * not unlike dtrace_unregister(), except that it doesn't remove the provider
8549 * -- just as many of its associated probes as it can.
8550 */
8551 int
8552 dtrace_condense(dtrace_provider_id_t id)
8553 {
8554 dtrace_provider_t *prov = (dtrace_provider_t *)id;
8555 int i;
8556 dtrace_probe_t *probe;
8557
8558 /*
8559 * Make sure this isn't the dtrace provider itself.
8560 */
8561 ASSERT(prov->dtpv_pops.dtps_enable !=
8562 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8563
8564 mutex_enter(&dtrace_provider_lock);
8565 mutex_enter(&dtrace_lock);
8566
8567 /*
8568 * Attempt to destroy the probes associated with this provider.
8569 */
8570 for (i = 0; i < dtrace_nprobes; i++) {
8571 if ((probe = dtrace_probes[i]) == NULL)
8572 continue;
8573
8574 if (probe->dtpr_provider != prov)
8575 continue;
8576
8577 if (probe->dtpr_ecb != NULL)
8578 continue;
8579
8580 dtrace_probes[i] = NULL;
8581
8582 dtrace_hash_remove(dtrace_bymod, probe);
8583 dtrace_hash_remove(dtrace_byfunc, probe);
8584 dtrace_hash_remove(dtrace_byname, probe);
8585
8586 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
8587 probe->dtpr_arg);
8588 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8589 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8590 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8591 kmem_free(probe, sizeof (dtrace_probe_t));
8592 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
8593 }
8594
8595 mutex_exit(&dtrace_lock);
8596 mutex_exit(&dtrace_provider_lock);
8597
8598 return (0);
8599 }
8600
8601 /*
8602 * DTrace Probe Management Functions
8603 *
8604 * The functions in this section perform the DTrace probe management,
8605 * including functions to create probes, look-up probes, and call into the
8606 * providers to request that probes be provided. Some of these functions are
8607 * in the Provider-to-Framework API; these functions can be identified by the
8608 * fact that they are not declared "static".
8609 */
8610
8611 /*
8612 * Create a probe with the specified module name, function name, and name.
8613 */
8614 dtrace_id_t
8615 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
8616 const char *func, const char *name, int aframes, void *arg)
8617 {
8618 dtrace_probe_t *probe, **probes;
8619 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
8620 dtrace_id_t id;
8621
8622 if (provider == dtrace_provider) {
8623 ASSERT(MUTEX_HELD(&dtrace_lock));
8624 } else {
8625 mutex_enter(&dtrace_lock);
8626 }
8627
8628 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
8629 VM_BESTFIT | VM_SLEEP);
8630 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8631
8632 probe->dtpr_id = id;
8633 probe->dtpr_gen = dtrace_probegen++;
8634 probe->dtpr_mod = dtrace_strdup(mod);
8635 probe->dtpr_func = dtrace_strdup(func);
8636 probe->dtpr_name = dtrace_strdup(name);
8637 probe->dtpr_arg = arg;
8638 probe->dtpr_aframes = aframes;
8639 probe->dtpr_provider = provider;
8640
8641 dtrace_hash_add(dtrace_bymod, probe);
8642 dtrace_hash_add(dtrace_byfunc, probe);
8643 dtrace_hash_add(dtrace_byname, probe);
8644
8645 if (id - 1 >= dtrace_nprobes) {
8646 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8647 size_t nsize = osize << 1;
8648
8649 if (nsize == 0) {
8650 ASSERT(osize == 0);
8651 ASSERT(dtrace_probes == NULL);
8652 nsize = sizeof (dtrace_probe_t *);
8653 }
8654
8655 probes = kmem_zalloc(nsize, KM_SLEEP);
8656
8657 if (dtrace_probes == NULL) {
8658 ASSERT(osize == 0);
8659 dtrace_probes = probes;
8660 dtrace_nprobes = 1;
8661 } else {
8662 dtrace_probe_t **oprobes = dtrace_probes;
8663
8664 bcopy(oprobes, probes, osize);
8665 dtrace_membar_producer();
8666 dtrace_probes = probes;
8667
8668 dtrace_sync();
8669
8670 /*
8671 * All CPUs are now seeing the new probes array; we can
8672 * safely free the old array.
8673 */
8674 kmem_free(oprobes, osize);
8675 dtrace_nprobes <<= 1;
8676 }
8677
8678 ASSERT(id - 1 < dtrace_nprobes);
8679 }
8680
8681 ASSERT(dtrace_probes[id - 1] == NULL);
8682 dtrace_probes[id - 1] = probe;
8683
8684 if (provider != dtrace_provider)
8685 mutex_exit(&dtrace_lock);
8686
8687 return (id);
8688 }
8689
8690 static dtrace_probe_t *
8691 dtrace_probe_lookup_id(dtrace_id_t id)
8692 {
8693 ASSERT(MUTEX_HELD(&dtrace_lock));
8694
8695 if (id == 0 || id > dtrace_nprobes)
8696 return (NULL);
8697
8698 return (dtrace_probes[id - 1]);
8699 }
8700
8701 static int
8702 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8703 {
8704 *((dtrace_id_t *)arg) = probe->dtpr_id;
8705
8706 return (DTRACE_MATCH_DONE);
8707 }
8708
8709 /*
8710 * Look up a probe based on provider and one or more of module name, function
8711 * name and probe name.
8712 */
8713 dtrace_id_t
8714 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
8715 const char *func, const char *name)
8716 {
8717 dtrace_probekey_t pkey;
8718 dtrace_id_t id;
8719 int match;
8720
8721 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8722 pkey.dtpk_pmatch = &dtrace_match_string;
8723 pkey.dtpk_mod = mod;
8724 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8725 pkey.dtpk_func = func;
8726 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8727 pkey.dtpk_name = name;
8728 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8729 pkey.dtpk_id = DTRACE_IDNONE;
8730
8731 mutex_enter(&dtrace_lock);
8732 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8733 dtrace_probe_lookup_match, &id);
8734 mutex_exit(&dtrace_lock);
8735
8736 ASSERT(match == 1 || match == 0);
8737 return (match ? id : 0);
8738 }
8739
8740 /*
8741 * Returns the probe argument associated with the specified probe.
8742 */
8743 void *
8744 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8745 {
8746 dtrace_probe_t *probe;
8747 void *rval = NULL;
8748
8749 mutex_enter(&dtrace_lock);
8750
8751 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8752 probe->dtpr_provider == (dtrace_provider_t *)id)
8753 rval = probe->dtpr_arg;
8754
8755 mutex_exit(&dtrace_lock);
8756
8757 return (rval);
8758 }
8759
8760 /*
8761 * Copy a probe into a probe description.
8762 */
8763 static void
8764 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8765 {
8766 bzero(pdp, sizeof (dtrace_probedesc_t));
8767 pdp->dtpd_id = prp->dtpr_id;
8768
8769 (void) strncpy(pdp->dtpd_provider,
8770 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8771
8772 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8773 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8774 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8775 }
8776
8777 /*
8778 * Called to indicate that a probe -- or probes -- should be provided by a
8779 * specfied provider. If the specified description is NULL, the provider will
8780 * be told to provide all of its probes. (This is done whenever a new
8781 * consumer comes along, or whenever a retained enabling is to be matched.) If
8782 * the specified description is non-NULL, the provider is given the
8783 * opportunity to dynamically provide the specified probe, allowing providers
8784 * to support the creation of probes on-the-fly. (So-called _autocreated_
8785 * probes.) If the provider is NULL, the operations will be applied to all
8786 * providers; if the provider is non-NULL the operations will only be applied
8787 * to the specified provider. The dtrace_provider_lock must be held, and the
8788 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8789 * will need to grab the dtrace_lock when it reenters the framework through
8790 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8791 */
8792 static void
8793 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8794 {
8795 struct modctl *ctl;
8796 int all = 0;
8797
8798 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8799
8800 if (prv == NULL) {
8801 all = 1;
8802 prv = dtrace_provider;
8803 }
8804
8805 do {
8806 /*
8807 * First, call the blanket provide operation.
8808 */
8809 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8810
8811 /*
8812 * Now call the per-module provide operation. We will grab
8813 * mod_lock to prevent the list from being modified. Note
8814 * that this also prevents the mod_busy bits from changing.
8815 * (mod_busy can only be changed with mod_lock held.)
8816 */
8817 mutex_enter(&mod_lock);
8818
8819 ctl = &modules;
8820 do {
8821 if (ctl->mod_busy || ctl->mod_mp == NULL)
8822 continue;
8823
8824 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8825
8826 } while ((ctl = ctl->mod_next) != &modules);
8827
8828 mutex_exit(&mod_lock);
8829 } while (all && (prv = prv->dtpv_next) != NULL);
8830 }
8831
8832 /*
8833 * Iterate over each probe, and call the Framework-to-Provider API function
8834 * denoted by offs.
8835 */
8836 static void
8837 dtrace_probe_foreach(uintptr_t offs)
8838 {
8839 dtrace_provider_t *prov;
8840 void (*func)(void *, dtrace_id_t, void *);
8841 dtrace_probe_t *probe;
8842 dtrace_icookie_t cookie;
8843 int i;
8844
8845 /*
8846 * We disable interrupts to walk through the probe array. This is
8847 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8848 * won't see stale data.
8849 */
8850 cookie = dtrace_interrupt_disable();
8851
8852 for (i = 0; i < dtrace_nprobes; i++) {
8853 if ((probe = dtrace_probes[i]) == NULL)
8854 continue;
8855
8856 if (probe->dtpr_ecb == NULL) {
8857 /*
8858 * This probe isn't enabled -- don't call the function.
8859 */
8860 continue;
8861 }
8862
8863 prov = probe->dtpr_provider;
8864 func = *((void(**)(void *, dtrace_id_t, void *))
8865 ((uintptr_t)&prov->dtpv_pops + offs));
8866
8867 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8868 }
8869
8870 dtrace_interrupt_enable(cookie);
8871 }
8872
8873 static int
8874 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8875 {
8876 dtrace_probekey_t pkey;
8877 uint32_t priv;
8878 uid_t uid;
8879 zoneid_t zoneid;
8880
8881 ASSERT(MUTEX_HELD(&dtrace_lock));
8882 dtrace_ecb_create_cache = NULL;
8883
8884 if (desc == NULL) {
8885 /*
8886 * If we're passed a NULL description, we're being asked to
8887 * create an ECB with a NULL probe.
8888 */
8889 (void) dtrace_ecb_create_enable(NULL, enab);
8890 return (0);
8891 }
8892
8893 dtrace_probekey(desc, &pkey);
8894 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8895 &priv, &uid, &zoneid);
8896
8897 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8898 enab));
8899 }
8900
8901 /*
8902 * DTrace Helper Provider Functions
8903 */
8904 static void
8905 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8906 {
8907 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8908 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8909 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8910 }
8911
8912 static void
8913 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8914 const dof_provider_t *dofprov, char *strtab)
8915 {
8916 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8917 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8918 dofprov->dofpv_provattr);
8919 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8920 dofprov->dofpv_modattr);
8921 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8922 dofprov->dofpv_funcattr);
8923 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8924 dofprov->dofpv_nameattr);
8925 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8926 dofprov->dofpv_argsattr);
8927 }
8928
8929 static void
8930 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8931 {
8932 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8933 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8934 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8935 dof_provider_t *provider;
8936 dof_probe_t *probe;
8937 uint32_t *off, *enoff;
8938 uint8_t *arg;
8939 char *strtab;
8940 uint_t i, nprobes;
8941 dtrace_helper_provdesc_t dhpv;
8942 dtrace_helper_probedesc_t dhpb;
8943 dtrace_meta_t *meta = dtrace_meta_pid;
8944 dtrace_mops_t *mops = &meta->dtm_mops;
8945 void *parg;
8946
8947 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8948 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8949 provider->dofpv_strtab * dof->dofh_secsize);
8950 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8951 provider->dofpv_probes * dof->dofh_secsize);
8952 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8953 provider->dofpv_prargs * dof->dofh_secsize);
8954 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8955 provider->dofpv_proffs * dof->dofh_secsize);
8956
8957 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8958 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8959 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8960 enoff = NULL;
8961
8962 /*
8963 * See dtrace_helper_provider_validate().
8964 */
8965 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8966 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8967 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8968 provider->dofpv_prenoffs * dof->dofh_secsize);
8969 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8970 }
8971
8972 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8973
8974 /*
8975 * Create the provider.
8976 */
8977 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8978
8979 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8980 return;
8981
8982 meta->dtm_count++;
8983
8984 /*
8985 * Create the probes.
8986 */
8987 for (i = 0; i < nprobes; i++) {
8988 probe = (dof_probe_t *)(uintptr_t)(daddr +
8989 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8990
8991 dhpb.dthpb_mod = dhp->dofhp_mod;
8992 dhpb.dthpb_func = strtab + probe->dofpr_func;
8993 dhpb.dthpb_name = strtab + probe->dofpr_name;
8994 dhpb.dthpb_base = probe->dofpr_addr;
8995 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8996 dhpb.dthpb_noffs = probe->dofpr_noffs;
8997 if (enoff != NULL) {
8998 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8999 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
9000 } else {
9001 dhpb.dthpb_enoffs = NULL;
9002 dhpb.dthpb_nenoffs = 0;
9003 }
9004 dhpb.dthpb_args = arg + probe->dofpr_argidx;
9005 dhpb.dthpb_nargc = probe->dofpr_nargc;
9006 dhpb.dthpb_xargc = probe->dofpr_xargc;
9007 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
9008 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
9009
9010 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
9011 }
9012 }
9013
9014 static void
9015 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
9016 {
9017 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9018 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9019 int i;
9020
9021 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9022
9023 for (i = 0; i < dof->dofh_secnum; i++) {
9024 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9025 dof->dofh_secoff + i * dof->dofh_secsize);
9026
9027 if (sec->dofs_type != DOF_SECT_PROVIDER)
9028 continue;
9029
9030 dtrace_helper_provide_one(dhp, sec, pid);
9031 }
9032
9033 /*
9034 * We may have just created probes, so we must now rematch against
9035 * any retained enablings. Note that this call will acquire both
9036 * cpu_lock and dtrace_lock; the fact that we are holding
9037 * dtrace_meta_lock now is what defines the ordering with respect to
9038 * these three locks.
9039 */
9040 dtrace_enabling_matchall();
9041 }
9042
9043 static void
9044 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9045 {
9046 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9047 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9048 dof_sec_t *str_sec;
9049 dof_provider_t *provider;
9050 char *strtab;
9051 dtrace_helper_provdesc_t dhpv;
9052 dtrace_meta_t *meta = dtrace_meta_pid;
9053 dtrace_mops_t *mops = &meta->dtm_mops;
9054
9055 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9056 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9057 provider->dofpv_strtab * dof->dofh_secsize);
9058
9059 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9060
9061 /*
9062 * Create the provider.
9063 */
9064 dtrace_dofprov2hprov(&dhpv, provider, strtab);
9065
9066 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
9067
9068 meta->dtm_count--;
9069 }
9070
9071 static void
9072 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
9073 {
9074 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9075 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9076 int i;
9077
9078 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9079
9080 for (i = 0; i < dof->dofh_secnum; i++) {
9081 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9082 dof->dofh_secoff + i * dof->dofh_secsize);
9083
9084 if (sec->dofs_type != DOF_SECT_PROVIDER)
9085 continue;
9086
9087 dtrace_helper_provider_remove_one(dhp, sec, pid);
9088 }
9089 }
9090
9091 /*
9092 * DTrace Meta Provider-to-Framework API Functions
9093 *
9094 * These functions implement the Meta Provider-to-Framework API, as described
9095 * in <sys/dtrace.h>.
9096 */
9097 int
9098 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
9099 dtrace_meta_provider_id_t *idp)
9100 {
9101 dtrace_meta_t *meta;
9102 dtrace_helpers_t *help, *next;
9103 int i;
9104
9105 *idp = DTRACE_METAPROVNONE;
9106
9107 /*
9108 * We strictly don't need the name, but we hold onto it for
9109 * debuggability. All hail error queues!
9110 */
9111 if (name == NULL) {
9112 cmn_err(CE_WARN, "failed to register meta-provider: "
9113 "invalid name");
9114 return (EINVAL);
9115 }
9116
9117 if (mops == NULL ||
9118 mops->dtms_create_probe == NULL ||
9119 mops->dtms_provide_pid == NULL ||
9120 mops->dtms_remove_pid == NULL) {
9121 cmn_err(CE_WARN, "failed to register meta-register %s: "
9122 "invalid ops", name);
9123 return (EINVAL);
9124 }
9125
9126 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
9127 meta->dtm_mops = *mops;
9128 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
9129 (void) strcpy(meta->dtm_name, name);
9130 meta->dtm_arg = arg;
9131
9132 mutex_enter(&dtrace_meta_lock);
9133 mutex_enter(&dtrace_lock);
9134
9135 if (dtrace_meta_pid != NULL) {
9136 mutex_exit(&dtrace_lock);
9137 mutex_exit(&dtrace_meta_lock);
9138 cmn_err(CE_WARN, "failed to register meta-register %s: "
9139 "user-land meta-provider exists", name);
9140 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
9141 kmem_free(meta, sizeof (dtrace_meta_t));
9142 return (EINVAL);
9143 }
9144
9145 dtrace_meta_pid = meta;
9146 *idp = (dtrace_meta_provider_id_t)meta;
9147
9148 /*
9149 * If there are providers and probes ready to go, pass them
9150 * off to the new meta provider now.
9151 */
9152
9153 help = dtrace_deferred_pid;
9154 dtrace_deferred_pid = NULL;
9155
9156 mutex_exit(&dtrace_lock);
9157
9158 while (help != NULL) {
9159 for (i = 0; i < help->dthps_nprovs; i++) {
9160 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
9161 help->dthps_pid);
9162 }
9163
9164 next = help->dthps_next;
9165 help->dthps_next = NULL;
9166 help->dthps_prev = NULL;
9167 help->dthps_deferred = 0;
9168 help = next;
9169 }
9170
9171 mutex_exit(&dtrace_meta_lock);
9172
9173 return (0);
9174 }
9175
9176 int
9177 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
9178 {
9179 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
9180
9181 mutex_enter(&dtrace_meta_lock);
9182 mutex_enter(&dtrace_lock);
9183
9184 if (old == dtrace_meta_pid) {
9185 pp = &dtrace_meta_pid;
9186 } else {
9187 panic("attempt to unregister non-existent "
9188 "dtrace meta-provider %p\n", (void *)old);
9189 }
9190
9191 if (old->dtm_count != 0) {
9192 mutex_exit(&dtrace_lock);
9193 mutex_exit(&dtrace_meta_lock);
9194 return (EBUSY);
9195 }
9196
9197 *pp = NULL;
9198
9199 mutex_exit(&dtrace_lock);
9200 mutex_exit(&dtrace_meta_lock);
9201
9202 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
9203 kmem_free(old, sizeof (dtrace_meta_t));
9204
9205 return (0);
9206 }
9207
9208
9209 /*
9210 * DTrace DIF Object Functions
9211 */
9212 static int
9213 dtrace_difo_err(uint_t pc, const char *format, ...)
9214 {
9215 if (dtrace_err_verbose) {
9216 va_list alist;
9217
9218 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
9219 va_start(alist, format);
9220 (void) vuprintf(format, alist);
9221 va_end(alist);
9222 }
9223
9224 #ifdef DTRACE_ERRDEBUG
9225 dtrace_errdebug(format);
9226 #endif
9227 return (1);
9228 }
9229
9230 /*
9231 * Validate a DTrace DIF object by checking the IR instructions. The following
9232 * rules are currently enforced by dtrace_difo_validate():
9233 *
9234 * 1. Each instruction must have a valid opcode
9235 * 2. Each register, string, variable, or subroutine reference must be valid
9236 * 3. No instruction can modify register %r0 (must be zero)
9237 * 4. All instruction reserved bits must be set to zero
9238 * 5. The last instruction must be a "ret" instruction
9239 * 6. All branch targets must reference a valid instruction _after_ the branch
9240 */
9241 static int
9242 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
9243 cred_t *cr)
9244 {
9245 int err = 0, i;
9246 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9247 int kcheckload;
9248 uint_t pc;
9249 int maxglobal = -1, maxlocal = -1, maxtlocal = -1;
9250
9251 kcheckload = cr == NULL ||
9252 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
9253
9254 dp->dtdo_destructive = 0;
9255
9256 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9257 dif_instr_t instr = dp->dtdo_buf[pc];
9258
9259 uint_t r1 = DIF_INSTR_R1(instr);
9260 uint_t r2 = DIF_INSTR_R2(instr);
9261 uint_t rd = DIF_INSTR_RD(instr);
9262 uint_t rs = DIF_INSTR_RS(instr);
9263 uint_t label = DIF_INSTR_LABEL(instr);
9264 uint_t v = DIF_INSTR_VAR(instr);
9265 uint_t subr = DIF_INSTR_SUBR(instr);
9266 uint_t type = DIF_INSTR_TYPE(instr);
9267 uint_t op = DIF_INSTR_OP(instr);
9268
9269 switch (op) {
9270 case DIF_OP_OR:
9271 case DIF_OP_XOR:
9272 case DIF_OP_AND:
9273 case DIF_OP_SLL:
9274 case DIF_OP_SRL:
9275 case DIF_OP_SRA:
9276 case DIF_OP_SUB:
9277 case DIF_OP_ADD:
9278 case DIF_OP_MUL:
9279 case DIF_OP_SDIV:
9280 case DIF_OP_UDIV:
9281 case DIF_OP_SREM:
9282 case DIF_OP_UREM:
9283 case DIF_OP_COPYS:
9284 if (r1 >= nregs)
9285 err += efunc(pc, "invalid register %u\n", r1);
9286 if (r2 >= nregs)
9287 err += efunc(pc, "invalid register %u\n", r2);
9288 if (rd >= nregs)
9289 err += efunc(pc, "invalid register %u\n", rd);
9290 if (rd == 0)
9291 err += efunc(pc, "cannot write to %r0\n");
9292 break;
9293 case DIF_OP_NOT:
9294 case DIF_OP_MOV:
9295 case DIF_OP_ALLOCS:
9296 if (r1 >= nregs)
9297 err += efunc(pc, "invalid register %u\n", r1);
9298 if (r2 != 0)
9299 err += efunc(pc, "non-zero reserved bits\n");
9300 if (rd >= nregs)
9301 err += efunc(pc, "invalid register %u\n", rd);
9302 if (rd == 0)
9303 err += efunc(pc, "cannot write to %r0\n");
9304 break;
9305 case DIF_OP_LDSB:
9306 case DIF_OP_LDSH:
9307 case DIF_OP_LDSW:
9308 case DIF_OP_LDUB:
9309 case DIF_OP_LDUH:
9310 case DIF_OP_LDUW:
9311 case DIF_OP_LDX:
9312 if (r1 >= nregs)
9313 err += efunc(pc, "invalid register %u\n", r1);
9314 if (r2 != 0)
9315 err += efunc(pc, "non-zero reserved bits\n");
9316 if (rd >= nregs)
9317 err += efunc(pc, "invalid register %u\n", rd);
9318 if (rd == 0)
9319 err += efunc(pc, "cannot write to %r0\n");
9320 if (kcheckload)
9321 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9322 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9323 break;
9324 case DIF_OP_RLDSB:
9325 case DIF_OP_RLDSH:
9326 case DIF_OP_RLDSW:
9327 case DIF_OP_RLDUB:
9328 case DIF_OP_RLDUH:
9329 case DIF_OP_RLDUW:
9330 case DIF_OP_RLDX:
9331 if (r1 >= nregs)
9332 err += efunc(pc, "invalid register %u\n", r1);
9333 if (r2 != 0)
9334 err += efunc(pc, "non-zero reserved bits\n");
9335 if (rd >= nregs)
9336 err += efunc(pc, "invalid register %u\n", rd);
9337 if (rd == 0)
9338 err += efunc(pc, "cannot write to %r0\n");
9339 break;
9340 case DIF_OP_ULDSB:
9341 case DIF_OP_ULDSH:
9342 case DIF_OP_ULDSW:
9343 case DIF_OP_ULDUB:
9344 case DIF_OP_ULDUH:
9345 case DIF_OP_ULDUW:
9346 case DIF_OP_ULDX:
9347 if (r1 >= nregs)
9348 err += efunc(pc, "invalid register %u\n", r1);
9349 if (r2 != 0)
9350 err += efunc(pc, "non-zero reserved bits\n");
9351 if (rd >= nregs)
9352 err += efunc(pc, "invalid register %u\n", rd);
9353 if (rd == 0)
9354 err += efunc(pc, "cannot write to %r0\n");
9355 break;
9356 case DIF_OP_STB:
9357 case DIF_OP_STH:
9358 case DIF_OP_STW:
9359 case DIF_OP_STX:
9360 if (r1 >= nregs)
9361 err += efunc(pc, "invalid register %u\n", r1);
9362 if (r2 != 0)
9363 err += efunc(pc, "non-zero reserved bits\n");
9364 if (rd >= nregs)
9365 err += efunc(pc, "invalid register %u\n", rd);
9366 if (rd == 0)
9367 err += efunc(pc, "cannot write to 0 address\n");
9368 break;
9369 case DIF_OP_CMP:
9370 case DIF_OP_SCMP:
9371 if (r1 >= nregs)
9372 err += efunc(pc, "invalid register %u\n", r1);
9373 if (r2 >= nregs)
9374 err += efunc(pc, "invalid register %u\n", r2);
9375 if (rd != 0)
9376 err += efunc(pc, "non-zero reserved bits\n");
9377 break;
9378 case DIF_OP_TST:
9379 if (r1 >= nregs)
9380 err += efunc(pc, "invalid register %u\n", r1);
9381 if (r2 != 0 || rd != 0)
9382 err += efunc(pc, "non-zero reserved bits\n");
9383 break;
9384 case DIF_OP_BA:
9385 case DIF_OP_BE:
9386 case DIF_OP_BNE:
9387 case DIF_OP_BG:
9388 case DIF_OP_BGU:
9389 case DIF_OP_BGE:
9390 case DIF_OP_BGEU:
9391 case DIF_OP_BL:
9392 case DIF_OP_BLU:
9393 case DIF_OP_BLE:
9394 case DIF_OP_BLEU:
9395 if (label >= dp->dtdo_len) {
9396 err += efunc(pc, "invalid branch target %u\n",
9397 label);
9398 }
9399 if (label <= pc) {
9400 err += efunc(pc, "backward branch to %u\n",
9401 label);
9402 }
9403 break;
9404 case DIF_OP_RET:
9405 if (r1 != 0 || r2 != 0)
9406 err += efunc(pc, "non-zero reserved bits\n");
9407 if (rd >= nregs)
9408 err += efunc(pc, "invalid register %u\n", rd);
9409 break;
9410 case DIF_OP_NOP:
9411 case DIF_OP_POPTS:
9412 case DIF_OP_FLUSHTS:
9413 if (r1 != 0 || r2 != 0 || rd != 0)
9414 err += efunc(pc, "non-zero reserved bits\n");
9415 break;
9416 case DIF_OP_SETX:
9417 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9418 err += efunc(pc, "invalid integer ref %u\n",
9419 DIF_INSTR_INTEGER(instr));
9420 }
9421 if (rd >= nregs)
9422 err += efunc(pc, "invalid register %u\n", rd);
9423 if (rd == 0)
9424 err += efunc(pc, "cannot write to %r0\n");
9425 break;
9426 case DIF_OP_SETS:
9427 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9428 err += efunc(pc, "invalid string ref %u\n",
9429 DIF_INSTR_STRING(instr));
9430 }
9431 if (rd >= nregs)
9432 err += efunc(pc, "invalid register %u\n", rd);
9433 if (rd == 0)
9434 err += efunc(pc, "cannot write to %r0\n");
9435 break;
9436 case DIF_OP_LDGA:
9437 case DIF_OP_LDTA:
9438 if (r1 > DIF_VAR_ARRAY_MAX)
9439 err += efunc(pc, "invalid array %u\n", r1);
9440 if (r2 >= nregs)
9441 err += efunc(pc, "invalid register %u\n", r2);
9442 if (rd >= nregs)
9443 err += efunc(pc, "invalid register %u\n", rd);
9444 if (rd == 0)
9445 err += efunc(pc, "cannot write to %r0\n");
9446 break;
9447 case DIF_OP_STGA:
9448 if (r1 > DIF_VAR_ARRAY_MAX)
9449 err += efunc(pc, "invalid array %u\n", r1);
9450 if (r2 >= nregs)
9451 err += efunc(pc, "invalid register %u\n", r2);
9452 if (rd >= nregs)
9453 err += efunc(pc, "invalid register %u\n", rd);
9454 dp->dtdo_destructive = 1;
9455 break;
9456 case DIF_OP_LDGS:
9457 case DIF_OP_LDTS:
9458 case DIF_OP_LDLS:
9459 case DIF_OP_LDGAA:
9460 case DIF_OP_LDTAA:
9461 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9462 err += efunc(pc, "invalid variable %u\n", v);
9463 if (rd >= nregs)
9464 err += efunc(pc, "invalid register %u\n", rd);
9465 if (rd == 0)
9466 err += efunc(pc, "cannot write to %r0\n");
9467 break;
9468 case DIF_OP_STGS:
9469 case DIF_OP_STTS:
9470 case DIF_OP_STLS:
9471 case DIF_OP_STGAA:
9472 case DIF_OP_STTAA:
9473 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9474 err += efunc(pc, "invalid variable %u\n", v);
9475 if (rs >= nregs)
9476 err += efunc(pc, "invalid register %u\n", rd);
9477 break;
9478 case DIF_OP_CALL:
9479 if (subr > DIF_SUBR_MAX)
9480 err += efunc(pc, "invalid subr %u\n", subr);
9481 if (rd >= nregs)
9482 err += efunc(pc, "invalid register %u\n", rd);
9483 if (rd == 0)
9484 err += efunc(pc, "cannot write to %r0\n");
9485
9486 if (subr == DIF_SUBR_COPYOUT ||
9487 subr == DIF_SUBR_COPYOUTSTR) {
9488 dp->dtdo_destructive = 1;
9489 }
9490
9491 if (subr == DIF_SUBR_GETF) {
9492 /*
9493 * If we have a getf() we need to record that
9494 * in our state. Note that our state can be
9495 * NULL if this is a helper -- but in that
9496 * case, the call to getf() is itself illegal,
9497 * and will be caught (slightly later) when
9498 * the helper is validated.
9499 */
9500 if (vstate->dtvs_state != NULL)
9501 vstate->dtvs_state->dts_getf++;
9502 }
9503
9504 break;
9505 case DIF_OP_PUSHTR:
9506 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
9507 err += efunc(pc, "invalid ref type %u\n", type);
9508 if (r2 >= nregs)
9509 err += efunc(pc, "invalid register %u\n", r2);
9510 if (rs >= nregs)
9511 err += efunc(pc, "invalid register %u\n", rs);
9512 break;
9513 case DIF_OP_PUSHTV:
9514 if (type != DIF_TYPE_CTF)
9515 err += efunc(pc, "invalid val type %u\n", type);
9516 if (r2 >= nregs)
9517 err += efunc(pc, "invalid register %u\n", r2);
9518 if (rs >= nregs)
9519 err += efunc(pc, "invalid register %u\n", rs);
9520 break;
9521 default:
9522 err += efunc(pc, "invalid opcode %u\n",
9523 DIF_INSTR_OP(instr));
9524 }
9525 }
9526
9527 if (dp->dtdo_len != 0 &&
9528 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
9529 err += efunc(dp->dtdo_len - 1,
9530 "expected 'ret' as last DIF instruction\n");
9531 }
9532
9533 if (!(dp->dtdo_rtype.dtdt_flags & (DIF_TF_BYREF | DIF_TF_BYUREF))) {
9534 /*
9535 * If we're not returning by reference, the size must be either
9536 * 0 or the size of one of the base types.
9537 */
9538 switch (dp->dtdo_rtype.dtdt_size) {
9539 case 0:
9540 case sizeof (uint8_t):
9541 case sizeof (uint16_t):
9542 case sizeof (uint32_t):
9543 case sizeof (uint64_t):
9544 break;
9545
9546 default:
9547 err += efunc(dp->dtdo_len - 1, "bad return size\n");
9548 }
9549 }
9550
9551 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
9552 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
9553 dtrace_diftype_t *vt, *et;
9554 uint_t id, ndx;
9555
9556 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
9557 v->dtdv_scope != DIFV_SCOPE_THREAD &&
9558 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
9559 err += efunc(i, "unrecognized variable scope %d\n",
9560 v->dtdv_scope);
9561 break;
9562 }
9563
9564 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
9565 v->dtdv_kind != DIFV_KIND_SCALAR) {
9566 err += efunc(i, "unrecognized variable type %d\n",
9567 v->dtdv_kind);
9568 break;
9569 }
9570
9571 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
9572 err += efunc(i, "%d exceeds variable id limit\n", id);
9573 break;
9574 }
9575
9576 if (id < DIF_VAR_OTHER_UBASE)
9577 continue;
9578
9579 /*
9580 * For user-defined variables, we need to check that this
9581 * definition is identical to any previous definition that we
9582 * encountered.
