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) 2012, Joyent, Inc. All rights reserved.
25 * Copyright (c) 2012 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_global_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 = 500 * (NANOSEC / MILLISEC); /* 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 __i386
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 /*
388 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
389 * alloc_sz on the righthand side of the comparison in order to avoid overflow
390 * or underflow in the comparison with it. This is simpler than the INRANGE
391 * check above, because we know that the dtms_scratch_ptr is valid in the
392 * range. Allocations of size zero are allowed.
393 */
394 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
395 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
396 (mstate)->dtms_scratch_ptr >= (alloc_sz))
397
398 #define DTRACE_LOADFUNC(bits) \
399 /*CSTYLED*/ \
400 uint##bits##_t \
401 dtrace_load##bits(uintptr_t addr) \
402 { \
403 size_t size = bits / NBBY; \
404 /*CSTYLED*/ \
405 uint##bits##_t rval; \
406 int i; \
407 volatile uint16_t *flags = (volatile uint16_t *) \
408 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; \
409 \
410 DTRACE_ALIGNCHECK(addr, size, flags); \
411 \
412 for (i = 0; i < dtrace_toxranges; i++) { \
413 if (addr >= dtrace_toxrange[i].dtt_limit) \
414 continue; \
415 \
416 if (addr + size <= dtrace_toxrange[i].dtt_base) \
417 continue; \
418 \
419 /* \
420 * This address falls within a toxic region; return 0. \
421 */ \
422 *flags |= CPU_DTRACE_BADADDR; \
423 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
424 return (0); \
425 } \
426 \
427 *flags |= CPU_DTRACE_NOFAULT; \
428 /*CSTYLED*/ \
429 rval = *((volatile uint##bits##_t *)addr); \
430 *flags &= ~CPU_DTRACE_NOFAULT; \
431 \
432 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
433 }
434
435 #ifdef _LP64
436 #define dtrace_loadptr dtrace_load64
437 #else
438 #define dtrace_loadptr dtrace_load32
439 #endif
440
441 #define DTRACE_DYNHASH_FREE 0
442 #define DTRACE_DYNHASH_SINK 1
443 #define DTRACE_DYNHASH_VALID 2
444
445 #define DTRACE_MATCH_FAIL -1
446 #define DTRACE_MATCH_NEXT 0
447 #define DTRACE_MATCH_DONE 1
448 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
449 #define DTRACE_STATE_ALIGN 64
450
451 #define DTRACE_FLAGS2FLT(flags) \
452 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
453 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
454 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
455 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
456 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
457 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
458 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
459 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
460 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
461 DTRACEFLT_UNKNOWN)
462
463 #define DTRACEACT_ISSTRING(act) \
464 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
465 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
466
467 static size_t dtrace_strlen(const char *, size_t);
468 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
469 static void dtrace_enabling_provide(dtrace_provider_t *);
470 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
471 static void dtrace_enabling_matchall(void);
472 static void dtrace_enabling_reap(void);
473 static dtrace_state_t *dtrace_anon_grab(void);
474 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
475 dtrace_state_t *, uint64_t, uint64_t);
476 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
477 static void dtrace_buffer_drop(dtrace_buffer_t *);
478 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
479 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
480 dtrace_state_t *, dtrace_mstate_t *);
481 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
482 dtrace_optval_t);
483 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
484 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
485 static int dtrace_priv_proc(dtrace_state_t *, dtrace_mstate_t *);
486 static void dtrace_getf_barrier(void);
487
488 /*
489 * DTrace Probe Context Functions
490 *
491 * These functions are called from probe context. Because probe context is
492 * any context in which C may be called, arbitrarily locks may be held,
493 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
494 * As a result, functions called from probe context may only call other DTrace
495 * support functions -- they may not interact at all with the system at large.
496 * (Note that the ASSERT macro is made probe-context safe by redefining it in
497 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
498 * loads are to be performed from probe context, they _must_ be in terms of
499 * the safe dtrace_load*() variants.
500 *
501 * Some functions in this block are not actually called from probe context;
502 * for these functions, there will be a comment above the function reading
503 * "Note: not called from probe context."
504 */
505 void
506 dtrace_panic(const char *format, ...)
507 {
508 va_list alist;
509
510 va_start(alist, format);
511 dtrace_vpanic(format, alist);
512 va_end(alist);
513 }
514
515 int
516 dtrace_assfail(const char *a, const char *f, int l)
517 {
518 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
519
520 /*
521 * We just need something here that even the most clever compiler
522 * cannot optimize away.
523 */
524 return (a[(uintptr_t)f]);
525 }
526
527 /*
528 * Atomically increment a specified error counter from probe context.
529 */
530 static void
531 dtrace_error(uint32_t *counter)
532 {
533 /*
534 * Most counters stored to in probe context are per-CPU counters.
535 * However, there are some error conditions that are sufficiently
536 * arcane that they don't merit per-CPU storage. If these counters
537 * are incremented concurrently on different CPUs, scalability will be
538 * adversely affected -- but we don't expect them to be white-hot in a
539 * correctly constructed enabling...
540 */
541 uint32_t oval, nval;
542
543 do {
544 oval = *counter;
545
546 if ((nval = oval + 1) == 0) {
547 /*
548 * If the counter would wrap, set it to 1 -- assuring
549 * that the counter is never zero when we have seen
550 * errors. (The counter must be 32-bits because we
551 * aren't guaranteed a 64-bit compare&swap operation.)
552 * To save this code both the infamy of being fingered
553 * by a priggish news story and the indignity of being
554 * the target of a neo-puritan witch trial, we're
555 * carefully avoiding any colorful description of the
556 * likelihood of this condition -- but suffice it to
557 * say that it is only slightly more likely than the
558 * overflow of predicate cache IDs, as discussed in
559 * dtrace_predicate_create().
560 */
561 nval = 1;
562 }
563 } while (dtrace_cas32(counter, oval, nval) != oval);
564 }
565
566 /*
567 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
568 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
569 */
570 DTRACE_LOADFUNC(8)
571 DTRACE_LOADFUNC(16)
572 DTRACE_LOADFUNC(32)
573 DTRACE_LOADFUNC(64)
574
575 static int
576 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
577 {
578 if (dest < mstate->dtms_scratch_base)
579 return (0);
580
581 if (dest + size < dest)
582 return (0);
583
584 if (dest + size > mstate->dtms_scratch_ptr)
585 return (0);
586
587 return (1);
588 }
589
590 static int
591 dtrace_canstore_statvar(uint64_t addr, size_t sz,
592 dtrace_statvar_t **svars, int nsvars)
593 {
594 int i;
595
596 for (i = 0; i < nsvars; i++) {
597 dtrace_statvar_t *svar = svars[i];
598
599 if (svar == NULL || svar->dtsv_size == 0)
600 continue;
601
602 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
603 return (1);
604 }
605
606 return (0);
607 }
608
609 /*
610 * Check to see if the address is within a memory region to which a store may
611 * be issued. This includes the DTrace scratch areas, and any DTrace variable
612 * region. The caller of dtrace_canstore() is responsible for performing any
613 * alignment checks that are needed before stores are actually executed.
614 */
615 static int
616 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
617 dtrace_vstate_t *vstate)
618 {
619 /*
620 * First, check to see if the address is in scratch space...
621 */
622 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
623 mstate->dtms_scratch_size))
624 return (1);
625
626 /*
627 * Now check to see if it's a dynamic variable. This check will pick
628 * up both thread-local variables and any global dynamically-allocated
629 * variables.
630 */
631 if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
632 vstate->dtvs_dynvars.dtds_size)) {
633 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
634 uintptr_t base = (uintptr_t)dstate->dtds_base +
635 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
636 uintptr_t chunkoffs;
637
638 /*
639 * Before we assume that we can store here, we need to make
640 * sure that it isn't in our metadata -- storing to our
641 * dynamic variable metadata would corrupt our state. For
642 * the range to not include any dynamic variable metadata,
643 * it must:
644 *
645 * (1) Start above the hash table that is at the base of
646 * the dynamic variable space
647 *
648 * (2) Have a starting chunk offset that is beyond the
649 * dtrace_dynvar_t that is at the base of every chunk
650 *
651 * (3) Not span a chunk boundary
652 *
653 */
654 if (addr < base)
655 return (0);
656
657 chunkoffs = (addr - base) % dstate->dtds_chunksize;
658
659 if (chunkoffs < sizeof (dtrace_dynvar_t))
660 return (0);
661
662 if (chunkoffs + sz > dstate->dtds_chunksize)
663 return (0);
664
665 return (1);
666 }
667
668 /*
669 * Finally, check the static local and global variables. These checks
670 * take the longest, so we perform them last.
671 */
672 if (dtrace_canstore_statvar(addr, sz,
673 vstate->dtvs_locals, vstate->dtvs_nlocals))
674 return (1);
675
676 if (dtrace_canstore_statvar(addr, sz,
677 vstate->dtvs_globals, vstate->dtvs_nglobals))
678 return (1);
679
680 return (0);
681 }
682
683
684 /*
685 * Convenience routine to check to see if the address is within a memory
686 * region in which a load may be issued given the user's privilege level;
687 * if not, it sets the appropriate error flags and loads 'addr' into the
688 * illegal value slot.
689 *
690 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
691 * appropriate memory access protection.
692 */
693 static int
694 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
695 dtrace_vstate_t *vstate)
696 {
697 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
698 file_t *fp;
699
700 /*
701 * If we hold the privilege to read from kernel memory, then
702 * everything is readable.
703 */
704 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
705 return (1);
706
707 /*
708 * You can obviously read that which you can store.
709 */
710 if (dtrace_canstore(addr, sz, mstate, vstate))
711 return (1);
712
713 /*
714 * We're allowed to read from our own string table.
715 */
716 if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
717 mstate->dtms_difo->dtdo_strlen))
718 return (1);
719
720 if (vstate->dtvs_state != NULL &&
721 dtrace_priv_proc(vstate->dtvs_state, mstate)) {
722 proc_t *p;
723
724 /*
725 * When we have privileges to the current process, there are
726 * several context-related kernel structures that are safe to
727 * read, even absent the privilege to read from kernel memory.
728 * These reads are safe because these structures contain only
729 * state that (1) we're permitted to read, (2) is harmless or
730 * (3) contains pointers to additional kernel state that we're
731 * not permitted to read (and as such, do not present an
732 * opportunity for privilege escalation). Finally (and
733 * critically), because of the nature of their relation with
734 * the current thread context, the memory associated with these
735 * structures cannot change over the duration of probe context,
736 * and it is therefore impossible for this memory to be
737 * deallocated and reallocated as something else while it's
738 * being operated upon.
739 */
740 if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t)))
741 return (1);
742
743 if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
744 sz, curthread->t_procp, sizeof (proc_t))) {
745 return (1);
746 }
747
748 if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
749 curthread->t_cred, sizeof (cred_t))) {
750 return (1);
751 }
752
753 if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
754 &(p->p_pidp->pid_id), sizeof (pid_t))) {
755 return (1);
756 }
757
758 if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
759 curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
760 return (1);
761 }
762 }
763
764 if ((fp = mstate->dtms_getf) != NULL) {
765 uintptr_t psz = sizeof (void *);
766 vnode_t *vp;
767 vnodeops_t *op;
768
769 /*
770 * When getf() returns a file_t, the enabling is implicitly
771 * granted the (transient) right to read the returned file_t
772 * as well as the v_path and v_op->vnop_name of the underlying
773 * vnode. These accesses are allowed after a successful
774 * getf() because the members that they refer to cannot change
775 * once set -- and the barrier logic in the kernel's closef()
776 * path assures that the file_t and its referenced vode_t
777 * cannot themselves be stale (that is, it impossible for
778 * either dtms_getf itself or its f_vnode member to reference
779 * freed memory).
780 */
781 if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t)))
782 return (1);
783
784 if ((vp = fp->f_vnode) != NULL) {
785 if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz))
786 return (1);
787
788 if (vp->v_path != NULL && DTRACE_INRANGE(addr, sz,
789 vp->v_path, strlen(vp->v_path) + 1)) {
790 return (1);
791 }
792
793 if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz))
794 return (1);
795
796 if ((op = vp->v_op) != NULL &&
797 DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
798 return (1);
799 }
800
801 if (op != NULL && op->vnop_name != NULL &&
802 DTRACE_INRANGE(addr, sz, op->vnop_name,
803 strlen(op->vnop_name) + 1)) {
804 return (1);
805 }
806 }
807 }
808
809 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
810 *illval = addr;
811 return (0);
812 }
813
814 /*
815 * Convenience routine to check to see if a given string is within a memory
816 * region in which a load may be issued given the user's privilege level;
817 * this exists so that we don't need to issue unnecessary dtrace_strlen()
818 * calls in the event that the user has all privileges.
819 */
820 static int
821 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
822 dtrace_vstate_t *vstate)
823 {
824 size_t strsz;
825
826 /*
827 * If we hold the privilege to read from kernel memory, then
828 * everything is readable.
829 */
830 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
831 return (1);
832
833 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
834 if (dtrace_canload(addr, strsz, mstate, vstate))
835 return (1);
836
837 return (0);
838 }
839
840 /*
841 * Convenience routine to check to see if a given variable is within a memory
842 * region in which a load may be issued given the user's privilege level.
843 */
844 static int
845 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
846 dtrace_vstate_t *vstate)
847 {
848 size_t sz, strsize;
849 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
850
851 /*
852 * If we hold the privilege to read from kernel memory, then
853 * everything is readable.
854 */
855 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
856 return (1);
857
858 if (type->dtdt_kind == DIF_TYPE_STRING) {
859 dtrace_state_t *state = vstate->dtvs_state;
860
861 if (state != NULL) {
862 strsize = state->dts_options[DTRACEOPT_STRSIZE];
863 } else {
864 /*
865 * In helper context, we have a NULL state; fall back
866 * to using the system-wide default for the string size
867 * in this case.
868 */
869 strsize = dtrace_strsize_default;
870 }
871
872 sz = dtrace_strlen(src, strsize) + 1;
873 } else {
874 sz = type->dtdt_size;
875 }
876
877 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
878 }
879
880 /*
881 * Convert a string to a signed integer using safe loads.
882 *
883 * NOTE: This function uses various macros from strtolctype.h to manipulate
884 * digit values, etc -- these have all been checked to ensure they make
885 * no additional function calls.