9583 */
9584 ndx = id - DIF_VAR_OTHER_UBASE;
9585
9586 switch (v->dtdv_scope) {
9587 case DIFV_SCOPE_GLOBAL:
9588 if (maxglobal == -1 || ndx > maxglobal)
9589 maxglobal = ndx;
9590
9591 if (ndx < vstate->dtvs_nglobals) {
9592 dtrace_statvar_t *svar;
9593
9594 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
9595 existing = &svar->dtsv_var;
9596 }
9597
9598 break;
9599
9600 case DIFV_SCOPE_THREAD:
9601 if (maxtlocal == -1 || ndx > maxtlocal)
9602 maxtlocal = ndx;
9603
9604 if (ndx < vstate->dtvs_ntlocals)
9605 existing = &vstate->dtvs_tlocals[ndx];
9606 break;
9607
9608 case DIFV_SCOPE_LOCAL:
9609 if (maxlocal == -1 || ndx > maxlocal)
9610 maxlocal = ndx;
9611
9612 if (ndx < vstate->dtvs_nlocals) {
9613 dtrace_statvar_t *svar;
9614
9615 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
9616 existing = &svar->dtsv_var;
9617 }
9618
9619 break;
9620 }
9621
9622 vt = &v->dtdv_type;
9623
9624 if (vt->dtdt_flags & DIF_TF_BYREF) {
9625 if (vt->dtdt_size == 0) {
9626 err += efunc(i, "zero-sized variable\n");
9627 break;
9628 }
9629
9630 if ((v->dtdv_scope == DIFV_SCOPE_GLOBAL ||
9631 v->dtdv_scope == DIFV_SCOPE_LOCAL) &&
9632 vt->dtdt_size > dtrace_statvar_maxsize) {
9633 err += efunc(i, "oversized by-ref static\n");
9634 break;
9635 }
9636 }
9637
9638 if (existing == NULL || existing->dtdv_id == 0)
9639 continue;
9640
9641 ASSERT(existing->dtdv_id == v->dtdv_id);
9642 ASSERT(existing->dtdv_scope == v->dtdv_scope);
9643
9644 if (existing->dtdv_kind != v->dtdv_kind)
9645 err += efunc(i, "%d changed variable kind\n", id);
9646
9647 et = &existing->dtdv_type;
9648
9649 if (vt->dtdt_flags != et->dtdt_flags) {
9650 err += efunc(i, "%d changed variable type flags\n", id);
9651 break;
9652 }
9653
9654 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
9655 err += efunc(i, "%d changed variable type size\n", id);
9656 break;
9657 }
9658 }
9659
9660 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9661 dif_instr_t instr = dp->dtdo_buf[pc];
9662
9663 uint_t v = DIF_INSTR_VAR(instr);
9664 uint_t op = DIF_INSTR_OP(instr);
9665
9666 switch (op) {
9667 case DIF_OP_LDGS:
9668 case DIF_OP_LDGAA:
9669 case DIF_OP_STGS:
9670 case DIF_OP_STGAA:
9671 if (v > DIF_VAR_OTHER_UBASE + maxglobal)
9672 err += efunc(pc, "invalid variable %u\n", v);
9673 break;
9674 case DIF_OP_LDTS:
9675 case DIF_OP_LDTAA:
9676 case DIF_OP_STTS:
9677 case DIF_OP_STTAA:
9678 if (v > DIF_VAR_OTHER_UBASE + maxtlocal)
9679 err += efunc(pc, "invalid variable %u\n", v);
9680 break;
9681 case DIF_OP_LDLS:
9682 case DIF_OP_STLS:
9683 if (v > DIF_VAR_OTHER_UBASE + maxlocal)
9684 err += efunc(pc, "invalid variable %u\n", v);
9685 break;
9686 default:
9687 break;
9688 }
9689 }
9690
9691 return (err);
9692 }
9693
9694 /*
9695 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
9696 * are much more constrained than normal DIFOs. Specifically, they may
9697 * not:
9698 *
9699 * 1. Make calls to subroutines other than copyin(), copyinstr() or
9700 * miscellaneous string routines
9701 * 2. Access DTrace variables other than the args[] array, and the
9702 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9703 * 3. Have thread-local variables.
9704 * 4. Have dynamic variables.
9705 */
9706 static int
9707 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9708 {
9709 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9710 int err = 0;
9711 uint_t pc;
9712
9713 for (pc = 0; pc < dp->dtdo_len; pc++) {
9714 dif_instr_t instr = dp->dtdo_buf[pc];
9715
9716 uint_t v = DIF_INSTR_VAR(instr);
9717 uint_t subr = DIF_INSTR_SUBR(instr);
9718 uint_t op = DIF_INSTR_OP(instr);
9719
9720 switch (op) {
9721 case DIF_OP_OR:
9722 case DIF_OP_XOR:
9723 case DIF_OP_AND:
9724 case DIF_OP_SLL:
9725 case DIF_OP_SRL:
9726 case DIF_OP_SRA:
9727 case DIF_OP_SUB:
9728 case DIF_OP_ADD:
9729 case DIF_OP_MUL:
9730 case DIF_OP_SDIV:
9731 case DIF_OP_UDIV:
9732 case DIF_OP_SREM:
9733 case DIF_OP_UREM:
9734 case DIF_OP_COPYS:
9735 case DIF_OP_NOT:
9736 case DIF_OP_MOV:
9737 case DIF_OP_RLDSB:
9738 case DIF_OP_RLDSH:
9739 case DIF_OP_RLDSW:
9740 case DIF_OP_RLDUB:
9741 case DIF_OP_RLDUH:
9742 case DIF_OP_RLDUW:
9743 case DIF_OP_RLDX:
9744 case DIF_OP_ULDSB:
9745 case DIF_OP_ULDSH:
9746 case DIF_OP_ULDSW:
9747 case DIF_OP_ULDUB:
9748 case DIF_OP_ULDUH:
9749 case DIF_OP_ULDUW:
9750 case DIF_OP_ULDX:
9751 case DIF_OP_STB:
9752 case DIF_OP_STH:
9753 case DIF_OP_STW:
9754 case DIF_OP_STX:
9755 case DIF_OP_ALLOCS:
9756 case DIF_OP_CMP:
9757 case DIF_OP_SCMP:
9758 case DIF_OP_TST:
9759 case DIF_OP_BA:
9760 case DIF_OP_BE:
9761 case DIF_OP_BNE:
9762 case DIF_OP_BG:
9763 case DIF_OP_BGU:
9764 case DIF_OP_BGE:
9765 case DIF_OP_BGEU:
9766 case DIF_OP_BL:
9767 case DIF_OP_BLU:
9768 case DIF_OP_BLE:
9769 case DIF_OP_BLEU:
9770 case DIF_OP_RET:
9771 case DIF_OP_NOP:
9772 case DIF_OP_POPTS:
9773 case DIF_OP_FLUSHTS:
9774 case DIF_OP_SETX:
9775 case DIF_OP_SETS:
9776 case DIF_OP_LDGA:
9777 case DIF_OP_LDLS:
9778 case DIF_OP_STGS:
9779 case DIF_OP_STLS:
9780 case DIF_OP_PUSHTR:
9781 case DIF_OP_PUSHTV:
9782 break;
9783
9784 case DIF_OP_LDGS:
9785 if (v >= DIF_VAR_OTHER_UBASE)
9786 break;
9787
9788 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9789 break;
9790
9791 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9792 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9793 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9794 v == DIF_VAR_UID || v == DIF_VAR_GID)
9795 break;
9796
9797 err += efunc(pc, "illegal variable %u\n", v);
9798 break;
9799
9800 case DIF_OP_LDTA:
9801 if (v < DIF_VAR_OTHER_UBASE) {
9802 err += efunc(pc, "illegal variable load\n");
9803 break;
9804 }
9805 /* FALLTHROUGH */
9806 case DIF_OP_LDTS:
9807 case DIF_OP_LDGAA:
9808 case DIF_OP_LDTAA:
9809 err += efunc(pc, "illegal dynamic variable load\n");
9810 break;
9811
9812 case DIF_OP_STGA:
9813 if (v < DIF_VAR_OTHER_UBASE) {
9814 err += efunc(pc, "illegal variable store\n");
9815 break;
9816 }
9817 /* FALLTHROUGH */
9818 case DIF_OP_STTS:
9819 case DIF_OP_STGAA:
9820 case DIF_OP_STTAA:
9821 err += efunc(pc, "illegal dynamic variable store\n");
9822 break;
9823
9824 case DIF_OP_CALL:
9825 if (subr == DIF_SUBR_ALLOCA ||
9826 subr == DIF_SUBR_BCOPY ||
9827 subr == DIF_SUBR_COPYIN ||
9828 subr == DIF_SUBR_COPYINTO ||
9829 subr == DIF_SUBR_COPYINSTR ||
9830 subr == DIF_SUBR_INDEX ||
9831 subr == DIF_SUBR_INET_NTOA ||
9832 subr == DIF_SUBR_INET_NTOA6 ||
9833 subr == DIF_SUBR_INET_NTOP ||
9834 subr == DIF_SUBR_JSON ||
9835 subr == DIF_SUBR_LLTOSTR ||
9836 subr == DIF_SUBR_STRTOLL ||
9837 subr == DIF_SUBR_RINDEX ||
9838 subr == DIF_SUBR_STRCHR ||
9839 subr == DIF_SUBR_STRJOIN ||
9840 subr == DIF_SUBR_STRRCHR ||
9841 subr == DIF_SUBR_STRSTR ||
9842 subr == DIF_SUBR_HTONS ||
9843 subr == DIF_SUBR_HTONL ||
9844 subr == DIF_SUBR_HTONLL ||
9845 subr == DIF_SUBR_NTOHS ||
9846 subr == DIF_SUBR_NTOHL ||
9847 subr == DIF_SUBR_NTOHLL)
9848 break;
9849
9850 err += efunc(pc, "invalid subr %u\n", subr);
9851 break;
9852
9853 default:
9854 err += efunc(pc, "invalid opcode %u\n",
9855 DIF_INSTR_OP(instr));
9856 }
9857 }
9858
9859 return (err);
9860 }
9861
9862 /*
9863 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9864 * basis; 0 if not.
9865 */
9866 static int
9867 dtrace_difo_cacheable(dtrace_difo_t *dp)
9868 {
9869 int i;
9870
9871 if (dp == NULL)
9872 return (0);
9873
9874 for (i = 0; i < dp->dtdo_varlen; i++) {
9875 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9876
9877 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9878 continue;
9879
9880 switch (v->dtdv_id) {
9881 case DIF_VAR_CURTHREAD:
9882 case DIF_VAR_PID:
9883 case DIF_VAR_TID:
9884 case DIF_VAR_EXECNAME:
9885 case DIF_VAR_ZONENAME:
9886 break;
9887
9888 default:
9889 return (0);
9890 }
9891 }
9892
9893 /*
9894 * This DIF object may be cacheable. Now we need to look for any
9895 * array loading instructions, any memory loading instructions, or
9896 * any stores to thread-local variables.
9897 */
9898 for (i = 0; i < dp->dtdo_len; i++) {
9899 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9900
9901 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9902 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9903 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9904 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9905 return (0);
9906 }
9907
9908 return (1);
9909 }
9910
9911 static void
9912 dtrace_difo_hold(dtrace_difo_t *dp)
9913 {
9914 int i;
9915
9916 ASSERT(MUTEX_HELD(&dtrace_lock));
9917
9918 dp->dtdo_refcnt++;
9919 ASSERT(dp->dtdo_refcnt != 0);
9920
9921 /*
9922 * We need to check this DIF object for references to the variable
9923 * DIF_VAR_VTIMESTAMP.
9924 */
9925 for (i = 0; i < dp->dtdo_varlen; i++) {
9926 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9927
9928 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9929 continue;
9930
9931 if (dtrace_vtime_references++ == 0)
9932 dtrace_vtime_enable();
9933 }
9934 }
9935
9936 /*
9937 * This routine calculates the dynamic variable chunksize for a given DIF
9938 * object. The calculation is not fool-proof, and can probably be tricked by
9939 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9940 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9941 * if a dynamic variable size exceeds the chunksize.
9942 */
9943 static void
9944 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9945 {
9946 uint64_t sval;
9947 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9948 const dif_instr_t *text = dp->dtdo_buf;
9949 uint_t pc, srd = 0;
9950 uint_t ttop = 0;
9951 size_t size, ksize;
9952 uint_t id, i;
9953
9954 for (pc = 0; pc < dp->dtdo_len; pc++) {
9955 dif_instr_t instr = text[pc];
9956 uint_t op = DIF_INSTR_OP(instr);
9957 uint_t rd = DIF_INSTR_RD(instr);
9958 uint_t r1 = DIF_INSTR_R1(instr);
9959 uint_t nkeys = 0;
9960 uchar_t scope;
9961
9962 dtrace_key_t *key = tupregs;
9963
9964 switch (op) {
9965 case DIF_OP_SETX:
9966 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9967 srd = rd;
9968 continue;
9969
9970 case DIF_OP_STTS:
9971 key = &tupregs[DIF_DTR_NREGS];
9972 key[0].dttk_size = 0;
9973 key[1].dttk_size = 0;
9974 nkeys = 2;
9975 scope = DIFV_SCOPE_THREAD;
9976 break;
9977
9978 case DIF_OP_STGAA:
9979 case DIF_OP_STTAA:
9980 nkeys = ttop;
9981
9982 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9983 key[nkeys++].dttk_size = 0;
9984
9985 key[nkeys++].dttk_size = 0;
9986
9987 if (op == DIF_OP_STTAA) {
9988 scope = DIFV_SCOPE_THREAD;
9989 } else {
9990 scope = DIFV_SCOPE_GLOBAL;
9991 }
9992
9993 break;
9994
9995 case DIF_OP_PUSHTR:
9996 if (ttop == DIF_DTR_NREGS)
9997 return;
9998
9999 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
10000 /*
10001 * If the register for the size of the "pushtr"
10002 * is %r0 (or the value is 0) and the type is
10003 * a string, we'll use the system-wide default
10004 * string size.
10005 */
10006 tupregs[ttop++].dttk_size =
10007 dtrace_strsize_default;
10008 } else {
10009 if (srd == 0)
10010 return;
10011
10012 if (sval > LONG_MAX)
10013 return;
10014
10015 tupregs[ttop++].dttk_size = sval;
10016 }
10017
10018 break;
10019
10020 case DIF_OP_PUSHTV:
10021 if (ttop == DIF_DTR_NREGS)
10022 return;
10023
10024 tupregs[ttop++].dttk_size = 0;
10025 break;
10026
10027 case DIF_OP_FLUSHTS:
10028 ttop = 0;
10029 break;
10030
10031 case DIF_OP_POPTS:
10032 if (ttop != 0)
10033 ttop--;
10034 break;
10035 }
10036
10037 sval = 0;
10038 srd = 0;
10039
10040 if (nkeys == 0)
10041 continue;
10042
10043 /*
10044 * We have a dynamic variable allocation; calculate its size.
10045 */
10046 for (ksize = 0, i = 0; i < nkeys; i++)
10047 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
10048
10049 size = sizeof (dtrace_dynvar_t);
10050 size += sizeof (dtrace_key_t) * (nkeys - 1);
10051 size += ksize;
10052
10053 /*
10054 * Now we need to determine the size of the stored data.
10055 */
10056 id = DIF_INSTR_VAR(instr);
10057
10058 for (i = 0; i < dp->dtdo_varlen; i++) {
10059 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10060
10061 if (v->dtdv_id == id && v->dtdv_scope == scope) {
10062 size += v->dtdv_type.dtdt_size;
10063 break;
10064 }
10065 }
10066
10067 if (i == dp->dtdo_varlen)
10068 return;
10069
10070 /*
10071 * We have the size. If this is larger than the chunk size
10072 * for our dynamic variable state, reset the chunk size.
10073 */
10074 size = P2ROUNDUP(size, sizeof (uint64_t));
10075
10076 /*
10077 * Before setting the chunk size, check that we're not going
10078 * to set it to a negative value...
10079 */
10080 if (size > LONG_MAX)
10081 return;
10082
10083 /*
10084 * ...and make certain that we didn't badly overflow.
10085 */
10086 if (size < ksize || size < sizeof (dtrace_dynvar_t))
10087 return;
10088
10089 if (size > vstate->dtvs_dynvars.dtds_chunksize)
10090 vstate->dtvs_dynvars.dtds_chunksize = size;
10091 }
10092 }
10093
10094 static void
10095 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10096 {
10097 int i, oldsvars, osz, nsz, otlocals, ntlocals;
10098 uint_t id;
10099
10100 ASSERT(MUTEX_HELD(&dtrace_lock));
10101 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
10102
10103 for (i = 0; i < dp->dtdo_varlen; i++) {
10104 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10105 dtrace_statvar_t *svar, ***svarp;
10106 size_t dsize = 0;
10107 uint8_t scope = v->dtdv_scope;
10108 int *np;
10109
10110 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10111 continue;
10112
10113 id -= DIF_VAR_OTHER_UBASE;
10114
10115 switch (scope) {
10116 case DIFV_SCOPE_THREAD:
10117 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
10118 dtrace_difv_t *tlocals;
10119
10120 if ((ntlocals = (otlocals << 1)) == 0)
10121 ntlocals = 1;
10122
10123 osz = otlocals * sizeof (dtrace_difv_t);
10124 nsz = ntlocals * sizeof (dtrace_difv_t);
10125
10126 tlocals = kmem_zalloc(nsz, KM_SLEEP);
10127
10128 if (osz != 0) {
10129 bcopy(vstate->dtvs_tlocals,
10130 tlocals, osz);
10131 kmem_free(vstate->dtvs_tlocals, osz);
10132 }
10133
10134 vstate->dtvs_tlocals = tlocals;
10135 vstate->dtvs_ntlocals = ntlocals;
10136 }
10137
10138 vstate->dtvs_tlocals[id] = *v;
10139 continue;
10140
10141 case DIFV_SCOPE_LOCAL:
10142 np = &vstate->dtvs_nlocals;
10143 svarp = &vstate->dtvs_locals;
10144
10145 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10146 dsize = NCPU * (v->dtdv_type.dtdt_size +
10147 sizeof (uint64_t));
10148 else
10149 dsize = NCPU * sizeof (uint64_t);
10150
10151 break;
10152
10153 case DIFV_SCOPE_GLOBAL:
10154 np = &vstate->dtvs_nglobals;
10155 svarp = &vstate->dtvs_globals;
10156
10157 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10158 dsize = v->dtdv_type.dtdt_size +
10159 sizeof (uint64_t);
10160
10161 break;
10162
10163 default:
10164 ASSERT(0);
10165 }
10166
10167 while (id >= (oldsvars = *np)) {
10168 dtrace_statvar_t **statics;
10169 int newsvars, oldsize, newsize;
10170
10171 if ((newsvars = (oldsvars << 1)) == 0)
10172 newsvars = 1;
10173
10174 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
10175 newsize = newsvars * sizeof (dtrace_statvar_t *);
10176
10177 statics = kmem_zalloc(newsize, KM_SLEEP);
10178
10179 if (oldsize != 0) {
10180 bcopy(*svarp, statics, oldsize);
10181 kmem_free(*svarp, oldsize);
10182 }
10183
10184 *svarp = statics;
10185 *np = newsvars;
10186 }
10187
10188 if ((svar = (*svarp)[id]) == NULL) {
10189 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
10190 svar->dtsv_var = *v;
10191
10192 if ((svar->dtsv_size = dsize) != 0) {
10193 svar->dtsv_data = (uint64_t)(uintptr_t)
10194 kmem_zalloc(dsize, KM_SLEEP);
10195 }
10196
10197 (*svarp)[id] = svar;
10198 }
10199
10200 svar->dtsv_refcnt++;
10201 }
10202
10203 dtrace_difo_chunksize(dp, vstate);
10204 dtrace_difo_hold(dp);
10205 }
10206
10207 static dtrace_difo_t *
10208 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10209 {
10210 dtrace_difo_t *new;
10211 size_t sz;
10212
10213 ASSERT(dp->dtdo_buf != NULL);
10214 ASSERT(dp->dtdo_refcnt != 0);
10215
10216 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
10217
10218 ASSERT(dp->dtdo_buf != NULL);
10219 sz = dp->dtdo_len * sizeof (dif_instr_t);
10220 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
10221 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
10222 new->dtdo_len = dp->dtdo_len;
10223
10224 if (dp->dtdo_strtab != NULL) {
10225 ASSERT(dp->dtdo_strlen != 0);
10226 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
10227 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
10228 new->dtdo_strlen = dp->dtdo_strlen;
10229 }
10230
10231 if (dp->dtdo_inttab != NULL) {
10232 ASSERT(dp->dtdo_intlen != 0);
10233 sz = dp->dtdo_intlen * sizeof (uint64_t);
10234 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
10235 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
10236 new->dtdo_intlen = dp->dtdo_intlen;
10237 }
10238
10239 if (dp->dtdo_vartab != NULL) {
10240 ASSERT(dp->dtdo_varlen != 0);
10241 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
10242 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
10243 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
10244 new->dtdo_varlen = dp->dtdo_varlen;
10245 }
10246
10247 dtrace_difo_init(new, vstate);
10248 return (new);
10249 }
10250
10251 static void
10252 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10253 {
10254 int i;
10255
10256 ASSERT(dp->dtdo_refcnt == 0);
10257
10258 for (i = 0; i < dp->dtdo_varlen; i++) {
10259 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10260 dtrace_statvar_t *svar, **svarp;
10261 uint_t id;
10262 uint8_t scope = v->dtdv_scope;
10263 int *np;
10264
10265 switch (scope) {
10266 case DIFV_SCOPE_THREAD:
10267 continue;
10268
10269 case DIFV_SCOPE_LOCAL:
10270 np = &vstate->dtvs_nlocals;
10271 svarp = vstate->dtvs_locals;
10272 break;
10273
10274 case DIFV_SCOPE_GLOBAL:
10275 np = &vstate->dtvs_nglobals;
10276 svarp = vstate->dtvs_globals;
10277 break;
10278
10279 default:
10280 ASSERT(0);
10281 }
10282
10283 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10284 continue;
10285
10286 id -= DIF_VAR_OTHER_UBASE;
10287 ASSERT(id < *np);
10288
10289 svar = svarp[id];
10290 ASSERT(svar != NULL);
10291 ASSERT(svar->dtsv_refcnt > 0);
10292
10293 if (--svar->dtsv_refcnt > 0)
10294 continue;
10295
10296 if (svar->dtsv_size != 0) {
10297 ASSERT(svar->dtsv_data != NULL);
10298 kmem_free((void *)(uintptr_t)svar->dtsv_data,
10299 svar->dtsv_size);
10300 }
10301
10302 kmem_free(svar, sizeof (dtrace_statvar_t));
10303 svarp[id] = NULL;
10304 }
10305
10306 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
10307 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
10308 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
10309 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
10310
10311 kmem_free(dp, sizeof (dtrace_difo_t));
10312 }
10313
10314 static void
10315 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10316 {
10317 int i;
10318
10319 ASSERT(MUTEX_HELD(&dtrace_lock));
10320 ASSERT(dp->dtdo_refcnt != 0);
10321
10322 for (i = 0; i < dp->dtdo_varlen; i++) {
10323 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10324
10325 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10326 continue;
10327
10328 ASSERT(dtrace_vtime_references > 0);
10329 if (--dtrace_vtime_references == 0)
10330 dtrace_vtime_disable();
10331 }
10332
10333 if (--dp->dtdo_refcnt == 0)
10334 dtrace_difo_destroy(dp, vstate);
10335 }
10336
10337 /*
10338 * DTrace Format Functions
10339 */
10340 static uint16_t
10341 dtrace_format_add(dtrace_state_t *state, char *str)
10342 {
10343 char *fmt, **new;
10344 uint16_t ndx, len = strlen(str) + 1;
10345
10346 fmt = kmem_zalloc(len, KM_SLEEP);
10347 bcopy(str, fmt, len);
10348
10349 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
10350 if (state->dts_formats[ndx] == NULL) {
10351 state->dts_formats[ndx] = fmt;
10352 return (ndx + 1);
10353 }
10354 }
10355
10356 if (state->dts_nformats == USHRT_MAX) {
10357 /*
10358 * This is only likely if a denial-of-service attack is being
10359 * attempted. As such, it's okay to fail silently here.
10360 */
10361 kmem_free(fmt, len);
10362 return (0);
10363 }
10364
10365 /*
10366 * For simplicity, we always resize the formats array to be exactly the
10367 * number of formats.
10368 */
10369 ndx = state->dts_nformats++;
10370 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10371
10372 if (state->dts_formats != NULL) {
10373 ASSERT(ndx != 0);
10374 bcopy(state->dts_formats, new, ndx * sizeof (char *));
10375 kmem_free(state->dts_formats, ndx * sizeof (char *));
10376 }
10377
10378 state->dts_formats = new;
10379 state->dts_formats[ndx] = fmt;
10380
10381 return (ndx + 1);
10382 }
10383
10384 static void
10385 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10386 {
10387 char *fmt;
10388
10389 ASSERT(state->dts_formats != NULL);
10390 ASSERT(format <= state->dts_nformats);
10391 ASSERT(state->dts_formats[format - 1] != NULL);
10392
10393 fmt = state->dts_formats[format - 1];
10394 kmem_free(fmt, strlen(fmt) + 1);
10395 state->dts_formats[format - 1] = NULL;
10396 }
10397
10398 static void
10399 dtrace_format_destroy(dtrace_state_t *state)
10400 {
10401 int i;
10402
10403 if (state->dts_nformats == 0) {
10404 ASSERT(state->dts_formats == NULL);
10405 return;
10406 }
10407
10408 ASSERT(state->dts_formats != NULL);
10409
10410 for (i = 0; i < state->dts_nformats; i++) {
10411 char *fmt = state->dts_formats[i];
10412
10413 if (fmt == NULL)
10414 continue;
10415
10416 kmem_free(fmt, strlen(fmt) + 1);
10417 }
10418
10419 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10420 state->dts_nformats = 0;
10421 state->dts_formats = NULL;
10422 }
10423
10424 /*
10425 * DTrace Predicate Functions
10426 */
10427 static dtrace_predicate_t *
10428 dtrace_predicate_create(dtrace_difo_t *dp)
10429 {
10430 dtrace_predicate_t *pred;
10431
10432 ASSERT(MUTEX_HELD(&dtrace_lock));
10433 ASSERT(dp->dtdo_refcnt != 0);
10434
10435 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10436 pred->dtp_difo = dp;
10437 pred->dtp_refcnt = 1;
10438
10439 if (!dtrace_difo_cacheable(dp))
10440 return (pred);
10441
10442 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10443 /*
10444 * This is only theoretically possible -- we have had 2^32
10445 * cacheable predicates on this machine. We cannot allow any
10446 * more predicates to become cacheable: as unlikely as it is,
10447 * there may be a thread caching a (now stale) predicate cache
10448 * ID. (N.B.: the temptation is being successfully resisted to
10449 * have this cmn_err() "Holy shit -- we executed this code!")
10450 */
10451 return (pred);
10452 }
10453
10454 pred->dtp_cacheid = dtrace_predcache_id++;
10455
10456 return (pred);
10457 }
10458
10459 static void
10460 dtrace_predicate_hold(dtrace_predicate_t *pred)
10461 {
10462 ASSERT(MUTEX_HELD(&dtrace_lock));
10463 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10464 ASSERT(pred->dtp_refcnt > 0);
10465
10466 pred->dtp_refcnt++;
10467 }
10468
10469 static void
10470 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10471 {
10472 dtrace_difo_t *dp = pred->dtp_difo;
10473
10474 ASSERT(MUTEX_HELD(&dtrace_lock));
10475 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
10476 ASSERT(pred->dtp_refcnt > 0);
10477
10478 if (--pred->dtp_refcnt == 0) {
10479 dtrace_difo_release(pred->dtp_difo, vstate);
10480 kmem_free(pred, sizeof (dtrace_predicate_t));
10481 }
10482 }
10483
10484 /*
10485 * DTrace Action Description Functions
10486 */
10487 static dtrace_actdesc_t *
10488 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
10489 uint64_t uarg, uint64_t arg)
10490 {
10491 dtrace_actdesc_t *act;
10492
10493 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
10494 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
10495
10496 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
10497 act->dtad_kind = kind;
10498 act->dtad_ntuple = ntuple;
10499 act->dtad_uarg = uarg;
10500 act->dtad_arg = arg;
10501 act->dtad_refcnt = 1;
10502
10503 return (act);
10504 }
10505
10506 static void
10507 dtrace_actdesc_hold(dtrace_actdesc_t *act)
10508 {
10509 ASSERT(act->dtad_refcnt >= 1);
10510 act->dtad_refcnt++;
10511 }
10512
10513 static void
10514 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
10515 {
10516 dtrace_actkind_t kind = act->dtad_kind;
10517 dtrace_difo_t *dp;
10518
10519 ASSERT(act->dtad_refcnt >= 1);
10520
10521 if (--act->dtad_refcnt != 0)
10522 return;
10523
10524 if ((dp = act->dtad_difo) != NULL)
10525 dtrace_difo_release(dp, vstate);
10526
10527 if (DTRACEACT_ISPRINTFLIKE(kind)) {
10528 char *str = (char *)(uintptr_t)act->dtad_arg;
10529
10530 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
10531 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
10532
10533 if (str != NULL)
10534 kmem_free(str, strlen(str) + 1);
10535 }
10536
10537 kmem_free(act, sizeof (dtrace_actdesc_t));
10538 }
10539
10540 /*
10541 * DTrace ECB Functions
10542 */
10543 static dtrace_ecb_t *
10544 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
10545 {
10546 dtrace_ecb_t *ecb;
10547 dtrace_epid_t epid;
10548
10549 ASSERT(MUTEX_HELD(&dtrace_lock));
10550
10551 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
10552 ecb->dte_predicate = NULL;
10553 ecb->dte_probe = probe;
10554
10555 /*
10556 * The default size is the size of the default action: recording
10557 * the header.
10558 */
10559 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
10560 ecb->dte_alignment = sizeof (dtrace_epid_t);
10561
10562 epid = state->dts_epid++;
10563
10564 if (epid - 1 >= state->dts_necbs) {
10565 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
10566 int necbs = state->dts_necbs << 1;
10567
10568 ASSERT(epid == state->dts_necbs + 1);
10569
10570 if (necbs == 0) {
10571 ASSERT(oecbs == NULL);
10572 necbs = 1;
10573 }
10574
10575 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
10576
10577 if (oecbs != NULL)
10578 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
10579
10580 dtrace_membar_producer();
10581 state->dts_ecbs = ecbs;
10582
10583 if (oecbs != NULL) {
10584 /*
10585 * If this state is active, we must dtrace_sync()
10586 * before we can free the old dts_ecbs array: we're
10587 * coming in hot, and there may be active ring
10588 * buffer processing (which indexes into the dts_ecbs
10589 * array) on another CPU.
10590 */
10591 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
10592 dtrace_sync();
10593
10594 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
10595 }
10596
10597 dtrace_membar_producer();
10598 state->dts_necbs = necbs;
10599 }
10600
10601 ecb->dte_state = state;
10602
10603 ASSERT(state->dts_ecbs[epid - 1] == NULL);
10604 dtrace_membar_producer();
10605 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
10606
10607 return (ecb);
10608 }
10609
10610 static int
10611 dtrace_ecb_enable(dtrace_ecb_t *ecb)
10612 {
10613 dtrace_probe_t *probe = ecb->dte_probe;
10614
10615 ASSERT(MUTEX_HELD(&cpu_lock));
10616 ASSERT(MUTEX_HELD(&dtrace_lock));
10617 ASSERT(ecb->dte_next == NULL);
10618
10619 if (probe == NULL) {
10620 /*
10621 * This is the NULL probe -- there's nothing to do.
10622 */
10623 return (0);
10624 }
10625
10626 if (probe->dtpr_ecb == NULL) {
10627 dtrace_provider_t *prov = probe->dtpr_provider;
10628
10629 /*
10630 * We're the first ECB on this probe.