886 */
887 static int64_t
888 dtrace_strtoll(char *input, int base, size_t limit)
889 {
890 uintptr_t pos = (uintptr_t)input;
891 int64_t val = 0;
892 int x;
893 boolean_t neg = B_FALSE;
894 char c, cc, ccc;
895 uintptr_t end = pos + limit;
896
897 /*
898 * Consume any whitespace preceding digits.
899 */
900 while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
901 pos++;
902
903 /*
904 * Handle an explicit sign if one is present.
905 */
906 if (c == '-' || c == '+') {
907 if (c == '-')
908 neg = B_TRUE;
909 c = dtrace_load8(++pos);
910 }
911
912 /*
913 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
914 * if present.
915 */
916 if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
917 cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
918 pos += 2;
919 c = ccc;
920 }
921
922 /*
923 * Read in contiguous digits until the first non-digit character.
924 */
925 for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
926 c = dtrace_load8(++pos))
927 val = val * base + x;
928
929 return (neg ? -val : val);
930 }
931
932 /*
933 * Compare two strings using safe loads.
934 */
935 static int
936 dtrace_strncmp(char *s1, char *s2, size_t limit)
937 {
938 uint8_t c1, c2;
939 volatile uint16_t *flags;
940
941 if (s1 == s2 || limit == 0)
942 return (0);
943
944 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
945
946 do {
947 if (s1 == NULL) {
948 c1 = '\0';
949 } else {
950 c1 = dtrace_load8((uintptr_t)s1++);
951 }
952
953 if (s2 == NULL) {
954 c2 = '\0';
955 } else {
956 c2 = dtrace_load8((uintptr_t)s2++);
957 }
958
959 if (c1 != c2)
960 return (c1 - c2);
961 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
962
963 return (0);
964 }
965
966 /*
967 * Compute strlen(s) for a string using safe memory accesses. The additional
968 * len parameter is used to specify a maximum length to ensure completion.
969 */
970 static size_t
971 dtrace_strlen(const char *s, size_t lim)
972 {
973 uint_t len;
974
975 for (len = 0; len != lim; len++) {
976 if (dtrace_load8((uintptr_t)s++) == '\0')
977 break;
978 }
979
980 return (len);
981 }
982
983 /*
984 * Check if an address falls within a toxic region.
985 */
986 static int
987 dtrace_istoxic(uintptr_t kaddr, size_t size)
988 {
989 uintptr_t taddr, tsize;
990 int i;
991
992 for (i = 0; i < dtrace_toxranges; i++) {
993 taddr = dtrace_toxrange[i].dtt_base;
994 tsize = dtrace_toxrange[i].dtt_limit - taddr;
995
996 if (kaddr - taddr < tsize) {
997 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
998 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
999 return (1);
1000 }
1001
1002 if (taddr - kaddr < size) {
1003 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1004 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
1005 return (1);
1006 }
1007 }
1008
1009 return (0);
1010 }
1011
1012 /*
1013 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
1014 * memory specified by the DIF program. The dst is assumed to be safe memory
1015 * that we can store to directly because it is managed by DTrace. As with
1016 * standard bcopy, overlapping copies are handled properly.
1017 */
1018 static void
1019 dtrace_bcopy(const void *src, void *dst, size_t len)
1020 {
1021 if (len != 0) {
1022 uint8_t *s1 = dst;
1023 const uint8_t *s2 = src;
1024
1025 if (s1 <= s2) {
1026 do {
1027 *s1++ = dtrace_load8((uintptr_t)s2++);
1028 } while (--len != 0);
1029 } else {
1030 s2 += len;
1031 s1 += len;
1032
1033 do {
1034 *--s1 = dtrace_load8((uintptr_t)--s2);
1035 } while (--len != 0);
1036 }
1037 }
1038 }
1039
1040 /*
1041 * Copy src to dst using safe memory accesses, up to either the specified
1042 * length, or the point that a nul byte is encountered. The src is assumed to
1043 * be unsafe memory specified by the DIF program. The dst is assumed to be
1044 * safe memory that we can store to directly because it is managed by DTrace.
1045 * Unlike dtrace_bcopy(), overlapping regions are not handled.
1046 */
1047 static void
1048 dtrace_strcpy(const void *src, void *dst, size_t len)
1049 {
1050 if (len != 0) {
1051 uint8_t *s1 = dst, c;
1052 const uint8_t *s2 = src;
1053
1054 do {
1055 *s1++ = c = dtrace_load8((uintptr_t)s2++);
1056 } while (--len != 0 && c != '\0');
1057 }
1058 }
1059
1060 /*
1061 * Copy src to dst, deriving the size and type from the specified (BYREF)
1062 * variable type. The src is assumed to be unsafe memory specified by the DIF
1063 * program. The dst is assumed to be DTrace variable memory that is of the
1064 * specified type; we assume that we can store to directly.
1065 */
1066 static void
1067 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1068 {
1069 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1070
1071 if (type->dtdt_kind == DIF_TYPE_STRING) {
1072 dtrace_strcpy(src, dst, type->dtdt_size);
1073 } else {
1074 dtrace_bcopy(src, dst, type->dtdt_size);
1075 }
1076 }
1077
1078 /*
1079 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1080 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1081 * safe memory that we can access directly because it is managed by DTrace.
1082 */
1083 static int
1084 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1085 {
1086 volatile uint16_t *flags;
1087
1088 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1089
1090 if (s1 == s2)
1091 return (0);
1092
1093 if (s1 == NULL || s2 == NULL)
1094 return (1);
1095
1096 if (s1 != s2 && len != 0) {
1097 const uint8_t *ps1 = s1;
1098 const uint8_t *ps2 = s2;
1099
1100 do {
1101 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1102 return (1);
1103 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1104 }
1105 return (0);
1106 }
1107
1108 /*
1109 * Zero the specified region using a simple byte-by-byte loop. Note that this
1110 * is for safe DTrace-managed memory only.
1111 */
1112 static void
1113 dtrace_bzero(void *dst, size_t len)
1114 {
1115 uchar_t *cp;
1116
1117 for (cp = dst; len != 0; len--)
1118 *cp++ = 0;
1119 }
1120
1121 static void
1122 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1123 {
1124 uint64_t result[2];
1125
1126 result[0] = addend1[0] + addend2[0];
1127 result[1] = addend1[1] + addend2[1] +
1128 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1129
1130 sum[0] = result[0];
1131 sum[1] = result[1];
1132 }
1133
1134 /*
1135 * Shift the 128-bit value in a by b. If b is positive, shift left.
1136 * If b is negative, shift right.
1137 */
1138 static void
1139 dtrace_shift_128(uint64_t *a, int b)
1140 {
1141 uint64_t mask;
1142
1143 if (b == 0)
1144 return;
1145
1146 if (b < 0) {
1147 b = -b;
1148 if (b >= 64) {
1149 a[0] = a[1] >> (b - 64);
1150 a[1] = 0;
1151 } else {
1152 a[0] >>= b;
1153 mask = 1LL << (64 - b);
1154 mask -= 1;
1155 a[0] |= ((a[1] & mask) << (64 - b));
1156 a[1] >>= b;
1157 }
1158 } else {
1159 if (b >= 64) {
1160 a[1] = a[0] << (b - 64);
1161 a[0] = 0;
1162 } else {
1163 a[1] <<= b;
1164 mask = a[0] >> (64 - b);
1165 a[1] |= mask;
1166 a[0] <<= b;
1167 }
1168 }
1169 }
1170
1171 /*
1172 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1173 * use native multiplication on those, and then re-combine into the
1174 * resulting 128-bit value.
1175 *
1176 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1177 * hi1 * hi2 << 64 +
1178 * hi1 * lo2 << 32 +
1179 * hi2 * lo1 << 32 +
1180 * lo1 * lo2
1181 */
1182 static void
1183 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1184 {
1185 uint64_t hi1, hi2, lo1, lo2;
1186 uint64_t tmp[2];
1187
1188 hi1 = factor1 >> 32;
1189 hi2 = factor2 >> 32;
1190
1191 lo1 = factor1 & DT_MASK_LO;
1192 lo2 = factor2 & DT_MASK_LO;
1193
1194 product[0] = lo1 * lo2;
1195 product[1] = hi1 * hi2;
1196
1197 tmp[0] = hi1 * lo2;
1198 tmp[1] = 0;
1199 dtrace_shift_128(tmp, 32);
1200 dtrace_add_128(product, tmp, product);
1201
1202 tmp[0] = hi2 * lo1;
1203 tmp[1] = 0;
1204 dtrace_shift_128(tmp, 32);
1205 dtrace_add_128(product, tmp, product);
1206 }
1207
1208 /*
1209 * This privilege check should be used by actions and subroutines to
1210 * verify that the user credentials of the process that enabled the
1211 * invoking ECB match the target credentials
1212 */
1213 static int
1214 dtrace_priv_proc_common_user(dtrace_state_t *state)
1215 {
1216 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1217
1218 /*
1219 * We should always have a non-NULL state cred here, since if cred
1220 * is null (anonymous tracing), we fast-path bypass this routine.
1221 */
1222 ASSERT(s_cr != NULL);
1223
1224 if ((cr = CRED()) != NULL &&
1225 s_cr->cr_uid == cr->cr_uid &&
1226 s_cr->cr_uid == cr->cr_ruid &&
1227 s_cr->cr_uid == cr->cr_suid &&
1228 s_cr->cr_gid == cr->cr_gid &&
1229 s_cr->cr_gid == cr->cr_rgid &&
1230 s_cr->cr_gid == cr->cr_sgid)
1231 return (1);
1232
1233 return (0);
1234 }
1235
1236 /*
1237 * This privilege check should be used by actions and subroutines to
1238 * verify that the zone of the process that enabled the invoking ECB
1239 * matches the target credentials
1240 */
1241 static int
1242 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1243 {
1244 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1245
1246 /*
1247 * We should always have a non-NULL state cred here, since if cred
1248 * is null (anonymous tracing), we fast-path bypass this routine.
1249 */
1250 ASSERT(s_cr != NULL);
1251
1252 if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1253 return (1);
1254
1255 return (0);
1256 }
1257
1258 /*
1259 * This privilege check should be used by actions and subroutines to
1260 * verify that the process has not setuid or changed credentials.
1261 */
1262 static int
1263 dtrace_priv_proc_common_nocd()
1264 {
1265 proc_t *proc;
1266
1267 if ((proc = ttoproc(curthread)) != NULL &&
1268 !(proc->p_flag & SNOCD))
1269 return (1);
1270
1271 return (0);
1272 }
1273
1274 static int
1275 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1276 {
1277 int action = state->dts_cred.dcr_action;
1278
1279 if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1280 goto bad;
1281
1282 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1283 dtrace_priv_proc_common_zone(state) == 0)
1284 goto bad;
1285
1286 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1287 dtrace_priv_proc_common_user(state) == 0)
1288 goto bad;
1289
1290 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1291 dtrace_priv_proc_common_nocd() == 0)
1292 goto bad;
1293
1294 return (1);
1295
1296 bad:
1297 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1298
1299 return (0);
1300 }
1301
1302 static int
1303 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1304 {
1305 if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1306 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1307 return (1);
1308
1309 if (dtrace_priv_proc_common_zone(state) &&
1310 dtrace_priv_proc_common_user(state) &&
1311 dtrace_priv_proc_common_nocd())
1312 return (1);
1313 }
1314
1315 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1316
1317 return (0);
1318 }
1319
1320 static int
1321 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1322 {
1323 if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1324 (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1325 return (1);
1326
1327 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1328
1329 return (0);
1330 }
1331
1332 static int
1333 dtrace_priv_kernel(dtrace_state_t *state)
1334 {
1335 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1336 return (1);
1337
1338 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1339
1340 return (0);
1341 }
1342
1343 static int
1344 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1345 {
1346 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1347 return (1);
1348
1349 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1350
1351 return (0);
1352 }
1353
1354 /*
1355 * Determine if the dte_cond of the specified ECB allows for processing of
1356 * the current probe to continue. Note that this routine may allow continued
1357 * processing, but with access(es) stripped from the mstate's dtms_access
1358 * field.
1359 */
1360 static int
1361 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1362 dtrace_ecb_t *ecb)
1363 {
1364 dtrace_probe_t *probe = ecb->dte_probe;
1365 dtrace_provider_t *prov = probe->dtpr_provider;
1366 dtrace_pops_t *pops = &prov->dtpv_pops;
1367 int mode = DTRACE_MODE_NOPRIV_DROP;
1368
1369 ASSERT(ecb->dte_cond);
1370
1371 if (pops->dtps_mode != NULL) {
1372 mode = pops->dtps_mode(prov->dtpv_arg,
1373 probe->dtpr_id, probe->dtpr_arg);
1374
1375 ASSERT(mode & (DTRACE_MODE_USER | DTRACE_MODE_KERNEL));
1376 ASSERT(mode & (DTRACE_MODE_NOPRIV_RESTRICT |
1377 DTRACE_MODE_NOPRIV_DROP));
1378 }
1379
1380 /*
1381 * If the dte_cond bits indicate that this consumer is only allowed to
1382 * see user-mode firings of this probe, check that the probe was fired
1383 * while in a user context. If that's not the case, use the policy
1384 * specified by the provider to determine if we drop the probe or
1385 * merely restrict operation.
1386 */
1387 if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1388 ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1389
1390 if (!(mode & DTRACE_MODE_USER)) {
1391 if (mode & DTRACE_MODE_NOPRIV_DROP)
1392 return (0);
1393
1394 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1395 }
1396 }
1397
1398 /*
1399 * This is more subtle than it looks. We have to be absolutely certain
1400 * that CRED() isn't going to change out from under us so it's only
1401 * legit to examine that structure if we're in constrained situations.
1402 * Currently, the only times we'll this check is if a non-super-user
1403 * has enabled the profile or syscall providers -- providers that
1404 * allow visibility of all processes. For the profile case, the check
1405 * above will ensure that we're examining a user context.
1406 */
1407 if (ecb->dte_cond & DTRACE_COND_OWNER) {
1408 cred_t *cr;
1409 cred_t *s_cr = state->dts_cred.dcr_cred;
1410 proc_t *proc;
1411
1412 ASSERT(s_cr != NULL);
1413
1414 if ((cr = CRED()) == NULL ||
1415 s_cr->cr_uid != cr->cr_uid ||
1416 s_cr->cr_uid != cr->cr_ruid ||
1417 s_cr->cr_uid != cr->cr_suid ||
1418 s_cr->cr_gid != cr->cr_gid ||
1419 s_cr->cr_gid != cr->cr_rgid ||
1420 s_cr->cr_gid != cr->cr_sgid ||
1421 (proc = ttoproc(curthread)) == NULL ||
1422 (proc->p_flag & SNOCD)) {
1423 if (mode & DTRACE_MODE_NOPRIV_DROP)
1424 return (0);
1425
1426 mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1427 }
1428 }
1429
1430 /*
1431 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1432 * in our zone, check to see if our mode policy is to restrict rather
1433 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1434 * and DTRACE_ACCESS_ARGS
1435 */
1436 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1437 cred_t *cr;
1438 cred_t *s_cr = state->dts_cred.dcr_cred;
1439
1440 ASSERT(s_cr != NULL);
1441
1442 if ((cr = CRED()) == NULL ||
1443 s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1444 if (mode & DTRACE_MODE_NOPRIV_DROP)
1445 return (0);
1446
1447 mstate->dtms_access &=
1448 ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1449 }
1450 }
1451
1452 /*
1453 * By merits of being in this code path at all, we have limited
1454 * privileges. If the provider has indicated that limited privileges
1455 * are to denote restricted operation, strip off the ability to access
1456 * arguments.
1457 */
1458 if (mode & DTRACE_MODE_LIMITEDPRIV_RESTRICT)
1459 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1460
1461 return (1);
1462 }
1463
1464 /*
1465 * Note: not called from probe context. This function is called
1466 * asynchronously (and at a regular interval) from outside of probe context to
1467 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1468 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1469 */
1470 void
1471 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1472 {
1473 dtrace_dynvar_t *dirty;
1474 dtrace_dstate_percpu_t *dcpu;
1475 dtrace_dynvar_t **rinsep;
1476 int i, j, work = 0;
1477
1478 for (i = 0; i < NCPU; i++) {
1479 dcpu = &dstate->dtds_percpu[i];
1480 rinsep = &dcpu->dtdsc_rinsing;
1481
1482 /*
1483 * If the dirty list is NULL, there is no dirty work to do.
1484 */
1485 if (dcpu->dtdsc_dirty == NULL)
1486 continue;
1487
1488 if (dcpu->dtdsc_rinsing != NULL) {
1489 /*
1490 * If the rinsing list is non-NULL, then it is because
1491 * this CPU was selected to accept another CPU's
1492 * dirty list -- and since that time, dirty buffers
1493 * have accumulated. This is a highly unlikely
1494 * condition, but we choose to ignore the dirty
1495 * buffers -- they'll be picked up a future cleanse.
1496 */
1497 continue;
1498 }
1499
1500 if (dcpu->dtdsc_clean != NULL) {
1501 /*
1502 * If the clean list is non-NULL, then we're in a
1503 * situation where a CPU has done deallocations (we
1504 * have a non-NULL dirty list) but no allocations (we
1505 * also have a non-NULL clean list). We can't simply
1506 * move the dirty list into the clean list on this
1507 * CPU, yet we also don't want to allow this condition
1508 * to persist, lest a short clean list prevent a
1509 * massive dirty list from being cleaned (which in
1510 * turn could lead to otherwise avoidable dynamic
1511 * drops). To deal with this, we look for some CPU
1512 * with a NULL clean list, NULL dirty list, and NULL
1513 * rinsing list -- and then we borrow this CPU to
1514 * rinse our dirty list.
1515 */
1516 for (j = 0; j < NCPU; j++) {
1517 dtrace_dstate_percpu_t *rinser;
1518
1519 rinser = &dstate->dtds_percpu[j];
1520
1521 if (rinser->dtdsc_rinsing != NULL)
1522 continue;
1523
1524 if (rinser->dtdsc_dirty != NULL)
1525 continue;
1526
1527 if (rinser->dtdsc_clean != NULL)
1528 continue;
1529
1530 rinsep = &rinser->dtdsc_rinsing;
1531 break;
1532 }
1533
1534 if (j == NCPU) {
1535 /*
1536 * We were unable to find another CPU that
1537 * could accept this dirty list -- we are
1538 * therefore unable to clean it now.
1539 */
1540 dtrace_dynvar_failclean++;
1541 continue;
1542 }
1543 }
1544
1545 work = 1;
1546
1547 /*
1548 * Atomically move the dirty list aside.
1549 */
1550 do {
1551 dirty = dcpu->dtdsc_dirty;
1552
1553 /*
1554 * Before we zap the dirty list, set the rinsing list.
1555 * (This allows for a potential assertion in
1556 * dtrace_dynvar(): if a free dynamic variable appears
1557 * on a hash chain, either the dirty list or the
1558 * rinsing list for some CPU must be non-NULL.)
1559 */
1560 *rinsep = dirty;
1561 dtrace_membar_producer();
1562 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1563 dirty, NULL) != dirty);
1564 }
1565
1566 if (!work) {
1567 /*
1568 * We have no work to do; we can simply return.
1569 */
1570 return;
1571 }
1572
1573 dtrace_sync();
1574
1575 for (i = 0; i < NCPU; i++) {
1576 dcpu = &dstate->dtds_percpu[i];
1577
1578 if (dcpu->dtdsc_rinsing == NULL)
1579 continue;
1580
1581 /*
1582 * We are now guaranteed that no hash chain contains a pointer
1583 * into this dirty list; we can make it clean.
1584 */
1585 ASSERT(dcpu->dtdsc_clean == NULL);
1586 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1587 dcpu->dtdsc_rinsing = NULL;
1588 }
1589
1590 /*
1591 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1592 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1593 * This prevents a race whereby a CPU incorrectly decides that
1594 * the state should be something other than DTRACE_DSTATE_CLEAN
1595 * after dtrace_dynvar_clean() has completed.
1596 */
1597 dtrace_sync();
1598
1599 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1600 }
1601
1602 /*
1603 * Depending on the value of the op parameter, this function looks-up,
1604 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1605 * allocation is requested, this function will return a pointer to a
1606 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1607 * variable can be allocated. If NULL is returned, the appropriate counter
1608 * will be incremented.
1609 */
1610 dtrace_dynvar_t *
1611 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1612 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1613 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1614 {
1615 uint64_t hashval = DTRACE_DYNHASH_VALID;
1616 dtrace_dynhash_t *hash = dstate->dtds_hash;
1617 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1618 processorid_t me = CPU->cpu_id, cpu = me;
1619 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1620 size_t bucket, ksize;
1621 size_t chunksize = dstate->dtds_chunksize;
1622 uintptr_t kdata, lock, nstate;
1623 uint_t i;
1624
1625 ASSERT(nkeys != 0);
1626
1627 /*
1628 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1629 * algorithm. For the by-value portions, we perform the algorithm in
1630 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1631 * bit, and seems to have only a minute effect on distribution. For
1632 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1633 * over each referenced byte. It's painful to do this, but it's much
1634 * better than pathological hash distribution. The efficacy of the
1635 * hashing algorithm (and a comparison with other algorithms) may be
1636 * found by running the ::dtrace_dynstat MDB dcmd.
1637 */
1638 for (i = 0; i < nkeys; i++) {
1639 if (key[i].dttk_size == 0) {
1640 uint64_t val = key[i].dttk_value;
1641
1642 hashval += (val >> 48) & 0xffff;
1643 hashval += (hashval << 10);
1644 hashval ^= (hashval >> 6);
1645
1646 hashval += (val >> 32) & 0xffff;
1647 hashval += (hashval << 10);
1648 hashval ^= (hashval >> 6);
1649
1650 hashval += (val >> 16) & 0xffff;
1651 hashval += (hashval << 10);
1652 hashval ^= (hashval >> 6);
1653
1654 hashval += val & 0xffff;
1655 hashval += (hashval << 10);
1656 hashval ^= (hashval >> 6);
1657 } else {
1658 /*
1659 * This is incredibly painful, but it beats the hell
1660 * out of the alternative.
1661 */
1662 uint64_t j, size = key[i].dttk_size;
1663 uintptr_t base = (uintptr_t)key[i].dttk_value;
1664
1665 if (!dtrace_canload(base, size, mstate, vstate))
1666 break;
1667
1668 for (j = 0; j < size; j++) {
1669 hashval += dtrace_load8(base + j);
1670 hashval += (hashval << 10);
1671 hashval ^= (hashval >> 6);
1672 }
1673 }
1674 }
1675
1676 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1677 return (NULL);
1678
1679 hashval += (hashval << 3);
1680 hashval ^= (hashval >> 11);
1681 hashval += (hashval << 15);
1682
1683 /*
1684 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1685 * comes out to be one of our two sentinel hash values. If this
1686 * actually happens, we set the hashval to be a value known to be a
1687 * non-sentinel value.
1688 */
1689 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1690 hashval = DTRACE_DYNHASH_VALID;
1691
1692 /*
1693 * Yes, it's painful to do a divide here. If the cycle count becomes
1694 * important here, tricks can be pulled to reduce it. (However, it's
1695 * critical that hash collisions be kept to an absolute minimum;
1696 * they're much more painful than a divide.) It's better to have a
1697 * solution that generates few collisions and still keeps things
1698 * relatively simple.
1699 */
1700 bucket = hashval % dstate->dtds_hashsize;
1701
1702 if (op == DTRACE_DYNVAR_DEALLOC) {
1703 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1704
1705 for (;;) {
1706 while ((lock = *lockp) & 1)
1707 continue;
1708
1709 if (dtrace_casptr((void *)lockp,
1710 (void *)lock, (void *)(lock + 1)) == (void *)lock)
1711 break;
1712 }
1713
1714 dtrace_membar_producer();
1715 }
1716
1717 top:
1718 prev = NULL;
1719 lock = hash[bucket].dtdh_lock;
1720
1721 dtrace_membar_consumer();
1722
1723 start = hash[bucket].dtdh_chain;
1724 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1725 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1726 op != DTRACE_DYNVAR_DEALLOC));
1727
1728 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1729 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1730 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1731
1732 if (dvar->dtdv_hashval != hashval) {
1733 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1734 /*
1735 * We've reached the sink, and therefore the
1736 * end of the hash chain; we can kick out of
1737 * the loop knowing that we have seen a valid
1738 * snapshot of state.
1739 */
1740 ASSERT(dvar->dtdv_next == NULL);
1741 ASSERT(dvar == &dtrace_dynhash_sink);
1742 break;
1743 }
1744
1745 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1746 /*
1747 * We've gone off the rails: somewhere along
1748 * the line, one of the members of this hash
1749 * chain was deleted. Note that we could also
1750 * detect this by simply letting this loop run
1751 * to completion, as we would eventually hit
1752 * the end of the dirty list. However, we
1753 * want to avoid running the length of the
1754 * dirty list unnecessarily (it might be quite
1755 * long), so we catch this as early as
1756 * possible by detecting the hash marker. In
1757 * this case, we simply set dvar to NULL and
1758 * break; the conditional after the loop will
1759 * send us back to top.
1760 */
1761 dvar = NULL;
1762 break;
1763 }
1764
1765 goto next;
1766 }
1767
1768 if (dtuple->dtt_nkeys != nkeys)
1769 goto next;
1770
1771 for (i = 0; i < nkeys; i++, dkey++) {
1772 if (dkey->dttk_size != key[i].dttk_size)
1773 goto next; /* size or type mismatch */
1774
1775 if (dkey->dttk_size != 0) {
1776 if (dtrace_bcmp(
1777 (void *)(uintptr_t)key[i].dttk_value,
1778 (void *)(uintptr_t)dkey->dttk_value,
1779 dkey->dttk_size))
1780 goto next;
1781 } else {
1782 if (dkey->dttk_value != key[i].dttk_value)
1783 goto next;
1784 }
1785 }
1786
1787 if (op != DTRACE_DYNVAR_DEALLOC)
1788 return (dvar);
1789
1790 ASSERT(dvar->dtdv_next == NULL ||
1791 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1792
1793 if (prev != NULL) {
1794 ASSERT(hash[bucket].dtdh_chain != dvar);
1795 ASSERT(start != dvar);
1796 ASSERT(prev->dtdv_next == dvar);
1797 prev->dtdv_next = dvar->dtdv_next;
1798 } else {
1799 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1800 start, dvar->dtdv_next) != start) {
1801 /*
1802 * We have failed to atomically swing the
1803 * hash table head pointer, presumably because
1804 * of a conflicting allocation on another CPU.
1805 * We need to reread the hash chain and try
1806 * again.
1807 */
1808 goto top;
1809 }
1810 }
1811
1812 dtrace_membar_producer();
1813
1814 /*
1815 * Now set the hash value to indicate that it's free.
1816 */
1817 ASSERT(hash[bucket].dtdh_chain != dvar);
1818 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1819
1820 dtrace_membar_producer();
1821
1822 /*
1823 * Set the next pointer to point at the dirty list, and
1824 * atomically swing the dirty pointer to the newly freed dvar.
1825 */
1826 do {
1827 next = dcpu->dtdsc_dirty;
1828 dvar->dtdv_next = next;
1829 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1830
1831 /*
1832 * Finally, unlock this hash bucket.
1833 */
1834 ASSERT(hash[bucket].dtdh_lock == lock);
1835 ASSERT(lock & 1);
1836 hash[bucket].dtdh_lock++;
1837
1838 return (NULL);
1839 next:
1840 prev = dvar;
1841 continue;
1842 }
1843
1844 if (dvar == NULL) {
1845 /*
1846 * If dvar is NULL, it is because we went off the rails:
1847 * one of the elements that we traversed in the hash chain
1848 * was deleted while we were traversing it. In this case,
1849 * we assert that we aren't doing a dealloc (deallocs lock
1850 * the hash bucket to prevent themselves from racing with
1851 * one another), and retry the hash chain traversal.
1852 */
1853 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1854 goto top;
1855 }
1856
1857 if (op != DTRACE_DYNVAR_ALLOC) {
1858 /*
1859 * If we are not to allocate a new variable, we want to
1860 * return NULL now. Before we return, check that the value
1861 * of the lock word hasn't changed. If it has, we may have
1862 * seen an inconsistent snapshot.
1863 */
1864 if (op == DTRACE_DYNVAR_NOALLOC) {
1865 if (hash[bucket].dtdh_lock != lock)
1866 goto top;
1867 } else {
1868 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1869 ASSERT(hash[bucket].dtdh_lock == lock);
1870 ASSERT(lock & 1);
1871 hash[bucket].dtdh_lock++;
1872 }
1873
1874 return (NULL);
1875 }
1876
1877 /*
1878 * We need to allocate a new dynamic variable. The size we need is the
1879 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1880 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1881 * the size of any referred-to data (dsize). We then round the final
1882 * size up to the chunksize for allocation.