10631 */
10632 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
10633
10634 if (ecb->dte_predicate != NULL)
10635 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
10636
10637 return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
10638 probe->dtpr_id, probe->dtpr_arg));
10639 } else {
10640 /*
10641 * This probe is already active. Swing the last pointer to
10642 * point to the new ECB, and issue a dtrace_sync() to assure
10643 * that all CPUs have seen the change.
10644 */
10645 ASSERT(probe->dtpr_ecb_last != NULL);
10646 probe->dtpr_ecb_last->dte_next = ecb;
10647 probe->dtpr_ecb_last = ecb;
10648 probe->dtpr_predcache = 0;
10649
10650 dtrace_sync();
10651 return (0);
10652 }
10653 }
10654
10655 static int
10656 dtrace_ecb_resize(dtrace_ecb_t *ecb)
10657 {
10658 dtrace_action_t *act;
10659 uint32_t curneeded = UINT32_MAX;
10660 uint32_t aggbase = UINT32_MAX;
10661
10662 /*
10663 * If we record anything, we always record the dtrace_rechdr_t. (And
10664 * we always record it first.)
10665 */
10666 ecb->dte_size = sizeof (dtrace_rechdr_t);
10667 ecb->dte_alignment = sizeof (dtrace_epid_t);
10668
10669 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10670 dtrace_recdesc_t *rec = &act->dta_rec;
10671 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
10672
10673 ecb->dte_alignment = MAX(ecb->dte_alignment,
10674 rec->dtrd_alignment);
10675
10676 if (DTRACEACT_ISAGG(act->dta_kind)) {
10677 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10678
10679 ASSERT(rec->dtrd_size != 0);
10680 ASSERT(agg->dtag_first != NULL);
10681 ASSERT(act->dta_prev->dta_intuple);
10682 ASSERT(aggbase != UINT32_MAX);
10683 ASSERT(curneeded != UINT32_MAX);
10684
10685 agg->dtag_base = aggbase;
10686
10687 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10688 rec->dtrd_offset = curneeded;
10689 if (curneeded + rec->dtrd_size < curneeded)
10690 return (EINVAL);
10691 curneeded += rec->dtrd_size;
10692 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
10693
10694 aggbase = UINT32_MAX;
10695 curneeded = UINT32_MAX;
10696 } else if (act->dta_intuple) {
10697 if (curneeded == UINT32_MAX) {
10698 /*
10699 * This is the first record in a tuple. Align
10700 * curneeded to be at offset 4 in an 8-byte
10701 * aligned block.
10702 */
10703 ASSERT(act->dta_prev == NULL ||
10704 !act->dta_prev->dta_intuple);
10705 ASSERT3U(aggbase, ==, UINT32_MAX);
10706 curneeded = P2PHASEUP(ecb->dte_size,
10707 sizeof (uint64_t), sizeof (dtrace_aggid_t));
10708
10709 aggbase = curneeded - sizeof (dtrace_aggid_t);
10710 ASSERT(IS_P2ALIGNED(aggbase,
10711 sizeof (uint64_t)));
10712 }
10713 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10714 rec->dtrd_offset = curneeded;
10715 if (curneeded + rec->dtrd_size < curneeded)
10716 return (EINVAL);
10717 curneeded += rec->dtrd_size;
10718 } else {
10719 /* tuples must be followed by an aggregation */
10720 ASSERT(act->dta_prev == NULL ||
10721 !act->dta_prev->dta_intuple);
10722
10723 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10724 rec->dtrd_alignment);
10725 rec->dtrd_offset = ecb->dte_size;
10726 if (ecb->dte_size + rec->dtrd_size < ecb->dte_size)
10727 return (EINVAL);
10728 ecb->dte_size += rec->dtrd_size;
10729 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10730 }
10731 }
10732
10733 if ((act = ecb->dte_action) != NULL &&
10734 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10735 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10736 /*
10737 * If the size is still sizeof (dtrace_rechdr_t), then all
10738 * actions store no data; set the size to 0.
10739 */
10740 ecb->dte_size = 0;
10741 }
10742
10743 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10744 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10745 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10746 ecb->dte_needed);
10747 return (0);
10748 }
10749
10750 static dtrace_action_t *
10751 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10752 {
10753 dtrace_aggregation_t *agg;
10754 size_t size = sizeof (uint64_t);
10755 int ntuple = desc->dtad_ntuple;
10756 dtrace_action_t *act;
10757 dtrace_recdesc_t *frec;
10758 dtrace_aggid_t aggid;
10759 dtrace_state_t *state = ecb->dte_state;
10760
10761 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10762 agg->dtag_ecb = ecb;
10763
10764 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10765
10766 switch (desc->dtad_kind) {
10767 case DTRACEAGG_MIN:
10768 agg->dtag_initial = INT64_MAX;
10769 agg->dtag_aggregate = dtrace_aggregate_min;
10770 break;
10771
10772 case DTRACEAGG_MAX:
10773 agg->dtag_initial = INT64_MIN;
10774 agg->dtag_aggregate = dtrace_aggregate_max;
10775 break;
10776
10777 case DTRACEAGG_COUNT:
10778 agg->dtag_aggregate = dtrace_aggregate_count;
10779 break;
10780
10781 case DTRACEAGG_QUANTIZE:
10782 agg->dtag_aggregate = dtrace_aggregate_quantize;
10783 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10784 sizeof (uint64_t);
10785 break;
10786
10787 case DTRACEAGG_LQUANTIZE: {
10788 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10789 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10790
10791 agg->dtag_initial = desc->dtad_arg;
10792 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10793
10794 if (step == 0 || levels == 0)
10795 goto err;
10796
10797 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10798 break;
10799 }
10800
10801 case DTRACEAGG_LLQUANTIZE: {
10802 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10803 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10804 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10805 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10806 int64_t v;
10807
10808 agg->dtag_initial = desc->dtad_arg;
10809 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10810
10811 if (factor < 2 || low >= high || nsteps < factor)
10812 goto err;
10813
10814 /*
10815 * Now check that the number of steps evenly divides a power
10816 * of the factor. (This assures both integer bucket size and
10817 * linearity within each magnitude.)
10818 */
10819 for (v = factor; v < nsteps; v *= factor)
10820 continue;
10821
10822 if ((v % nsteps) || (nsteps % factor))
10823 goto err;
10824
10825 size = (dtrace_aggregate_llquantize_bucket(factor,
10826 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10827 break;
10828 }
10829
10830 case DTRACEAGG_AVG:
10831 agg->dtag_aggregate = dtrace_aggregate_avg;
10832 size = sizeof (uint64_t) * 2;
10833 break;
10834
10835 case DTRACEAGG_STDDEV:
10836 agg->dtag_aggregate = dtrace_aggregate_stddev;
10837 size = sizeof (uint64_t) * 4;
10838 break;
10839
10840 case DTRACEAGG_SUM:
10841 agg->dtag_aggregate = dtrace_aggregate_sum;
10842 break;
10843
10844 default:
10845 goto err;
10846 }
10847
10848 agg->dtag_action.dta_rec.dtrd_size = size;
10849
10850 if (ntuple == 0)
10851 goto err;
10852
10853 /*
10854 * We must make sure that we have enough actions for the n-tuple.
10855 */
10856 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10857 if (DTRACEACT_ISAGG(act->dta_kind))
10858 break;
10859
10860 if (--ntuple == 0) {
10861 /*
10862 * This is the action with which our n-tuple begins.
10863 */
10864 agg->dtag_first = act;
10865 goto success;
10866 }
10867 }
10868
10869 /*
10870 * This n-tuple is short by ntuple elements. Return failure.
10871 */
10872 ASSERT(ntuple != 0);
10873 err:
10874 kmem_free(agg, sizeof (dtrace_aggregation_t));
10875 return (NULL);
10876
10877 success:
10878 /*
10879 * If the last action in the tuple has a size of zero, it's actually
10880 * an expression argument for the aggregating action.
10881 */
10882 ASSERT(ecb->dte_action_last != NULL);
10883 act = ecb->dte_action_last;
10884
10885 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10886 ASSERT(act->dta_difo != NULL);
10887
10888 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10889 agg->dtag_hasarg = 1;
10890 }
10891
10892 /*
10893 * We need to allocate an id for this aggregation.
10894 */
10895 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10896 VM_BESTFIT | VM_SLEEP);
10897
10898 if (aggid - 1 >= state->dts_naggregations) {
10899 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10900 dtrace_aggregation_t **aggs;
10901 int naggs = state->dts_naggregations << 1;
10902 int onaggs = state->dts_naggregations;
10903
10904 ASSERT(aggid == state->dts_naggregations + 1);
10905
10906 if (naggs == 0) {
10907 ASSERT(oaggs == NULL);
10908 naggs = 1;
10909 }
10910
10911 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10912
10913 if (oaggs != NULL) {
10914 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10915 kmem_free(oaggs, onaggs * sizeof (*aggs));
10916 }
10917
10918 state->dts_aggregations = aggs;
10919 state->dts_naggregations = naggs;
10920 }
10921
10922 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10923 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10924
10925 frec = &agg->dtag_first->dta_rec;
10926 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10927 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10928
10929 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10930 ASSERT(!act->dta_intuple);
10931 act->dta_intuple = 1;
10932 }
10933
10934 return (&agg->dtag_action);
10935 }
10936
10937 static void
10938 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10939 {
10940 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10941 dtrace_state_t *state = ecb->dte_state;
10942 dtrace_aggid_t aggid = agg->dtag_id;
10943
10944 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10945 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10946
10947 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10948 state->dts_aggregations[aggid - 1] = NULL;
10949
10950 kmem_free(agg, sizeof (dtrace_aggregation_t));
10951 }
10952
10953 static int
10954 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10955 {
10956 dtrace_action_t *action, *last;
10957 dtrace_difo_t *dp = desc->dtad_difo;
10958 uint32_t size = 0, align = sizeof (uint8_t), mask;
10959 uint16_t format = 0;
10960 dtrace_recdesc_t *rec;
10961 dtrace_state_t *state = ecb->dte_state;
10962 dtrace_optval_t *opt = state->dts_options, nframes, strsize;
10963 uint64_t arg = desc->dtad_arg;
10964
10965 ASSERT(MUTEX_HELD(&dtrace_lock));
10966 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10967
10968 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10969 /*
10970 * If this is an aggregating action, there must be neither
10971 * a speculate nor a commit on the action chain.
10972 */
10973 dtrace_action_t *act;
10974
10975 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10976 if (act->dta_kind == DTRACEACT_COMMIT)
10977 return (EINVAL);
10978
10979 if (act->dta_kind == DTRACEACT_SPECULATE)
10980 return (EINVAL);
10981 }
10982
10983 action = dtrace_ecb_aggregation_create(ecb, desc);
10984
10985 if (action == NULL)
10986 return (EINVAL);
10987 } else {
10988 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10989 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10990 dp != NULL && dp->dtdo_destructive)) {
10991 state->dts_destructive = 1;
10992 }
10993
10994 switch (desc->dtad_kind) {
10995 case DTRACEACT_PRINTF:
10996 case DTRACEACT_PRINTA:
10997 case DTRACEACT_SYSTEM:
10998 case DTRACEACT_FREOPEN:
10999 case DTRACEACT_DIFEXPR:
11000 /*
11001 * We know that our arg is a string -- turn it into a
11002 * format.
11003 */
11004 if (arg == NULL) {
11005 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
11006 desc->dtad_kind == DTRACEACT_DIFEXPR);
11007 format = 0;
11008 } else {
11009 ASSERT(arg != NULL);
11010 ASSERT(arg > KERNELBASE);
11011 format = dtrace_format_add(state,
11012 (char *)(uintptr_t)arg);
11013 }
11014
11015 /*FALLTHROUGH*/
11016 case DTRACEACT_LIBACT:
11017 case DTRACEACT_TRACEMEM:
11018 case DTRACEACT_TRACEMEM_DYNSIZE:
11019 if (dp == NULL)
11020 return (EINVAL);
11021
11022 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
11023 break;
11024
11025 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
11026 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11027 return (EINVAL);
11028
11029 size = opt[DTRACEOPT_STRSIZE];
11030 }
11031
11032 break;
11033
11034 case DTRACEACT_STACK:
11035 if ((nframes = arg) == 0) {
11036 nframes = opt[DTRACEOPT_STACKFRAMES];
11037 ASSERT(nframes > 0);
11038 arg = nframes;
11039 }
11040
11041 size = nframes * sizeof (pc_t);
11042 break;
11043
11044 case DTRACEACT_JSTACK:
11045 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
11046 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
11047
11048 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
11049 nframes = opt[DTRACEOPT_JSTACKFRAMES];
11050
11051 arg = DTRACE_USTACK_ARG(nframes, strsize);
11052
11053 /*FALLTHROUGH*/
11054 case DTRACEACT_USTACK:
11055 if (desc->dtad_kind != DTRACEACT_JSTACK &&
11056 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
11057 strsize = DTRACE_USTACK_STRSIZE(arg);
11058 nframes = opt[DTRACEOPT_USTACKFRAMES];
11059 ASSERT(nframes > 0);
11060 arg = DTRACE_USTACK_ARG(nframes, strsize);
11061 }
11062
11063 /*
11064 * Save a slot for the pid.
11065 */
11066 size = (nframes + 1) * sizeof (uint64_t);
11067 size += DTRACE_USTACK_STRSIZE(arg);
11068 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
11069
11070 break;
11071
11072 case DTRACEACT_SYM:
11073 case DTRACEACT_MOD:
11074 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
11075 sizeof (uint64_t)) ||
11076 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11077 return (EINVAL);
11078 break;
11079
11080 case DTRACEACT_USYM:
11081 case DTRACEACT_UMOD:
11082 case DTRACEACT_UADDR:
11083 if (dp == NULL ||
11084 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
11085 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11086 return (EINVAL);
11087
11088 /*
11089 * We have a slot for the pid, plus a slot for the
11090 * argument. To keep things simple (aligned with
11091 * bitness-neutral sizing), we store each as a 64-bit
11092 * quantity.
11093 */
11094 size = 2 * sizeof (uint64_t);
11095 break;
11096
11097 case DTRACEACT_STOP:
11098 case DTRACEACT_BREAKPOINT:
11099 case DTRACEACT_PANIC:
11100 break;
11101
11102 case DTRACEACT_CHILL:
11103 case DTRACEACT_DISCARD:
11104 case DTRACEACT_RAISE:
11105 if (dp == NULL)
11106 return (EINVAL);
11107 break;
11108
11109 case DTRACEACT_EXIT:
11110 if (dp == NULL ||
11111 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
11112 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11113 return (EINVAL);
11114 break;
11115
11116 case DTRACEACT_SPECULATE:
11117 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
11118 return (EINVAL);
11119
11120 if (dp == NULL)
11121 return (EINVAL);
11122
11123 state->dts_speculates = 1;
11124 break;
11125
11126 case DTRACEACT_COMMIT: {
11127 dtrace_action_t *act = ecb->dte_action;
11128
11129 for (; act != NULL; act = act->dta_next) {
11130 if (act->dta_kind == DTRACEACT_COMMIT)
11131 return (EINVAL);
11132 }
11133
11134 if (dp == NULL)
11135 return (EINVAL);
11136 break;
11137 }
11138
11139 default:
11140 return (EINVAL);
11141 }
11142
11143 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
11144 /*
11145 * If this is a data-storing action or a speculate,
11146 * we must be sure that there isn't a commit on the
11147 * action chain.
11148 */
11149 dtrace_action_t *act = ecb->dte_action;
11150
11151 for (; act != NULL; act = act->dta_next) {
11152 if (act->dta_kind == DTRACEACT_COMMIT)
11153 return (EINVAL);
11154 }
11155 }
11156
11157 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
11158 action->dta_rec.dtrd_size = size;
11159 }
11160
11161 action->dta_refcnt = 1;
11162 rec = &action->dta_rec;
11163 size = rec->dtrd_size;
11164
11165 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
11166 if (!(size & mask)) {
11167 align = mask + 1;
11168 break;
11169 }
11170 }
11171
11172 action->dta_kind = desc->dtad_kind;
11173
11174 if ((action->dta_difo = dp) != NULL)
11175 dtrace_difo_hold(dp);
11176
11177 rec->dtrd_action = action->dta_kind;
11178 rec->dtrd_arg = arg;
11179 rec->dtrd_uarg = desc->dtad_uarg;
11180 rec->dtrd_alignment = (uint16_t)align;
11181 rec->dtrd_format = format;
11182
11183 if ((last = ecb->dte_action_last) != NULL) {
11184 ASSERT(ecb->dte_action != NULL);
11185 action->dta_prev = last;
11186 last->dta_next = action;
11187 } else {
11188 ASSERT(ecb->dte_action == NULL);
11189 ecb->dte_action = action;
11190 }
11191
11192 ecb->dte_action_last = action;
11193
11194 return (0);
11195 }
11196
11197 static void
11198 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
11199 {
11200 dtrace_action_t *act = ecb->dte_action, *next;
11201 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
11202 dtrace_difo_t *dp;
11203 uint16_t format;
11204
11205 if (act != NULL && act->dta_refcnt > 1) {
11206 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
11207 act->dta_refcnt--;
11208 } else {
11209 for (; act != NULL; act = next) {
11210 next = act->dta_next;
11211 ASSERT(next != NULL || act == ecb->dte_action_last);
11212 ASSERT(act->dta_refcnt == 1);
11213
11214 if ((format = act->dta_rec.dtrd_format) != 0)
11215 dtrace_format_remove(ecb->dte_state, format);
11216
11217 if ((dp = act->dta_difo) != NULL)
11218 dtrace_difo_release(dp, vstate);
11219
11220 if (DTRACEACT_ISAGG(act->dta_kind)) {
11221 dtrace_ecb_aggregation_destroy(ecb, act);
11222 } else {
11223 kmem_free(act, sizeof (dtrace_action_t));
11224 }
11225 }
11226 }
11227
11228 ecb->dte_action = NULL;
11229 ecb->dte_action_last = NULL;
11230 ecb->dte_size = 0;
11231 }
11232
11233 static void
11234 dtrace_ecb_disable(dtrace_ecb_t *ecb)
11235 {
11236 /*
11237 * We disable the ECB by removing it from its probe.
11238 */
11239 dtrace_ecb_t *pecb, *prev = NULL;
11240 dtrace_probe_t *probe = ecb->dte_probe;
11241
11242 ASSERT(MUTEX_HELD(&dtrace_lock));
11243
11244 if (probe == NULL) {
11245 /*
11246 * This is the NULL probe; there is nothing to disable.
11247 */
11248 return;
11249 }
11250
11251 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
11252 if (pecb == ecb)
11253 break;
11254 prev = pecb;
11255 }
11256
11257 ASSERT(pecb != NULL);
11258
11259 if (prev == NULL) {
11260 probe->dtpr_ecb = ecb->dte_next;
11261 } else {
11262 prev->dte_next = ecb->dte_next;
11263 }
11264
11265 if (ecb == probe->dtpr_ecb_last) {
11266 ASSERT(ecb->dte_next == NULL);
11267 probe->dtpr_ecb_last = prev;
11268 }
11269
11270 /*
11271 * The ECB has been disconnected from the probe; now sync to assure
11272 * that all CPUs have seen the change before returning.
11273 */
11274 dtrace_sync();
11275
11276 if (probe->dtpr_ecb == NULL) {
11277 /*
11278 * That was the last ECB on the probe; clear the predicate
11279 * cache ID for the probe, disable it and sync one more time
11280 * to assure that we'll never hit it again.
11281 */
11282 dtrace_provider_t *prov = probe->dtpr_provider;
11283
11284 ASSERT(ecb->dte_next == NULL);
11285 ASSERT(probe->dtpr_ecb_last == NULL);
11286 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
11287 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
11288 probe->dtpr_id, probe->dtpr_arg);
11289 dtrace_sync();
11290 } else {
11291 /*
11292 * There is at least one ECB remaining on the probe. If there
11293 * is _exactly_ one, set the probe's predicate cache ID to be
11294 * the predicate cache ID of the remaining ECB.
11295 */
11296 ASSERT(probe->dtpr_ecb_last != NULL);
11297 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
11298
11299 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
11300 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
11301
11302 ASSERT(probe->dtpr_ecb->dte_next == NULL);
11303
11304 if (p != NULL)
11305 probe->dtpr_predcache = p->dtp_cacheid;
11306 }
11307
11308 ecb->dte_next = NULL;
11309 }
11310 }
11311
11312 static void
11313 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
11314 {
11315 dtrace_state_t *state = ecb->dte_state;
11316 dtrace_vstate_t *vstate = &state->dts_vstate;
11317 dtrace_predicate_t *pred;
11318 dtrace_epid_t epid = ecb->dte_epid;
11319
11320 ASSERT(MUTEX_HELD(&dtrace_lock));
11321 ASSERT(ecb->dte_next == NULL);
11322 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
11323
11324 if ((pred = ecb->dte_predicate) != NULL)
11325 dtrace_predicate_release(pred, vstate);
11326
11327 dtrace_ecb_action_remove(ecb);
11328
11329 ASSERT(state->dts_ecbs[epid - 1] == ecb);
11330 state->dts_ecbs[epid - 1] = NULL;
11331
11332 kmem_free(ecb, sizeof (dtrace_ecb_t));
11333 }
11334
11335 static dtrace_ecb_t *
11336 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
11337 dtrace_enabling_t *enab)
11338 {
11339 dtrace_ecb_t *ecb;
11340 dtrace_predicate_t *pred;
11341 dtrace_actdesc_t *act;
11342 dtrace_provider_t *prov;
11343 dtrace_ecbdesc_t *desc = enab->dten_current;
11344
11345 ASSERT(MUTEX_HELD(&dtrace_lock));
11346 ASSERT(state != NULL);
11347
11348 ecb = dtrace_ecb_add(state, probe);
11349 ecb->dte_uarg = desc->dted_uarg;
11350
11351 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
11352 dtrace_predicate_hold(pred);
11353 ecb->dte_predicate = pred;
11354 }
11355
11356 if (probe != NULL) {
11357 /*
11358 * If the provider shows more leg than the consumer is old
11359 * enough to see, we need to enable the appropriate implicit
11360 * predicate bits to prevent the ecb from activating at
11361 * revealing times.
11362 *
11363 * Providers specifying DTRACE_PRIV_USER at register time
11364 * are stating that they need the /proc-style privilege
11365 * model to be enforced, and this is what DTRACE_COND_OWNER
11366 * and DTRACE_COND_ZONEOWNER will then do at probe time.
11367 */
11368 prov = probe->dtpr_provider;
11369 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
11370 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11371 ecb->dte_cond |= DTRACE_COND_OWNER;
11372
11373 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11374 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11375 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11376
11377 /*
11378 * If the provider shows us kernel innards and the user
11379 * is lacking sufficient privilege, enable the
11380 * DTRACE_COND_USERMODE implicit predicate.
11381 */
11382 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11383 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11384 ecb->dte_cond |= DTRACE_COND_USERMODE;
11385 }
11386
11387 if (dtrace_ecb_create_cache != NULL) {
11388 /*
11389 * If we have a cached ecb, we'll use its action list instead
11390 * of creating our own (saving both time and space).
11391 */
11392 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11393 dtrace_action_t *act = cached->dte_action;
11394
11395 if (act != NULL) {
11396 ASSERT(act->dta_refcnt > 0);
11397 act->dta_refcnt++;
11398 ecb->dte_action = act;
11399 ecb->dte_action_last = cached->dte_action_last;
11400 ecb->dte_needed = cached->dte_needed;
11401 ecb->dte_size = cached->dte_size;
11402 ecb->dte_alignment = cached->dte_alignment;
11403 }
11404
11405 return (ecb);
11406 }
11407
11408 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11409 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11410 dtrace_ecb_destroy(ecb);
11411 return (NULL);
11412 }
11413 }
11414
11415 if ((enab->dten_error = dtrace_ecb_resize(ecb)) != 0) {
11416 dtrace_ecb_destroy(ecb);
11417 return (NULL);
11418 }
11419
11420 return (dtrace_ecb_create_cache = ecb);
11421 }
11422
11423 static int
11424 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11425 {
11426 dtrace_ecb_t *ecb;
11427 dtrace_enabling_t *enab = arg;
11428 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11429
11430 ASSERT(state != NULL);
11431
11432 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11433 /*
11434 * This probe was created in a generation for which this
11435 * enabling has previously created ECBs; we don't want to
11436 * enable it again, so just kick out.
11437 */
11438 return (DTRACE_MATCH_NEXT);
11439 }
11440
11441 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11442 return (DTRACE_MATCH_DONE);
11443
11444 if (dtrace_ecb_enable(ecb) < 0)
11445 return (DTRACE_MATCH_FAIL);
11446
11447 return (DTRACE_MATCH_NEXT);
11448 }
11449
11450 static dtrace_ecb_t *
11451 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11452 {
11453 dtrace_ecb_t *ecb;
11454
11455 ASSERT(MUTEX_HELD(&dtrace_lock));
11456
11457 if (id == 0 || id > state->dts_necbs)
11458 return (NULL);
11459
11460 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
11461 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
11462
11463 return (state->dts_ecbs[id - 1]);
11464 }
11465
11466 static dtrace_aggregation_t *
11467 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
11468 {
11469 dtrace_aggregation_t *agg;
11470
11471 ASSERT(MUTEX_HELD(&dtrace_lock));
11472
11473 if (id == 0 || id > state->dts_naggregations)
11474 return (NULL);
11475
11476 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
11477 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
11478 agg->dtag_id == id);
11479
11480 return (state->dts_aggregations[id - 1]);
11481 }
11482
11483 /*
11484 * DTrace Buffer Functions
11485 *
11486 * The following functions manipulate DTrace buffers. Most of these functions
11487 * are called in the context of establishing or processing consumer state;
11488 * exceptions are explicitly noted.
11489 */
11490
11491 /*
11492 * Note: called from cross call context. This function switches the two
11493 * buffers on a given CPU. The atomicity of this operation is assured by
11494 * disabling interrupts while the actual switch takes place; the disabling of
11495 * interrupts serializes the execution with any execution of dtrace_probe() on
11496 * the same CPU.
11497 */
11498 static void
11499 dtrace_buffer_switch(dtrace_buffer_t *buf)
11500 {
11501 caddr_t tomax = buf->dtb_tomax;
11502 caddr_t xamot = buf->dtb_xamot;
11503 dtrace_icookie_t cookie;
11504 hrtime_t now;
11505
11506 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11507 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
11508
11509 cookie = dtrace_interrupt_disable();
11510 now = dtrace_gethrtime();
11511 buf->dtb_tomax = xamot;
11512 buf->dtb_xamot = tomax;
11513 buf->dtb_xamot_drops = buf->dtb_drops;
11514 buf->dtb_xamot_offset = buf->dtb_offset;
11515 buf->dtb_xamot_errors = buf->dtb_errors;
11516 buf->dtb_xamot_flags = buf->dtb_flags;
11517 buf->dtb_offset = 0;
11518 buf->dtb_drops = 0;
11519 buf->dtb_errors = 0;
11520 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
11521 buf->dtb_interval = now - buf->dtb_switched;
11522 buf->dtb_switched = now;
11523 dtrace_interrupt_enable(cookie);
11524 }
11525
11526 /*
11527 * Note: called from cross call context. This function activates a buffer
11528 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
11529 * is guaranteed by the disabling of interrupts.
11530 */
11531 static void
11532 dtrace_buffer_activate(dtrace_state_t *state)
11533 {
11534 dtrace_buffer_t *buf;
11535 dtrace_icookie_t cookie = dtrace_interrupt_disable();
11536
11537 buf = &state->dts_buffer[CPU->cpu_id];
11538
11539 if (buf->dtb_tomax != NULL) {
11540 /*
11541 * We might like to assert that the buffer is marked inactive,
11542 * but this isn't necessarily true: the buffer for the CPU
11543 * that processes the BEGIN probe has its buffer activated
11544 * manually. In this case, we take the (harmless) action
11545 * re-clearing the bit INACTIVE bit.
11546 */
11547 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
11548 }
11549
11550 dtrace_interrupt_enable(cookie);
11551 }
11552
11553 static int
11554 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
11555 processorid_t cpu, int *factor)
11556 {
11557 cpu_t *cp;
11558 dtrace_buffer_t *buf;
11559 int allocated = 0, desired = 0;
11560
11561 ASSERT(MUTEX_HELD(&cpu_lock));
11562 ASSERT(MUTEX_HELD(&dtrace_lock));
11563
11564 *factor = 1;
11565
11566 if (size > dtrace_nonroot_maxsize &&
11567 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
11568 return (EFBIG);
11569
11570 cp = cpu_list;
11571
11572 do {
11573 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11574 continue;
11575
11576 buf = &bufs[cp->cpu_id];
11577
11578 /*
11579 * If there is already a buffer allocated for this CPU, it
11580 * is only possible that this is a DR event. In this case,
11581 * the buffer size must match our specified size.
11582 */
11583 if (buf->dtb_tomax != NULL) {
11584 ASSERT(buf->dtb_size == size);
11585 continue;
11586 }
11587
11588 ASSERT(buf->dtb_xamot == NULL);
11589
11590 if ((buf->dtb_tomax = kmem_zalloc(size,
11591 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11592 goto err;
11593
11594 buf->dtb_size = size;
11595 buf->dtb_flags = flags;
11596 buf->dtb_offset = 0;
11597 buf->dtb_drops = 0;
11598
11599 if (flags & DTRACEBUF_NOSWITCH)
11600 continue;
11601
11602 if ((buf->dtb_xamot = kmem_zalloc(size,
11603 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11604 goto err;
11605 } while ((cp = cp->cpu_next) != cpu_list);
11606
11607 return (0);
11608
11609 err:
11610 cp = cpu_list;
11611
11612 do {
11613 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11614 continue;
11615
11616 buf = &bufs[cp->cpu_id];
11617 desired += 2;
11618
11619 if (buf->dtb_xamot != NULL) {
11620 ASSERT(buf->dtb_tomax != NULL);
11621 ASSERT(buf->dtb_size == size);
11622 kmem_free(buf->dtb_xamot, size);
11623 allocated++;
11624 }
11625
11626 if (buf->dtb_tomax != NULL) {
11627 ASSERT(buf->dtb_size == size);
11628 kmem_free(buf->dtb_tomax, size);
11629 allocated++;
11630 }
11631
11632 buf->dtb_tomax = NULL;
11633 buf->dtb_xamot = NULL;
11634 buf->dtb_size = 0;
11635 } while ((cp = cp->cpu_next) != cpu_list);
11636
11637 *factor = desired / (allocated > 0 ? allocated : 1);
11638
11639 return (ENOMEM);
11640 }
11641
11642 /*
11643 * Note: called from probe context. This function just increments the drop
11644 * count on a buffer. It has been made a function to allow for the
11645 * possibility of understanding the source of mysterious drop counts. (A
11646 * problem for which one may be particularly disappointed that DTrace cannot
11647 * be used to understand DTrace.)
11648 */
11649 static void
11650 dtrace_buffer_drop(dtrace_buffer_t *buf)
11651 {
11652 buf->dtb_drops++;
11653 }
11654
11655 /*
11656 * Note: called from probe context. This function is called to reserve space
11657 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
11658 * mstate. Returns the new offset in the buffer, or a negative value if an
11659 * error has occurred.
11660 */
11661 static intptr_t
11662 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11663 dtrace_state_t *state, dtrace_mstate_t *mstate)
11664 {
11665 intptr_t offs = buf->dtb_offset, soffs;
11666 intptr_t woffs;
11667 caddr_t tomax;
11668 size_t total;
11669
11670 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11671 return (-1);
11672
11673 if ((tomax = buf->dtb_tomax) == NULL) {
11674 dtrace_buffer_drop(buf);
11675 return (-1);
11676 }
11677
11678 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11679 while (offs & (align - 1)) {
11680 /*
11681 * Assert that our alignment is off by a number which
11682 * is itself sizeof (uint32_t) aligned.
11683 */
11684 ASSERT(!((align - (offs & (align - 1))) &
11685 (sizeof (uint32_t) - 1)));
11686 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11687 offs += sizeof (uint32_t);
11688 }
11689
11690 if ((soffs = offs + needed) > buf->dtb_size) {
11691 dtrace_buffer_drop(buf);
11692 return (-1);
11693 }
11694
11695 if (mstate == NULL)
11696 return (offs);
11697
11698 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11699 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11700 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11701
11702 return (offs);
11703 }
11704
11705 if (buf->dtb_flags & DTRACEBUF_FILL) {
11706 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11707 (buf->dtb_flags & DTRACEBUF_FULL))
11708 return (-1);
11709 goto out;
11710 }
11711
11712 total = needed + (offs & (align - 1));
11713
11714 /*
11715 * For a ring buffer, life is quite a bit more complicated. Before
11716 * we can store any padding, we need to adjust our wrapping offset.