1883 */
1884 for (ksize = 0, i = 0; i < nkeys; i++)
1885 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1886
1887 /*
1888 * This should be pretty much impossible, but could happen if, say,
1889 * strange DIF specified the tuple. Ideally, this should be an
1890 * assertion and not an error condition -- but that requires that the
1891 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1892 * bullet-proof. (That is, it must not be able to be fooled by
1893 * malicious DIF.) Given the lack of backwards branches in DIF,
1894 * solving this would presumably not amount to solving the Halting
1895 * Problem -- but it still seems awfully hard.
1896 */
1897 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1898 ksize + dsize > chunksize) {
1899 dcpu->dtdsc_drops++;
1900 return (NULL);
1901 }
1902
1903 nstate = DTRACE_DSTATE_EMPTY;
1904
1905 do {
1906 retry:
1907 free = dcpu->dtdsc_free;
1908
1909 if (free == NULL) {
1910 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1911 void *rval;
1912
1913 if (clean == NULL) {
1914 /*
1915 * We're out of dynamic variable space on
1916 * this CPU. Unless we have tried all CPUs,
1917 * we'll try to allocate from a different
1918 * CPU.
1919 */
1920 switch (dstate->dtds_state) {
1921 case DTRACE_DSTATE_CLEAN: {
1922 void *sp = &dstate->dtds_state;
1923
1924 if (++cpu >= NCPU)
1925 cpu = 0;
1926
1927 if (dcpu->dtdsc_dirty != NULL &&
1928 nstate == DTRACE_DSTATE_EMPTY)
1929 nstate = DTRACE_DSTATE_DIRTY;
1930
1931 if (dcpu->dtdsc_rinsing != NULL)
1932 nstate = DTRACE_DSTATE_RINSING;
1933
1934 dcpu = &dstate->dtds_percpu[cpu];
1935
1936 if (cpu != me)
1937 goto retry;
1938
1939 (void) dtrace_cas32(sp,
1940 DTRACE_DSTATE_CLEAN, nstate);
1941
1942 /*
1943 * To increment the correct bean
1944 * counter, take another lap.
1945 */
1946 goto retry;
1947 }
1948
1949 case DTRACE_DSTATE_DIRTY:
1950 dcpu->dtdsc_dirty_drops++;
1951 break;
1952
1953 case DTRACE_DSTATE_RINSING:
1954 dcpu->dtdsc_rinsing_drops++;
1955 break;
1956
1957 case DTRACE_DSTATE_EMPTY:
1958 dcpu->dtdsc_drops++;
1959 break;
1960 }
1961
1962 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1963 return (NULL);
1964 }
1965
1966 /*
1967 * The clean list appears to be non-empty. We want to
1968 * move the clean list to the free list; we start by
1969 * moving the clean pointer aside.
1970 */
1971 if (dtrace_casptr(&dcpu->dtdsc_clean,
1972 clean, NULL) != clean) {
1973 /*
1974 * We are in one of two situations:
1975 *
1976 * (a) The clean list was switched to the
1977 * free list by another CPU.
1978 *
1979 * (b) The clean list was added to by the
1980 * cleansing cyclic.
1981 *
1982 * In either of these situations, we can
1983 * just reattempt the free list allocation.
1984 */
1985 goto retry;
1986 }
1987
1988 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1989
1990 /*
1991 * Now we'll move the clean list to our free list.
1992 * It's impossible for this to fail: the only way
1993 * the free list can be updated is through this
1994 * code path, and only one CPU can own the clean list.
1995 * Thus, it would only be possible for this to fail if
1996 * this code were racing with dtrace_dynvar_clean().
1997 * (That is, if dtrace_dynvar_clean() updated the clean
1998 * list, and we ended up racing to update the free
1999 * list.) This race is prevented by the dtrace_sync()
2000 * in dtrace_dynvar_clean() -- which flushes the
2001 * owners of the clean lists out before resetting
2002 * the clean lists.
2003 */
2004 dcpu = &dstate->dtds_percpu[me];
2005 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
2006 ASSERT(rval == NULL);
2007 goto retry;
2008 }
2009
2010 dvar = free;
2011 new_free = dvar->dtdv_next;
2012 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2013
2014 /*
2015 * We have now allocated a new chunk. We copy the tuple keys into the
2016 * tuple array and copy any referenced key data into the data space
2017 * following the tuple array. As we do this, we relocate dttk_value
2018 * in the final tuple to point to the key data address in the chunk.
2019 */
2020 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2021 dvar->dtdv_data = (void *)(kdata + ksize);
2022 dvar->dtdv_tuple.dtt_nkeys = nkeys;
2023
2024 for (i = 0; i < nkeys; i++) {
2025 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2026 size_t kesize = key[i].dttk_size;
2027
2028 if (kesize != 0) {
2029 dtrace_bcopy(
2030 (const void *)(uintptr_t)key[i].dttk_value,
2031 (void *)kdata, kesize);
2032 dkey->dttk_value = kdata;
2033 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2034 } else {
2035 dkey->dttk_value = key[i].dttk_value;
2036 }
2037
2038 dkey->dttk_size = kesize;
2039 }
2040
2041 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2042 dvar->dtdv_hashval = hashval;
2043 dvar->dtdv_next = start;
2044
2045 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2046 return (dvar);
2047
2048 /*
2049 * The cas has failed. Either another CPU is adding an element to
2050 * this hash chain, or another CPU is deleting an element from this
2051 * hash chain. The simplest way to deal with both of these cases
2052 * (though not necessarily the most efficient) is to free our
2053 * allocated block and tail-call ourselves. Note that the free is
2054 * to the dirty list and _not_ to the free list. This is to prevent
2055 * races with allocators, above.
2056 */
2057 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2058
2059 dtrace_membar_producer();
2060
2061 do {
2062 free = dcpu->dtdsc_dirty;
2063 dvar->dtdv_next = free;
2064 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2065
2066 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
2067 }
2068
2069 /*ARGSUSED*/
2070 static void
2071 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2072 {
2073 if ((int64_t)nval < (int64_t)*oval)
2074 *oval = nval;
2075 }
2076
2077 /*ARGSUSED*/
2078 static void
2079 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2080 {
2081 if ((int64_t)nval > (int64_t)*oval)
2082 *oval = nval;
2083 }
2084
2085 static void
2086 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2087 {
2088 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2089 int64_t val = (int64_t)nval;
2090
2091 if (val < 0) {
2092 for (i = 0; i < zero; i++) {
2093 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2094 quanta[i] += incr;
2095 return;
2096 }
2097 }
2098 } else {
2099 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2100 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2101 quanta[i - 1] += incr;
2102 return;
2103 }
2104 }
2105
2106 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2107 return;
2108 }
2109
2110 ASSERT(0);
2111 }
2112
2113 static void
2114 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2115 {
2116 uint64_t arg = *lquanta++;
2117 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2118 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2119 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2120 int32_t val = (int32_t)nval, level;
2121
2122 ASSERT(step != 0);
2123 ASSERT(levels != 0);
2124
2125 if (val < base) {
2126 /*
2127 * This is an underflow.
2128 */
2129 lquanta[0] += incr;
2130 return;
2131 }
2132
2133 level = (val - base) / step;
2134
2135 if (level < levels) {
2136 lquanta[level + 1] += incr;
2137 return;
2138 }
2139
2140 /*
2141 * This is an overflow.
2142 */
2143 lquanta[levels + 1] += incr;
2144 }
2145
2146 static int
2147 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2148 uint16_t high, uint16_t nsteps, int64_t value)
2149 {
2150 int64_t this = 1, last, next;
2151 int base = 1, order;
2152
2153 ASSERT(factor <= nsteps);
2154 ASSERT(nsteps % factor == 0);
2155
2156 for (order = 0; order < low; order++)
2157 this *= factor;
2158
2159 /*
2160 * If our value is less than our factor taken to the power of the
2161 * low order of magnitude, it goes into the zeroth bucket.
2162 */
2163 if (value < (last = this))
2164 return (0);
2165
2166 for (this *= factor; order <= high; order++) {
2167 int nbuckets = this > nsteps ? nsteps : this;
2168
2169 if ((next = this * factor) < this) {
2170 /*
2171 * We should not generally get log/linear quantizations
2172 * with a high magnitude that allows 64-bits to
2173 * overflow, but we nonetheless protect against this
2174 * by explicitly checking for overflow, and clamping
2175 * our value accordingly.
2176 */
2177 value = this - 1;
2178 }
2179
2180 if (value < this) {
2181 /*
2182 * If our value lies within this order of magnitude,
2183 * determine its position by taking the offset within
2184 * the order of magnitude, dividing by the bucket
2185 * width, and adding to our (accumulated) base.
2186 */
2187 return (base + (value - last) / (this / nbuckets));
2188 }
2189
2190 base += nbuckets - (nbuckets / factor);
2191 last = this;
2192 this = next;
2193 }
2194
2195 /*
2196 * Our value is greater than or equal to our factor taken to the
2197 * power of one plus the high magnitude -- return the top bucket.
2198 */
2199 return (base);
2200 }
2201
2202 static void
2203 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2204 {
2205 uint64_t arg = *llquanta++;
2206 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2207 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2208 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2209 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2210
2211 llquanta[dtrace_aggregate_llquantize_bucket(factor,
2212 low, high, nsteps, nval)] += incr;
2213 }
2214
2215 /*ARGSUSED*/
2216 static void
2217 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2218 {
2219 data[0]++;
2220 data[1] += nval;
2221 }
2222
2223 /*ARGSUSED*/
2224 static void
2225 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2226 {
2227 int64_t snval = (int64_t)nval;
2228 uint64_t tmp[2];
2229
2230 data[0]++;
2231 data[1] += nval;
2232
2233 /*
2234 * What we want to say here is:
2235 *
2236 * data[2] += nval * nval;
2237 *
2238 * But given that nval is 64-bit, we could easily overflow, so
2239 * we do this as 128-bit arithmetic.
2240 */
2241 if (snval < 0)
2242 snval = -snval;
2243
2244 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2245 dtrace_add_128(data + 2, tmp, data + 2);
2246 }
2247
2248 /*ARGSUSED*/
2249 static void
2250 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2251 {
2252 *oval = *oval + 1;
2253 }
2254
2255 /*ARGSUSED*/
2256 static void
2257 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2258 {
2259 *oval += nval;
2260 }
2261
2262 /*
2263 * Aggregate given the tuple in the principal data buffer, and the aggregating
2264 * action denoted by the specified dtrace_aggregation_t. The aggregation
2265 * buffer is specified as the buf parameter. This routine does not return
2266 * failure; if there is no space in the aggregation buffer, the data will be
2267 * dropped, and a corresponding counter incremented.
2268 */
2269 static void
2270 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2271 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2272 {
2273 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2274 uint32_t i, ndx, size, fsize;
2275 uint32_t align = sizeof (uint64_t) - 1;
2276 dtrace_aggbuffer_t *agb;
2277 dtrace_aggkey_t *key;
2278 uint32_t hashval = 0, limit, isstr;
2279 caddr_t tomax, data, kdata;
2280 dtrace_actkind_t action;
2281 dtrace_action_t *act;
2282 uintptr_t offs;
2283
2284 if (buf == NULL)
2285 return;
2286
2287 if (!agg->dtag_hasarg) {
2288 /*
2289 * Currently, only quantize() and lquantize() take additional
2290 * arguments, and they have the same semantics: an increment
2291 * value that defaults to 1 when not present. If additional
2292 * aggregating actions take arguments, the setting of the
2293 * default argument value will presumably have to become more
2294 * sophisticated...
2295 */
2296 arg = 1;
2297 }
2298
2299 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2300 size = rec->dtrd_offset - agg->dtag_base;
2301 fsize = size + rec->dtrd_size;
2302
2303 ASSERT(dbuf->dtb_tomax != NULL);
2304 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2305
2306 if ((tomax = buf->dtb_tomax) == NULL) {
2307 dtrace_buffer_drop(buf);
2308 return;
2309 }
2310
2311 /*
2312 * The metastructure is always at the bottom of the buffer.
2313 */
2314 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2315 sizeof (dtrace_aggbuffer_t));
2316
2317 if (buf->dtb_offset == 0) {
2318 /*
2319 * We just kludge up approximately 1/8th of the size to be
2320 * buckets. If this guess ends up being routinely
2321 * off-the-mark, we may need to dynamically readjust this
2322 * based on past performance.
2323 */
2324 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2325
2326 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2327 (uintptr_t)tomax || hashsize == 0) {
2328 /*
2329 * We've been given a ludicrously small buffer;
2330 * increment our drop count and leave.
2331 */
2332 dtrace_buffer_drop(buf);
2333 return;
2334 }
2335
2336 /*
2337 * And now, a pathetic attempt to try to get a an odd (or
2338 * perchance, a prime) hash size for better hash distribution.
2339 */
2340 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2341 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2342
2343 agb->dtagb_hashsize = hashsize;
2344 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2345 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2346 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2347
2348 for (i = 0; i < agb->dtagb_hashsize; i++)
2349 agb->dtagb_hash[i] = NULL;
2350 }
2351
2352 ASSERT(agg->dtag_first != NULL);
2353 ASSERT(agg->dtag_first->dta_intuple);
2354
2355 /*
2356 * Calculate the hash value based on the key. Note that we _don't_
2357 * include the aggid in the hashing (but we will store it as part of
2358 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2359 * algorithm: a simple, quick algorithm that has no known funnels, and
2360 * gets good distribution in practice. The efficacy of the hashing
2361 * algorithm (and a comparison with other algorithms) may be found by
2362 * running the ::dtrace_aggstat MDB dcmd.
2363 */
2364 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2365 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2366 limit = i + act->dta_rec.dtrd_size;
2367 ASSERT(limit <= size);
2368 isstr = DTRACEACT_ISSTRING(act);
2369
2370 for (; i < limit; i++) {
2371 hashval += data[i];
2372 hashval += (hashval << 10);
2373 hashval ^= (hashval >> 6);
2374
2375 if (isstr && data[i] == '\0')
2376 break;
2377 }
2378 }
2379
2380 hashval += (hashval << 3);
2381 hashval ^= (hashval >> 11);
2382 hashval += (hashval << 15);
2383
2384 /*
2385 * Yes, the divide here is expensive -- but it's generally the least
2386 * of the performance issues given the amount of data that we iterate
2387 * over to compute hash values, compare data, etc.
2388 */
2389 ndx = hashval % agb->dtagb_hashsize;
2390
2391 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2392 ASSERT((caddr_t)key >= tomax);
2393 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2394
2395 if (hashval != key->dtak_hashval || key->dtak_size != size)
2396 continue;
2397
2398 kdata = key->dtak_data;
2399 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2400
2401 for (act = agg->dtag_first; act->dta_intuple;
2402 act = act->dta_next) {
2403 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2404 limit = i + act->dta_rec.dtrd_size;
2405 ASSERT(limit <= size);
2406 isstr = DTRACEACT_ISSTRING(act);
2407
2408 for (; i < limit; i++) {
2409 if (kdata[i] != data[i])
2410 goto next;
2411
2412 if (isstr && data[i] == '\0')
2413 break;
2414 }
2415 }
2416
2417 if (action != key->dtak_action) {
2418 /*
2419 * We are aggregating on the same value in the same
2420 * aggregation with two different aggregating actions.
2421 * (This should have been picked up in the compiler,
2422 * so we may be dealing with errant or devious DIF.)
2423 * This is an error condition; we indicate as much,
2424 * and return.
2425 */
2426 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2427 return;
2428 }
2429
2430 /*
2431 * This is a hit: we need to apply the aggregator to
2432 * the value at this key.
2433 */
2434 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2435 return;
2436 next:
2437 continue;
2438 }
2439
2440 /*
2441 * We didn't find it. We need to allocate some zero-filled space,
2442 * link it into the hash table appropriately, and apply the aggregator
2443 * to the (zero-filled) value.
2444 */
2445 offs = buf->dtb_offset;
2446 while (offs & (align - 1))
2447 offs += sizeof (uint32_t);
2448
2449 /*
2450 * If we don't have enough room to both allocate a new key _and_
2451 * its associated data, increment the drop count and return.
2452 */
2453 if ((uintptr_t)tomax + offs + fsize >
2454 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2455 dtrace_buffer_drop(buf);
2456 return;
2457 }
2458
2459 /*CONSTCOND*/
2460 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2461 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2462 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2463
2464 key->dtak_data = kdata = tomax + offs;
2465 buf->dtb_offset = offs + fsize;
2466
2467 /*
2468 * Now copy the data across.
2469 */
2470 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2471
2472 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2473 kdata[i] = data[i];
2474
2475 /*
2476 * Because strings are not zeroed out by default, we need to iterate
2477 * looking for actions that store strings, and we need to explicitly
2478 * pad these strings out with zeroes.
2479 */
2480 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2481 int nul;
2482
2483 if (!DTRACEACT_ISSTRING(act))
2484 continue;
2485
2486 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2487 limit = i + act->dta_rec.dtrd_size;
2488 ASSERT(limit <= size);
2489
2490 for (nul = 0; i < limit; i++) {
2491 if (nul) {
2492 kdata[i] = '\0';
2493 continue;
2494 }
2495
2496 if (data[i] != '\0')
2497 continue;
2498
2499 nul = 1;
2500 }
2501 }
2502
2503 for (i = size; i < fsize; i++)
2504 kdata[i] = 0;
2505
2506 key->dtak_hashval = hashval;
2507 key->dtak_size = size;
2508 key->dtak_action = action;
2509 key->dtak_next = agb->dtagb_hash[ndx];
2510 agb->dtagb_hash[ndx] = key;
2511
2512 /*
2513 * Finally, apply the aggregator.
2514 */
2515 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2516 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2517 }
2518
2519 /*
2520 * Given consumer state, this routine finds a speculation in the INACTIVE
2521 * state and transitions it into the ACTIVE state. If there is no speculation
2522 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2523 * incremented -- it is up to the caller to take appropriate action.
2524 */
2525 static int
2526 dtrace_speculation(dtrace_state_t *state)
2527 {
2528 int i = 0;
2529 dtrace_speculation_state_t current;
2530 uint32_t *stat = &state->dts_speculations_unavail, count;
2531
2532 while (i < state->dts_nspeculations) {
2533 dtrace_speculation_t *spec = &state->dts_speculations[i];
2534
2535 current = spec->dtsp_state;
2536
2537 if (current != DTRACESPEC_INACTIVE) {
2538 if (current == DTRACESPEC_COMMITTINGMANY ||
2539 current == DTRACESPEC_COMMITTING ||
2540 current == DTRACESPEC_DISCARDING)
2541 stat = &state->dts_speculations_busy;
2542 i++;
2543 continue;
2544 }
2545
2546 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2547 current, DTRACESPEC_ACTIVE) == current)
2548 return (i + 1);
2549 }
2550
2551 /*
2552 * We couldn't find a speculation. If we found as much as a single
2553 * busy speculation buffer, we'll attribute this failure as "busy"
2554 * instead of "unavail".
2555 */
2556 do {
2557 count = *stat;
2558 } while (dtrace_cas32(stat, count, count + 1) != count);
2559
2560 return (0);
2561 }
2562
2563 /*
2564 * This routine commits an active speculation. If the specified speculation
2565 * is not in a valid state to perform a commit(), this routine will silently do
2566 * nothing. The state of the specified speculation is transitioned according
2567 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2568 */
2569 static void
2570 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2571 dtrace_specid_t which)
2572 {
2573 dtrace_speculation_t *spec;
2574 dtrace_buffer_t *src, *dest;
2575 uintptr_t daddr, saddr, dlimit, slimit;
2576 dtrace_speculation_state_t current, new;
2577 intptr_t offs;
2578 uint64_t timestamp;
2579
2580 if (which == 0)
2581 return;
2582
2583 if (which > state->dts_nspeculations) {
2584 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2585 return;
2586 }
2587
2588 spec = &state->dts_speculations[which - 1];
2589 src = &spec->dtsp_buffer[cpu];
2590 dest = &state->dts_buffer[cpu];
2591
2592 do {
2593 current = spec->dtsp_state;
2594
2595 if (current == DTRACESPEC_COMMITTINGMANY)
2596 break;
2597
2598 switch (current) {
2599 case DTRACESPEC_INACTIVE:
2600 case DTRACESPEC_DISCARDING:
2601 return;
2602
2603 case DTRACESPEC_COMMITTING:
2604 /*
2605 * This is only possible if we are (a) commit()'ing
2606 * without having done a prior speculate() on this CPU
2607 * and (b) racing with another commit() on a different
2608 * CPU. There's nothing to do -- we just assert that
2609 * our offset is 0.
2610 */
2611 ASSERT(src->dtb_offset == 0);
2612 return;
2613
2614 case DTRACESPEC_ACTIVE:
2615 new = DTRACESPEC_COMMITTING;
2616 break;
2617
2618 case DTRACESPEC_ACTIVEONE:
2619 /*
2620 * This speculation is active on one CPU. If our
2621 * buffer offset is non-zero, we know that the one CPU
2622 * must be us. Otherwise, we are committing on a
2623 * different CPU from the speculate(), and we must
2624 * rely on being asynchronously cleaned.
2625 */
2626 if (src->dtb_offset != 0) {
2627 new = DTRACESPEC_COMMITTING;
2628 break;
2629 }
2630 /*FALLTHROUGH*/
2631
2632 case DTRACESPEC_ACTIVEMANY:
2633 new = DTRACESPEC_COMMITTINGMANY;
2634 break;
2635
2636 default:
2637 ASSERT(0);
2638 }
2639 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2640 current, new) != current);
2641
2642 /*
2643 * We have set the state to indicate that we are committing this
2644 * speculation. Now reserve the necessary space in the destination
2645 * buffer.
2646 */
2647 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2648 sizeof (uint64_t), state, NULL)) < 0) {
2649 dtrace_buffer_drop(dest);
2650 goto out;
2651 }
2652
2653 /*
2654 * We have sufficient space to copy the speculative buffer into the
2655 * primary buffer. First, modify the speculative buffer, filling
2656 * in the timestamp of all entries with the current time. The data
2657 * must have the commit() time rather than the time it was traced,
2658 * so that all entries in the primary buffer are in timestamp order.
2659 */
2660 timestamp = dtrace_gethrtime();
2661 saddr = (uintptr_t)src->dtb_tomax;
2662 slimit = saddr + src->dtb_offset;
2663 while (saddr < slimit) {
2664 size_t size;
2665 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2666
2667 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2668 saddr += sizeof (dtrace_epid_t);
2669 continue;
2670 }
2671 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2672 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2673
2674 ASSERT3U(saddr + size, <=, slimit);
2675 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2676 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2677
2678 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2679
2680 saddr += size;
2681 }
2682
2683 /*
2684 * Copy the buffer across. (Note that this is a
2685 * highly subobtimal bcopy(); in the unlikely event that this becomes
2686 * a serious performance issue, a high-performance DTrace-specific
2687 * bcopy() should obviously be invented.)
2688 */
2689 daddr = (uintptr_t)dest->dtb_tomax + offs;
2690 dlimit = daddr + src->dtb_offset;
2691 saddr = (uintptr_t)src->dtb_tomax;
2692
2693 /*
2694 * First, the aligned portion.
2695 */
2696 while (dlimit - daddr >= sizeof (uint64_t)) {
2697 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2698
2699 daddr += sizeof (uint64_t);
2700 saddr += sizeof (uint64_t);
2701 }
2702
2703 /*
2704 * Now any left-over bit...
2705 */
2706 while (dlimit - daddr)
2707 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2708
2709 /*
2710 * Finally, commit the reserved space in the destination buffer.
2711 */
2712 dest->dtb_offset = offs + src->dtb_offset;
2713
2714 out:
2715 /*
2716 * If we're lucky enough to be the only active CPU on this speculation
2717 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2718 */
2719 if (current == DTRACESPEC_ACTIVE ||
2720 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2721 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2722 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2723
2724 ASSERT(rval == DTRACESPEC_COMMITTING);
2725 }
2726
2727 src->dtb_offset = 0;
2728 src->dtb_xamot_drops += src->dtb_drops;
2729 src->dtb_drops = 0;
2730 }
2731
2732 /*
2733 * This routine discards an active speculation. If the specified speculation
2734 * is not in a valid state to perform a discard(), this routine will silently
2735 * do nothing. The state of the specified speculation is transitioned
2736 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2737 */
2738 static void
2739 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2740 dtrace_specid_t which)
2741 {
2742 dtrace_speculation_t *spec;
2743 dtrace_speculation_state_t current, new;
2744 dtrace_buffer_t *buf;
2745
2746 if (which == 0)
2747 return;
2748
2749 if (which > state->dts_nspeculations) {
2750 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2751 return;
2752 }
2753
2754 spec = &state->dts_speculations[which - 1];
2755 buf = &spec->dtsp_buffer[cpu];
2756
2757 do {
2758 current = spec->dtsp_state;
2759
2760 switch (current) {
2761 case DTRACESPEC_INACTIVE:
2762 case DTRACESPEC_COMMITTINGMANY:
2763 case DTRACESPEC_COMMITTING:
2764 case DTRACESPEC_DISCARDING:
2765 return;
2766
2767 case DTRACESPEC_ACTIVE:
2768 case DTRACESPEC_ACTIVEMANY:
2769 new = DTRACESPEC_DISCARDING;
2770 break;
2771
2772 case DTRACESPEC_ACTIVEONE:
2773 if (buf->dtb_offset != 0) {
2774 new = DTRACESPEC_INACTIVE;
2775 } else {
2776 new = DTRACESPEC_DISCARDING;
2777 }
2778 break;
2779
2780 default:
2781 ASSERT(0);
2782 }
2783 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2784 current, new) != current);
2785
2786 buf->dtb_offset = 0;
2787 buf->dtb_drops = 0;
2788 }
2789
2790 /*
2791 * Note: not called from probe context. This function is called
2792 * asynchronously from cross call context to clean any speculations that are
2793 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2794 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2795 * speculation.
2796 */
2797 static void
2798 dtrace_speculation_clean_here(dtrace_state_t *state)
2799 {
2800 dtrace_icookie_t cookie;
2801 processorid_t cpu = CPU->cpu_id;
2802 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2803 dtrace_specid_t i;
2804
2805 cookie = dtrace_interrupt_disable();
2806
2807 if (dest->dtb_tomax == NULL) {
2808 dtrace_interrupt_enable(cookie);
2809 return;
2810 }
2811
2812 for (i = 0; i < state->dts_nspeculations; i++) {
2813 dtrace_speculation_t *spec = &state->dts_speculations[i];
2814 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2815
2816 if (src->dtb_tomax == NULL)
2817 continue;
2818
2819 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2820 src->dtb_offset = 0;
2821 continue;
2822 }
2823
2824 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2825 continue;
2826
2827 if (src->dtb_offset == 0)
2828 continue;
2829
2830 dtrace_speculation_commit(state, cpu, i + 1);
2831 }
2832
2833 dtrace_interrupt_enable(cookie);
2834 }
2835
2836 /*
2837 * Note: not called from probe context. This function is called
2838 * asynchronously (and at a regular interval) to clean any speculations that
2839 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2840 * is work to be done, it cross calls all CPUs to perform that work;
2841 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2842 * INACTIVE state until they have been cleaned by all CPUs.
2843 */
2844 static void
2845 dtrace_speculation_clean(dtrace_state_t *state)
2846 {
2847 int work = 0, rv;
2848 dtrace_specid_t i;
2849
2850 for (i = 0; i < state->dts_nspeculations; i++) {
2851 dtrace_speculation_t *spec = &state->dts_speculations[i];
2852
2853 ASSERT(!spec->dtsp_cleaning);
2854
2855 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2856 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2857 continue;
2858
2859 work++;
2860 spec->dtsp_cleaning = 1;
2861 }
2862
2863 if (!work)
2864 return;
2865
2866 dtrace_xcall(DTRACE_CPUALL,
2867 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2868
2869 /*
2870 * We now know that all CPUs have committed or discarded their
2871 * speculation buffers, as appropriate. We can now set the state
2872 * to inactive.