11717 * (If we've never before wrapped or we're not about to, no adjustment
11718 * is required.)
11719 */
11720 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11721 offs + total > buf->dtb_size) {
11722 woffs = buf->dtb_xamot_offset;
11723
11724 if (offs + total > buf->dtb_size) {
11725 /*
11726 * We can't fit in the end of the buffer. First, a
11727 * sanity check that we can fit in the buffer at all.
11728 */
11729 if (total > buf->dtb_size) {
11730 dtrace_buffer_drop(buf);
11731 return (-1);
11732 }
11733
11734 /*
11735 * We're going to be storing at the top of the buffer,
11736 * so now we need to deal with the wrapped offset. We
11737 * only reset our wrapped offset to 0 if it is
11738 * currently greater than the current offset. If it
11739 * is less than the current offset, it is because a
11740 * previous allocation induced a wrap -- but the
11741 * allocation didn't subsequently take the space due
11742 * to an error or false predicate evaluation. In this
11743 * case, we'll just leave the wrapped offset alone: if
11744 * the wrapped offset hasn't been advanced far enough
11745 * for this allocation, it will be adjusted in the
11746 * lower loop.
11747 */
11748 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11749 if (woffs >= offs)
11750 woffs = 0;
11751 } else {
11752 woffs = 0;
11753 }
11754
11755 /*
11756 * Now we know that we're going to be storing to the
11757 * top of the buffer and that there is room for us
11758 * there. We need to clear the buffer from the current
11759 * offset to the end (there may be old gunk there).
11760 */
11761 while (offs < buf->dtb_size)
11762 tomax[offs++] = 0;
11763
11764 /*
11765 * We need to set our offset to zero. And because we
11766 * are wrapping, we need to set the bit indicating as
11767 * much. We can also adjust our needed space back
11768 * down to the space required by the ECB -- we know
11769 * that the top of the buffer is aligned.
11770 */
11771 offs = 0;
11772 total = needed;
11773 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11774 } else {
11775 /*
11776 * There is room for us in the buffer, so we simply
11777 * need to check the wrapped offset.
11778 */
11779 if (woffs < offs) {
11780 /*
11781 * The wrapped offset is less than the offset.
11782 * This can happen if we allocated buffer space
11783 * that induced a wrap, but then we didn't
11784 * subsequently take the space due to an error
11785 * or false predicate evaluation. This is
11786 * okay; we know that _this_ allocation isn't
11787 * going to induce a wrap. We still can't
11788 * reset the wrapped offset to be zero,
11789 * however: the space may have been trashed in
11790 * the previous failed probe attempt. But at
11791 * least the wrapped offset doesn't need to
11792 * be adjusted at all...
11793 */
11794 goto out;
11795 }
11796 }
11797
11798 while (offs + total > woffs) {
11799 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11800 size_t size;
11801
11802 if (epid == DTRACE_EPIDNONE) {
11803 size = sizeof (uint32_t);
11804 } else {
11805 ASSERT3U(epid, <=, state->dts_necbs);
11806 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11807
11808 size = state->dts_ecbs[epid - 1]->dte_size;
11809 }
11810
11811 ASSERT(woffs + size <= buf->dtb_size);
11812 ASSERT(size != 0);
11813
11814 if (woffs + size == buf->dtb_size) {
11815 /*
11816 * We've reached the end of the buffer; we want
11817 * to set the wrapped offset to 0 and break
11818 * out. However, if the offs is 0, then we're
11819 * in a strange edge-condition: the amount of
11820 * space that we want to reserve plus the size
11821 * of the record that we're overwriting is
11822 * greater than the size of the buffer. This
11823 * is problematic because if we reserve the
11824 * space but subsequently don't consume it (due
11825 * to a failed predicate or error) the wrapped
11826 * offset will be 0 -- yet the EPID at offset 0
11827 * will not be committed. This situation is
11828 * relatively easy to deal with: if we're in
11829 * this case, the buffer is indistinguishable
11830 * from one that hasn't wrapped; we need only
11831 * finish the job by clearing the wrapped bit,
11832 * explicitly setting the offset to be 0, and
11833 * zero'ing out the old data in the buffer.
11834 */
11835 if (offs == 0) {
11836 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11837 buf->dtb_offset = 0;
11838 woffs = total;
11839
11840 while (woffs < buf->dtb_size)
11841 tomax[woffs++] = 0;
11842 }
11843
11844 woffs = 0;
11845 break;
11846 }
11847
11848 woffs += size;
11849 }
11850
11851 /*
11852 * We have a wrapped offset. It may be that the wrapped offset
11853 * has become zero -- that's okay.
11854 */
11855 buf->dtb_xamot_offset = woffs;
11856 }
11857
11858 out:
11859 /*
11860 * Now we can plow the buffer with any necessary padding.
11861 */
11862 while (offs & (align - 1)) {
11863 /*
11864 * Assert that our alignment is off by a number which
11865 * is itself sizeof (uint32_t) aligned.
11866 */
11867 ASSERT(!((align - (offs & (align - 1))) &
11868 (sizeof (uint32_t) - 1)));
11869 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11870 offs += sizeof (uint32_t);
11871 }
11872
11873 if (buf->dtb_flags & DTRACEBUF_FILL) {
11874 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11875 buf->dtb_flags |= DTRACEBUF_FULL;
11876 return (-1);
11877 }
11878 }
11879
11880 if (mstate == NULL)
11881 return (offs);
11882
11883 /*
11884 * For ring buffers and fill buffers, the scratch space is always
11885 * the inactive buffer.
11886 */
11887 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11888 mstate->dtms_scratch_size = buf->dtb_size;
11889 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11890
11891 return (offs);
11892 }
11893
11894 static void
11895 dtrace_buffer_polish(dtrace_buffer_t *buf)
11896 {
11897 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11898 ASSERT(MUTEX_HELD(&dtrace_lock));
11899
11900 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11901 return;
11902
11903 /*
11904 * We need to polish the ring buffer. There are three cases:
11905 *
11906 * - The first (and presumably most common) is that there is no gap
11907 * between the buffer offset and the wrapped offset. In this case,
11908 * there is nothing in the buffer that isn't valid data; we can
11909 * mark the buffer as polished and return.
11910 *
11911 * - The second (less common than the first but still more common
11912 * than the third) is that there is a gap between the buffer offset
11913 * and the wrapped offset, and the wrapped offset is larger than the
11914 * buffer offset. This can happen because of an alignment issue, or
11915 * can happen because of a call to dtrace_buffer_reserve() that
11916 * didn't subsequently consume the buffer space. In this case,
11917 * we need to zero the data from the buffer offset to the wrapped
11918 * offset.
11919 *
11920 * - The third (and least common) is that there is a gap between the
11921 * buffer offset and the wrapped offset, but the wrapped offset is
11922 * _less_ than the buffer offset. This can only happen because a
11923 * call to dtrace_buffer_reserve() induced a wrap, but the space
11924 * was not subsequently consumed. In this case, we need to zero the
11925 * space from the offset to the end of the buffer _and_ from the
11926 * top of the buffer to the wrapped offset.
11927 */
11928 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11929 bzero(buf->dtb_tomax + buf->dtb_offset,
11930 buf->dtb_xamot_offset - buf->dtb_offset);
11931 }
11932
11933 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11934 bzero(buf->dtb_tomax + buf->dtb_offset,
11935 buf->dtb_size - buf->dtb_offset);
11936 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11937 }
11938 }
11939
11940 /*
11941 * This routine determines if data generated at the specified time has likely
11942 * been entirely consumed at user-level. This routine is called to determine
11943 * if an ECB on a defunct probe (but for an active enabling) can be safely
11944 * disabled and destroyed.
11945 */
11946 static int
11947 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11948 {
11949 int i;
11950
11951 for (i = 0; i < NCPU; i++) {
11952 dtrace_buffer_t *buf = &bufs[i];
11953
11954 if (buf->dtb_size == 0)
11955 continue;
11956
11957 if (buf->dtb_flags & DTRACEBUF_RING)
11958 return (0);
11959
11960 if (!buf->dtb_switched && buf->dtb_offset != 0)
11961 return (0);
11962
11963 if (buf->dtb_switched - buf->dtb_interval < when)
11964 return (0);
11965 }
11966
11967 return (1);
11968 }
11969
11970 static void
11971 dtrace_buffer_free(dtrace_buffer_t *bufs)
11972 {
11973 int i;
11974
11975 for (i = 0; i < NCPU; i++) {
11976 dtrace_buffer_t *buf = &bufs[i];
11977
11978 if (buf->dtb_tomax == NULL) {
11979 ASSERT(buf->dtb_xamot == NULL);
11980 ASSERT(buf->dtb_size == 0);
11981 continue;
11982 }
11983
11984 if (buf->dtb_xamot != NULL) {
11985 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11986 kmem_free(buf->dtb_xamot, buf->dtb_size);
11987 }
11988
11989 kmem_free(buf->dtb_tomax, buf->dtb_size);
11990 buf->dtb_size = 0;
11991 buf->dtb_tomax = NULL;
11992 buf->dtb_xamot = NULL;
11993 }
11994 }
11995
11996 /*
11997 * DTrace Enabling Functions
11998 */
11999 static dtrace_enabling_t *
12000 dtrace_enabling_create(dtrace_vstate_t *vstate)
12001 {
12002 dtrace_enabling_t *enab;
12003
12004 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
12005 enab->dten_vstate = vstate;
12006
12007 return (enab);
12008 }
12009
12010 static void
12011 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
12012 {
12013 dtrace_ecbdesc_t **ndesc;
12014 size_t osize, nsize;
12015
12016 /*
12017 * We can't add to enablings after we've enabled them, or after we've
12018 * retained them.
12019 */
12020 ASSERT(enab->dten_probegen == 0);
12021 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12022
12023 if (enab->dten_ndesc < enab->dten_maxdesc) {
12024 enab->dten_desc[enab->dten_ndesc++] = ecb;
12025 return;
12026 }
12027
12028 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12029
12030 if (enab->dten_maxdesc == 0) {
12031 enab->dten_maxdesc = 1;
12032 } else {
12033 enab->dten_maxdesc <<= 1;
12034 }
12035
12036 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
12037
12038 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12039 ndesc = kmem_zalloc(nsize, KM_SLEEP);
12040 bcopy(enab->dten_desc, ndesc, osize);
12041 kmem_free(enab->dten_desc, osize);
12042
12043 enab->dten_desc = ndesc;
12044 enab->dten_desc[enab->dten_ndesc++] = ecb;
12045 }
12046
12047 static void
12048 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
12049 dtrace_probedesc_t *pd)
12050 {
12051 dtrace_ecbdesc_t *new;
12052 dtrace_predicate_t *pred;
12053 dtrace_actdesc_t *act;
12054
12055 /*
12056 * We're going to create a new ECB description that matches the
12057 * specified ECB in every way, but has the specified probe description.
12058 */
12059 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12060
12061 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
12062 dtrace_predicate_hold(pred);
12063
12064 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
12065 dtrace_actdesc_hold(act);
12066
12067 new->dted_action = ecb->dted_action;
12068 new->dted_pred = ecb->dted_pred;
12069 new->dted_probe = *pd;
12070 new->dted_uarg = ecb->dted_uarg;
12071
12072 dtrace_enabling_add(enab, new);
12073 }
12074
12075 static void
12076 dtrace_enabling_dump(dtrace_enabling_t *enab)
12077 {
12078 int i;
12079
12080 for (i = 0; i < enab->dten_ndesc; i++) {
12081 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
12082
12083 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
12084 desc->dtpd_provider, desc->dtpd_mod,
12085 desc->dtpd_func, desc->dtpd_name);
12086 }
12087 }
12088
12089 static void
12090 dtrace_enabling_destroy(dtrace_enabling_t *enab)
12091 {
12092 int i;
12093 dtrace_ecbdesc_t *ep;
12094 dtrace_vstate_t *vstate = enab->dten_vstate;
12095
12096 ASSERT(MUTEX_HELD(&dtrace_lock));
12097
12098 for (i = 0; i < enab->dten_ndesc; i++) {
12099 dtrace_actdesc_t *act, *next;
12100 dtrace_predicate_t *pred;
12101
12102 ep = enab->dten_desc[i];
12103
12104 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
12105 dtrace_predicate_release(pred, vstate);
12106
12107 for (act = ep->dted_action; act != NULL; act = next) {
12108 next = act->dtad_next;
12109 dtrace_actdesc_release(act, vstate);
12110 }
12111
12112 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12113 }
12114
12115 kmem_free(enab->dten_desc,
12116 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
12117
12118 /*
12119 * If this was a retained enabling, decrement the dts_nretained count
12120 * and take it off of the dtrace_retained list.
12121 */
12122 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
12123 dtrace_retained == enab) {
12124 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12125 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
12126 enab->dten_vstate->dtvs_state->dts_nretained--;
12127 dtrace_retained_gen++;
12128 }
12129
12130 if (enab->dten_prev == NULL) {
12131 if (dtrace_retained == enab) {
12132 dtrace_retained = enab->dten_next;
12133
12134 if (dtrace_retained != NULL)
12135 dtrace_retained->dten_prev = NULL;
12136 }
12137 } else {
12138 ASSERT(enab != dtrace_retained);
12139 ASSERT(dtrace_retained != NULL);
12140 enab->dten_prev->dten_next = enab->dten_next;
12141 }
12142
12143 if (enab->dten_next != NULL) {
12144 ASSERT(dtrace_retained != NULL);
12145 enab->dten_next->dten_prev = enab->dten_prev;
12146 }
12147
12148 kmem_free(enab, sizeof (dtrace_enabling_t));
12149 }
12150
12151 static int
12152 dtrace_enabling_retain(dtrace_enabling_t *enab)
12153 {
12154 dtrace_state_t *state;
12155
12156 ASSERT(MUTEX_HELD(&dtrace_lock));
12157 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12158 ASSERT(enab->dten_vstate != NULL);
12159
12160 state = enab->dten_vstate->dtvs_state;
12161 ASSERT(state != NULL);
12162
12163 /*
12164 * We only allow each state to retain dtrace_retain_max enablings.
12165 */
12166 if (state->dts_nretained >= dtrace_retain_max)
12167 return (ENOSPC);
12168
12169 state->dts_nretained++;
12170 dtrace_retained_gen++;
12171
12172 if (dtrace_retained == NULL) {
12173 dtrace_retained = enab;
12174 return (0);
12175 }
12176
12177 enab->dten_next = dtrace_retained;
12178 dtrace_retained->dten_prev = enab;
12179 dtrace_retained = enab;
12180
12181 return (0);
12182 }
12183
12184 static int
12185 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
12186 dtrace_probedesc_t *create)
12187 {
12188 dtrace_enabling_t *new, *enab;
12189 int found = 0, err = ENOENT;
12190
12191 ASSERT(MUTEX_HELD(&dtrace_lock));
12192 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
12193 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
12194 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
12195 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
12196
12197 new = dtrace_enabling_create(&state->dts_vstate);
12198
12199 /*
12200 * Iterate over all retained enablings, looking for enablings that
12201 * match the specified state.
12202 */
12203 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12204 int i;
12205
12206 /*
12207 * dtvs_state can only be NULL for helper enablings -- and
12208 * helper enablings can't be retained.
12209 */
12210 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12211
12212 if (enab->dten_vstate->dtvs_state != state)
12213 continue;
12214
12215 /*
12216 * Now iterate over each probe description; we're looking for
12217 * an exact match to the specified probe description.
12218 */
12219 for (i = 0; i < enab->dten_ndesc; i++) {
12220 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12221 dtrace_probedesc_t *pd = &ep->dted_probe;
12222
12223 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
12224 continue;
12225
12226 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
12227 continue;
12228
12229 if (strcmp(pd->dtpd_func, match->dtpd_func))
12230 continue;
12231
12232 if (strcmp(pd->dtpd_name, match->dtpd_name))
12233 continue;
12234
12235 /*
12236 * We have a winning probe! Add it to our growing
12237 * enabling.
12238 */
12239 found = 1;
12240 dtrace_enabling_addlike(new, ep, create);
12241 }
12242 }
12243
12244 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
12245 dtrace_enabling_destroy(new);
12246 return (err);
12247 }
12248
12249 return (0);
12250 }
12251
12252 static void
12253 dtrace_enabling_retract(dtrace_state_t *state)
12254 {
12255 dtrace_enabling_t *enab, *next;
12256
12257 ASSERT(MUTEX_HELD(&dtrace_lock));
12258
12259 /*
12260 * Iterate over all retained enablings, destroy the enablings retained
12261 * for the specified state.
12262 */
12263 for (enab = dtrace_retained; enab != NULL; enab = next) {
12264 next = enab->dten_next;
12265
12266 /*
12267 * dtvs_state can only be NULL for helper enablings -- and
12268 * helper enablings can't be retained.
12269 */
12270 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12271
12272 if (enab->dten_vstate->dtvs_state == state) {
12273 ASSERT(state->dts_nretained > 0);
12274 dtrace_enabling_destroy(enab);
12275 }
12276 }
12277
12278 ASSERT(state->dts_nretained == 0);
12279 }
12280
12281 static int
12282 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
12283 {
12284 int i = 0;
12285 int total_matched = 0, matched = 0;
12286
12287 ASSERT(MUTEX_HELD(&cpu_lock));
12288 ASSERT(MUTEX_HELD(&dtrace_lock));
12289
12290 for (i = 0; i < enab->dten_ndesc; i++) {
12291 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12292
12293 enab->dten_current = ep;
12294 enab->dten_error = 0;
12295
12296 /*
12297 * If a provider failed to enable a probe then get out and
12298 * let the consumer know we failed.
12299 */
12300 if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
12301 return (EBUSY);
12302
12303 total_matched += matched;
12304
12305 if (enab->dten_error != 0) {
12306 /*
12307 * If we get an error half-way through enabling the
12308 * probes, we kick out -- perhaps with some number of
12309 * them enabled. Leaving enabled probes enabled may
12310 * be slightly confusing for user-level, but we expect
12311 * that no one will attempt to actually drive on in
12312 * the face of such errors. If this is an anonymous
12313 * enabling (indicated with a NULL nmatched pointer),
12314 * we cmn_err() a message. We aren't expecting to
12315 * get such an error -- such as it can exist at all,
12316 * it would be a result of corrupted DOF in the driver
12317 * properties.
12318 */
12319 if (nmatched == NULL) {
12320 cmn_err(CE_WARN, "dtrace_enabling_match() "
12321 "error on %p: %d", (void *)ep,
12322 enab->dten_error);
12323 }
12324
12325 return (enab->dten_error);
12326 }
12327 }
12328
12329 enab->dten_probegen = dtrace_probegen;
12330 if (nmatched != NULL)
12331 *nmatched = total_matched;
12332
12333 return (0);
12334 }
12335
12336 static void
12337 dtrace_enabling_matchall(void)
12338 {
12339 dtrace_enabling_t *enab;
12340
12341 mutex_enter(&cpu_lock);
12342 mutex_enter(&dtrace_lock);
12343
12344 /*
12345 * Iterate over all retained enablings to see if any probes match
12346 * against them. We only perform this operation on enablings for which
12347 * we have sufficient permissions by virtue of being in the global zone
12348 * or in the same zone as the DTrace client. Because we can be called
12349 * after dtrace_detach() has been called, we cannot assert that there
12350 * are retained enablings. We can safely load from dtrace_retained,
12351 * however: the taskq_destroy() at the end of dtrace_detach() will
12352 * block pending our completion.
12353 */
12354 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12355 dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
12356 cred_t *cr = dcr->dcr_cred;
12357 zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
12358
12359 if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
12360 (zone == GLOBAL_ZONEID || getzoneid() == zone)))
12361 (void) dtrace_enabling_match(enab, NULL);
12362 }
12363
12364 mutex_exit(&dtrace_lock);
12365 mutex_exit(&cpu_lock);
12366 }
12367
12368 /*
12369 * If an enabling is to be enabled without having matched probes (that is, if
12370 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
12371 * enabling must be _primed_ by creating an ECB for every ECB description.
12372 * This must be done to assure that we know the number of speculations, the
12373 * number of aggregations, the minimum buffer size needed, etc. before we
12374 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
12375 * enabling any probes, we create ECBs for every ECB decription, but with a
12376 * NULL probe -- which is exactly what this function does.
12377 */
12378 static void
12379 dtrace_enabling_prime(dtrace_state_t *state)
12380 {
12381 dtrace_enabling_t *enab;
12382 int i;
12383
12384 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12385 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12386
12387 if (enab->dten_vstate->dtvs_state != state)
12388 continue;
12389
12390 /*
12391 * We don't want to prime an enabling more than once, lest
12392 * we allow a malicious user to induce resource exhaustion.
12393 * (The ECBs that result from priming an enabling aren't
12394 * leaked -- but they also aren't deallocated until the
12395 * consumer state is destroyed.)
12396 */
12397 if (enab->dten_primed)
12398 continue;
12399
12400 for (i = 0; i < enab->dten_ndesc; i++) {
12401 enab->dten_current = enab->dten_desc[i];
12402 (void) dtrace_probe_enable(NULL, enab);
12403 }
12404
12405 enab->dten_primed = 1;
12406 }
12407 }
12408
12409 /*
12410 * Called to indicate that probes should be provided due to retained
12411 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
12412 * must take an initial lap through the enabling calling the dtps_provide()
12413 * entry point explicitly to allow for autocreated probes.
12414 */
12415 static void
12416 dtrace_enabling_provide(dtrace_provider_t *prv)
12417 {
12418 int i, all = 0;
12419 dtrace_probedesc_t desc;
12420 dtrace_genid_t gen;
12421
12422 ASSERT(MUTEX_HELD(&dtrace_lock));
12423 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
12424
12425 if (prv == NULL) {
12426 all = 1;
12427 prv = dtrace_provider;
12428 }
12429
12430 do {
12431 dtrace_enabling_t *enab;
12432 void *parg = prv->dtpv_arg;
12433
12434 retry:
12435 gen = dtrace_retained_gen;
12436 for (enab = dtrace_retained; enab != NULL;
12437 enab = enab->dten_next) {
12438 for (i = 0; i < enab->dten_ndesc; i++) {
12439 desc = enab->dten_desc[i]->dted_probe;
12440 mutex_exit(&dtrace_lock);
12441 prv->dtpv_pops.dtps_provide(parg, &desc);
12442 mutex_enter(&dtrace_lock);
12443 /*
12444 * Process the retained enablings again if
12445 * they have changed while we weren't holding
12446 * dtrace_lock.
12447 */
12448 if (gen != dtrace_retained_gen)
12449 goto retry;
12450 }
12451 }
12452 } while (all && (prv = prv->dtpv_next) != NULL);
12453
12454 mutex_exit(&dtrace_lock);
12455 dtrace_probe_provide(NULL, all ? NULL : prv);
12456 mutex_enter(&dtrace_lock);
12457 }
12458
12459 /*
12460 * Called to reap ECBs that are attached to probes from defunct providers.
12461 */
12462 static void
12463 dtrace_enabling_reap(void)
12464 {
12465 dtrace_provider_t *prov;
12466 dtrace_probe_t *probe;
12467 dtrace_ecb_t *ecb;
12468 hrtime_t when;
12469 int i;
12470
12471 mutex_enter(&cpu_lock);
12472 mutex_enter(&dtrace_lock);
12473
12474 for (i = 0; i < dtrace_nprobes; i++) {
12475 if ((probe = dtrace_probes[i]) == NULL)
12476 continue;
12477
12478 if (probe->dtpr_ecb == NULL)
12479 continue;
12480
12481 prov = probe->dtpr_provider;
12482
12483 if ((when = prov->dtpv_defunct) == 0)
12484 continue;
12485
12486 /*
12487 * We have ECBs on a defunct provider: we want to reap these
12488 * ECBs to allow the provider to unregister. The destruction
12489 * of these ECBs must be done carefully: if we destroy the ECB
12490 * and the consumer later wishes to consume an EPID that
12491 * corresponds to the destroyed ECB (and if the EPID metadata
12492 * has not been previously consumed), the consumer will abort
12493 * processing on the unknown EPID. To reduce (but not, sadly,
12494 * eliminate) the possibility of this, we will only destroy an
12495 * ECB for a defunct provider if, for the state that
12496 * corresponds to the ECB:
12497 *
12498 * (a) There is no speculative tracing (which can effectively
12499 * cache an EPID for an arbitrary amount of time).
12500 *
12501 * (b) The principal buffers have been switched twice since the
12502 * provider became defunct.
12503 *
12504 * (c) The aggregation buffers are of zero size or have been
12505 * switched twice since the provider became defunct.
12506 *
12507 * We use dts_speculates to determine (a) and call a function
12508 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
12509 * that as soon as we've been unable to destroy one of the ECBs
12510 * associated with the probe, we quit trying -- reaping is only
12511 * fruitful in as much as we can destroy all ECBs associated
12512 * with the defunct provider's probes.
12513 */
12514 while ((ecb = probe->dtpr_ecb) != NULL) {
12515 dtrace_state_t *state = ecb->dte_state;
12516 dtrace_buffer_t *buf = state->dts_buffer;
12517 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
12518
12519 if (state->dts_speculates)
12520 break;
12521
12522 if (!dtrace_buffer_consumed(buf, when))
12523 break;
12524
12525 if (!dtrace_buffer_consumed(aggbuf, when))
12526 break;
12527
12528 dtrace_ecb_disable(ecb);
12529 ASSERT(probe->dtpr_ecb != ecb);
12530 dtrace_ecb_destroy(ecb);
12531 }
12532 }
12533
12534 mutex_exit(&dtrace_lock);
12535 mutex_exit(&cpu_lock);
12536 }
12537
12538 /*
12539 * DTrace DOF Functions
12540 */
12541 /*ARGSUSED*/
12542 static void
12543 dtrace_dof_error(dof_hdr_t *dof, const char *str)
12544 {
12545 if (dtrace_err_verbose)
12546 cmn_err(CE_WARN, "failed to process DOF: %s", str);
12547
12548 #ifdef DTRACE_ERRDEBUG
12549 dtrace_errdebug(str);
12550 #endif
12551 }
12552
12553 /*
12554 * Create DOF out of a currently enabled state. Right now, we only create
12555 * DOF containing the run-time options -- but this could be expanded to create
12556 * complete DOF representing the enabled state.
12557 */
12558 static dof_hdr_t *
12559 dtrace_dof_create(dtrace_state_t *state)
12560 {
12561 dof_hdr_t *dof;
12562 dof_sec_t *sec;
12563 dof_optdesc_t *opt;
12564 int i, len = sizeof (dof_hdr_t) +
12565 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
12566 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12567
12568 ASSERT(MUTEX_HELD(&dtrace_lock));
12569
12570 dof = kmem_zalloc(len, KM_SLEEP);
12571 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
12572 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
12573 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
12574 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
12575
12576 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
12577 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
12578 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
12579 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
12580 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
12581 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
12582
12583 dof->dofh_flags = 0;
12584 dof->dofh_hdrsize = sizeof (dof_hdr_t);
12585 dof->dofh_secsize = sizeof (dof_sec_t);
12586 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
12587 dof->dofh_secoff = sizeof (dof_hdr_t);
12588 dof->dofh_loadsz = len;
12589 dof->dofh_filesz = len;
12590 dof->dofh_pad = 0;
12591
12592 /*
12593 * Fill in the option section header...
12594 */
12595 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
12596 sec->dofs_type = DOF_SECT_OPTDESC;
12597 sec->dofs_align = sizeof (uint64_t);
12598 sec->dofs_flags = DOF_SECF_LOAD;
12599 sec->dofs_entsize = sizeof (dof_optdesc_t);
12600
12601 opt = (dof_optdesc_t *)((uintptr_t)sec +
12602 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
12603
12604 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
12605 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12606
12607 for (i = 0; i < DTRACEOPT_MAX; i++) {
12608 opt[i].dofo_option = i;
12609 opt[i].dofo_strtab = DOF_SECIDX_NONE;
12610 opt[i].dofo_value = state->dts_options[i];
12611 }
12612
12613 return (dof);
12614 }
12615
12616 static dof_hdr_t *
12617 dtrace_dof_copyin(uintptr_t uarg, int *errp)
12618 {
12619 dof_hdr_t hdr, *dof;
12620
12621 ASSERT(!MUTEX_HELD(&dtrace_lock));
12622
12623 /*
12624 * First, we're going to copyin() the sizeof (dof_hdr_t).
12625 */
12626 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
12627 dtrace_dof_error(NULL, "failed to copyin DOF header");
12628 *errp = EFAULT;
12629 return (NULL);
12630 }
12631
12632 /*
12633 * Now we'll allocate the entire DOF and copy it in -- provided
12634 * that the length isn't outrageous.
12635 */
12636 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
12637 dtrace_dof_error(&hdr, "load size exceeds maximum");
12638 *errp = E2BIG;
12639 return (NULL);
12640 }
12641
12642 if (hdr.dofh_loadsz < sizeof (hdr)) {
12643 dtrace_dof_error(&hdr, "invalid load size");
12644 *errp = EINVAL;
12645 return (NULL);
12646 }
12647
12648 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12649
12650 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
12651 dof->dofh_loadsz != hdr.dofh_loadsz) {
12652 kmem_free(dof, hdr.dofh_loadsz);
12653 *errp = EFAULT;
12654 return (NULL);
12655 }
12656
12657 return (dof);
12658 }
12659
12660 static dof_hdr_t *
12661 dtrace_dof_property(const char *name)
12662 {
12663 uchar_t *buf;
12664 uint64_t loadsz;
12665 unsigned int len, i;
12666 dof_hdr_t *dof;
12667
12668 /*
12669 * Unfortunately, array of values in .conf files are always (and
12670 * only) interpreted to be integer arrays. We must read our DOF
12671 * as an integer array, and then squeeze it into a byte array.
12672 */
12673 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12674 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12675 return (NULL);
12676
12677 for (i = 0; i < len; i++)
12678 buf[i] = (uchar_t)(((int *)buf)[i]);
12679
12680 if (len < sizeof (dof_hdr_t)) {
12681 ddi_prop_free(buf);
12682 dtrace_dof_error(NULL, "truncated header");
12683 return (NULL);
12684 }
12685
12686 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12687 ddi_prop_free(buf);
12688 dtrace_dof_error(NULL, "truncated DOF");
12689 return (NULL);
12690 }
12691
12692 if (loadsz >= dtrace_dof_maxsize) {
12693 ddi_prop_free(buf);
12694 dtrace_dof_error(NULL, "oversized DOF");
12695 return (NULL);
12696 }
12697
12698 dof = kmem_alloc(loadsz, KM_SLEEP);
12699 bcopy(buf, dof, loadsz);
12700 ddi_prop_free(buf);
12701
12702 return (dof);
12703 }
12704
12705 static void
12706 dtrace_dof_destroy(dof_hdr_t *dof)
12707 {
12708 kmem_free(dof, dof->dofh_loadsz);
12709 }
12710
12711 /*
12712 * Return the dof_sec_t pointer corresponding to a given section index. If the
12713 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12714 * a type other than DOF_SECT_NONE is specified, the header is checked against
12715 * this type and NULL is returned if the types do not match.