2873 */
2874 for (i = 0; i < state->dts_nspeculations; i++) {
2875 dtrace_speculation_t *spec = &state->dts_speculations[i];
2876 dtrace_speculation_state_t current, new;
2877
2878 if (!spec->dtsp_cleaning)
2879 continue;
2880
2881 current = spec->dtsp_state;
2882 ASSERT(current == DTRACESPEC_DISCARDING ||
2883 current == DTRACESPEC_COMMITTINGMANY);
2884
2885 new = DTRACESPEC_INACTIVE;
2886
2887 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2888 ASSERT(rv == current);
2889 spec->dtsp_cleaning = 0;
2890 }
2891 }
2892
2893 /*
2894 * Called as part of a speculate() to get the speculative buffer associated
2895 * with a given speculation. Returns NULL if the specified speculation is not
2896 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2897 * the active CPU is not the specified CPU -- the speculation will be
2898 * atomically transitioned into the ACTIVEMANY state.
2899 */
2900 static dtrace_buffer_t *
2901 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2902 dtrace_specid_t which)
2903 {
2904 dtrace_speculation_t *spec;
2905 dtrace_speculation_state_t current, new;
2906 dtrace_buffer_t *buf;
2907
2908 if (which == 0)
2909 return (NULL);
2910
2911 if (which > state->dts_nspeculations) {
2912 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2913 return (NULL);
2914 }
2915
2916 spec = &state->dts_speculations[which - 1];
2917 buf = &spec->dtsp_buffer[cpuid];
2918
2919 do {
2920 current = spec->dtsp_state;
2921
2922 switch (current) {
2923 case DTRACESPEC_INACTIVE:
2924 case DTRACESPEC_COMMITTINGMANY:
2925 case DTRACESPEC_DISCARDING:
2926 return (NULL);
2927
2928 case DTRACESPEC_COMMITTING:
2929 ASSERT(buf->dtb_offset == 0);
2930 return (NULL);
2931
2932 case DTRACESPEC_ACTIVEONE:
2933 /*
2934 * This speculation is currently active on one CPU.
2935 * Check the offset in the buffer; if it's non-zero,
2936 * that CPU must be us (and we leave the state alone).
2937 * If it's zero, assume that we're starting on a new
2938 * CPU -- and change the state to indicate that the
2939 * speculation is active on more than one CPU.
2940 */
2941 if (buf->dtb_offset != 0)
2942 return (buf);
2943
2944 new = DTRACESPEC_ACTIVEMANY;
2945 break;
2946
2947 case DTRACESPEC_ACTIVEMANY:
2948 return (buf);
2949
2950 case DTRACESPEC_ACTIVE:
2951 new = DTRACESPEC_ACTIVEONE;
2952 break;
2953
2954 default:
2955 ASSERT(0);
2956 }
2957 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2958 current, new) != current);
2959
2960 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2961 return (buf);
2962 }
2963
2964 /*
2965 * Return a string. In the event that the user lacks the privilege to access
2966 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2967 * don't fail access checking.
2968 *
2969 * dtrace_dif_variable() uses this routine as a helper for various
2970 * builtin values such as 'execname' and 'probefunc.'
2971 */
2972 uintptr_t
2973 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2974 dtrace_mstate_t *mstate)
2975 {
2976 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2977 uintptr_t ret;
2978 size_t strsz;
2979
2980 /*
2981 * The easy case: this probe is allowed to read all of memory, so
2982 * we can just return this as a vanilla pointer.
2983 */
2984 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2985 return (addr);
2986
2987 /*
2988 * This is the tougher case: we copy the string in question from
2989 * kernel memory into scratch memory and return it that way: this
2990 * ensures that we won't trip up when access checking tests the
2991 * BYREF return value.
2992 */
2993 strsz = dtrace_strlen((char *)addr, size) + 1;
2994
2995 if (mstate->dtms_scratch_ptr + strsz >
2996 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2997 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2998 return (NULL);
2999 }
3000
3001 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3002 strsz);
3003 ret = mstate->dtms_scratch_ptr;
3004 mstate->dtms_scratch_ptr += strsz;
3005 return (ret);
3006 }
3007
3008 /*
3009 * This function implements the DIF emulator's variable lookups. The emulator
3010 * passes a reserved variable identifier and optional built-in array index.
3011 */
3012 static uint64_t
3013 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3014 uint64_t ndx)
3015 {
3016 /*
3017 * If we're accessing one of the uncached arguments, we'll turn this
3018 * into a reference in the args array.
3019 */
3020 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3021 ndx = v - DIF_VAR_ARG0;
3022 v = DIF_VAR_ARGS;
3023 }
3024
3025 switch (v) {
3026 case DIF_VAR_ARGS:
3027 if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
3028 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
3029 CPU_DTRACE_KPRIV;
3030 return (0);
3031 }
3032
3033 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3034 if (ndx >= sizeof (mstate->dtms_arg) /
3035 sizeof (mstate->dtms_arg[0])) {
3036 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3037 dtrace_provider_t *pv;
3038 uint64_t val;
3039
3040 pv = mstate->dtms_probe->dtpr_provider;
3041 if (pv->dtpv_pops.dtps_getargval != NULL)
3042 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3043 mstate->dtms_probe->dtpr_id,
3044 mstate->dtms_probe->dtpr_arg, ndx, aframes);
3045 else
3046 val = dtrace_getarg(ndx, aframes);
3047
3048 /*
3049 * This is regrettably required to keep the compiler
3050 * from tail-optimizing the call to dtrace_getarg().
3051 * The condition always evaluates to true, but the
3052 * compiler has no way of figuring that out a priori.
3053 * (None of this would be necessary if the compiler
3054 * could be relied upon to _always_ tail-optimize
3055 * the call to dtrace_getarg() -- but it can't.)
3056 */
3057 if (mstate->dtms_probe != NULL)
3058 return (val);
3059
3060 ASSERT(0);
3061 }
3062
3063 return (mstate->dtms_arg[ndx]);
3064
3065 case DIF_VAR_UREGS: {
3066 klwp_t *lwp;
3067
3068 if (!dtrace_priv_proc(state, mstate))
3069 return (0);
3070
3071 if ((lwp = curthread->t_lwp) == NULL) {
3072 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3073 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
3074 return (0);
3075 }
3076
3077 return (dtrace_getreg(lwp->lwp_regs, ndx));
3078 }
3079
3080 case DIF_VAR_VMREGS: {
3081 uint64_t rval;
3082
3083 if (!dtrace_priv_kernel(state))
3084 return (0);
3085
3086 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3087
3088 rval = dtrace_getvmreg(ndx,
3089 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
3090
3091 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3092
3093 return (rval);
3094 }
3095
3096 case DIF_VAR_CURTHREAD:
3097 if (!dtrace_priv_proc(state, mstate))
3098 return (0);
3099 return ((uint64_t)(uintptr_t)curthread);
3100
3101 case DIF_VAR_TIMESTAMP:
3102 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3103 mstate->dtms_timestamp = dtrace_gethrtime();
3104 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3105 }
3106 return (mstate->dtms_timestamp);
3107
3108 case DIF_VAR_VTIMESTAMP:
3109 ASSERT(dtrace_vtime_references != 0);
3110 return (curthread->t_dtrace_vtime);
3111
3112 case DIF_VAR_WALLTIMESTAMP:
3113 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3114 mstate->dtms_walltimestamp = dtrace_gethrestime();
3115 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3116 }
3117 return (mstate->dtms_walltimestamp);
3118
3119 case DIF_VAR_IPL:
3120 if (!dtrace_priv_kernel(state))
3121 return (0);
3122 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3123 mstate->dtms_ipl = dtrace_getipl();
3124 mstate->dtms_present |= DTRACE_MSTATE_IPL;
3125 }
3126 return (mstate->dtms_ipl);
3127
3128 case DIF_VAR_EPID:
3129 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3130 return (mstate->dtms_epid);
3131
3132 case DIF_VAR_ID:
3133 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3134 return (mstate->dtms_probe->dtpr_id);
3135
3136 case DIF_VAR_STACKDEPTH:
3137 if (!dtrace_priv_kernel(state))
3138 return (0);
3139 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3140 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3141
3142 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3143 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3144 }
3145 return (mstate->dtms_stackdepth);
3146
3147 case DIF_VAR_USTACKDEPTH:
3148 if (!dtrace_priv_proc(state, mstate))
3149 return (0);
3150 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3151 /*
3152 * See comment in DIF_VAR_PID.
3153 */
3154 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3155 CPU_ON_INTR(CPU)) {
3156 mstate->dtms_ustackdepth = 0;
3157 } else {
3158 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3159 mstate->dtms_ustackdepth =
3160 dtrace_getustackdepth();
3161 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3162 }
3163 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3164 }
3165 return (mstate->dtms_ustackdepth);
3166
3167 case DIF_VAR_CALLER:
3168 if (!dtrace_priv_kernel(state))
3169 return (0);
3170 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3171 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3172
3173 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3174 /*
3175 * If this is an unanchored probe, we are
3176 * required to go through the slow path:
3177 * dtrace_caller() only guarantees correct
3178 * results for anchored probes.
3179 */
3180 pc_t caller[2];
3181
3182 dtrace_getpcstack(caller, 2, aframes,
3183 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3184 mstate->dtms_caller = caller[1];
3185 } else if ((mstate->dtms_caller =
3186 dtrace_caller(aframes)) == -1) {
3187 /*
3188 * We have failed to do this the quick way;
3189 * we must resort to the slower approach of
3190 * calling dtrace_getpcstack().
3191 */
3192 pc_t caller;
3193
3194 dtrace_getpcstack(&caller, 1, aframes, NULL);
3195 mstate->dtms_caller = caller;
3196 }
3197
3198 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3199 }
3200 return (mstate->dtms_caller);
3201
3202 case DIF_VAR_UCALLER:
3203 if (!dtrace_priv_proc(state, mstate))
3204 return (0);
3205
3206 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3207 uint64_t ustack[3];
3208
3209 /*
3210 * dtrace_getupcstack() fills in the first uint64_t
3211 * with the current PID. The second uint64_t will
3212 * be the program counter at user-level. The third
3213 * uint64_t will contain the caller, which is what
3214 * we're after.
3215 */
3216 ustack[2] = NULL;
3217 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3218 dtrace_getupcstack(ustack, 3);
3219 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3220 mstate->dtms_ucaller = ustack[2];
3221 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3222 }
3223
3224 return (mstate->dtms_ucaller);
3225
3226 case DIF_VAR_PROBEPROV:
3227 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3228 return (dtrace_dif_varstr(
3229 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3230 state, mstate));
3231
3232 case DIF_VAR_PROBEMOD:
3233 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3234 return (dtrace_dif_varstr(
3235 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3236 state, mstate));
3237
3238 case DIF_VAR_PROBEFUNC:
3239 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3240 return (dtrace_dif_varstr(
3241 (uintptr_t)mstate->dtms_probe->dtpr_func,
3242 state, mstate));
3243
3244 case DIF_VAR_PROBENAME:
3245 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3246 return (dtrace_dif_varstr(
3247 (uintptr_t)mstate->dtms_probe->dtpr_name,
3248 state, mstate));
3249
3250 case DIF_VAR_PID:
3251 if (!dtrace_priv_proc(state, mstate))
3252 return (0);
3253
3254 /*
3255 * Note that we are assuming that an unanchored probe is
3256 * always due to a high-level interrupt. (And we're assuming
3257 * that there is only a single high level interrupt.)
3258 */
3259 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3260 return (pid0.pid_id);
3261
3262 /*
3263 * It is always safe to dereference one's own t_procp pointer:
3264 * it always points to a valid, allocated proc structure.
3265 * Further, it is always safe to dereference the p_pidp member
3266 * of one's own proc structure. (These are truisms becuase
3267 * threads and processes don't clean up their own state --
3268 * they leave that task to whomever reaps them.)
3269 */
3270 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3271
3272 case DIF_VAR_PPID:
3273 if (!dtrace_priv_proc(state, mstate))
3274 return (0);
3275
3276 /*
3277 * See comment in DIF_VAR_PID.
3278 */
3279 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3280 return (pid0.pid_id);
3281
3282 /*
3283 * It is always safe to dereference one's own t_procp pointer:
3284 * it always points to a valid, allocated proc structure.
3285 * (This is true because threads don't clean up their own
3286 * state -- they leave that task to whomever reaps them.)
3287 */
3288 return ((uint64_t)curthread->t_procp->p_ppid);
3289
3290 case DIF_VAR_TID:
3291 /*
3292 * See comment in DIF_VAR_PID.
3293 */
3294 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3295 return (0);
3296
3297 return ((uint64_t)curthread->t_tid);
3298
3299 case DIF_VAR_EXECNAME:
3300 if (!dtrace_priv_proc(state, mstate))
3301 return (0);
3302
3303 /*
3304 * See comment in DIF_VAR_PID.
3305 */
3306 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3307 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3308
3309 /*
3310 * It is always safe to dereference one's own t_procp pointer:
3311 * it always points to a valid, allocated proc structure.
3312 * (This is true because threads don't clean up their own
3313 * state -- they leave that task to whomever reaps them.)
3314 */
3315 return (dtrace_dif_varstr(
3316 (uintptr_t)curthread->t_procp->p_user.u_comm,
3317 state, mstate));
3318
3319 case DIF_VAR_ZONENAME:
3320 if (!dtrace_priv_proc(state, mstate))
3321 return (0);
3322
3323 /*
3324 * See comment in DIF_VAR_PID.
3325 */
3326 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3327 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3328
3329 /*
3330 * It is always safe to dereference one's own t_procp pointer:
3331 * it always points to a valid, allocated proc structure.
3332 * (This is true because threads don't clean up their own
3333 * state -- they leave that task to whomever reaps them.)
3334 */
3335 return (dtrace_dif_varstr(
3336 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3337 state, mstate));
3338
3339 case DIF_VAR_UID:
3340 if (!dtrace_priv_proc(state, mstate))
3341 return (0);
3342
3343 /*
3344 * See comment in DIF_VAR_PID.
3345 */
3346 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3347 return ((uint64_t)p0.p_cred->cr_uid);
3348
3349 /*
3350 * It is always safe to dereference one's own t_procp pointer:
3351 * it always points to a valid, allocated proc structure.
3352 * (This is true because threads don't clean up their own
3353 * state -- they leave that task to whomever reaps them.)
3354 *
3355 * Additionally, it is safe to dereference one's own process
3356 * credential, since this is never NULL after process birth.
3357 */
3358 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3359
3360 case DIF_VAR_GID:
3361 if (!dtrace_priv_proc(state, mstate))
3362 return (0);
3363
3364 /*
3365 * See comment in DIF_VAR_PID.