12716 */
12717 static dof_sec_t *
12718 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12719 {
12720 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12721 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12722
12723 if (i >= dof->dofh_secnum) {
12724 dtrace_dof_error(dof, "referenced section index is invalid");
12725 return (NULL);
12726 }
12727
12728 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12729 dtrace_dof_error(dof, "referenced section is not loadable");
12730 return (NULL);
12731 }
12732
12733 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12734 dtrace_dof_error(dof, "referenced section is the wrong type");
12735 return (NULL);
12736 }
12737
12738 return (sec);
12739 }
12740
12741 static dtrace_probedesc_t *
12742 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12743 {
12744 dof_probedesc_t *probe;
12745 dof_sec_t *strtab;
12746 uintptr_t daddr = (uintptr_t)dof;
12747 uintptr_t str;
12748 size_t size;
12749
12750 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12751 dtrace_dof_error(dof, "invalid probe section");
12752 return (NULL);
12753 }
12754
12755 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12756 dtrace_dof_error(dof, "bad alignment in probe description");
12757 return (NULL);
12758 }
12759
12760 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12761 dtrace_dof_error(dof, "truncated probe description");
12762 return (NULL);
12763 }
12764
12765 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12766 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12767
12768 if (strtab == NULL)
12769 return (NULL);
12770
12771 str = daddr + strtab->dofs_offset;
12772 size = strtab->dofs_size;
12773
12774 if (probe->dofp_provider >= strtab->dofs_size) {
12775 dtrace_dof_error(dof, "corrupt probe provider");
12776 return (NULL);
12777 }
12778
12779 (void) strncpy(desc->dtpd_provider,
12780 (char *)(str + probe->dofp_provider),
12781 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12782
12783 if (probe->dofp_mod >= strtab->dofs_size) {
12784 dtrace_dof_error(dof, "corrupt probe module");
12785 return (NULL);
12786 }
12787
12788 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12789 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12790
12791 if (probe->dofp_func >= strtab->dofs_size) {
12792 dtrace_dof_error(dof, "corrupt probe function");
12793 return (NULL);
12794 }
12795
12796 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12797 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12798
12799 if (probe->dofp_name >= strtab->dofs_size) {
12800 dtrace_dof_error(dof, "corrupt probe name");
12801 return (NULL);
12802 }
12803
12804 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12805 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12806
12807 return (desc);
12808 }
12809
12810 static dtrace_difo_t *
12811 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12812 cred_t *cr)
12813 {
12814 dtrace_difo_t *dp;
12815 size_t ttl = 0;
12816 dof_difohdr_t *dofd;
12817 uintptr_t daddr = (uintptr_t)dof;
12818 size_t max = dtrace_difo_maxsize;
12819 int i, l, n;
12820
12821 static const struct {
12822 int section;
12823 int bufoffs;
12824 int lenoffs;
12825 int entsize;
12826 int align;
12827 const char *msg;
12828 } difo[] = {
12829 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12830 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12831 sizeof (dif_instr_t), "multiple DIF sections" },
12832
12833 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12834 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12835 sizeof (uint64_t), "multiple integer tables" },
12836
12837 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12838 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12839 sizeof (char), "multiple string tables" },
12840
12841 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12842 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12843 sizeof (uint_t), "multiple variable tables" },
12844
12845 { DOF_SECT_NONE, 0, 0, 0, NULL }
12846 };
12847
12848 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12849 dtrace_dof_error(dof, "invalid DIFO header section");
12850 return (NULL);
12851 }
12852
12853 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12854 dtrace_dof_error(dof, "bad alignment in DIFO header");
12855 return (NULL);
12856 }
12857
12858 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12859 sec->dofs_size % sizeof (dof_secidx_t)) {
12860 dtrace_dof_error(dof, "bad size in DIFO header");
12861 return (NULL);
12862 }
12863
12864 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12865 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12866
12867 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12868 dp->dtdo_rtype = dofd->dofd_rtype;
12869
12870 for (l = 0; l < n; l++) {
12871 dof_sec_t *subsec;
12872 void **bufp;
12873 uint32_t *lenp;
12874
12875 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12876 dofd->dofd_links[l])) == NULL)
12877 goto err; /* invalid section link */
12878
12879 if (ttl + subsec->dofs_size > max) {
12880 dtrace_dof_error(dof, "exceeds maximum size");
12881 goto err;
12882 }
12883
12884 ttl += subsec->dofs_size;
12885
12886 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12887 if (subsec->dofs_type != difo[i].section)
12888 continue;
12889
12890 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12891 dtrace_dof_error(dof, "section not loaded");
12892 goto err;
12893 }
12894
12895 if (subsec->dofs_align != difo[i].align) {
12896 dtrace_dof_error(dof, "bad alignment");
12897 goto err;
12898 }
12899
12900 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12901 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12902
12903 if (*bufp != NULL) {
12904 dtrace_dof_error(dof, difo[i].msg);
12905 goto err;
12906 }
12907
12908 if (difo[i].entsize != subsec->dofs_entsize) {
12909 dtrace_dof_error(dof, "entry size mismatch");
12910 goto err;
12911 }
12912
12913 if (subsec->dofs_entsize != 0 &&
12914 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12915 dtrace_dof_error(dof, "corrupt entry size");
12916 goto err;
12917 }
12918
12919 *lenp = subsec->dofs_size;
12920 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12921 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12922 *bufp, subsec->dofs_size);
12923
12924 if (subsec->dofs_entsize != 0)
12925 *lenp /= subsec->dofs_entsize;
12926
12927 break;
12928 }
12929
12930 /*
12931 * If we encounter a loadable DIFO sub-section that is not
12932 * known to us, assume this is a broken program and fail.
12933 */
12934 if (difo[i].section == DOF_SECT_NONE &&
12935 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12936 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12937 goto err;
12938 }
12939 }
12940
12941 if (dp->dtdo_buf == NULL) {
12942 /*
12943 * We can't have a DIF object without DIF text.
12944 */
12945 dtrace_dof_error(dof, "missing DIF text");
12946 goto err;
12947 }
12948
12949 /*
12950 * Before we validate the DIF object, run through the variable table
12951 * looking for the strings -- if any of their size are under, we'll set
12952 * their size to be the system-wide default string size. Note that
12953 * this should _not_ happen if the "strsize" option has been set --
12954 * in this case, the compiler should have set the size to reflect the
12955 * setting of the option.
12956 */
12957 for (i = 0; i < dp->dtdo_varlen; i++) {
12958 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12959 dtrace_diftype_t *t = &v->dtdv_type;
12960
12961 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12962 continue;
12963
12964 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12965 t->dtdt_size = dtrace_strsize_default;
12966 }
12967
12968 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12969 goto err;
12970
12971 dtrace_difo_init(dp, vstate);
12972 return (dp);
12973
12974 err:
12975 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12976 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12977 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12978 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12979
12980 kmem_free(dp, sizeof (dtrace_difo_t));
12981 return (NULL);
12982 }
12983
12984 static dtrace_predicate_t *
12985 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12986 cred_t *cr)
12987 {
12988 dtrace_difo_t *dp;
12989
12990 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12991 return (NULL);
12992
12993 return (dtrace_predicate_create(dp));
12994 }
12995
12996 static dtrace_actdesc_t *
12997 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12998 cred_t *cr)
12999 {
13000 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
13001 dof_actdesc_t *desc;
13002 dof_sec_t *difosec;
13003 size_t offs;
13004 uintptr_t daddr = (uintptr_t)dof;
13005 uint64_t arg;
13006 dtrace_actkind_t kind;
13007
13008 if (sec->dofs_type != DOF_SECT_ACTDESC) {
13009 dtrace_dof_error(dof, "invalid action section");
13010 return (NULL);
13011 }
13012
13013 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
13014 dtrace_dof_error(dof, "truncated action description");
13015 return (NULL);
13016 }
13017
13018 if (sec->dofs_align != sizeof (uint64_t)) {
13019 dtrace_dof_error(dof, "bad alignment in action description");
13020 return (NULL);
13021 }
13022
13023 if (sec->dofs_size < sec->dofs_entsize) {
13024 dtrace_dof_error(dof, "section entry size exceeds total size");
13025 return (NULL);
13026 }
13027
13028 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
13029 dtrace_dof_error(dof, "bad entry size in action description");
13030 return (NULL);
13031 }
13032
13033 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
13034 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
13035 return (NULL);
13036 }
13037
13038 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
13039 desc = (dof_actdesc_t *)(daddr +
13040 (uintptr_t)sec->dofs_offset + offs);
13041 kind = (dtrace_actkind_t)desc->dofa_kind;
13042
13043 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
13044 (kind != DTRACEACT_PRINTA ||
13045 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
13046 (kind == DTRACEACT_DIFEXPR &&
13047 desc->dofa_strtab != DOF_SECIDX_NONE)) {
13048 dof_sec_t *strtab;
13049 char *str, *fmt;
13050 uint64_t i;
13051
13052 /*
13053 * The argument to these actions is an index into the
13054 * DOF string table. For printf()-like actions, this
13055 * is the format string. For print(), this is the
13056 * CTF type of the expression result.
13057 */
13058 if ((strtab = dtrace_dof_sect(dof,
13059 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
13060 goto err;
13061
13062 str = (char *)((uintptr_t)dof +
13063 (uintptr_t)strtab->dofs_offset);
13064
13065 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
13066 if (str[i] == '\0')
13067 break;
13068 }
13069
13070 if (i >= strtab->dofs_size) {
13071 dtrace_dof_error(dof, "bogus format string");
13072 goto err;
13073 }
13074
13075 if (i == desc->dofa_arg) {
13076 dtrace_dof_error(dof, "empty format string");
13077 goto err;
13078 }
13079
13080 i -= desc->dofa_arg;
13081 fmt = kmem_alloc(i + 1, KM_SLEEP);
13082 bcopy(&str[desc->dofa_arg], fmt, i + 1);
13083 arg = (uint64_t)(uintptr_t)fmt;
13084 } else {
13085 if (kind == DTRACEACT_PRINTA) {
13086 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
13087 arg = 0;
13088 } else {
13089 arg = desc->dofa_arg;
13090 }
13091 }
13092
13093 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
13094 desc->dofa_uarg, arg);
13095
13096 if (last != NULL) {
13097 last->dtad_next = act;
13098 } else {
13099 first = act;
13100 }
13101
13102 last = act;
13103
13104 if (desc->dofa_difo == DOF_SECIDX_NONE)
13105 continue;
13106
13107 if ((difosec = dtrace_dof_sect(dof,
13108 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
13109 goto err;
13110
13111 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
13112
13113 if (act->dtad_difo == NULL)
13114 goto err;
13115 }
13116
13117 ASSERT(first != NULL);
13118 return (first);
13119
13120 err:
13121 for (act = first; act != NULL; act = next) {
13122 next = act->dtad_next;
13123 dtrace_actdesc_release(act, vstate);
13124 }
13125
13126 return (NULL);
13127 }
13128
13129 static dtrace_ecbdesc_t *
13130 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13131 cred_t *cr)
13132 {
13133 dtrace_ecbdesc_t *ep;
13134 dof_ecbdesc_t *ecb;
13135 dtrace_probedesc_t *desc;
13136 dtrace_predicate_t *pred = NULL;
13137
13138 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
13139 dtrace_dof_error(dof, "truncated ECB description");
13140 return (NULL);
13141 }
13142
13143 if (sec->dofs_align != sizeof (uint64_t)) {
13144 dtrace_dof_error(dof, "bad alignment in ECB description");
13145 return (NULL);
13146 }
13147
13148 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
13149 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
13150
13151 if (sec == NULL)
13152 return (NULL);
13153
13154 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
13155 ep->dted_uarg = ecb->dofe_uarg;
13156 desc = &ep->dted_probe;
13157
13158 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
13159 goto err;
13160
13161 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
13162 if ((sec = dtrace_dof_sect(dof,
13163 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
13164 goto err;
13165
13166 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
13167 goto err;
13168
13169 ep->dted_pred.dtpdd_predicate = pred;
13170 }
13171
13172 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
13173 if ((sec = dtrace_dof_sect(dof,
13174 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
13175 goto err;
13176
13177 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
13178
13179 if (ep->dted_action == NULL)
13180 goto err;
13181 }
13182
13183 return (ep);
13184
13185 err:
13186 if (pred != NULL)
13187 dtrace_predicate_release(pred, vstate);
13188 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
13189 return (NULL);
13190 }
13191
13192 /*
13193 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
13194 * specified DOF. At present, this amounts to simply adding 'ubase' to the
13195 * site of any user SETX relocations to account for load object base address.
13196 * In the future, if we need other relocations, this function can be extended.
13197 */
13198 static int
13199 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
13200 {
13201 uintptr_t daddr = (uintptr_t)dof;
13202 uintptr_t ts_end;
13203 dof_relohdr_t *dofr =
13204 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13205 dof_sec_t *ss, *rs, *ts;
13206 dof_relodesc_t *r;
13207 uint_t i, n;
13208
13209 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
13210 sec->dofs_align != sizeof (dof_secidx_t)) {
13211 dtrace_dof_error(dof, "invalid relocation header");
13212 return (-1);
13213 }
13214
13215 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
13216 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
13217 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
13218 ts_end = (uintptr_t)ts + sizeof (dof_sec_t);
13219
13220 if (ss == NULL || rs == NULL || ts == NULL)
13221 return (-1); /* dtrace_dof_error() has been called already */
13222
13223 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
13224 rs->dofs_align != sizeof (uint64_t)) {
13225 dtrace_dof_error(dof, "invalid relocation section");
13226 return (-1);
13227 }
13228
13229 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
13230 n = rs->dofs_size / rs->dofs_entsize;
13231
13232 for (i = 0; i < n; i++) {
13233 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
13234
13235 switch (r->dofr_type) {
13236 case DOF_RELO_NONE:
13237 break;
13238 case DOF_RELO_SETX:
13239 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
13240 sizeof (uint64_t) > ts->dofs_size) {
13241 dtrace_dof_error(dof, "bad relocation offset");
13242 return (-1);
13243 }
13244
13245 if (taddr >= (uintptr_t)ts && taddr < ts_end) {
13246 dtrace_dof_error(dof, "bad relocation offset");
13247 return (-1);
13248 }
13249
13250 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
13251 dtrace_dof_error(dof, "misaligned setx relo");
13252 return (-1);
13253 }
13254
13255 *(uint64_t *)taddr += ubase;
13256 break;
13257 default:
13258 dtrace_dof_error(dof, "invalid relocation type");
13259 return (-1);
13260 }
13261
13262 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
13263 }
13264
13265 return (0);
13266 }
13267
13268 /*
13269 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
13270 * header: it should be at the front of a memory region that is at least
13271 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
13272 * size. It need not be validated in any other way.
13273 */
13274 static int
13275 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
13276 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
13277 {
13278 uint64_t len = dof->dofh_loadsz, seclen;
13279 uintptr_t daddr = (uintptr_t)dof;
13280 dtrace_ecbdesc_t *ep;
13281 dtrace_enabling_t *enab;
13282 uint_t i;
13283
13284 ASSERT(MUTEX_HELD(&dtrace_lock));
13285 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
13286
13287 /*
13288 * Check the DOF header identification bytes. In addition to checking
13289 * valid settings, we also verify that unused bits/bytes are zeroed so
13290 * we can use them later without fear of regressing existing binaries.
13291 */
13292 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
13293 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
13294 dtrace_dof_error(dof, "DOF magic string mismatch");
13295 return (-1);
13296 }
13297
13298 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
13299 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
13300 dtrace_dof_error(dof, "DOF has invalid data model");
13301 return (-1);
13302 }
13303
13304 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
13305 dtrace_dof_error(dof, "DOF encoding mismatch");
13306 return (-1);
13307 }
13308
13309 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
13310 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
13311 dtrace_dof_error(dof, "DOF version mismatch");
13312 return (-1);
13313 }
13314
13315 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
13316 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
13317 return (-1);
13318 }
13319
13320 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
13321 dtrace_dof_error(dof, "DOF uses too many integer registers");
13322 return (-1);
13323 }
13324
13325 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
13326 dtrace_dof_error(dof, "DOF uses too many tuple registers");
13327 return (-1);
13328 }
13329
13330 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
13331 if (dof->dofh_ident[i] != 0) {
13332 dtrace_dof_error(dof, "DOF has invalid ident byte set");
13333 return (-1);
13334 }
13335 }
13336
13337 if (dof->dofh_flags & ~DOF_FL_VALID) {
13338 dtrace_dof_error(dof, "DOF has invalid flag bits set");
13339 return (-1);
13340 }
13341
13342 if (dof->dofh_secsize == 0) {
13343 dtrace_dof_error(dof, "zero section header size");
13344 return (-1);
13345 }
13346
13347 /*
13348 * Check that the section headers don't exceed the amount of DOF
13349 * data. Note that we cast the section size and number of sections
13350 * to uint64_t's to prevent possible overflow in the multiplication.
13351 */
13352 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
13353
13354 if (dof->dofh_secoff > len || seclen > len ||
13355 dof->dofh_secoff + seclen > len) {
13356 dtrace_dof_error(dof, "truncated section headers");
13357 return (-1);
13358 }
13359
13360 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
13361 dtrace_dof_error(dof, "misaligned section headers");
13362 return (-1);
13363 }
13364
13365 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
13366 dtrace_dof_error(dof, "misaligned section size");
13367 return (-1);
13368 }
13369
13370 /*
13371 * Take an initial pass through the section headers to be sure that
13372 * the headers don't have stray offsets. If the 'noprobes' flag is
13373 * set, do not permit sections relating to providers, probes, or args.
13374 */
13375 for (i = 0; i < dof->dofh_secnum; i++) {
13376 dof_sec_t *sec = (dof_sec_t *)(daddr +
13377 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13378
13379 if (noprobes) {
13380 switch (sec->dofs_type) {
13381 case DOF_SECT_PROVIDER:
13382 case DOF_SECT_PROBES:
13383 case DOF_SECT_PRARGS:
13384 case DOF_SECT_PROFFS:
13385 dtrace_dof_error(dof, "illegal sections "
13386 "for enabling");
13387 return (-1);
13388 }
13389 }
13390
13391 if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
13392 !(sec->dofs_flags & DOF_SECF_LOAD)) {
13393 dtrace_dof_error(dof, "loadable section with load "
13394 "flag unset");
13395 return (-1);
13396 }
13397
13398 if (!(sec->dofs_flags & DOF_SECF_LOAD))
13399 continue; /* just ignore non-loadable sections */
13400
13401 if (!ISP2(sec->dofs_align)) {
13402 dtrace_dof_error(dof, "bad section alignment");
13403 return (-1);
13404 }
13405
13406 if (sec->dofs_offset & (sec->dofs_align - 1)) {
13407 dtrace_dof_error(dof, "misaligned section");
13408 return (-1);
13409 }
13410
13411 if (sec->dofs_offset > len || sec->dofs_size > len ||
13412 sec->dofs_offset + sec->dofs_size > len) {
13413 dtrace_dof_error(dof, "corrupt section header");
13414 return (-1);
13415 }
13416
13417 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
13418 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
13419 dtrace_dof_error(dof, "non-terminating string table");
13420 return (-1);
13421 }
13422 }
13423
13424 /*
13425 * Take a second pass through the sections and locate and perform any
13426 * relocations that are present. We do this after the first pass to
13427 * be sure that all sections have had their headers validated.
13428 */
13429 for (i = 0; i < dof->dofh_secnum; i++) {
13430 dof_sec_t *sec = (dof_sec_t *)(daddr +
13431 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13432
13433 if (!(sec->dofs_flags & DOF_SECF_LOAD))
13434 continue; /* skip sections that are not loadable */
13435
13436 switch (sec->dofs_type) {
13437 case DOF_SECT_URELHDR:
13438 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
13439 return (-1);
13440 break;
13441 }
13442 }
13443
13444 if ((enab = *enabp) == NULL)
13445 enab = *enabp = dtrace_enabling_create(vstate);
13446
13447 for (i = 0; i < dof->dofh_secnum; i++) {
13448 dof_sec_t *sec = (dof_sec_t *)(daddr +
13449 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13450
13451 if (sec->dofs_type != DOF_SECT_ECBDESC)
13452 continue;
13453
13454 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
13455 dtrace_enabling_destroy(enab);
13456 *enabp = NULL;
13457 return (-1);
13458 }
13459
13460 dtrace_enabling_add(enab, ep);
13461 }
13462
13463 return (0);
13464 }
13465
13466 /*
13467 * Process DOF for any options. This routine assumes that the DOF has been
13468 * at least processed by dtrace_dof_slurp().
13469 */
13470 static int
13471 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
13472 {
13473 int i, rval;
13474 uint32_t entsize;
13475 size_t offs;
13476 dof_optdesc_t *desc;
13477
13478 for (i = 0; i < dof->dofh_secnum; i++) {
13479 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
13480 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13481
13482 if (sec->dofs_type != DOF_SECT_OPTDESC)
13483 continue;
13484
13485 if (sec->dofs_align != sizeof (uint64_t)) {
13486 dtrace_dof_error(dof, "bad alignment in "
13487 "option description");
13488 return (EINVAL);
13489 }
13490
13491 if ((entsize = sec->dofs_entsize) == 0) {
13492 dtrace_dof_error(dof, "zeroed option entry size");
13493 return (EINVAL);
13494 }
13495
13496 if (entsize < sizeof (dof_optdesc_t)) {
13497 dtrace_dof_error(dof, "bad option entry size");
13498 return (EINVAL);
13499 }
13500
13501 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
13502 desc = (dof_optdesc_t *)((uintptr_t)dof +
13503 (uintptr_t)sec->dofs_offset + offs);
13504
13505 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
13506 dtrace_dof_error(dof, "non-zero option string");
13507 return (EINVAL);
13508 }
13509
13510 if (desc->dofo_value == DTRACEOPT_UNSET) {
13511 dtrace_dof_error(dof, "unset option");
13512 return (EINVAL);
13513 }
13514
13515 if ((rval = dtrace_state_option(state,
13516 desc->dofo_option, desc->dofo_value)) != 0) {
13517 dtrace_dof_error(dof, "rejected option");
13518 return (rval);
13519 }
13520 }
13521 }
13522
13523 return (0);
13524 }
13525
13526 /*
13527 * DTrace Consumer State Functions
13528 */
13529 int
13530 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
13531 {
13532 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
13533 void *base;
13534 uintptr_t limit;
13535 dtrace_dynvar_t *dvar, *next, *start;
13536 int i;
13537
13538 ASSERT(MUTEX_HELD(&dtrace_lock));
13539 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
13540
13541 bzero(dstate, sizeof (dtrace_dstate_t));
13542
13543 if ((dstate->dtds_chunksize = chunksize) == 0)
13544 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
13545
13546 VERIFY(dstate->dtds_chunksize < LONG_MAX);
13547
13548 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
13549 size = min;
13550
13551 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
13552 return (ENOMEM);
13553
13554 dstate->dtds_size = size;
13555 dstate->dtds_base = base;
13556 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
13557 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
13558
13559 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
13560
13561 if (hashsize != 1 && (hashsize & 1))
13562 hashsize--;
13563
13564 dstate->dtds_hashsize = hashsize;
13565 dstate->dtds_hash = dstate->dtds_base;
13566
13567 /*
13568 * Set all of our hash buckets to point to the single sink, and (if
13569 * it hasn't already been set), set the sink's hash value to be the
13570 * sink sentinel value. The sink is needed for dynamic variable
13571 * lookups to know that they have iterated over an entire, valid hash
13572 * chain.
13573 */
13574 for (i = 0; i < hashsize; i++)
13575 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
13576
13577 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
13578 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
13579
13580 /*
13581 * Determine number of active CPUs. Divide free list evenly among
13582 * active CPUs.
13583 */
13584 start = (dtrace_dynvar_t *)
13585 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
13586 limit = (uintptr_t)base + size;
13587
13588 VERIFY((uintptr_t)start < limit);
13589 VERIFY((uintptr_t)start >= (uintptr_t)base);
13590
13591 maxper = (limit - (uintptr_t)start) / NCPU;
13592 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13593
13594 for (i = 0; i < NCPU; i++) {
13595 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13596
13597 /*
13598 * If we don't even have enough chunks to make it once through
13599 * NCPUs, we're just going to allocate everything to the first
13600 * CPU. And if we're on the last CPU, we're going to allocate
13601 * whatever is left over. In either case, we set the limit to
13602 * be the limit of the dynamic variable space.
13603 */
13604 if (maxper == 0 || i == NCPU - 1) {
13605 limit = (uintptr_t)base + size;
13606 start = NULL;
13607 } else {
13608 limit = (uintptr_t)start + maxper;
13609 start = (dtrace_dynvar_t *)limit;
13610 }
13611
13612 VERIFY(limit <= (uintptr_t)base + size);
13613
13614 for (;;) {
13615 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13616 dstate->dtds_chunksize);
13617
13618 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13619 break;
13620
13621 VERIFY((uintptr_t)dvar >= (uintptr_t)base &&
13622 (uintptr_t)dvar <= (uintptr_t)base + size);
13623 dvar->dtdv_next = next;
13624 dvar = next;
13625 }
13626
13627 if (maxper == 0)
13628 break;
13629 }
13630
13631 return (0);
13632 }
13633
13634 void
13635 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13636 {
13637 ASSERT(MUTEX_HELD(&cpu_lock));
13638
13639 if (dstate->dtds_base == NULL)
13640 return;
13641
13642 kmem_free(dstate->dtds_base, dstate->dtds_size);
13643 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13644 }
13645
13646 static void
13647 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13648 {
13649 /*
13650 * Logical XOR, where are you?
13651 */
13652 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13653
13654 if (vstate->dtvs_nglobals > 0) {
13655 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13656 sizeof (dtrace_statvar_t *));
13657 }
13658
13659 if (vstate->dtvs_ntlocals > 0) {
13660 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13661 sizeof (dtrace_difv_t));
13662 }
13663
13664 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13665
13666 if (vstate->dtvs_nlocals > 0) {
13667 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13668 sizeof (dtrace_statvar_t *));
13669 }
13670 }
13671
13672 static void
13673 dtrace_state_clean(dtrace_state_t *state)
13674 {
13675 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13676 return;
13677
13678 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13679 dtrace_speculation_clean(state);
13680 }
13681
13682 static void
13683 dtrace_state_deadman(dtrace_state_t *state)
13684 {
13685 hrtime_t now;
13686
13687 dtrace_sync();
13688
13689 now = dtrace_gethrtime();
13690
13691 if (state != dtrace_anon.dta_state &&
13692 now - state->dts_laststatus >= dtrace_deadman_user)
13693 return;
13694
13695 /*
13696 * We must be sure that dts_alive never appears to be less than the
13697 * value upon entry to dtrace_state_deadman(), and because we lack a
13698 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13699 * store INT64_MAX to it, followed by a memory barrier, followed by
13700 * the new value. This assures that dts_alive never appears to be
13701 * less than its true value, regardless of the order in which the
13702 * stores to the underlying storage are issued.
13703 */
13704 state->dts_alive = INT64_MAX;
13705 dtrace_membar_producer();
13706 state->dts_alive = now;
13707 }
13708
13709 dtrace_state_t *
13710 dtrace_state_create(dev_t *devp, cred_t *cr)
13711 {
13712 minor_t minor;
13713 major_t major;
13714 char c[30];
13715 dtrace_state_t *state;
13716 dtrace_optval_t *opt;
13717 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13718
13719 ASSERT(MUTEX_HELD(&dtrace_lock));
13720 ASSERT(MUTEX_HELD(&cpu_lock));
13721
13722 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13723 VM_BESTFIT | VM_SLEEP);
13724
13725 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13726 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13727 return (NULL);
13728 }
13729
13730 state = ddi_get_soft_state(dtrace_softstate, minor);
13731 state->dts_epid = DTRACE_EPIDNONE + 1;
13732
13733 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
13734 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13735 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13736
13737 if (devp != NULL) {
13738 major = getemajor(*devp);
13739 } else {
13740 major = ddi_driver_major(dtrace_devi);
13741 }
13742
13743 state->dts_dev = makedevice(major, minor);
13744
13745 if (devp != NULL)
13746 *devp = state->dts_dev;
13747
13748 /*
13749 * We allocate NCPU buffers. On the one hand, this can be quite
13750 * a bit of memory per instance (nearly 36K on a Starcat). On the
13751 * other hand, it saves an additional memory reference in the probe
13752 * path.
13753 */
13754 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13755 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13756 state->dts_cleaner = CYCLIC_NONE;
13757 state->dts_deadman = CYCLIC_NONE;
13758 state->dts_vstate.dtvs_state = state;
13759
13760 for (i = 0; i < DTRACEOPT_MAX; i++)
13761 state->dts_options[i] = DTRACEOPT_UNSET;
13762
13763 /*
13764 * Set the default options.
13765 */
13766 opt = state->dts_options;
13767 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13768 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13769 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13770 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13771 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13772 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13773 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13774 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13775 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13776 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13777 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13778 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13779 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13780 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13781
13782 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13783
13784 /*
13785 * Depending on the user credentials, we set flag bits which alter probe
13786 * visibility or the amount of destructiveness allowed. In the case of
13787 * actual anonymous tracing, or the possession of all privileges, all of
13788 * the normal checks are bypassed.
13789 */
13790 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13791 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13792 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13793 } else {
13794 /*
13795 * Set up the credentials for this instantiation. We take a
13796 * hold on the credential to prevent it from disappearing on
13797 * us; this in turn prevents the zone_t referenced by this
13798 * credential from disappearing. This means that we can
13799 * examine the credential and the zone from probe context.
13800 */
13801 crhold(cr);
13802 state->dts_cred.dcr_cred = cr;
13803
13804 /*
13805 * CRA_PROC means "we have *some* privilege for dtrace" and
13806 * unlocks the use of variables like pid, zonename, etc.
13807 */
13808 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13809 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13810 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13811 }
13812
13813 /*
13814 * dtrace_user allows use of syscall and profile providers.
13815 * If the user also has proc_owner and/or proc_zone, we
13816 * extend the scope to include additional visibility and
13817 * destructive power.
13818 */
13819 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13820 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13821 state->dts_cred.dcr_visible |=
13822 DTRACE_CRV_ALLPROC;
13823
13824 state->dts_cred.dcr_action |=
13825 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13826 }
13827
13828 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13829 state->dts_cred.dcr_visible |=
13830 DTRACE_CRV_ALLZONE;
13831
13832 state->dts_cred.dcr_action |=
13833 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13834 }
13835
13836 /*
13837 * If we have all privs in whatever zone this is,
13838 * we can do destructive things to processes which
13839 * have altered credentials.
13840 */
13841 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13842 cr->cr_zone->zone_privset)) {
13843 state->dts_cred.dcr_action |=
13844 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13845 }
13846 }
13847
13848 /*
13849 * Holding the dtrace_kernel privilege also implies that
13850 * the user has the dtrace_user privilege from a visibility
13851 * perspective. But without further privileges, some
13852 * destructive actions are not available.
13853 */
13854 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13855 /*
13856 * Make all probes in all zones visible. However,
13857 * this doesn't mean that all actions become available
13858 * to all zones.
13859 */
13860 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13861 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13862
13863 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13864 DTRACE_CRA_PROC;
13865 /*
13866 * Holding proc_owner means that destructive actions
13867 * for *this* zone are allowed.
13868 */
13869 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13870 state->dts_cred.dcr_action |=
13871 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13872
13873 /*
13874 * Holding proc_zone means that destructive actions
13875 * for this user/group ID in all zones is allowed.
13876 */
13877 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13878 state->dts_cred.dcr_action |=
13879 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13880
13881 /*
13882 * If we have all privs in whatever zone this is,
13883 * we can do destructive things to processes which
13884 * have altered credentials.
13885 */
13886 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13887 cr->cr_zone->zone_privset)) {
13888 state->dts_cred.dcr_action |=
13889 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13890 }
13891 }
13892
13893 /*
13894 * Holding the dtrace_proc privilege gives control over fasttrap
13895 * and pid providers. We need to grant wider destructive
13896 * privileges in the event that the user has proc_owner and/or
13897 * proc_zone.
13898 */
13899 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
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 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13905 state->dts_cred.dcr_action |=
13906 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13907 }
13908 }
13909
13910 return (state);
13911 }
13912
13913 static int
13914 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13915 {
13916 dtrace_optval_t *opt = state->dts_options, size;
13917 processorid_t cpu;
13918 int flags = 0, rval, factor, divisor = 1;
13919
13920 ASSERT(MUTEX_HELD(&dtrace_lock));
13921 ASSERT(MUTEX_HELD(&cpu_lock));
13922 ASSERT(which < DTRACEOPT_MAX);
13923 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13924 (state == dtrace_anon.dta_state &&
13925 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13926
13927 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13928 return (0);
13929
13930 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13931 cpu = opt[DTRACEOPT_CPU];
13932
13933 if (which == DTRACEOPT_SPECSIZE)
13934 flags |= DTRACEBUF_NOSWITCH;
13935
13936 if (which == DTRACEOPT_BUFSIZE) {
13937 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13938 flags |= DTRACEBUF_RING;
13939
13940 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13941 flags |= DTRACEBUF_FILL;
13942
13943 if (state != dtrace_anon.dta_state ||
13944 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13945 flags |= DTRACEBUF_INACTIVE;
13946 }
13947
13948 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
13949 /*
13950 * The size must be 8-byte aligned. If the size is not 8-byte
13951 * aligned, drop it down by the difference.
13952 */
13953 if (size & (sizeof (uint64_t) - 1))
13954 size -= size & (sizeof (uint64_t) - 1);
13955
13956 if (size < state->dts_reserve) {
13957 /*
13958 * Buffers always must be large enough to accommodate
13959 * their prereserved space. We return E2BIG instead
13960 * of ENOMEM in this case to allow for user-level
13961 * software to differentiate the cases.