3366 */
3367 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3368 return ((uint64_t)p0.p_cred->cr_gid);
3369
3370 /*
3371 * It is always safe to dereference one's own t_procp pointer:
3372 * it always points to a valid, allocated proc structure.
3373 * (This is true because threads don't clean up their own
3374 * state -- they leave that task to whomever reaps them.)
3375 *
3376 * Additionally, it is safe to dereference one's own process
3377 * credential, since this is never NULL after process birth.
3378 */
3379 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3380
3381 case DIF_VAR_ERRNO: {
3382 klwp_t *lwp;
3383 if (!dtrace_priv_proc(state, mstate))
3384 return (0);
3385
3386 /*
3387 * See comment in DIF_VAR_PID.
3388 */
3389 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3390 return (0);
3391
3392 /*
3393 * It is always safe to dereference one's own t_lwp pointer in
3394 * the event that this pointer is non-NULL. (This is true
3395 * because threads and lwps don't clean up their own state --
3396 * they leave that task to whomever reaps them.)
3397 */
3398 if ((lwp = curthread->t_lwp) == NULL)
3399 return (0);
3400
3401 return ((uint64_t)lwp->lwp_errno);
3402 }
3403 default:
3404 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3405 return (0);
3406 }
3407 }
3408
3409
3410 typedef enum dtrace_json_state {
3411 DTRACE_JSON_REST = 1,
3412 DTRACE_JSON_OBJECT,
3413 DTRACE_JSON_STRING,
3414 DTRACE_JSON_STRING_ESCAPE,
3415 DTRACE_JSON_STRING_ESCAPE_UNICODE,
3416 DTRACE_JSON_COLON,
3417 DTRACE_JSON_COMMA,
3418 DTRACE_JSON_VALUE,
3419 DTRACE_JSON_IDENTIFIER,
3420 DTRACE_JSON_NUMBER,
3421 DTRACE_JSON_NUMBER_FRAC,
3422 DTRACE_JSON_NUMBER_EXP,
3423 DTRACE_JSON_COLLECT_OBJECT
3424 } dtrace_json_state_t;
3425
3426 /*
3427 * This function possesses just enough knowledge about JSON to extract a single
3428 * value from a JSON string and store it in the scratch buffer. It is able
3429 * to extract nested object values, and members of arrays by index.
3430 *
3431 * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3432 * be looked up as we descend into the object tree. e.g.
3433 *
3434 * foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3435 * with nelems = 5.
3436 *
3437 * The run time of this function must be bounded above by strsize to limit the
3438 * amount of work done in probe context. As such, it is implemented as a
3439 * simple state machine, reading one character at a time using safe loads
3440 * until we find the requested element, hit a parsing error or run off the
3441 * end of the object or string.
3442 *
3443 * As there is no way for a subroutine to return an error without interrupting
3444 * clause execution, we simply return NULL in the event of a missing key or any
3445 * other error condition. Each NULL return in this function is commented with
3446 * the error condition it represents -- parsing or otherwise.
3447 *
3448 * The set of states for the state machine closely matches the JSON
3449 * specification (http://json.org/). Briefly:
3450 *
3451 * DTRACE_JSON_REST:
3452 * Skip whitespace until we find either a top-level Object, moving
3453 * to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3454 *
3455 * DTRACE_JSON_OBJECT:
3456 * Locate the next key String in an Object. Sets a flag to denote
3457 * the next String as a key string and moves to DTRACE_JSON_STRING.
3458 *
3459 * DTRACE_JSON_COLON:
3460 * Skip whitespace until we find the colon that separates key Strings
3461 * from their values. Once found, move to DTRACE_JSON_VALUE.
3462 *
3463 * DTRACE_JSON_VALUE:
3464 * Detects the type of the next value (String, Number, Identifier, Object
3465 * or Array) and routes to the states that process that type. Here we also
3466 * deal with the element selector list if we are requested to traverse down
3467 * into the object tree.
3468 *
3469 * DTRACE_JSON_COMMA:
3470 * Skip whitespace until we find the comma that separates key-value pairs
3471 * in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3472 * (similarly DTRACE_JSON_VALUE). All following literal value processing
3473 * states return to this state at the end of their value, unless otherwise
3474 * noted.
3475 *
3476 * DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3477 * Processes a Number literal from the JSON, including any exponent
3478 * component that may be present. Numbers are returned as strings, which
3479 * may be passed to strtoll() if an integer is required.
3480 *
3481 * DTRACE_JSON_IDENTIFIER:
3482 * Processes a "true", "false" or "null" literal in the JSON.
3483 *
3484 * DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3485 * DTRACE_JSON_STRING_ESCAPE_UNICODE:
3486 * Processes a String literal from the JSON, whether the String denotes
3487 * a key, a value or part of a larger Object. Handles all escape sequences
3488 * present in the specification, including four-digit unicode characters,
3489 * but merely includes the escape sequence without converting it to the
3490 * actual escaped character. If the String is flagged as a key, we
3491 * move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3492 *
3493 * DTRACE_JSON_COLLECT_OBJECT:
3494 * This state collects an entire Object (or Array), correctly handling
3495 * embedded strings. If the full element selector list matches this nested
3496 * object, we return the Object in full as a string. If not, we use this
3497 * state to skip to the next value at this level and continue processing.
3498 *
3499 * NOTE: This function uses various macros from strtolctype.h to manipulate
3500 * digit values, etc -- these have all been checked to ensure they make
3501 * no additional function calls.
3502 */
3503 static char *
3504 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3505 char *dest)
3506 {
3507 dtrace_json_state_t state = DTRACE_JSON_REST;
3508 int64_t array_elem = INT64_MIN;
3509 int64_t array_pos = 0;
3510 uint8_t escape_unicount = 0;
3511 boolean_t string_is_key = B_FALSE;
3512 boolean_t collect_object = B_FALSE;
3513 boolean_t found_key = B_FALSE;
3514 boolean_t in_array = B_FALSE;
3515 uint32_t braces = 0, brackets = 0;
3516 char *elem = elemlist;
3517 char *dd = dest;
3518 uintptr_t cur;
3519
3520 for (cur = json; cur < json + size; cur++) {
3521 char cc = dtrace_load8(cur);
3522 if (cc == '\0')
3523 return (NULL);
3524
3525 switch (state) {
3526 case DTRACE_JSON_REST:
3527 if (isspace(cc))
3528 break;
3529
3530 if (cc == '{') {
3531 state = DTRACE_JSON_OBJECT;
3532 break;
3533 }
3534
3535 if (cc == '[') {
3536 in_array = B_TRUE;
3537 array_pos = 0;
3538 array_elem = dtrace_strtoll(elem, 10, size);
3539 found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3540 state = DTRACE_JSON_VALUE;
3541 break;
3542 }
3543
3544 /*
3545 * ERROR: expected to find a top-level object or array.
3546 */
3547 return (NULL);
3548 case DTRACE_JSON_OBJECT:
3549 if (isspace(cc))
3550 break;
3551
3552 if (cc == '"') {
3553 state = DTRACE_JSON_STRING;
3554 string_is_key = B_TRUE;
3555 break;
3556 }
3557
3558 /*
3559 * ERROR: either the object did not start with a key
3560 * string, or we've run off the end of the object
3561 * without finding the requested key.
3562 */
3563 return (NULL);
3564 case DTRACE_JSON_STRING:
3565 if (cc == '\\') {
3566 *dd++ = '\\';
3567 state = DTRACE_JSON_STRING_ESCAPE;
3568 break;
3569 }
3570
3571 if (cc == '"') {
3572 if (collect_object) {
3573 /*
3574 * We don't reset the dest here, as
3575 * the string is part of a larger
3576 * object being collected.
3577 */
3578 *dd++ = cc;
3579 collect_object = B_FALSE;
3580 state = DTRACE_JSON_COLLECT_OBJECT;
3581 break;
3582 }
3583 *dd = '\0';
3584 dd = dest; /* reset string buffer */
3585 if (string_is_key) {
3586 if (dtrace_strncmp(dest, elem,
3587 size) == 0)
3588 found_key = B_TRUE;
3589 } else if (found_key) {
3590 if (nelems > 1) {
3591 /*
3592 * We expected an object, not
3593 * this string.
3594 */
3595 return (NULL);
3596 }
3597 return (dest);
3598 }
3599 state = string_is_key ? DTRACE_JSON_COLON :
3600 DTRACE_JSON_COMMA;
3601 string_is_key = B_FALSE;
3602 break;
3603 }
3604
3605 *dd++ = cc;
3606 break;
3607 case DTRACE_JSON_STRING_ESCAPE:
3608 *dd++ = cc;
3609 if (cc == 'u') {
3610 escape_unicount = 0;
3611 state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3612 } else {
3613 state = DTRACE_JSON_STRING;
3614 }
3615 break;
3616 case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3617 if (!isxdigit(cc)) {
3618 /*
3619 * ERROR: invalid unicode escape, expected
3620 * four valid hexidecimal digits.
3621 */
3622 return (NULL);
3623 }
3624
3625 *dd++ = cc;
3626 if (++escape_unicount == 4)
3627 state = DTRACE_JSON_STRING;
3628 break;
3629 case DTRACE_JSON_COLON:
3630 if (isspace(cc))
3631 break;
3632
3633 if (cc == ':') {
3634 state = DTRACE_JSON_VALUE;
3635 break;
3636 }
3637
3638 /*
3639 * ERROR: expected a colon.
3640 */
3641 return (NULL);
3642 case DTRACE_JSON_COMMA:
3643 if (isspace(cc))
3644 break;
3645
3646 if (cc == ',') {
3647 if (in_array) {
3648 state = DTRACE_JSON_VALUE;
3649 if (++array_pos == array_elem)
3650 found_key = B_TRUE;
3651 } else {
3652 state = DTRACE_JSON_OBJECT;
3653 }
3654 break;
3655 }
3656
3657 /*
3658 * ERROR: either we hit an unexpected character, or
3659 * we reached the end of the object or array without
3660 * finding the requested key.
3661 */
3662 return (NULL);
3663 case DTRACE_JSON_IDENTIFIER:
3664 if (islower(cc)) {
3665 *dd++ = cc;
3666 break;
3667 }
3668
3669 *dd = '\0';
3670 dd = dest; /* reset string buffer */
3671
3672 if (dtrace_strncmp(dest, "true", 5) == 0 ||
3673 dtrace_strncmp(dest, "false", 6) == 0 ||
3674 dtrace_strncmp(dest, "null", 5) == 0) {
3675 if (found_key) {
3676 if (nelems > 1) {
3677 /*
3678 * ERROR: We expected an object,
3679 * not this identifier.
3680 */
3681 return (NULL);
3682 }
3683 return (dest);
3684 } else {
3685 cur--;
3686 state = DTRACE_JSON_COMMA;
3687 break;
3688 }
3689 }
3690
3691 /*
3692 * ERROR: we did not recognise the identifier as one
3693 * of those in the JSON specification.
3694 */
3695 return (NULL);
3696 case DTRACE_JSON_NUMBER:
3697 if (cc == '.') {
3698 *dd++ = cc;
3699 state = DTRACE_JSON_NUMBER_FRAC;
3700 break;
3701 }
3702
3703 if (cc == 'x' || cc == 'X') {
3704 /*
3705 * ERROR: specification explicitly excludes
3706 * hexidecimal or octal numbers.
3707 */
3708 return (NULL);
3709 }
3710
3711 /* FALLTHRU */
3712 case DTRACE_JSON_NUMBER_FRAC:
3713 if (cc == 'e' || cc == 'E') {
3714 *dd++ = cc;
3715 state = DTRACE_JSON_NUMBER_EXP;
3716 break;
3717 }
3718
3719 if (cc == '+' || cc == '-') {
3720 /*
3721 * ERROR: expect sign as part of exponent only.
3722 */
3723 return (NULL);
3724 }
3725 /* FALLTHRU */
3726 case DTRACE_JSON_NUMBER_EXP:
3727 if (isdigit(cc) || cc == '+' || cc == '-') {
3728 *dd++ = cc;
3729 break;
3730 }
3731
3732 *dd = '\0';
3733 dd = dest; /* reset string buffer */
3734 if (found_key) {
3735 if (nelems > 1) {
3736 /*
3737 * ERROR: We expected an object, not
3738 * this number.
3739 */
3740 return (NULL);
3741 }
3742 return (dest);
3743 }
3744
3745 cur--;
3746 state = DTRACE_JSON_COMMA;
3747 break;
3748 case DTRACE_JSON_VALUE:
3749 if (isspace(cc))
3750 break;
3751
3752 if (cc == '{' || cc == '[') {
3753 if (nelems > 1 && found_key) {
3754 in_array = cc == '[' ? B_TRUE : B_FALSE;
3755 /*
3756 * If our element selector directs us
3757 * to descend into this nested object,
3758 * then move to the next selector
3759 * element in the list and restart the
3760 * state machine.
3761 */
3762 while (*elem != '\0')
3763 elem++;
3764 elem++; /* skip the inter-element NUL */
3765 nelems--;
3766 dd = dest;
3767 if (in_array) {
3768 state = DTRACE_JSON_VALUE;
3769 array_pos = 0;
3770 array_elem = dtrace_strtoll(
3771 elem, 10, size);
3772 found_key = array_elem == 0 ?
3773 B_TRUE : B_FALSE;
3774 } else {
3775 found_key = B_FALSE;
3776 state = DTRACE_JSON_OBJECT;
3777 }
3778 break;
3779 }
3780
3781 /*
3782 * Otherwise, we wish to either skip this
3783 * nested object or return it in full.
3784 */
3785 if (cc == '[')
3786 brackets = 1;
3787 else
3788 braces = 1;
3789 *dd++ = cc;
3790 state = DTRACE_JSON_COLLECT_OBJECT;
3791 break;
3792 }
3793
3794 if (cc == '"') {
3795 state = DTRACE_JSON_STRING;
3796 break;
3797 }
3798
3799 if (islower(cc)) {
3800 /*
3801 * Here we deal with true, false and null.
3802 */
3803 *dd++ = cc;
3804 state = DTRACE_JSON_IDENTIFIER;
3805 break;
3806 }
3807
3808 if (cc == '-' || isdigit(cc)) {
3809 *dd++ = cc;
3810 state = DTRACE_JSON_NUMBER;
3811 break;
3812 }
3813
3814 /*
3815 * ERROR: unexpected character at start of value.