13962 */
13963 return (E2BIG);
13964 }
13965
13966 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
13967
13968 if (rval != ENOMEM) {
13969 opt[which] = size;
13970 return (rval);
13971 }
13972
13973 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13974 return (rval);
13975
13976 for (divisor = 2; divisor < factor; divisor <<= 1)
13977 continue;
13978 }
13979
13980 return (ENOMEM);
13981 }
13982
13983 static int
13984 dtrace_state_buffers(dtrace_state_t *state)
13985 {
13986 dtrace_speculation_t *spec = state->dts_speculations;
13987 int rval, i;
13988
13989 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13990 DTRACEOPT_BUFSIZE)) != 0)
13991 return (rval);
13992
13993 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13994 DTRACEOPT_AGGSIZE)) != 0)
13995 return (rval);
13996
13997 for (i = 0; i < state->dts_nspeculations; i++) {
13998 if ((rval = dtrace_state_buffer(state,
13999 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
14000 return (rval);
14001 }
14002
14003 return (0);
14004 }
14005
14006 static void
14007 dtrace_state_prereserve(dtrace_state_t *state)
14008 {
14009 dtrace_ecb_t *ecb;
14010 dtrace_probe_t *probe;
14011
14012 state->dts_reserve = 0;
14013
14014 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
14015 return;
14016
14017 /*
14018 * If our buffer policy is a "fill" buffer policy, we need to set the
14019 * prereserved space to be the space required by the END probes.
14020 */
14021 probe = dtrace_probes[dtrace_probeid_end - 1];
14022 ASSERT(probe != NULL);
14023
14024 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
14025 if (ecb->dte_state != state)
14026 continue;
14027
14028 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
14029 }
14030 }
14031
14032 static int
14033 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
14034 {
14035 dtrace_optval_t *opt = state->dts_options, sz, nspec;
14036 dtrace_speculation_t *spec;
14037 dtrace_buffer_t *buf;
14038 cyc_handler_t hdlr;
14039 cyc_time_t when;
14040 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14041 dtrace_icookie_t cookie;
14042
14043 mutex_enter(&cpu_lock);
14044 mutex_enter(&dtrace_lock);
14045
14046 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
14047 rval = EBUSY;
14048 goto out;
14049 }
14050
14051 /*
14052 * Before we can perform any checks, we must prime all of the
14053 * retained enablings that correspond to this state.
14054 */
14055 dtrace_enabling_prime(state);
14056
14057 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
14058 rval = EACCES;
14059 goto out;
14060 }
14061
14062 dtrace_state_prereserve(state);
14063
14064 /*
14065 * Now we want to do is try to allocate our speculations.
14066 * We do not automatically resize the number of speculations; if
14067 * this fails, we will fail the operation.
14068 */
14069 nspec = opt[DTRACEOPT_NSPEC];
14070 ASSERT(nspec != DTRACEOPT_UNSET);
14071
14072 if (nspec > INT_MAX) {
14073 rval = ENOMEM;
14074 goto out;
14075 }
14076
14077 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
14078 KM_NOSLEEP | KM_NORMALPRI);
14079
14080 if (spec == NULL) {
14081 rval = ENOMEM;
14082 goto out;
14083 }
14084
14085 state->dts_speculations = spec;
14086 state->dts_nspeculations = (int)nspec;
14087
14088 for (i = 0; i < nspec; i++) {
14089 if ((buf = kmem_zalloc(bufsize,
14090 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
14091 rval = ENOMEM;
14092 goto err;
14093 }
14094
14095 spec[i].dtsp_buffer = buf;
14096 }
14097
14098 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
14099 if (dtrace_anon.dta_state == NULL) {
14100 rval = ENOENT;
14101 goto out;
14102 }
14103
14104 if (state->dts_necbs != 0) {
14105 rval = EALREADY;
14106 goto out;
14107 }
14108
14109 state->dts_anon = dtrace_anon_grab();
14110 ASSERT(state->dts_anon != NULL);
14111 state = state->dts_anon;
14112
14113 /*
14114 * We want "grabanon" to be set in the grabbed state, so we'll
14115 * copy that option value from the grabbing state into the
14116 * grabbed state.
14117 */
14118 state->dts_options[DTRACEOPT_GRABANON] =
14119 opt[DTRACEOPT_GRABANON];
14120
14121 *cpu = dtrace_anon.dta_beganon;
14122
14123 /*
14124 * If the anonymous state is active (as it almost certainly
14125 * is if the anonymous enabling ultimately matched anything),
14126 * we don't allow any further option processing -- but we
14127 * don't return failure.
14128 */
14129 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14130 goto out;
14131 }
14132
14133 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
14134 opt[DTRACEOPT_AGGSIZE] != 0) {
14135 if (state->dts_aggregations == NULL) {
14136 /*
14137 * We're not going to create an aggregation buffer
14138 * because we don't have any ECBs that contain
14139 * aggregations -- set this option to 0.
14140 */
14141 opt[DTRACEOPT_AGGSIZE] = 0;
14142 } else {
14143 /*
14144 * If we have an aggregation buffer, we must also have
14145 * a buffer to use as scratch.
14146 */
14147 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
14148 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
14149 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
14150 }
14151 }
14152 }
14153
14154 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
14155 opt[DTRACEOPT_SPECSIZE] != 0) {
14156 if (!state->dts_speculates) {
14157 /*
14158 * We're not going to create speculation buffers
14159 * because we don't have any ECBs that actually
14160 * speculate -- set the speculation size to 0.
14161 */
14162 opt[DTRACEOPT_SPECSIZE] = 0;
14163 }
14164 }
14165
14166 /*
14167 * The bare minimum size for any buffer that we're actually going to
14168 * do anything to is sizeof (uint64_t).
14169 */
14170 sz = sizeof (uint64_t);
14171
14172 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
14173 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
14174 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
14175 /*
14176 * A buffer size has been explicitly set to 0 (or to a size
14177 * that will be adjusted to 0) and we need the space -- we
14178 * need to return failure. We return ENOSPC to differentiate
14179 * it from failing to allocate a buffer due to failure to meet
14180 * the reserve (for which we return E2BIG).
14181 */
14182 rval = ENOSPC;
14183 goto out;
14184 }
14185
14186 if ((rval = dtrace_state_buffers(state)) != 0)
14187 goto err;
14188
14189 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
14190 sz = dtrace_dstate_defsize;
14191
14192 do {
14193 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
14194
14195 if (rval == 0)
14196 break;
14197
14198 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14199 goto err;
14200 } while (sz >>= 1);
14201
14202 opt[DTRACEOPT_DYNVARSIZE] = sz;
14203
14204 if (rval != 0)
14205 goto err;
14206
14207 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
14208 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
14209
14210 if (opt[DTRACEOPT_CLEANRATE] == 0)
14211 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14212
14213 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
14214 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
14215
14216 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
14217 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14218
14219 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
14220 hdlr.cyh_arg = state;
14221 hdlr.cyh_level = CY_LOW_LEVEL;
14222
14223 when.cyt_when = 0;
14224 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
14225
14226 state->dts_cleaner = cyclic_add(&hdlr, &when);
14227
14228 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
14229 hdlr.cyh_arg = state;
14230 hdlr.cyh_level = CY_LOW_LEVEL;
14231
14232 when.cyt_when = 0;
14233 when.cyt_interval = dtrace_deadman_interval;
14234
14235 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
14236 state->dts_deadman = cyclic_add(&hdlr, &when);
14237
14238 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
14239
14240 if (state->dts_getf != 0 &&
14241 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14242 /*
14243 * We don't have kernel privs but we have at least one call
14244 * to getf(); we need to bump our zone's count, and (if
14245 * this is the first enabling to have an unprivileged call
14246 * to getf()) we need to hook into closef().
14247 */
14248 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
14249
14250 if (dtrace_getf++ == 0) {
14251 ASSERT(dtrace_closef == NULL);
14252 dtrace_closef = dtrace_getf_barrier;
14253 }
14254 }
14255
14256 /*
14257 * Now it's time to actually fire the BEGIN probe. We need to disable
14258 * interrupts here both to record the CPU on which we fired the BEGIN
14259 * probe (the data from this CPU will be processed first at user
14260 * level) and to manually activate the buffer for this CPU.
14261 */
14262 cookie = dtrace_interrupt_disable();
14263 *cpu = CPU->cpu_id;
14264 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
14265 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
14266
14267 dtrace_probe(dtrace_probeid_begin,
14268 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
14269 dtrace_interrupt_enable(cookie);
14270 /*
14271 * We may have had an exit action from a BEGIN probe; only change our
14272 * state to ACTIVE if we're still in WARMUP.
14273 */
14274 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
14275 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
14276
14277 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
14278 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
14279
14280 /*
14281 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
14282 * want each CPU to transition its principal buffer out of the
14283 * INACTIVE state. Doing this assures that no CPU will suddenly begin
14284 * processing an ECB halfway down a probe's ECB chain; all CPUs will
14285 * atomically transition from processing none of a state's ECBs to
14286 * processing all of them.
14287 */
14288 dtrace_xcall(DTRACE_CPUALL,
14289 (dtrace_xcall_t)dtrace_buffer_activate, state);
14290 goto out;
14291
14292 err:
14293 dtrace_buffer_free(state->dts_buffer);
14294 dtrace_buffer_free(state->dts_aggbuffer);
14295
14296 if ((nspec = state->dts_nspeculations) == 0) {
14297 ASSERT(state->dts_speculations == NULL);
14298 goto out;
14299 }
14300
14301 spec = state->dts_speculations;
14302 ASSERT(spec != NULL);
14303
14304 for (i = 0; i < state->dts_nspeculations; i++) {
14305 if ((buf = spec[i].dtsp_buffer) == NULL)
14306 break;
14307
14308 dtrace_buffer_free(buf);
14309 kmem_free(buf, bufsize);
14310 }
14311
14312 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14313 state->dts_nspeculations = 0;
14314 state->dts_speculations = NULL;
14315
14316 out:
14317 mutex_exit(&dtrace_lock);
14318 mutex_exit(&cpu_lock);
14319
14320 return (rval);
14321 }
14322
14323 static int
14324 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
14325 {
14326 dtrace_icookie_t cookie;
14327
14328 ASSERT(MUTEX_HELD(&dtrace_lock));
14329
14330 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
14331 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
14332 return (EINVAL);
14333
14334 /*
14335 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
14336 * to be sure that every CPU has seen it. See below for the details
14337 * on why this is done.
14338 */
14339 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
14340 dtrace_sync();
14341
14342 /*
14343 * By this point, it is impossible for any CPU to be still processing
14344 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
14345 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
14346 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
14347 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
14348 * iff we're in the END probe.
14349 */
14350 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
14351 dtrace_sync();
14352 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
14353
14354 /*
14355 * Finally, we can release the reserve and call the END probe. We
14356 * disable interrupts across calling the END probe to allow us to
14357 * return the CPU on which we actually called the END probe. This
14358 * allows user-land to be sure that this CPU's principal buffer is
14359 * processed last.
14360 */
14361 state->dts_reserve = 0;
14362
14363 cookie = dtrace_interrupt_disable();
14364 *cpu = CPU->cpu_id;
14365 dtrace_probe(dtrace_probeid_end,
14366 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
14367 dtrace_interrupt_enable(cookie);
14368
14369 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
14370 dtrace_sync();
14371
14372 if (state->dts_getf != 0 &&
14373 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14374 /*
14375 * We don't have kernel privs but we have at least one call
14376 * to getf(); we need to lower our zone's count, and (if
14377 * this is the last enabling to have an unprivileged call
14378 * to getf()) we need to clear the closef() hook.
14379 */
14380 ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
14381 ASSERT(dtrace_closef == dtrace_getf_barrier);
14382 ASSERT(dtrace_getf > 0);
14383
14384 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
14385
14386 if (--dtrace_getf == 0)
14387 dtrace_closef = NULL;
14388 }
14389
14390 return (0);
14391 }
14392
14393 static int
14394 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
14395 dtrace_optval_t val)
14396 {
14397 ASSERT(MUTEX_HELD(&dtrace_lock));
14398
14399 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14400 return (EBUSY);
14401
14402 if (option >= DTRACEOPT_MAX)
14403 return (EINVAL);
14404
14405 if (option != DTRACEOPT_CPU && val < 0)
14406 return (EINVAL);
14407
14408 switch (option) {
14409 case DTRACEOPT_DESTRUCTIVE:
14410 if (dtrace_destructive_disallow)
14411 return (EACCES);
14412
14413 state->dts_cred.dcr_destructive = 1;
14414 break;
14415
14416 case DTRACEOPT_BUFSIZE:
14417 case DTRACEOPT_DYNVARSIZE:
14418 case DTRACEOPT_AGGSIZE:
14419 case DTRACEOPT_SPECSIZE:
14420 case DTRACEOPT_STRSIZE:
14421 if (val < 0)
14422 return (EINVAL);
14423
14424 if (val >= LONG_MAX) {
14425 /*
14426 * If this is an otherwise negative value, set it to
14427 * the highest multiple of 128m less than LONG_MAX.
14428 * Technically, we're adjusting the size without
14429 * regard to the buffer resizing policy, but in fact,
14430 * this has no effect -- if we set the buffer size to
14431 * ~LONG_MAX and the buffer policy is ultimately set to
14432 * be "manual", the buffer allocation is guaranteed to
14433 * fail, if only because the allocation requires two
14434 * buffers. (We set the the size to the highest
14435 * multiple of 128m because it ensures that the size
14436 * will remain a multiple of a megabyte when
14437 * repeatedly halved -- all the way down to 15m.)
14438 */
14439 val = LONG_MAX - (1 << 27) + 1;
14440 }
14441 }
14442
14443 state->dts_options[option] = val;
14444
14445 return (0);
14446 }
14447
14448 static void
14449 dtrace_state_destroy(dtrace_state_t *state)
14450 {
14451 dtrace_ecb_t *ecb;
14452 dtrace_vstate_t *vstate = &state->dts_vstate;
14453 minor_t minor = getminor(state->dts_dev);
14454 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14455 dtrace_speculation_t *spec = state->dts_speculations;
14456 int nspec = state->dts_nspeculations;
14457 uint32_t match;
14458
14459 ASSERT(MUTEX_HELD(&dtrace_lock));
14460 ASSERT(MUTEX_HELD(&cpu_lock));
14461
14462 /*
14463 * First, retract any retained enablings for this state.
14464 */
14465 dtrace_enabling_retract(state);
14466 ASSERT(state->dts_nretained == 0);
14467
14468 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
14469 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
14470 /*
14471 * We have managed to come into dtrace_state_destroy() on a
14472 * hot enabling -- almost certainly because of a disorderly
14473 * shutdown of a consumer. (That is, a consumer that is
14474 * exiting without having called dtrace_stop().) In this case,
14475 * we're going to set our activity to be KILLED, and then
14476 * issue a sync to be sure that everyone is out of probe
14477 * context before we start blowing away ECBs.
14478 */
14479 state->dts_activity = DTRACE_ACTIVITY_KILLED;
14480 dtrace_sync();
14481 }
14482
14483 /*
14484 * Release the credential hold we took in dtrace_state_create().
14485 */
14486 if (state->dts_cred.dcr_cred != NULL)
14487 crfree(state->dts_cred.dcr_cred);
14488
14489 /*
14490 * Now we can safely disable and destroy any enabled probes. Because
14491 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
14492 * (especially if they're all enabled), we take two passes through the
14493 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
14494 * in the second we disable whatever is left over.
14495 */
14496 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
14497 for (i = 0; i < state->dts_necbs; i++) {
14498 if ((ecb = state->dts_ecbs[i]) == NULL)
14499 continue;
14500
14501 if (match && ecb->dte_probe != NULL) {
14502 dtrace_probe_t *probe = ecb->dte_probe;
14503 dtrace_provider_t *prov = probe->dtpr_provider;
14504
14505 if (!(prov->dtpv_priv.dtpp_flags & match))
14506 continue;
14507 }
14508
14509 dtrace_ecb_disable(ecb);
14510 dtrace_ecb_destroy(ecb);
14511 }
14512
14513 if (!match)
14514 break;
14515 }
14516
14517 /*
14518 * Before we free the buffers, perform one more sync to assure that
14519 * every CPU is out of probe context.
14520 */
14521 dtrace_sync();
14522
14523 dtrace_buffer_free(state->dts_buffer);
14524 dtrace_buffer_free(state->dts_aggbuffer);
14525
14526 for (i = 0; i < nspec; i++)
14527 dtrace_buffer_free(spec[i].dtsp_buffer);
14528
14529 if (state->dts_cleaner != CYCLIC_NONE)
14530 cyclic_remove(state->dts_cleaner);
14531
14532 if (state->dts_deadman != CYCLIC_NONE)
14533 cyclic_remove(state->dts_deadman);
14534
14535 dtrace_dstate_fini(&vstate->dtvs_dynvars);
14536 dtrace_vstate_fini(vstate);
14537 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14538
14539 if (state->dts_aggregations != NULL) {
14540 #ifdef DEBUG
14541 for (i = 0; i < state->dts_naggregations; i++)
14542 ASSERT(state->dts_aggregations[i] == NULL);
14543 #endif
14544 ASSERT(state->dts_naggregations > 0);
14545 kmem_free(state->dts_aggregations,
14546 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14547 }
14548
14549 kmem_free(state->dts_buffer, bufsize);
14550 kmem_free(state->dts_aggbuffer, bufsize);
14551
14552 for (i = 0; i < nspec; i++)
14553 kmem_free(spec[i].dtsp_buffer, bufsize);
14554
14555 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14556
14557 dtrace_format_destroy(state);
14558
14559 vmem_destroy(state->dts_aggid_arena);
14560 ddi_soft_state_free(dtrace_softstate, minor);
14561 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14562 }
14563
14564 /*
14565 * DTrace Anonymous Enabling Functions
14566 */
14567 static dtrace_state_t *
14568 dtrace_anon_grab(void)
14569 {
14570 dtrace_state_t *state;
14571
14572 ASSERT(MUTEX_HELD(&dtrace_lock));
14573
14574 if ((state = dtrace_anon.dta_state) == NULL) {
14575 ASSERT(dtrace_anon.dta_enabling == NULL);
14576 return (NULL);
14577 }
14578
14579 ASSERT(dtrace_anon.dta_enabling != NULL);
14580 ASSERT(dtrace_retained != NULL);
14581
14582 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14583 dtrace_anon.dta_enabling = NULL;
14584 dtrace_anon.dta_state = NULL;
14585
14586 return (state);
14587 }
14588
14589 static void
14590 dtrace_anon_property(void)
14591 {
14592 int i, rv;
14593 dtrace_state_t *state;
14594 dof_hdr_t *dof;
14595 char c[32]; /* enough for "dof-data-" + digits */
14596
14597 ASSERT(MUTEX_HELD(&dtrace_lock));
14598 ASSERT(MUTEX_HELD(&cpu_lock));
14599
14600 for (i = 0; ; i++) {
14601 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14602
14603 dtrace_err_verbose = 1;
14604
14605 if ((dof = dtrace_dof_property(c)) == NULL) {
14606 dtrace_err_verbose = 0;
14607 break;
14608 }
14609
14610 /*
14611 * We want to create anonymous state, so we need to transition
14612 * the kernel debugger to indicate that DTrace is active. If
14613 * this fails (e.g. because the debugger has modified text in
14614 * some way), we won't continue with the processing.
14615 */
14616 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14617 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14618 "enabling ignored.");
14619 dtrace_dof_destroy(dof);
14620 break;
14621 }
14622
14623 /*
14624 * If we haven't allocated an anonymous state, we'll do so now.
14625 */
14626 if ((state = dtrace_anon.dta_state) == NULL) {
14627 state = dtrace_state_create(NULL, NULL);
14628 dtrace_anon.dta_state = state;
14629
14630 if (state == NULL) {
14631 /*
14632 * This basically shouldn't happen: the only
14633 * failure mode from dtrace_state_create() is a
14634 * failure of ddi_soft_state_zalloc() that
14635 * itself should never happen. Still, the
14636 * interface allows for a failure mode, and
14637 * we want to fail as gracefully as possible:
14638 * we'll emit an error message and cease
14639 * processing anonymous state in this case.
14640 */
14641 cmn_err(CE_WARN, "failed to create "
14642 "anonymous state");
14643 dtrace_dof_destroy(dof);
14644 break;
14645 }
14646 }
14647
14648 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14649 &dtrace_anon.dta_enabling, 0, B_TRUE);
14650
14651 if (rv == 0)
14652 rv = dtrace_dof_options(dof, state);
14653
14654 dtrace_err_verbose = 0;
14655 dtrace_dof_destroy(dof);
14656
14657 if (rv != 0) {
14658 /*
14659 * This is malformed DOF; chuck any anonymous state
14660 * that we created.
14661 */
14662 ASSERT(dtrace_anon.dta_enabling == NULL);
14663 dtrace_state_destroy(state);
14664 dtrace_anon.dta_state = NULL;
14665 break;
14666 }
14667
14668 ASSERT(dtrace_anon.dta_enabling != NULL);
14669 }
14670
14671 if (dtrace_anon.dta_enabling != NULL) {
14672 int rval;
14673
14674 /*
14675 * dtrace_enabling_retain() can only fail because we are
14676 * trying to retain more enablings than are allowed -- but
14677 * we only have one anonymous enabling, and we are guaranteed
14678 * to be allowed at least one retained enabling; we assert
14679 * that dtrace_enabling_retain() returns success.
14680 */
14681 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14682 ASSERT(rval == 0);
14683
14684 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14685 }
14686 }
14687
14688 /*
14689 * DTrace Helper Functions
14690 */
14691 static void
14692 dtrace_helper_trace(dtrace_helper_action_t *helper,
14693 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14694 {
14695 uint32_t size, next, nnext, i;
14696 dtrace_helptrace_t *ent, *buffer;
14697 uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14698
14699 if ((buffer = dtrace_helptrace_buffer) == NULL)
14700 return;
14701
14702 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14703
14704 /*
14705 * What would a tracing framework be without its own tracing
14706 * framework? (Well, a hell of a lot simpler, for starters...)
14707 */
14708 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14709 sizeof (uint64_t) - sizeof (uint64_t);
14710
14711 /*
14712 * Iterate until we can allocate a slot in the trace buffer.
14713 */
14714 do {
14715 next = dtrace_helptrace_next;
14716
14717 if (next + size < dtrace_helptrace_bufsize) {
14718 nnext = next + size;
14719 } else {
14720 nnext = size;
14721 }
14722 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14723
14724 /*
14725 * We have our slot; fill it in.
14726 */
14727 if (nnext == size) {
14728 dtrace_helptrace_wrapped++;
14729 next = 0;
14730 }
14731
14732 ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
14733 ent->dtht_helper = helper;
14734 ent->dtht_where = where;
14735 ent->dtht_nlocals = vstate->dtvs_nlocals;
14736
14737 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14738 mstate->dtms_fltoffs : -1;
14739 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14740 ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
14741
14742 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14743 dtrace_statvar_t *svar;
14744
14745 if ((svar = vstate->dtvs_locals[i]) == NULL)
14746 continue;
14747
14748 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14749 ent->dtht_locals[i] =
14750 ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
14751 }
14752 }
14753
14754 static uint64_t
14755 dtrace_helper(int which, dtrace_mstate_t *mstate,
14756 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14757 {
14758 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14759 uint64_t sarg0 = mstate->dtms_arg[0];
14760 uint64_t sarg1 = mstate->dtms_arg[1];
14761 uint64_t rval;
14762 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14763 dtrace_helper_action_t *helper;
14764 dtrace_vstate_t *vstate;
14765 dtrace_difo_t *pred;
14766 int i, trace = dtrace_helptrace_buffer != NULL;
14767
14768 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14769
14770 if (helpers == NULL)
14771 return (0);
14772
14773 if ((helper = helpers->dthps_actions[which]) == NULL)
14774 return (0);
14775
14776 vstate = &helpers->dthps_vstate;
14777 mstate->dtms_arg[0] = arg0;
14778 mstate->dtms_arg[1] = arg1;
14779
14780 /*
14781 * Now iterate over each helper. If its predicate evaluates to 'true',
14782 * we'll call the corresponding actions. Note that the below calls
14783 * to dtrace_dif_emulate() may set faults in machine state. This is
14784 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14785 * the stored DIF offset with its own (which is the desired behavior).
14786 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14787 * from machine state; this is okay, too.
14788 */
14789 for (; helper != NULL; helper = helper->dtha_next) {
14790 if ((pred = helper->dtha_predicate) != NULL) {
14791 if (trace)
14792 dtrace_helper_trace(helper, mstate, vstate, 0);
14793
14794 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14795 goto next;
14796
14797 if (*flags & CPU_DTRACE_FAULT)
14798 goto err;
14799 }
14800
14801 for (i = 0; i < helper->dtha_nactions; i++) {
14802 if (trace)
14803 dtrace_helper_trace(helper,
14804 mstate, vstate, i + 1);
14805
14806 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14807 mstate, vstate, state);
14808
14809 if (*flags & CPU_DTRACE_FAULT)
14810 goto err;
14811 }
14812
14813 next:
14814 if (trace)
14815 dtrace_helper_trace(helper, mstate, vstate,
14816 DTRACE_HELPTRACE_NEXT);
14817 }
14818
14819 if (trace)
14820 dtrace_helper_trace(helper, mstate, vstate,
14821 DTRACE_HELPTRACE_DONE);
14822
14823 /*
14824 * Restore the arg0 that we saved upon entry.
14825 */
14826 mstate->dtms_arg[0] = sarg0;
14827 mstate->dtms_arg[1] = sarg1;
14828
14829 return (rval);
14830
14831 err:
14832 if (trace)
14833 dtrace_helper_trace(helper, mstate, vstate,
14834 DTRACE_HELPTRACE_ERR);
14835
14836 /*
14837 * Restore the arg0 that we saved upon entry.
14838 */
14839 mstate->dtms_arg[0] = sarg0;
14840 mstate->dtms_arg[1] = sarg1;
14841
14842 return (NULL);
14843 }
14844
14845 static void
14846 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14847 dtrace_vstate_t *vstate)
14848 {
14849 int i;
14850
14851 if (helper->dtha_predicate != NULL)
14852 dtrace_difo_release(helper->dtha_predicate, vstate);
14853
14854 for (i = 0; i < helper->dtha_nactions; i++) {
14855 ASSERT(helper->dtha_actions[i] != NULL);
14856 dtrace_difo_release(helper->dtha_actions[i], vstate);
14857 }
14858
14859 kmem_free(helper->dtha_actions,
14860 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14861 kmem_free(helper, sizeof (dtrace_helper_action_t));
14862 }
14863
14864 static int
14865 dtrace_helper_destroygen(int gen)
14866 {
14867 proc_t *p = curproc;
14868 dtrace_helpers_t *help = p->p_dtrace_helpers;
14869 dtrace_vstate_t *vstate;
14870 int i;
14871
14872 ASSERT(MUTEX_HELD(&dtrace_lock));
14873
14874 if (help == NULL || gen > help->dthps_generation)
14875 return (EINVAL);
14876
14877 vstate = &help->dthps_vstate;
14878
14879 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14880 dtrace_helper_action_t *last = NULL, *h, *next;
14881
14882 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14883 next = h->dtha_next;
14884
14885 if (h->dtha_generation == gen) {
14886 if (last != NULL) {
14887 last->dtha_next = next;
14888 } else {
14889 help->dthps_actions[i] = next;
14890 }
14891
14892 dtrace_helper_action_destroy(h, vstate);
14893 } else {
14894 last = h;
14895 }
14896 }
14897 }
14898
14899 /*
14900 * Interate until we've cleared out all helper providers with the
14901 * given generation number.
14902 */
14903 for (;;) {
14904 dtrace_helper_provider_t *prov;
14905
14906 /*
14907 * Look for a helper provider with the right generation. We
14908 * have to start back at the beginning of the list each time
14909 * because we drop dtrace_lock. It's unlikely that we'll make
14910 * more than two passes.
14911 */
14912 for (i = 0; i < help->dthps_nprovs; i++) {
14913 prov = help->dthps_provs[i];
14914
14915 if (prov->dthp_generation == gen)
14916 break;
14917 }
14918
14919 /*
14920 * If there were no matches, we're done.
14921 */
14922 if (i == help->dthps_nprovs)
14923 break;
14924
14925 /*
14926 * Move the last helper provider into this slot.
14927 */
14928 help->dthps_nprovs--;
14929 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14930 help->dthps_provs[help->dthps_nprovs] = NULL;
14931
14932 mutex_exit(&dtrace_lock);
14933
14934 /*
14935 * If we have a meta provider, remove this helper provider.
14936 */
14937 mutex_enter(&dtrace_meta_lock);
14938 if (dtrace_meta_pid != NULL) {
14939 ASSERT(dtrace_deferred_pid == NULL);
14940 dtrace_helper_provider_remove(&prov->dthp_prov,
14941 p->p_pid);
14942 }
14943 mutex_exit(&dtrace_meta_lock);
14944
14945 dtrace_helper_provider_destroy(prov);
14946
14947 mutex_enter(&dtrace_lock);
14948 }
14949
14950 return (0);
14951 }
14952
14953 static int
14954 dtrace_helper_validate(dtrace_helper_action_t *helper)
14955 {
14956 int err = 0, i;
14957 dtrace_difo_t *dp;
14958
14959 if ((dp = helper->dtha_predicate) != NULL)
14960 err += dtrace_difo_validate_helper(dp);
14961
14962 for (i = 0; i < helper->dtha_nactions; i++)
14963 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14964
14965 return (err == 0);
14966 }
14967
14968 static int
14969 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14970 {
14971 dtrace_helpers_t *help;
14972 dtrace_helper_action_t *helper, *last;
14973 dtrace_actdesc_t *act;
14974 dtrace_vstate_t *vstate;
14975 dtrace_predicate_t *pred;
14976 int count = 0, nactions = 0, i;
14977
14978 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14979 return (EINVAL);
14980
14981 help = curproc->p_dtrace_helpers;
14982 last = help->dthps_actions[which];
14983 vstate = &help->dthps_vstate;
14984
14985 for (count = 0; last != NULL; last = last->dtha_next) {
14986 count++;
14987 if (last->dtha_next == NULL)
14988 break;
14989 }
14990
14991 /*
14992 * If we already have dtrace_helper_actions_max helper actions for this
14993 * helper action type, we'll refuse to add a new one.
14994 */
14995 if (count >= dtrace_helper_actions_max)
14996 return (ENOSPC);
14997
14998 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14999 helper->dtha_generation = help->dthps_generation;
15000
15001 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
15002 ASSERT(pred->dtp_difo != NULL);
15003 dtrace_difo_hold(pred->dtp_difo);
15004 helper->dtha_predicate = pred->dtp_difo;
15005 }
15006
15007 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
15008 if (act->dtad_kind != DTRACEACT_DIFEXPR)
15009 goto err;
15010
15011 if (act->dtad_difo == NULL)
15012 goto err;
15013
15014 nactions++;
15015 }
15016
15017 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
15018 (helper->dtha_nactions = nactions), KM_SLEEP);
15019
15020 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
15021 dtrace_difo_hold(act->dtad_difo);
15022 helper->dtha_actions[i++] = act->dtad_difo;
15023 }
15024
15025 if (!dtrace_helper_validate(helper))
15026 goto err;
15027
15028 if (last == NULL) {
15029 help->dthps_actions[which] = helper;
15030 } else {
15031 last->dtha_next = helper;
15032 }
15033
15034 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
15035 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
15036 dtrace_helptrace_next = 0;
15037 }
15038
15039 return (0);
15040 err:
15041 dtrace_helper_action_destroy(helper, vstate);
15042 return (EINVAL);
15043 }
15044
15045 static void
15046 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
15047 dof_helper_t *dofhp)
15048 {
15049 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
15050
15051 mutex_enter(&dtrace_meta_lock);
15052 mutex_enter(&dtrace_lock);
15053
15054 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
15055 /*
15056 * If the dtrace module is loaded but not attached, or if
15057 * there aren't isn't a meta provider registered to deal with
15058 * these provider descriptions, we need to postpone creating
15059 * the actual providers until later.
15060 */
15061
15062 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
15063 dtrace_deferred_pid != help) {
15064 help->dthps_deferred = 1;
15065 help->dthps_pid = p->p_pid;
15066 help->dthps_next = dtrace_deferred_pid;
15067 help->dthps_prev = NULL;
15068 if (dtrace_deferred_pid != NULL)
15069 dtrace_deferred_pid->dthps_prev = help;
15070 dtrace_deferred_pid = help;
15071 }
15072
15073 mutex_exit(&dtrace_lock);
15074
15075 } else if (dofhp != NULL) {
15076 /*
15077 * If the dtrace module is loaded and we have a particular
15078 * helper provider description, pass that off to the
15079 * meta provider.
15080 */
15081
15082 mutex_exit(&dtrace_lock);
15083
15084 dtrace_helper_provide(dofhp, p->p_pid);
15085
15086 } else {
15087 /*
15088 * Otherwise, just pass all the helper provider descriptions
15089 * off to the meta provider.
15090 */
15091
15092 int i;
15093 mutex_exit(&dtrace_lock);
15094
15095 for (i = 0; i < help->dthps_nprovs; i++) {
15096 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
15097 p->p_pid);
15098 }
15099 }
15100
15101 mutex_exit(&dtrace_meta_lock);
15102 }
15103
15104 static int
15105 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
15106 {
15107 dtrace_helpers_t *help;
15108 dtrace_helper_provider_t *hprov, **tmp_provs;
15109 uint_t tmp_maxprovs, i;
15110
15111 ASSERT(MUTEX_HELD(&dtrace_lock));
15112
15113 help = curproc->p_dtrace_helpers;
15114 ASSERT(help != NULL);
15115
15116 /*
15117 * If we already have dtrace_helper_providers_max helper providers,
15118 * we're refuse to add a new one.
15119 */
15120 if (help->dthps_nprovs >= dtrace_helper_providers_max)
15121 return (ENOSPC);
15122
15123 /*
15124 * Check to make sure this isn't a duplicate.
15125 */
15126 for (i = 0; i < help->dthps_nprovs; i++) {
15127 if (dofhp->dofhp_addr ==
15128 help->dthps_provs[i]->dthp_prov.dofhp_addr)
15129 return (EALREADY);
15130 }
15131
15132 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
15133 hprov->dthp_prov = *dofhp;
15134 hprov->dthp_ref = 1;
15135 hprov->dthp_generation = gen;
15136
15137 /*
15138 * Allocate a bigger table for helper providers if it's already full.
15139 */
15140 if (help->dthps_maxprovs == help->dthps_nprovs) {
15141 tmp_maxprovs = help->dthps_maxprovs;
15142 tmp_provs = help->dthps_provs;
15143
15144 if (help->dthps_maxprovs == 0)
15145 help->dthps_maxprovs = 2;
15146 else
15147 help->dthps_maxprovs *= 2;
15148 if (help->dthps_maxprovs > dtrace_helper_providers_max)
15149 help->dthps_maxprovs = dtrace_helper_providers_max;
15150
15151 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
15152
15153 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
15154 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15155
15156 if (tmp_provs != NULL) {
15157 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
15158 sizeof (dtrace_helper_provider_t *));
15159 kmem_free(tmp_provs, tmp_maxprovs *
15160 sizeof (dtrace_helper_provider_t *));
15161 }
15162 }
15163
15164 help->dthps_provs[help->dthps_nprovs] = hprov;
15165 help->dthps_nprovs++;
15166
15167 return (0);
15168 }
15169
15170 static void
15171 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
15172 {
15173 mutex_enter(&dtrace_lock);
15174
15175 if (--hprov->dthp_ref == 0) {
15176 dof_hdr_t *dof;
15177 mutex_exit(&dtrace_lock);
15178 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
15179 dtrace_dof_destroy(dof);
15180 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
15181 } else {
15182 mutex_exit(&dtrace_lock);
15183 }
15184 }
15185
15186 static int
15187 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
15188 {
15189 uintptr_t daddr = (uintptr_t)dof;
15190 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
15191 dof_provider_t *provider;
15192 dof_probe_t *probe;
15193 uint8_t *arg;
15194 char *strtab, *typestr;
15195 dof_stridx_t typeidx;
15196 size_t typesz;
15197 uint_t nprobes, j, k;
15198
15199 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
15200
15201 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
15202 dtrace_dof_error(dof, "misaligned section offset");
15203 return (-1);
15204 }
15205
15206 /*
15207 * The section needs to be large enough to contain the DOF provider
15208 * structure appropriate for the given version.
15209 */
15210 if (sec->dofs_size <
15211 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
15212 offsetof(dof_provider_t, dofpv_prenoffs) :
15213 sizeof (dof_provider_t))) {
15214 dtrace_dof_error(dof, "provider section too small");
15215 return (-1);
15216 }
15217
15218 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
15219 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
15220 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
15221 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
15222 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
15223
15224 if (str_sec == NULL || prb_sec == NULL ||
15225 arg_sec == NULL || off_sec == NULL)
15226 return (-1);
15227
15228 enoff_sec = NULL;
15229
15230 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
15231 provider->dofpv_prenoffs != DOF_SECT_NONE &&
15232 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
15233 provider->dofpv_prenoffs)) == NULL)
15234 return (-1);
15235
15236 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
15237
15238 if (provider->dofpv_name >= str_sec->dofs_size ||
15239 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
15240 dtrace_dof_error(dof, "invalid provider name");
15241 return (-1);
15242 }
15243
15244 if (prb_sec->dofs_entsize == 0 ||
15245 prb_sec->dofs_entsize > prb_sec->dofs_size) {
15246 dtrace_dof_error(dof, "invalid entry size");
15247 return (-1);
15248 }
15249
15250 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
15251 dtrace_dof_error(dof, "misaligned entry size");
15252 return (-1);
15253 }
15254
15255 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
15256 dtrace_dof_error(dof, "invalid entry size");
15257 return (-1);
15258 }
15259
15260 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
15261 dtrace_dof_error(dof, "misaligned section offset");
15262 return (-1);
15263 }
15264
15265 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
15266 dtrace_dof_error(dof, "invalid entry size");
15267 return (-1);
15268 }
15269
15270 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
15271
15272 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
15273
15274 /*
15275 * Take a pass through the probes to check for errors.
15276 */
15277 for (j = 0; j < nprobes; j++) {
15278 probe = (dof_probe_t *)(uintptr_t)(daddr +
15279 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
15280
15281 if (probe->dofpr_func >= str_sec->dofs_size) {
15282 dtrace_dof_error(dof, "invalid function name");
15283 return (-1);
15284 }
15285
15286 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
15287 dtrace_dof_error(dof, "function name too long");
15288 return (-1);
15289 }
15290
15291 if (probe->dofpr_name >= str_sec->dofs_size ||
15292 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
15293 dtrace_dof_error(dof, "invalid probe name");
15294 return (-1);
15295 }
15296
15297 /*
15298 * The offset count must not wrap the index, and the offsets
15299 * must also not overflow the section's data.
15300 */
15301 if (probe->dofpr_offidx + probe->dofpr_noffs <
15302 probe->dofpr_offidx ||
15303 (probe->dofpr_offidx + probe->dofpr_noffs) *
15304 off_sec->dofs_entsize > off_sec->dofs_size) {
15305 dtrace_dof_error(dof, "invalid probe offset");
15306 return (-1);
15307 }
15308
15309 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
15310 /*
15311 * If there's no is-enabled offset section, make sure
15312 * there aren't any is-enabled offsets. Otherwise
15313 * perform the same checks as for probe offsets
15314 * (immediately above).
15315 */
15316 if (enoff_sec == NULL) {
15317 if (probe->dofpr_enoffidx != 0 ||
15318 probe->dofpr_nenoffs != 0) {
15319 dtrace_dof_error(dof, "is-enabled "
15320 "offsets with null section");
15321 return (-1);
15322 }
15323 } else if (probe->dofpr_enoffidx +
15324 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
15325 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
15326 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
15327 dtrace_dof_error(dof, "invalid is-enabled "
15328 "offset");
15329 return (-1);
15330 }
15331
15332 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
15333 dtrace_dof_error(dof, "zero probe and "
15334 "is-enabled offsets");
15335 return (-1);
15336 }
15337 } else if (probe->dofpr_noffs == 0) {
15338 dtrace_dof_error(dof, "zero probe offsets");
15339 return (-1);
15340 }
15341
15342 if (probe->dofpr_argidx + probe->dofpr_xargc <
15343 probe->dofpr_argidx ||
15344 (probe->dofpr_argidx + probe->dofpr_xargc) *
15345 arg_sec->dofs_entsize > arg_sec->dofs_size) {
15346 dtrace_dof_error(dof, "invalid args");
15347 return (-1);
15348 }
15349
15350 typeidx = probe->dofpr_nargv;
15351 typestr = strtab + probe->dofpr_nargv;
15352 for (k = 0; k < probe->dofpr_nargc; k++) {
15353 if (typeidx >= str_sec->dofs_size) {
15354 dtrace_dof_error(dof, "bad "
15355 "native argument type");
15356 return (-1);
15357 }
15358
15359 typesz = strlen(typestr) + 1;
15360 if (typesz > DTRACE_ARGTYPELEN) {
15361 dtrace_dof_error(dof, "native "
15362 "argument type too long");
15363 return (-1);
15364 }
15365 typeidx += typesz;
15366 typestr += typesz;
15367 }
15368
15369 typeidx = probe->dofpr_xargv;
15370 typestr = strtab + probe->dofpr_xargv;
15371 for (k = 0; k < probe->dofpr_xargc; k++) {
15372 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
15373 dtrace_dof_error(dof, "bad "
15374 "native argument index");
15375 return (-1);
15376 }
15377
15378 if (typeidx >= str_sec->dofs_size) {
15379 dtrace_dof_error(dof, "bad "
15380 "translated argument type");
15381 return (-1);
15382 }
15383
15384 typesz = strlen(typestr) + 1;
15385 if (typesz > DTRACE_ARGTYPELEN) {
15386 dtrace_dof_error(dof, "translated argument "
15387 "type too long");
15388 return (-1);
15389 }
15390
15391 typeidx += typesz;
15392 typestr += typesz;
15393 }
15394 }
15395
15396 return (0);
15397 }
15398
15399 static int
15400 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
15401 {
15402 dtrace_helpers_t *help;
15403 dtrace_vstate_t *vstate;
15404 dtrace_enabling_t *enab = NULL;
15405 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
15406 uintptr_t daddr = (uintptr_t)dof;
15407
15408 ASSERT(MUTEX_HELD(&dtrace_lock));
15409
15410 if ((help = curproc->p_dtrace_helpers) == NULL)
15411 help = dtrace_helpers_create(curproc);
15412
15413 vstate = &help->dthps_vstate;
15414
15415 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
15416 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
15417 dtrace_dof_destroy(dof);
15418 return (rv);
15419 }
15420
15421 /*
15422 * Look for helper providers and validate their descriptions.
15423 */
15424 if (dhp != NULL) {
15425 for (i = 0; i < dof->dofh_secnum; i++) {
15426 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
15427 dof->dofh_secoff + i * dof->dofh_secsize);
15428
15429 if (sec->dofs_type != DOF_SECT_PROVIDER)
15430 continue;
15431
15432 if (dtrace_helper_provider_validate(dof, sec) != 0) {
15433 dtrace_enabling_destroy(enab);
15434 dtrace_dof_destroy(dof);
15435 return (-1);
15436 }
15437
15438 nprovs++;
15439 }
15440 }
15441
15442 /*
15443 * Now we need to walk through the ECB descriptions in the enabling.
15444 */
15445 for (i = 0; i < enab->dten_ndesc; i++) {
15446 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
15447 dtrace_probedesc_t *desc = &ep->dted_probe;
15448
15449 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
15450 continue;
15451
15452 if (strcmp(desc->dtpd_mod, "helper") != 0)
15453 continue;
15454
15455 if (strcmp(desc->dtpd_func, "ustack") != 0)
15456 continue;
15457
15458 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
15459 ep)) != 0) {
15460 /*
15461 * Adding this helper action failed -- we are now going
15462 * to rip out the entire generation and return failure.
15463 */
15464 (void) dtrace_helper_destroygen(help->dthps_generation);
15465 dtrace_enabling_destroy(enab);
15466 dtrace_dof_destroy(dof);
15467 return (-1);
15468 }
15469
15470 nhelpers++;
15471 }
15472
15473 if (nhelpers < enab->dten_ndesc)
15474 dtrace_dof_error(dof, "unmatched helpers");
15475
15476 gen = help->dthps_generation++;
15477 dtrace_enabling_destroy(enab);
15478
15479 if (dhp != NULL && nprovs > 0) {
15480 /*
15481 * Now that this is in-kernel, we change the sense of the
15482 * members: dofhp_dof denotes the in-kernel copy of the DOF
15483 * and dofhp_addr denotes the address at user-level.
15484 */
15485 dhp->dofhp_addr = dhp->dofhp_dof;
15486 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
15487
15488 if (dtrace_helper_provider_add(dhp, gen) == 0) {
15489 mutex_exit(&dtrace_lock);
15490 dtrace_helper_provider_register(curproc, help, dhp);
15491 mutex_enter(&dtrace_lock);
15492
15493 destroy = 0;
15494 }
15495 }
15496
15497 if (destroy)
15498 dtrace_dof_destroy(dof);
15499
15500 return (gen);
15501 }
15502
15503 static dtrace_helpers_t *
15504 dtrace_helpers_create(proc_t *p)
15505 {
15506 dtrace_helpers_t *help;
15507
15508 ASSERT(MUTEX_HELD(&dtrace_lock));
15509 ASSERT(p->p_dtrace_helpers == NULL);
15510
15511 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
15512 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
15513 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
15514
15515 p->p_dtrace_helpers = help;
15516 dtrace_helpers++;
15517
15518 return (help);
15519 }
15520
15521 static void
15522 dtrace_helpers_destroy(proc_t *p)
15523 {
15524 dtrace_helpers_t *help;
15525 dtrace_vstate_t *vstate;
15526 int i;
15527
15528 mutex_enter(&dtrace_lock);
15529
15530 ASSERT(p->p_dtrace_helpers != NULL);
15531 ASSERT(dtrace_helpers > 0);
15532
15533 help = p->p_dtrace_helpers;
15534 vstate = &help->dthps_vstate;
15535
15536 /*
15537 * We're now going to lose the help from this process.
15538 */
15539 p->p_dtrace_helpers = NULL;
15540 if (p == curproc) {
15541 dtrace_sync();
15542 } else {
15543 /*
15544 * It is sometimes necessary to clean up dtrace helpers from a
15545 * an incomplete child process as part of a failed fork
15546 * operation. In such situations, a dtrace_sync() call should
15547 * be unnecessary as the process should be devoid of threads,
15548 * much less any in probe context.
15549 */
15550 VERIFY(p->p_stat == SIDL);
15551 }
15552
15553 /*
15554 * Destroy the helper actions.
15555 */
15556 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15557 dtrace_helper_action_t *h, *next;
15558
15559 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15560 next = h->dtha_next;
15561 dtrace_helper_action_destroy(h, vstate);
15562 h = next;
15563 }
15564 }
15565
15566 mutex_exit(&dtrace_lock);
15567
15568 /*
15569 * Destroy the helper providers.
15570 */
15571 if (help->dthps_maxprovs > 0) {
15572 mutex_enter(&dtrace_meta_lock);
15573 if (dtrace_meta_pid != NULL) {
15574 ASSERT(dtrace_deferred_pid == NULL);
15575
15576 for (i = 0; i < help->dthps_nprovs; i++) {
15577 dtrace_helper_provider_remove(
15578 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15579 }
15580 } else {
15581 mutex_enter(&dtrace_lock);
15582 ASSERT(help->dthps_deferred == 0 ||
15583 help->dthps_next != NULL ||
15584 help->dthps_prev != NULL ||
15585 help == dtrace_deferred_pid);
15586
15587 /*
15588 * Remove the helper from the deferred list.
15589 */
15590 if (help->dthps_next != NULL)
15591 help->dthps_next->dthps_prev = help->dthps_prev;
15592 if (help->dthps_prev != NULL)
15593 help->dthps_prev->dthps_next = help->dthps_next;
15594 if (dtrace_deferred_pid == help) {
15595 dtrace_deferred_pid = help->dthps_next;
15596 ASSERT(help->dthps_prev == NULL);
15597 }
15598
15599 mutex_exit(&dtrace_lock);
15600 }
15601
15602 mutex_exit(&dtrace_meta_lock);
15603
15604 for (i = 0; i < help->dthps_nprovs; i++) {
15605 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15606 }
15607
15608 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15609 sizeof (dtrace_helper_provider_t *));
15610 }
15611
15612 mutex_enter(&dtrace_lock);
15613
15614 dtrace_vstate_fini(&help->dthps_vstate);
15615 kmem_free(help->dthps_actions,
15616 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15617 kmem_free(help, sizeof (dtrace_helpers_t));
15618
15619 --dtrace_helpers;
15620 mutex_exit(&dtrace_lock);
15621 }
15622
15623 static void
15624 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15625 {
15626 dtrace_helpers_t *help, *newhelp;
15627 dtrace_helper_action_t *helper, *new, *last;
15628 dtrace_difo_t *dp;
15629 dtrace_vstate_t *vstate;
15630 int i, j, sz, hasprovs = 0;
15631
15632 mutex_enter(&dtrace_lock);
15633 ASSERT(from->p_dtrace_helpers != NULL);
15634 ASSERT(dtrace_helpers > 0);
15635
15636 help = from->p_dtrace_helpers;
15637 newhelp = dtrace_helpers_create(to);
15638 ASSERT(to->p_dtrace_helpers != NULL);
15639
15640 newhelp->dthps_generation = help->dthps_generation;
15641 vstate = &newhelp->dthps_vstate;
15642
15643 /*
15644 * Duplicate the helper actions.
15645 */
15646 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15647 if ((helper = help->dthps_actions[i]) == NULL)
15648 continue;
15649
15650 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15651 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15652 KM_SLEEP);
15653 new->dtha_generation = helper->dtha_generation;
15654
15655 if ((dp = helper->dtha_predicate) != NULL) {
15656 dp = dtrace_difo_duplicate(dp, vstate);
15657 new->dtha_predicate = dp;
15658 }
15659
15660 new->dtha_nactions = helper->dtha_nactions;
15661 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15662 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15663
15664 for (j = 0; j < new->dtha_nactions; j++) {
15665 dtrace_difo_t *dp = helper->dtha_actions[j];
15666
15667 ASSERT(dp != NULL);
15668 dp = dtrace_difo_duplicate(dp, vstate);
15669 new->dtha_actions[j] = dp;
15670 }
15671
15672 if (last != NULL) {
15673 last->dtha_next = new;
15674 } else {
15675 newhelp->dthps_actions[i] = new;
15676 }
15677
15678 last = new;
15679 }
15680 }
15681
15682 /*
15683 * Duplicate the helper providers and register them with the
15684 * DTrace framework.
15685 */
15686 if (help->dthps_nprovs > 0) {
15687 newhelp->dthps_nprovs = help->dthps_nprovs;
15688 newhelp->dthps_maxprovs = help->dthps_nprovs;
15689 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15690 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15691 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15692 newhelp->dthps_provs[i] = help->dthps_provs[i];
15693 newhelp->dthps_provs[i]->dthp_ref++;
15694 }
15695
15696 hasprovs = 1;
15697 }
15698
15699 mutex_exit(&dtrace_lock);
15700
15701 if (hasprovs)
15702 dtrace_helper_provider_register(to, newhelp, NULL);
15703 }
15704
15705 /*
15706 * DTrace Hook Functions
15707 */
15708 static void
15709 dtrace_module_loaded(struct modctl *ctl)
15710 {
15711 dtrace_provider_t *prv;
15712
15713 mutex_enter(&dtrace_provider_lock);
15714 mutex_enter(&mod_lock);
15715
15716 ASSERT(ctl->mod_busy);
15717
15718 /*
15719 * We're going to call each providers per-module provide operation
15720 * specifying only this module.
15721 */
15722 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15723 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15724
15725 mutex_exit(&mod_lock);
15726 mutex_exit(&dtrace_provider_lock);
15727
15728 /*
15729 * If we have any retained enablings, we need to match against them.
15730 * Enabling probes requires that cpu_lock be held, and we cannot hold
15731 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15732 * module. (In particular, this happens when loading scheduling
15733 * classes.) So if we have any retained enablings, we need to dispatch
15734 * our task queue to do the match for us.
15735 */
15736 mutex_enter(&dtrace_lock);
15737
15738 if (dtrace_retained == NULL) {
15739 mutex_exit(&dtrace_lock);
15740 return;
15741 }
15742
15743 (void) taskq_dispatch(dtrace_taskq,
15744 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15745
15746 mutex_exit(&dtrace_lock);
15747
15748 /*
15749 * And now, for a little heuristic sleaze: in general, we want to
15750 * match modules as soon as they load. However, we cannot guarantee
15751 * this, because it would lead us to the lock ordering violation
15752 * outlined above. The common case, of course, is that cpu_lock is
15753 * _not_ held -- so we delay here for a clock tick, hoping that that's
15754 * long enough for the task queue to do its work. If it's not, it's
15755 * not a serious problem -- it just means that the module that we
15756 * just loaded may not be immediately instrumentable.
15757 */
15758 delay(1);
15759 }
15760
15761 static void
15762 dtrace_module_unloaded(struct modctl *ctl)
15763 {
15764 dtrace_probe_t template, *probe, *first, *next;
15765 dtrace_provider_t *prov;
15766
15767 template.dtpr_mod = ctl->mod_modname;
15768
15769 mutex_enter(&dtrace_provider_lock);
15770 mutex_enter(&mod_lock);
15771 mutex_enter(&dtrace_lock);
15772
15773 if (dtrace_bymod == NULL) {
15774 /*
15775 * The DTrace module is loaded (obviously) but not attached;
15776 * we don't have any work to do.
15777 */
15778 mutex_exit(&dtrace_provider_lock);
15779 mutex_exit(&mod_lock);
15780 mutex_exit(&dtrace_lock);
15781 return;
15782 }
15783
15784 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15785 probe != NULL; probe = probe->dtpr_nextmod) {
15786 if (probe->dtpr_ecb != NULL) {
15787 mutex_exit(&dtrace_provider_lock);
15788 mutex_exit(&mod_lock);
15789 mutex_exit(&dtrace_lock);
15790
15791 /*
15792 * This shouldn't _actually_ be possible -- we're
15793 * unloading a module that has an enabled probe in it.
15794 * (It's normally up to the provider to make sure that
15795 * this can't happen.) However, because dtps_enable()
15796 * doesn't have a failure mode, there can be an
15797 * enable/unload race. Upshot: we don't want to
15798 * assert, but we're not going to disable the
15799 * probe, either.
15800 */
15801 if (dtrace_err_verbose) {
15802 cmn_err(CE_WARN, "unloaded module '%s' had "
15803 "enabled probes", ctl->mod_modname);
15804 }
15805
15806 return;
15807 }
15808 }
15809
15810 probe = first;
15811
15812 for (first = NULL; probe != NULL; probe = next) {
15813 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15814
15815 dtrace_probes[probe->dtpr_id - 1] = NULL;
15816
15817 next = probe->dtpr_nextmod;
15818 dtrace_hash_remove(dtrace_bymod, probe);
15819 dtrace_hash_remove(dtrace_byfunc, probe);
15820 dtrace_hash_remove(dtrace_byname, probe);
15821
15822 if (first == NULL) {
15823 first = probe;
15824 probe->dtpr_nextmod = NULL;
15825 } else {
15826 probe->dtpr_nextmod = first;
15827 first = probe;
15828 }
15829 }
15830
15831 /*
15832 * We've removed all of the module's probes from the hash chains and
15833 * from the probe array. Now issue a dtrace_sync() to be sure that
15834 * everyone has cleared out from any probe array processing.
15835 */
15836 dtrace_sync();
15837
15838 for (probe = first; probe != NULL; probe = first) {
15839 first = probe->dtpr_nextmod;
15840 prov = probe->dtpr_provider;
15841 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15842 probe->dtpr_arg);
15843 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15844 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15845 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15846 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15847 kmem_free(probe, sizeof (dtrace_probe_t));
15848 }
15849
15850 mutex_exit(&dtrace_lock);
15851 mutex_exit(&mod_lock);
15852 mutex_exit(&dtrace_provider_lock);
15853 }
15854
15855 void
15856 dtrace_suspend(void)
15857 {
15858 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15859 }
15860
15861 void
15862 dtrace_resume(void)
15863 {
15864 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15865 }
15866
15867 static int
15868 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15869 {
15870 ASSERT(MUTEX_HELD(&cpu_lock));
15871 mutex_enter(&dtrace_lock);
15872
15873 switch (what) {
15874 case CPU_CONFIG: {
15875 dtrace_state_t *state;
15876 dtrace_optval_t *opt, rs, c;
15877
15878 /*
15879 * For now, we only allocate a new buffer for anonymous state.
15880 */
15881 if ((state = dtrace_anon.dta_state) == NULL)
15882 break;
15883
15884 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15885 break;
15886
15887 opt = state->dts_options;
15888 c = opt[DTRACEOPT_CPU];
15889
15890 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15891 break;
15892
15893 /*
15894 * Regardless of what the actual policy is, we're going to
15895 * temporarily set our resize policy to be manual. We're
15896 * also going to temporarily set our CPU option to denote
15897 * the newly configured CPU.
15898 */
15899 rs = opt[DTRACEOPT_BUFRESIZE];
15900 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15901 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15902
15903 (void) dtrace_state_buffers(state);
15904
15905 opt[DTRACEOPT_BUFRESIZE] = rs;
15906 opt[DTRACEOPT_CPU] = c;
15907
15908 break;
15909 }
15910
15911 case CPU_UNCONFIG:
15912 /*
15913 * We don't free the buffer in the CPU_UNCONFIG case. (The
15914 * buffer will be freed when the consumer exits.)
15915 */
15916 break;
15917
15918 default:
15919 break;
15920 }
15921
15922 mutex_exit(&dtrace_lock);
15923 return (0);
15924 }
15925
15926 static void
15927 dtrace_cpu_setup_initial(processorid_t cpu)
15928 {
15929 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15930 }
15931
15932 static void
15933 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15934 {
15935 if (dtrace_toxranges >= dtrace_toxranges_max) {
15936 int osize, nsize;
15937 dtrace_toxrange_t *range;
15938
15939 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15940
15941 if (osize == 0) {
15942 ASSERT(dtrace_toxrange == NULL);
15943 ASSERT(dtrace_toxranges_max == 0);
15944 dtrace_toxranges_max = 1;
15945 } else {
15946 dtrace_toxranges_max <<= 1;
15947 }
15948
15949 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15950 range = kmem_zalloc(nsize, KM_SLEEP);
15951
15952 if (dtrace_toxrange != NULL) {
15953 ASSERT(osize != 0);
15954 bcopy(dtrace_toxrange, range, osize);
15955 kmem_free(dtrace_toxrange, osize);
15956 }
15957
15958 dtrace_toxrange = range;
15959 }
15960
15961 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
15962 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
15963
15964 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15965 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15966 dtrace_toxranges++;
15967 }
15968
15969 static void
15970 dtrace_getf_barrier()
15971 {
15972 /*
15973 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
15974 * that contain calls to getf(), this routine will be called on every
15975 * closef() before either the underlying vnode is released or the
15976 * file_t itself is freed. By the time we are here, it is essential
15977 * that the file_t can no longer be accessed from a call to getf()
15978 * in probe context -- that assures that a dtrace_sync() can be used
15979 * to clear out any enablings referring to the old structures.
15980 */
15981 if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
15982 kcred->cr_zone->zone_dtrace_getf != 0)
15983 dtrace_sync();
15984 }
15985
15986 /*
15987 * DTrace Driver Cookbook Functions
15988 */
15989 /*ARGSUSED*/
15990 static int
15991 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15992 {
15993 dtrace_provider_id_t id;
15994 dtrace_state_t *state = NULL;
15995 dtrace_enabling_t *enab;
15996
15997 mutex_enter(&cpu_lock);
15998 mutex_enter(&dtrace_provider_lock);
15999 mutex_enter(&dtrace_lock);
16000
16001 if (ddi_soft_state_init(&dtrace_softstate,
16002 sizeof (dtrace_state_t), 0) != 0) {
16003 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
16004 mutex_exit(&cpu_lock);
16005 mutex_exit(&dtrace_provider_lock);
16006 mutex_exit(&dtrace_lock);
16007 return (DDI_FAILURE);
16008 }
16009
16010 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
16011 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
16012 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
16013 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
16014 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
16015 ddi_remove_minor_node(devi, NULL);
16016 ddi_soft_state_fini(&dtrace_softstate);
16017 mutex_exit(&cpu_lock);
16018 mutex_exit(&dtrace_provider_lock);
16019 mutex_exit(&dtrace_lock);
16020 return (DDI_FAILURE);
16021 }
16022
16023 ddi_report_dev(devi);
16024 dtrace_devi = devi;
16025
16026 dtrace_modload = dtrace_module_loaded;
16027 dtrace_modunload = dtrace_module_unloaded;
16028 dtrace_cpu_init = dtrace_cpu_setup_initial;
16029 dtrace_helpers_cleanup = dtrace_helpers_destroy;
16030 dtrace_helpers_fork = dtrace_helpers_duplicate;
16031 dtrace_cpustart_init = dtrace_suspend;
16032 dtrace_cpustart_fini = dtrace_resume;
16033 dtrace_debugger_init = dtrace_suspend;
16034 dtrace_debugger_fini = dtrace_resume;
16035
16036 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16037
16038 ASSERT(MUTEX_HELD(&cpu_lock));
16039
16040 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
16041 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
16042 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
16043 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
16044 VM_SLEEP | VMC_IDENTIFIER);
16045 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
16046 1, INT_MAX, 0);
16047
16048 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
16049 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
16050 NULL, NULL, NULL, NULL, NULL, 0);
16051
16052 ASSERT(MUTEX_HELD(&cpu_lock));
16053 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
16054 offsetof(dtrace_probe_t, dtpr_nextmod),
16055 offsetof(dtrace_probe_t, dtpr_prevmod));
16056
16057 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
16058 offsetof(dtrace_probe_t, dtpr_nextfunc),
16059 offsetof(dtrace_probe_t, dtpr_prevfunc));
16060
16061 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
16062 offsetof(dtrace_probe_t, dtpr_nextname),
16063 offsetof(dtrace_probe_t, dtpr_prevname));
16064
16065 if (dtrace_retain_max < 1) {
16066 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
16067 "setting to 1", dtrace_retain_max);
16068 dtrace_retain_max = 1;
16069 }
16070
16071 /*
16072 * Now discover our toxic ranges.
16073 */
16074 dtrace_toxic_ranges(dtrace_toxrange_add);
16075
16076 /*
16077 * Before we register ourselves as a provider to our own framework,
16078 * we would like to assert that dtrace_provider is NULL -- but that's
16079 * not true if we were loaded as a dependency of a DTrace provider.
16080 * Once we've registered, we can assert that dtrace_provider is our
16081 * pseudo provider.
16082 */
16083 (void) dtrace_register("dtrace", &dtrace_provider_attr,
16084 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
16085
16086 ASSERT(dtrace_provider != NULL);
16087 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
16088
16089 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
16090 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
16091 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
16092 dtrace_provider, NULL, NULL, "END", 0, NULL);
16093 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
16094 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
16095
16096 dtrace_anon_property();
16097 mutex_exit(&cpu_lock);
16098
16099 /*
16100 * If there are already providers, we must ask them to provide their
16101 * probes, and then match any anonymous enabling against them. Note
16102 * that there should be no other retained enablings at this time:
16103 * the only retained enablings at this time should be the anonymous
16104 * enabling.
16105 */
16106 if (dtrace_anon.dta_enabling != NULL) {
16107 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
16108
16109 dtrace_enabling_provide(NULL);
16110 state = dtrace_anon.dta_state;
16111
16112 /*
16113 * We couldn't hold cpu_lock across the above call to
16114 * dtrace_enabling_provide(), but we must hold it to actually
16115 * enable the probes. We have to drop all of our locks, pick
16116 * up cpu_lock, and regain our locks before matching the
16117 * retained anonymous enabling.
16118 */
16119 mutex_exit(&dtrace_lock);
16120 mutex_exit(&dtrace_provider_lock);
16121
16122 mutex_enter(&cpu_lock);
16123 mutex_enter(&dtrace_provider_lock);
16124 mutex_enter(&dtrace_lock);
16125
16126 if ((enab = dtrace_anon.dta_enabling) != NULL)
16127 (void) dtrace_enabling_match(enab, NULL);
16128
16129 mutex_exit(&cpu_lock);
16130 }
16131
16132 mutex_exit(&dtrace_lock);
16133 mutex_exit(&dtrace_provider_lock);
16134
16135 if (state != NULL) {
16136 /*
16137 * If we created any anonymous state, set it going now.
16138 */
16139 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
16140 }
16141
16142 return (DDI_SUCCESS);
16143 }
16144
16145 /*ARGSUSED*/
16146 static int
16147 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
16148 {
16149 dtrace_state_t *state;
16150 uint32_t priv;
16151 uid_t uid;
16152 zoneid_t zoneid;
16153
16154 if (getminor(*devp) == DTRACEMNRN_HELPER)
16155 return (0);
16156
16157 /*
16158 * If this wasn't an open with the "helper" minor, then it must be
16159 * the "dtrace" minor.
16160 */
16161 if (getminor(*devp) != DTRACEMNRN_DTRACE)
16162 return (ENXIO);
16163
16164 /*
16165 * If no DTRACE_PRIV_* bits are set in the credential, then the
16166 * caller lacks sufficient permission to do anything with DTrace.
16167 */
16168 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
16169 if (priv == DTRACE_PRIV_NONE)
16170 return (EACCES);
16171
16172 /*
16173 * Ask all providers to provide all their probes.
16174 */
16175 mutex_enter(&dtrace_provider_lock);
16176 dtrace_probe_provide(NULL, NULL);
16177 mutex_exit(&dtrace_provider_lock);
16178
16179 mutex_enter(&cpu_lock);
16180 mutex_enter(&dtrace_lock);
16181 dtrace_opens++;
16182 dtrace_membar_producer();
16183
16184 /*
16185 * If the kernel debugger is active (that is, if the kernel debugger
16186 * modified text in some way), we won't allow the open.
16187 */
16188 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
16189 dtrace_opens--;
16190 mutex_exit(&cpu_lock);
16191 mutex_exit(&dtrace_lock);
16192 return (EBUSY);
16193 }
16194
16195 if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
16196 /*
16197 * If DTrace helper tracing is enabled, we need to allocate the
16198 * trace buffer and initialize the values.
16199 */
16200 dtrace_helptrace_buffer =
16201 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
16202 dtrace_helptrace_next = 0;
16203 dtrace_helptrace_wrapped = 0;
16204 dtrace_helptrace_enable = 0;
16205 }
16206
16207 state = dtrace_state_create(devp, cred_p);
16208 mutex_exit(&cpu_lock);
16209
16210 if (state == NULL) {
16211 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
16212 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16213 mutex_exit(&dtrace_lock);
16214 return (EAGAIN);
16215 }
16216
16217 mutex_exit(&dtrace_lock);
16218
16219 return (0);
16220 }
16221
16222 /*ARGSUSED*/
16223 static int
16224 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
16225 {
16226 minor_t minor = getminor(dev);
16227 dtrace_state_t *state;
16228 dtrace_helptrace_t *buf = NULL;
16229
16230 if (minor == DTRACEMNRN_HELPER)
16231 return (0);
16232
16233 state = ddi_get_soft_state(dtrace_softstate, minor);
16234
16235 mutex_enter(&cpu_lock);
16236 mutex_enter(&dtrace_lock);
16237
16238 if (state->dts_anon) {
16239 /*
16240 * There is anonymous state. Destroy that first.
16241 */
16242 ASSERT(dtrace_anon.dta_state == NULL);
16243 dtrace_state_destroy(state->dts_anon);
16244 }
16245
16246 if (dtrace_helptrace_disable) {
16247 /*
16248 * If we have been told to disable helper tracing, set the
16249 * buffer to NULL before calling into dtrace_state_destroy();
16250 * we take advantage of its dtrace_sync() to know that no
16251 * CPU is in probe context with enabled helper tracing
16252 * after it returns.
16253 */
16254 buf = dtrace_helptrace_buffer;
16255 dtrace_helptrace_buffer = NULL;
16256 }
16257
16258 dtrace_state_destroy(state);
16259 ASSERT(dtrace_opens > 0);
16260
16261 /*
16262 * Only relinquish control of the kernel debugger interface when there
16263 * are no consumers and no anonymous enablings.
16264 */
16265 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
16266 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16267
16268 if (buf != NULL) {
16269 kmem_free(buf, dtrace_helptrace_bufsize);
16270 dtrace_helptrace_disable = 0;
16271 }
16272
16273 mutex_exit(&dtrace_lock);
16274 mutex_exit(&cpu_lock);
16275
16276 return (0);
16277 }
16278
16279 /*ARGSUSED*/
16280 static int
16281 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
16282 {
16283 int rval;
16284 dof_helper_t help, *dhp = NULL;
16285
16286 switch (cmd) {
16287 case DTRACEHIOC_ADDDOF:
16288 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
16289 dtrace_dof_error(NULL, "failed to copyin DOF helper");
16290 return (EFAULT);
16291 }
16292
16293 dhp = &help;
16294 arg = (intptr_t)help.dofhp_dof;
16295 /*FALLTHROUGH*/
16296
16297 case DTRACEHIOC_ADD: {
16298 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
16299
16300 if (dof == NULL)
16301 return (rval);
16302
16303 mutex_enter(&dtrace_lock);
16304
16305 /*
16306 * dtrace_helper_slurp() takes responsibility for the dof --
16307 * it may free it now or it may save it and free it later.
16308 */
16309 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
16310 *rv = rval;
16311 rval = 0;
16312 } else {
16313 rval = EINVAL;
16314 }
16315
16316 mutex_exit(&dtrace_lock);
16317 return (rval);
16318 }
16319
16320 case DTRACEHIOC_REMOVE: {
16321 mutex_enter(&dtrace_lock);
16322 rval = dtrace_helper_destroygen(arg);
16323 mutex_exit(&dtrace_lock);
16324
16325 return (rval);
16326 }
16327
16328 default:
16329 break;
16330 }
16331
16332 return (ENOTTY);
16333 }
16334
16335 /*ARGSUSED*/
16336 static int
16337 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
16338 {
16339 minor_t minor = getminor(dev);
16340 dtrace_state_t *state;
16341 int rval;
16342
16343 if (minor == DTRACEMNRN_HELPER)
16344 return (dtrace_ioctl_helper(cmd, arg, rv));
16345
16346 state = ddi_get_soft_state(dtrace_softstate, minor);
16347
16348 if (state->dts_anon) {
16349 ASSERT(dtrace_anon.dta_state == NULL);
16350 state = state->dts_anon;
16351 }
16352
16353 switch (cmd) {
16354 case DTRACEIOC_PROVIDER: {
16355 dtrace_providerdesc_t pvd;
16356 dtrace_provider_t *pvp;
16357
16358 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
16359 return (EFAULT);
16360
16361 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
16362 mutex_enter(&dtrace_provider_lock);
16363
16364 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
16365 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
16366 break;
16367 }
16368
16369 mutex_exit(&dtrace_provider_lock);
16370
16371 if (pvp == NULL)
16372 return (ESRCH);
16373
16374 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
16375 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
16376 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
16377 return (EFAULT);
16378
16379 return (0);
16380 }
16381
16382 case DTRACEIOC_EPROBE: {
16383 dtrace_eprobedesc_t epdesc;
16384 dtrace_ecb_t *ecb;
16385 dtrace_action_t *act;
16386 void *buf;
16387 size_t size;
16388 uintptr_t dest;
16389 int nrecs;
16390
16391 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
16392 return (EFAULT);
16393
16394 mutex_enter(&dtrace_lock);
16395
16396 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
16397 mutex_exit(&dtrace_lock);
16398 return (EINVAL);
16399 }
16400
16401 if (ecb->dte_probe == NULL) {
16402 mutex_exit(&dtrace_lock);
16403 return (EINVAL);
16404 }
16405
16406 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
16407 epdesc.dtepd_uarg = ecb->dte_uarg;
16408 epdesc.dtepd_size = ecb->dte_size;
16409
16410 nrecs = epdesc.dtepd_nrecs;
16411 epdesc.dtepd_nrecs = 0;
16412 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16413 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16414 continue;
16415
16416 epdesc.dtepd_nrecs++;
16417 }
16418
16419 /*
16420 * Now that we have the size, we need to allocate a temporary
16421 * buffer in which to store the complete description. We need
16422 * the temporary buffer to be able to drop dtrace_lock()
16423 * across the copyout(), below.
16424 */
16425 size = sizeof (dtrace_eprobedesc_t) +
16426 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
16427
16428 buf = kmem_alloc(size, KM_SLEEP);
16429 dest = (uintptr_t)buf;
16430
16431 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
16432 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
16433
16434 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16435 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16436 continue;
16437
16438 if (nrecs-- == 0)
16439 break;
16440
16441 bcopy(&act->dta_rec, (void *)dest,
16442 sizeof (dtrace_recdesc_t));
16443 dest += sizeof (dtrace_recdesc_t);
16444 }
16445
16446 mutex_exit(&dtrace_lock);
16447
16448 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16449 kmem_free(buf, size);
16450 return (EFAULT);
16451 }
16452
16453 kmem_free(buf, size);
16454 return (0);
16455 }
16456
16457 case DTRACEIOC_AGGDESC: {
16458 dtrace_aggdesc_t aggdesc;
16459 dtrace_action_t *act;
16460 dtrace_aggregation_t *agg;
16461 int nrecs;
16462 uint32_t offs;
16463 dtrace_recdesc_t *lrec;
16464 void *buf;
16465 size_t size;
16466 uintptr_t dest;
16467
16468 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16469 return (EFAULT);
16470
16471 mutex_enter(&dtrace_lock);
16472
16473 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16474 mutex_exit(&dtrace_lock);
16475 return (EINVAL);
16476 }
16477
16478 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16479
16480 nrecs = aggdesc.dtagd_nrecs;
16481 aggdesc.dtagd_nrecs = 0;
16482
16483 offs = agg->dtag_base;
16484 lrec = &agg->dtag_action.dta_rec;
16485 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16486
16487 for (act = agg->dtag_first; ; act = act->dta_next) {
16488 ASSERT(act->dta_intuple ||
16489 DTRACEACT_ISAGG(act->dta_kind));
16490
16491 /*
16492 * If this action has a record size of zero, it
16493 * denotes an argument to the aggregating action.
16494 * Because the presence of this record doesn't (or
16495 * shouldn't) affect the way the data is interpreted,
16496 * we don't copy it out to save user-level the
16497 * confusion of dealing with a zero-length record.
16498 */
16499 if (act->dta_rec.dtrd_size == 0) {
16500 ASSERT(agg->dtag_hasarg);
16501 continue;
16502 }
16503
16504 aggdesc.dtagd_nrecs++;
16505
16506 if (act == &agg->dtag_action)
16507 break;
16508 }
16509
16510 /*
16511 * Now that we have the size, we need to allocate a temporary
16512 * buffer in which to store the complete description. We need
16513 * the temporary buffer to be able to drop dtrace_lock()
16514 * across the copyout(), below.
16515 */
16516 size = sizeof (dtrace_aggdesc_t) +
16517 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16518
16519 buf = kmem_alloc(size, KM_SLEEP);
16520 dest = (uintptr_t)buf;
16521
16522 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16523 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16524
16525 for (act = agg->dtag_first; ; act = act->dta_next) {
16526 dtrace_recdesc_t rec = act->dta_rec;
16527
16528 /*
16529 * See the comment in the above loop for why we pass
16530 * over zero-length records.
16531 */
16532 if (rec.dtrd_size == 0) {
16533 ASSERT(agg->dtag_hasarg);
16534 continue;
16535 }
16536
16537 if (nrecs-- == 0)
16538 break;
16539
16540 rec.dtrd_offset -= offs;
16541 bcopy(&rec, (void *)dest, sizeof (rec));
16542 dest += sizeof (dtrace_recdesc_t);
16543
16544 if (act == &agg->dtag_action)
16545 break;
16546 }
16547
16548 mutex_exit(&dtrace_lock);
16549
16550 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16551 kmem_free(buf, size);
16552 return (EFAULT);
16553 }
16554
16555 kmem_free(buf, size);
16556 return (0);
16557 }
16558
16559 case DTRACEIOC_ENABLE: {
16560 dof_hdr_t *dof;
16561 dtrace_enabling_t *enab = NULL;
16562 dtrace_vstate_t *vstate;
16563 int err = 0;
16564
16565 *rv = 0;
16566
16567 /*
16568 * If a NULL argument has been passed, we take this as our
16569 * cue to reevaluate our enablings.
16570 */
16571 if (arg == NULL) {
16572 dtrace_enabling_matchall();
16573
16574 return (0);
16575 }
16576
16577 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16578 return (rval);
16579
16580 mutex_enter(&cpu_lock);
16581 mutex_enter(&dtrace_lock);
16582 vstate = &state->dts_vstate;
16583
16584 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16585 mutex_exit(&dtrace_lock);
16586 mutex_exit(&cpu_lock);
16587 dtrace_dof_destroy(dof);
16588 return (EBUSY);
16589 }
16590
16591 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16592 mutex_exit(&dtrace_lock);
16593 mutex_exit(&cpu_lock);
16594 dtrace_dof_destroy(dof);
16595 return (EINVAL);
16596 }
16597
16598 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16599 dtrace_enabling_destroy(enab);
16600 mutex_exit(&dtrace_lock);
16601 mutex_exit(&cpu_lock);
16602 dtrace_dof_destroy(dof);
16603 return (rval);
16604 }
16605
16606 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16607 err = dtrace_enabling_retain(enab);
16608 } else {
16609 dtrace_enabling_destroy(enab);
16610 }
16611
16612 mutex_exit(&cpu_lock);
16613 mutex_exit(&dtrace_lock);
16614 dtrace_dof_destroy(dof);
16615
16616 return (err);
16617 }
16618
16619 case DTRACEIOC_REPLICATE: {
16620 dtrace_repldesc_t desc;
16621 dtrace_probedesc_t *match = &desc.dtrpd_match;
16622 dtrace_probedesc_t *create = &desc.dtrpd_create;
16623 int err;
16624
16625 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16626 return (EFAULT);
16627
16628 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16629 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16630 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16631 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16632
16633 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16634 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16635 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16636 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16637
16638 mutex_enter(&dtrace_lock);
16639 err = dtrace_enabling_replicate(state, match, create);
16640 mutex_exit(&dtrace_lock);
16641
16642 return (err);
16643 }
16644
16645 case DTRACEIOC_PROBEMATCH:
16646 case DTRACEIOC_PROBES: {
16647 dtrace_probe_t *probe = NULL;
16648 dtrace_probedesc_t desc;
16649 dtrace_probekey_t pkey;
16650 dtrace_id_t i;
16651 int m = 0;
16652 uint32_t priv;
16653 uid_t uid;
16654 zoneid_t zoneid;
16655
16656 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16657 return (EFAULT);
16658
16659 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16660 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16661 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16662 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16663
16664 /*
16665 * Before we attempt to match this probe, we want to give
16666 * all providers the opportunity to provide it.
16667 */
16668 if (desc.dtpd_id == DTRACE_IDNONE) {
16669 mutex_enter(&dtrace_provider_lock);
16670 dtrace_probe_provide(&desc, NULL);
16671 mutex_exit(&dtrace_provider_lock);
16672 desc.dtpd_id++;
16673 }
16674
16675 if (cmd == DTRACEIOC_PROBEMATCH) {
16676 dtrace_probekey(&desc, &pkey);
16677 pkey.dtpk_id = DTRACE_IDNONE;
16678 }
16679
16680 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16681
16682 mutex_enter(&dtrace_lock);
16683
16684 if (cmd == DTRACEIOC_PROBEMATCH) {
16685 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16686 if ((probe = dtrace_probes[i - 1]) != NULL &&
16687 (m = dtrace_match_probe(probe, &pkey,
16688 priv, uid, zoneid)) != 0)
16689 break;
16690 }
16691
16692 if (m < 0) {
16693 mutex_exit(&dtrace_lock);
16694 return (EINVAL);
16695 }
16696
16697 } else {
16698 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16699 if ((probe = dtrace_probes[i - 1]) != NULL &&
16700 dtrace_match_priv(probe, priv, uid, zoneid))
16701 break;
16702 }
16703 }
16704
16705 if (probe == NULL) {
16706 mutex_exit(&dtrace_lock);
16707 return (ESRCH);
16708 }
16709
16710 dtrace_probe_description(probe, &desc);
16711 mutex_exit(&dtrace_lock);
16712
16713 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16714 return (EFAULT);
16715
16716 return (0);
16717 }
16718
16719 case DTRACEIOC_PROBEARG: {
16720 dtrace_argdesc_t desc;
16721 dtrace_probe_t *probe;
16722 dtrace_provider_t *prov;
16723
16724 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16725 return (EFAULT);
16726
16727 if (desc.dtargd_id == DTRACE_IDNONE)
16728 return (EINVAL);
16729
16730 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16731 return (EINVAL);
16732
16733 mutex_enter(&dtrace_provider_lock);
16734 mutex_enter(&mod_lock);
16735 mutex_enter(&dtrace_lock);
16736
16737 if (desc.dtargd_id > dtrace_nprobes) {
16738 mutex_exit(&dtrace_lock);
16739 mutex_exit(&mod_lock);
16740 mutex_exit(&dtrace_provider_lock);
16741 return (EINVAL);
16742 }
16743
16744 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16745 mutex_exit(&dtrace_lock);
16746 mutex_exit(&mod_lock);
16747 mutex_exit(&dtrace_provider_lock);
16748 return (EINVAL);
16749 }
16750
16751 mutex_exit(&dtrace_lock);
16752
16753 prov = probe->dtpr_provider;
16754
16755 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16756 /*
16757 * There isn't any typed information for this probe.
16758 * Set the argument number to DTRACE_ARGNONE.
16759 */
16760 desc.dtargd_ndx = DTRACE_ARGNONE;
16761 } else {
16762 desc.dtargd_native[0] = '\0';
16763 desc.dtargd_xlate[0] = '\0';
16764 desc.dtargd_mapping = desc.dtargd_ndx;
16765
16766 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16767 probe->dtpr_id, probe->dtpr_arg, &desc);
16768 }
16769
16770 mutex_exit(&mod_lock);
16771 mutex_exit(&dtrace_provider_lock);
16772
16773 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16774 return (EFAULT);
16775
16776 return (0);
16777 }
16778
16779 case DTRACEIOC_GO: {
16780 processorid_t cpuid;
16781 rval = dtrace_state_go(state, &cpuid);
16782
16783 if (rval != 0)
16784 return (rval);
16785
16786 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16787 return (EFAULT);
16788
16789 return (0);
16790 }
16791
16792 case DTRACEIOC_STOP: {
16793 processorid_t cpuid;
16794
16795 mutex_enter(&dtrace_lock);
16796 rval = dtrace_state_stop(state, &cpuid);
16797 mutex_exit(&dtrace_lock);
16798
16799 if (rval != 0)
16800 return (rval);
16801
16802 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16803 return (EFAULT);
16804
16805 return (0);
16806 }
16807
16808 case DTRACEIOC_DOFGET: {
16809 dof_hdr_t hdr, *dof;
16810 uint64_t len;
16811
16812 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16813 return (EFAULT);
16814
16815 mutex_enter(&dtrace_lock);
16816 dof = dtrace_dof_create(state);
16817 mutex_exit(&dtrace_lock);
16818
16819 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16820 rval = copyout(dof, (void *)arg, len);
16821 dtrace_dof_destroy(dof);
16822
16823 return (rval == 0 ? 0 : EFAULT);
16824 }
16825
16826 case DTRACEIOC_AGGSNAP:
16827 case DTRACEIOC_BUFSNAP: {
16828 dtrace_bufdesc_t desc;
16829 caddr_t cached;
16830 dtrace_buffer_t *buf;
16831
16832 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16833 return (EFAULT);
16834
16835 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16836 return (EINVAL);
16837
16838 mutex_enter(&dtrace_lock);
16839
16840 if (cmd == DTRACEIOC_BUFSNAP) {
16841 buf = &state->dts_buffer[desc.dtbd_cpu];
16842 } else {
16843 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16844 }
16845
16846 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16847 size_t sz = buf->dtb_offset;
16848
16849 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16850 mutex_exit(&dtrace_lock);
16851 return (EBUSY);
16852 }
16853
16854 /*
16855 * If this buffer has already been consumed, we're
16856 * going to indicate that there's nothing left here
16857 * to consume.
16858 */
16859 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16860 mutex_exit(&dtrace_lock);
16861
16862 desc.dtbd_size = 0;
16863 desc.dtbd_drops = 0;
16864 desc.dtbd_errors = 0;
16865 desc.dtbd_oldest = 0;
16866 sz = sizeof (desc);
16867
16868 if (copyout(&desc, (void *)arg, sz) != 0)
16869 return (EFAULT);
16870
16871 return (0);
16872 }
16873
16874 /*
16875 * If this is a ring buffer that has wrapped, we want
16876 * to copy the whole thing out.
16877 */
16878 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16879 dtrace_buffer_polish(buf);
16880 sz = buf->dtb_size;
16881 }
16882
16883 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16884 mutex_exit(&dtrace_lock);
16885 return (EFAULT);
16886 }
16887
16888 desc.dtbd_size = sz;
16889 desc.dtbd_drops = buf->dtb_drops;
16890 desc.dtbd_errors = buf->dtb_errors;
16891 desc.dtbd_oldest = buf->dtb_xamot_offset;
16892 desc.dtbd_timestamp = dtrace_gethrtime();
16893
16894 mutex_exit(&dtrace_lock);
16895
16896 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16897 return (EFAULT);
16898
16899 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16900
16901 return (0);
16902 }
16903
16904 if (buf->dtb_tomax == NULL) {
16905 ASSERT(buf->dtb_xamot == NULL);
16906 mutex_exit(&dtrace_lock);
16907 return (ENOENT);
16908 }
16909
16910 cached = buf->dtb_tomax;
16911 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16912
16913 dtrace_xcall(desc.dtbd_cpu,
16914 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16915
16916 state->dts_errors += buf->dtb_xamot_errors;
16917
16918 /*
16919 * If the buffers did not actually switch, then the cross call
16920 * did not take place -- presumably because the given CPU is
16921 * not in the ready set. If this is the case, we'll return
16922 * ENOENT.
16923 */
16924 if (buf->dtb_tomax == cached) {
16925 ASSERT(buf->dtb_xamot != cached);
16926 mutex_exit(&dtrace_lock);
16927 return (ENOENT);
16928 }
16929
16930 ASSERT(cached == buf->dtb_xamot);
16931
16932 /*
16933 * We have our snapshot; now copy it out.
16934 */
16935 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16936 buf->dtb_xamot_offset) != 0) {
16937 mutex_exit(&dtrace_lock);
16938 return (EFAULT);
16939 }
16940
16941 desc.dtbd_size = buf->dtb_xamot_offset;
16942 desc.dtbd_drops = buf->dtb_xamot_drops;
16943 desc.dtbd_errors = buf->dtb_xamot_errors;
16944 desc.dtbd_oldest = 0;
16945 desc.dtbd_timestamp = buf->dtb_switched;
16946
16947 mutex_exit(&dtrace_lock);
16948
16949 /*
16950 * Finally, copy out the buffer description.
16951 */
16952 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16953 return (EFAULT);
16954
16955 return (0);
16956 }
16957
16958 case DTRACEIOC_CONF: {
16959 dtrace_conf_t conf;
16960
16961 bzero(&conf, sizeof (conf));
16962 conf.dtc_difversion = DIF_VERSION;
16963 conf.dtc_difintregs = DIF_DIR_NREGS;
16964 conf.dtc_diftupregs = DIF_DTR_NREGS;
16965 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16966
16967 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16968 return (EFAULT);
16969
16970 return (0);
16971 }
16972
16973 case DTRACEIOC_STATUS: {
16974 dtrace_status_t stat;
16975 dtrace_dstate_t *dstate;
16976 int i, j;
16977 uint64_t nerrs;
16978
16979 /*
16980 * See the comment in dtrace_state_deadman() for the reason
16981 * for setting dts_laststatus to INT64_MAX before setting
16982 * it to the correct value.
16983 */
16984 state->dts_laststatus = INT64_MAX;
16985 dtrace_membar_producer();
16986 state->dts_laststatus = dtrace_gethrtime();
16987
16988 bzero(&stat, sizeof (stat));
16989
16990 mutex_enter(&dtrace_lock);
16991
16992 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16993 mutex_exit(&dtrace_lock);
16994 return (ENOENT);
16995 }
16996
16997 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16998 stat.dtst_exiting = 1;
16999
17000 nerrs = state->dts_errors;
17001 dstate = &state->dts_vstate.dtvs_dynvars;
17002
17003 for (i = 0; i < NCPU; i++) {
17004 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
17005
17006 stat.dtst_dyndrops += dcpu->dtdsc_drops;
17007 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
17008 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
17009
17010 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
17011 stat.dtst_filled++;
17012
17013 nerrs += state->dts_buffer[i].dtb_errors;
17014
17015 for (j = 0; j < state->dts_nspeculations; j++) {
17016 dtrace_speculation_t *spec;
17017 dtrace_buffer_t *buf;
17018
17019 spec = &state->dts_speculations[j];
17020 buf = &spec->dtsp_buffer[i];
17021 stat.dtst_specdrops += buf->dtb_xamot_drops;
17022 }
17023 }
17024
17025 stat.dtst_specdrops_busy = state->dts_speculations_busy;
17026 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
17027 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
17028 stat.dtst_dblerrors = state->dts_dblerrors;
17029 stat.dtst_killed =
17030 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
17031 stat.dtst_errors = nerrs;
17032
17033 mutex_exit(&dtrace_lock);
17034
17035 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
17036 return (EFAULT);
17037
17038 return (0);
17039 }
17040
17041 case DTRACEIOC_FORMAT: {
17042 dtrace_fmtdesc_t fmt;
17043 char *str;
17044 int len;
17045
17046 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
17047 return (EFAULT);
17048
17049 mutex_enter(&dtrace_lock);
17050
17051 if (fmt.dtfd_format == 0 ||
17052 fmt.dtfd_format > state->dts_nformats) {
17053 mutex_exit(&dtrace_lock);
17054 return (EINVAL);
17055 }
17056
17057 /*
17058 * Format strings are allocated contiguously and they are
17059 * never freed; if a format index is less than the number
17060 * of formats, we can assert that the format map is non-NULL
17061 * and that the format for the specified index is non-NULL.
17062 */
17063 ASSERT(state->dts_formats != NULL);
17064 str = state->dts_formats[fmt.dtfd_format - 1];
17065 ASSERT(str != NULL);
17066
17067 len = strlen(str) + 1;
17068
17069 if (len > fmt.dtfd_length) {
17070 fmt.dtfd_length = len;
17071
17072 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
17073 mutex_exit(&dtrace_lock);
17074 return (EINVAL);
17075 }
17076 } else {
17077 if (copyout(str, fmt.dtfd_string, len) != 0) {
17078 mutex_exit(&dtrace_lock);
17079 return (EINVAL);
17080 }
17081 }
17082
17083 mutex_exit(&dtrace_lock);
17084 return (0);
17085 }
17086
17087 default:
17088 break;
17089 }
17090
17091 return (ENOTTY);
17092 }
17093
17094 /*ARGSUSED*/
17095 static int
17096 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
17097 {
17098 dtrace_state_t *state;
17099
17100 switch (cmd) {
17101 case DDI_DETACH:
17102 break;
17103
17104 case DDI_SUSPEND:
17105 return (DDI_SUCCESS);
17106
17107 default:
17108 return (DDI_FAILURE);
17109 }
17110
17111 mutex_enter(&cpu_lock);
17112 mutex_enter(&dtrace_provider_lock);
17113 mutex_enter(&dtrace_lock);
17114
17115 ASSERT(dtrace_opens == 0);
17116
17117 if (dtrace_helpers > 0) {
17118 mutex_exit(&dtrace_provider_lock);
17119 mutex_exit(&dtrace_lock);
17120 mutex_exit(&cpu_lock);
17121 return (DDI_FAILURE);
17122 }
17123
17124 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
17125 mutex_exit(&dtrace_provider_lock);
17126 mutex_exit(&dtrace_lock);
17127 mutex_exit(&cpu_lock);
17128 return (DDI_FAILURE);
17129 }
17130
17131 dtrace_provider = NULL;
17132
17133 if ((state = dtrace_anon_grab()) != NULL) {
17134 /*
17135 * If there were ECBs on this state, the provider should
17136 * have not been allowed to detach; assert that there is
17137 * none.
17138 */
17139 ASSERT(state->dts_necbs == 0);
17140 dtrace_state_destroy(state);
17141
17142 /*
17143 * If we're being detached with anonymous state, we need to
17144 * indicate to the kernel debugger that DTrace is now inactive.
17145 */
17146 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17147 }
17148
17149 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
17150 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
17151 dtrace_cpu_init = NULL;
17152 dtrace_helpers_cleanup = NULL;
17153 dtrace_helpers_fork = NULL;
17154 dtrace_cpustart_init = NULL;
17155 dtrace_cpustart_fini = NULL;
17156 dtrace_debugger_init = NULL;
17157 dtrace_debugger_fini = NULL;
17158 dtrace_modload = NULL;
17159 dtrace_modunload = NULL;
17160
17161 ASSERT(dtrace_getf == 0);
17162 ASSERT(dtrace_closef == NULL);
17163
17164 mutex_exit(&cpu_lock);
17165
17166 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
17167 dtrace_probes = NULL;
17168 dtrace_nprobes = 0;
17169
17170 dtrace_hash_destroy(dtrace_bymod);
17171 dtrace_hash_destroy(dtrace_byfunc);
17172 dtrace_hash_destroy(dtrace_byname);
17173 dtrace_bymod = NULL;
17174 dtrace_byfunc = NULL;
17175 dtrace_byname = NULL;
17176
17177 kmem_cache_destroy(dtrace_state_cache);
17178 vmem_destroy(dtrace_minor);
17179 vmem_destroy(dtrace_arena);
17180
17181 if (dtrace_toxrange != NULL) {
17182 kmem_free(dtrace_toxrange,
17183 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
17184 dtrace_toxrange = NULL;
17185 dtrace_toxranges = 0;
17186 dtrace_toxranges_max = 0;
17187 }
17188
17189 ddi_remove_minor_node(dtrace_devi, NULL);
17190 dtrace_devi = NULL;
17191
17192 ddi_soft_state_fini(&dtrace_softstate);
17193
17194 ASSERT(dtrace_vtime_references == 0);
17195 ASSERT(dtrace_opens == 0);
17196 ASSERT(dtrace_retained == NULL);
17197
17198 mutex_exit(&dtrace_lock);
17199 mutex_exit(&dtrace_provider_lock);
17200
17201 /*
17202 * We don't destroy the task queue until after we have dropped our
17203 * locks (taskq_destroy() may block on running tasks). To prevent
17204 * attempting to do work after we have effectively detached but before
17205 * the task queue has been destroyed, all tasks dispatched via the
17206 * task queue must check that DTrace is still attached before
17207 * performing any operation.
17208 */
17209 taskq_destroy(dtrace_taskq);
17210 dtrace_taskq = NULL;
17211
17212 return (DDI_SUCCESS);
17213 }
17214
17215 /*ARGSUSED*/
17216 static int
17217 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
17218 {
17219 int error;
17220
17221 switch (infocmd) {
17222 case DDI_INFO_DEVT2DEVINFO:
17223 *result = (void *)dtrace_devi;
17224 error = DDI_SUCCESS;
17225 break;
17226 case DDI_INFO_DEVT2INSTANCE:
17227 *result = (void *)0;
17228 error = DDI_SUCCESS;
17229 break;
17230 default:
17231 error = DDI_FAILURE;
17232 }
17233 return (error);
17234 }
17235
17236 static struct cb_ops dtrace_cb_ops = {
17237 dtrace_open, /* open */
17238 dtrace_close, /* close */
17239 nulldev, /* strategy */
17240 nulldev, /* print */
17241 nodev, /* dump */
17242 nodev, /* read */
17243 nodev, /* write */
17244 dtrace_ioctl, /* ioctl */
17245 nodev, /* devmap */
17246 nodev, /* mmap */
17247 nodev, /* segmap */
17248 nochpoll, /* poll */
17249 ddi_prop_op, /* cb_prop_op */
17250 0, /* streamtab */
17251 D_NEW | D_MP /* Driver compatibility flag */
17252 };
17253
17254 static struct dev_ops dtrace_ops = {
17255 DEVO_REV, /* devo_rev */
17256 0, /* refcnt */
17257 dtrace_info, /* get_dev_info */
17258 nulldev, /* identify */
17259 nulldev, /* probe */
17260 dtrace_attach, /* attach */
17261 dtrace_detach, /* detach */
17262 nodev, /* reset */
17263 &dtrace_cb_ops, /* driver operations */
17264 NULL, /* bus operations */
17265 nodev, /* dev power */
17266 ddi_quiesce_not_needed, /* quiesce */
17267 };
17268
17269 static struct modldrv modldrv = {
17270 &mod_driverops, /* module type (this is a pseudo driver) */
17271 "Dynamic Tracing", /* name of module */
17272 &dtrace_ops, /* driver ops */
17273 };
17274
17275 static struct modlinkage modlinkage = {
17276 MODREV_1,
17277 (void *)&modldrv,
17278 NULL
17279 };
17280
17281 int
17282 _init(void)
17283 {
17284 return (mod_install(&modlinkage));
17285 }
17286
17287 int
17288 _info(struct modinfo *modinfop)
17289 {
17290 return (mod_info(&modlinkage, modinfop));
17291 }
17292
17293 int
17294 _fini(void)
17295 {
17296 return (mod_remove(&modlinkage));
17297 }