3816 */
3817 return (NULL);
3818 case DTRACE_JSON_COLLECT_OBJECT:
3819 if (cc == '\0')
3820 /*
3821 * ERROR: unexpected end of input.
3822 */
3823 return (NULL);
3824
3825 *dd++ = cc;
3826 if (cc == '"') {
3827 collect_object = B_TRUE;
3828 state = DTRACE_JSON_STRING;
3829 break;
3830 }
3831
3832 if (cc == ']') {
3833 if (brackets-- == 0) {
3834 /*
3835 * ERROR: unbalanced brackets.
3836 */
3837 return (NULL);
3838 }
3839 } else if (cc == '}') {
3840 if (braces-- == 0) {
3841 /*
3842 * ERROR: unbalanced braces.
3843 */
3844 return (NULL);
3845 }
3846 } else if (cc == '{') {
3847 braces++;
3848 } else if (cc == '[') {
3849 brackets++;
3850 }
3851
3852 if (brackets == 0 && braces == 0) {
3853 if (found_key) {
3854 *dd = '\0';
3855 return (dest);
3856 }
3857 dd = dest; /* reset string buffer */
3858 state = DTRACE_JSON_COMMA;
3859 }
3860 break;
3861 }
3862 }
3863 return (NULL);
3864 }
3865
3866 /*
3867 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3868 * Notice that we don't bother validating the proper number of arguments or
3869 * their types in the tuple stack. This isn't needed because all argument
3870 * interpretation is safe because of our load safety -- the worst that can
3871 * happen is that a bogus program can obtain bogus results.
3872 */
3873 static void
3874 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3875 dtrace_key_t *tupregs, int nargs,
3876 dtrace_mstate_t *mstate, dtrace_state_t *state)
3877 {
3878 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
3879 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
3880 dtrace_vstate_t *vstate = &state->dts_vstate;
3881
3882 union {
3883 mutex_impl_t mi;
3884 uint64_t mx;
3885 } m;
3886
3887 union {
3888 krwlock_t ri;
3889 uintptr_t rw;
3890 } r;
3891
3892 switch (subr) {
3893 case DIF_SUBR_RAND:
3894 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3895 break;
3896
3897 case DIF_SUBR_MUTEX_OWNED:
3898 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3899 mstate, vstate)) {
3900 regs[rd] = NULL;
3901 break;
3902 }
3903
3904 m.mx = dtrace_load64(tupregs[0].dttk_value);
3905 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3906 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3907 else
3908 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3909 break;
3910
3911 case DIF_SUBR_MUTEX_OWNER:
3912 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3913 mstate, vstate)) {
3914 regs[rd] = NULL;
3915 break;
3916 }
3917
3918 m.mx = dtrace_load64(tupregs[0].dttk_value);
3919 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3920 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3921 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3922 else
3923 regs[rd] = 0;
3924 break;
3925
3926 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3927 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3928 mstate, vstate)) {
3929 regs[rd] = NULL;
3930 break;
3931 }
3932
3933 m.mx = dtrace_load64(tupregs[0].dttk_value);
3934 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3935 break;
3936
3937 case DIF_SUBR_MUTEX_TYPE_SPIN:
3938 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3939 mstate, vstate)) {
3940 regs[rd] = NULL;
3941 break;
3942 }
3943
3944 m.mx = dtrace_load64(tupregs[0].dttk_value);
3945 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3946 break;
3947
3948 case DIF_SUBR_RW_READ_HELD: {
3949 uintptr_t tmp;
3950
3951 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3952 mstate, vstate)) {
3953 regs[rd] = NULL;
3954 break;
3955 }
3956
3957 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3958 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3959 break;
3960 }
3961
3962 case DIF_SUBR_RW_WRITE_HELD:
3963 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3964 mstate, vstate)) {
3965 regs[rd] = NULL;
3966 break;
3967 }
3968
3969 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3970 regs[rd] = _RW_WRITE_HELD(&r.ri);
3971 break;
3972
3973 case DIF_SUBR_RW_ISWRITER:
3974 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3975 mstate, vstate)) {
3976 regs[rd] = NULL;
3977 break;
3978 }
3979
3980 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3981 regs[rd] = _RW_ISWRITER(&r.ri);
3982 break;
3983
3984 case DIF_SUBR_BCOPY: {
3985 /*
3986 * We need to be sure that the destination is in the scratch
3987 * region -- no other region is allowed.
3988 */
3989 uintptr_t src = tupregs[0].dttk_value;
3990 uintptr_t dest = tupregs[1].dttk_value;
3991 size_t size = tupregs[2].dttk_value;
3992
3993 if (!dtrace_inscratch(dest, size, mstate)) {
3994 *flags |= CPU_DTRACE_BADADDR;
3995 *illval = regs[rd];
3996 break;
3997 }
3998
3999 if (!dtrace_canload(src, size, mstate, vstate)) {
4000 regs[rd] = NULL;
4001 break;
4002 }
4003
4004 dtrace_bcopy((void *)src, (void *)dest, size);
4005 break;
4006 }
4007
4008 case DIF_SUBR_ALLOCA:
4009 case DIF_SUBR_COPYIN: {
4010 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4011 uint64_t size =
4012 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
4013 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
4014
4015 /*
4016 * This action doesn't require any credential checks since
4017 * probes will not activate in user contexts to which the
4018 * enabling user does not have permissions.
4019 */
4020
4021 /*
4022 * Rounding up the user allocation size could have overflowed
4023 * a large, bogus allocation (like -1ULL) to 0.
4024 */
4025 if (scratch_size < size ||
4026 !DTRACE_INSCRATCH(mstate, scratch_size)) {
4027 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4028 regs[rd] = NULL;
4029 break;
4030 }
4031
4032 if (subr == DIF_SUBR_COPYIN) {
4033 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4034 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4035 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4036 }
4037
4038 mstate->dtms_scratch_ptr += scratch_size;
4039 regs[rd] = dest;
4040 break;
4041 }
4042
4043 case DIF_SUBR_COPYINTO: {
4044 uint64_t size = tupregs[1].dttk_value;
4045 uintptr_t dest = tupregs[2].dttk_value;
4046
4047 /*
4048 * This action doesn't require any credential checks since
4049 * probes will not activate in user contexts to which the
4050 * enabling user does not have permissions.
4051 */
4052 if (!dtrace_inscratch(dest, size, mstate)) {
4053 *flags |= CPU_DTRACE_BADADDR;
4054 *illval = regs[rd];
4055 break;
4056 }
4057
4058 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4059 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4060 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4061 break;
4062 }
4063
4064 case DIF_SUBR_COPYINSTR: {
4065 uintptr_t dest = mstate->dtms_scratch_ptr;
4066 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4067
4068 if (nargs > 1 && tupregs[1].dttk_value < size)
4069 size = tupregs[1].dttk_value + 1;
4070
4071 /*
4072 * This action doesn't require any credential checks since
4073 * probes will not activate in user contexts to which the
4074 * enabling user does not have permissions.
4075 */
4076 if (!DTRACE_INSCRATCH(mstate, size)) {
4077 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4078 regs[rd] = NULL;
4079 break;
4080 }
4081
4082 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4083 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4084 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4085
4086 ((char *)dest)[size - 1] = '\0';
4087 mstate->dtms_scratch_ptr += size;
4088 regs[rd] = dest;
4089 break;
4090 }
4091
4092 case DIF_SUBR_MSGSIZE:
4093 case DIF_SUBR_MSGDSIZE: {
4094 uintptr_t baddr = tupregs[0].dttk_value, daddr;
4095 uintptr_t wptr, rptr;
4096 size_t count = 0;
4097 int cont = 0;
4098
4099 while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4100
4101 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4102 vstate)) {
4103 regs[rd] = NULL;
4104 break;
4105 }
4106
4107 wptr = dtrace_loadptr(baddr +
4108 offsetof(mblk_t, b_wptr));
4109
4110 rptr = dtrace_loadptr(baddr +
4111 offsetof(mblk_t, b_rptr));
4112
4113 if (wptr < rptr) {
4114 *flags |= CPU_DTRACE_BADADDR;
4115 *illval = tupregs[0].dttk_value;
4116 break;
4117 }
4118
4119 daddr = dtrace_loadptr(baddr +
4120 offsetof(mblk_t, b_datap));
4121
4122 baddr = dtrace_loadptr(baddr +
4123 offsetof(mblk_t, b_cont));
4124
4125 /*
4126 * We want to prevent against denial-of-service here,
4127 * so we're only going to search the list for
4128 * dtrace_msgdsize_max mblks.
4129 */
4130 if (cont++ > dtrace_msgdsize_max) {
4131 *flags |= CPU_DTRACE_ILLOP;
4132 break;
4133 }
4134
4135 if (subr == DIF_SUBR_MSGDSIZE) {
4136 if (dtrace_load8(daddr +
4137 offsetof(dblk_t, db_type)) != M_DATA)
4138 continue;
4139 }
4140
4141 count += wptr - rptr;
4142 }
4143
4144 if (!(*flags & CPU_DTRACE_FAULT))
4145 regs[rd] = count;
4146
4147 break;
4148 }
4149
4150 case DIF_SUBR_PROGENYOF: {
4151 pid_t pid = tupregs[0].dttk_value;
4152 proc_t *p;
4153 int rval = 0;
4154
4155 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4156
4157 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4158 if (p->p_pidp->pid_id == pid) {
4159 rval = 1;
4160 break;
4161 }
4162 }
4163
4164 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4165
4166 regs[rd] = rval;
4167 break;
4168 }
4169
4170 case DIF_SUBR_SPECULATION:
4171 regs[rd] = dtrace_speculation(state);
4172 break;
4173
4174 case DIF_SUBR_COPYOUT: {
4175 uintptr_t kaddr = tupregs[0].dttk_value;
4176 uintptr_t uaddr = tupregs[1].dttk_value;
4177 uint64_t size = tupregs[2].dttk_value;
4178
4179 if (!dtrace_destructive_disallow &&
4180 dtrace_priv_proc_control(state, mstate) &&
4181 !dtrace_istoxic(kaddr, size)) {
4182 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4183 dtrace_copyout(kaddr, uaddr, size, flags);
4184 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4185 }
4186 break;
4187 }
4188
4189 case DIF_SUBR_COPYOUTSTR: {
4190 uintptr_t kaddr = tupregs[0].dttk_value;
4191 uintptr_t uaddr = tupregs[1].dttk_value;
4192 uint64_t size = tupregs[2].dttk_value;
4193
4194 if (!dtrace_destructive_disallow &&
4195 dtrace_priv_proc_control(state, mstate) &&
4196 !dtrace_istoxic(kaddr, size)) {
4197 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4198 dtrace_copyoutstr(kaddr, uaddr, size, flags);
4199 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4200 }
4201 break;
4202 }
4203
4204 case DIF_SUBR_STRLEN: {
4205 size_t sz;
4206 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4207 sz = dtrace_strlen((char *)addr,
4208 state->dts_options[DTRACEOPT_STRSIZE]);
4209
4210 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
4211 regs[rd] = NULL;
4212 break;
4213 }
4214
4215 regs[rd] = sz;
4216
4217 break;
4218 }
4219
4220 case DIF_SUBR_STRCHR:
4221 case DIF_SUBR_STRRCHR: {
4222 /*
4223 * We're going to iterate over the string looking for the
4224 * specified character. We will iterate until we have reached
4225 * the string length or we have found the character. If this
4226 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4227 * of the specified character instead of the first.
4228 */
4229 uintptr_t saddr = tupregs[0].dttk_value;
4230 uintptr_t addr = tupregs[0].dttk_value;
4231 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
4232 char c, target = (char)tupregs[1].dttk_value;
4233
4234 for (regs[rd] = NULL; addr < limit; addr++) {
4235 if ((c = dtrace_load8(addr)) == target) {
4236 regs[rd] = addr;
4237
4238 if (subr == DIF_SUBR_STRCHR)
4239 break;
4240 }
4241
4242 if (c == '\0')
4243 break;
4244 }
4245
4246 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
4247 regs[rd] = NULL;
4248 break;
4249 }
4250
4251 break;
4252 }
4253
4254 case DIF_SUBR_STRSTR:
4255 case DIF_SUBR_INDEX:
4256 case DIF_SUBR_RINDEX: {
4257 /*
4258 * We're going to iterate over the string looking for the
4259 * specified string. We will iterate until we have reached
4260 * the string length or we have found the string. (Yes, this
4261 * is done in the most naive way possible -- but considering
4262 * that the string we're searching for is likely to be
4263 * relatively short, the complexity of Rabin-Karp or similar
4264 * hardly seems merited.)
4265 */
4266 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4267 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4268 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4269 size_t len = dtrace_strlen(addr, size);
4270 size_t sublen = dtrace_strlen(substr, size);
4271 char *limit = addr + len, *orig = addr;
4272 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4273 int inc = 1;
4274
4275 regs[rd] = notfound;
4276
4277 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4278 regs[rd] = NULL;
4279 break;
4280 }
4281
4282 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4283 vstate)) {
4284 regs[rd] = NULL;
4285 break;
4286 }
4287
4288 /*
4289 * strstr() and index()/rindex() have similar semantics if
4290 * both strings are the empty string: strstr() returns a
4291 * pointer to the (empty) string, and index() and rindex()
4292 * both return index 0 (regardless of any position argument).
4293 */
4294 if (sublen == 0 && len == 0) {
4295 if (subr == DIF_SUBR_STRSTR)
4296 regs[rd] = (uintptr_t)addr;
4297 else
4298 regs[rd] = 0;
4299 break;
4300 }
4301
4302 if (subr != DIF_SUBR_STRSTR) {
4303 if (subr == DIF_SUBR_RINDEX) {
4304 limit = orig - 1;
4305 addr += len;
4306 inc = -1;
4307 }
4308
4309 /*
4310 * Both index() and rindex() take an optional position
4311 * argument that denotes the starting position.
4312 */
4313 if (nargs == 3) {
4314 int64_t pos = (int64_t)tupregs[2].dttk_value;
4315
4316 /*
4317 * If the position argument to index() is
4318 * negative, Perl implicitly clamps it at
4319 * zero. This semantic is a little surprising
4320 * given the special meaning of negative
4321 * positions to similar Perl functions like
4322 * substr(), but it appears to reflect a
4323 * notion that index() can start from a
4324 * negative index and increment its way up to
4325 * the string. Given this notion, Perl's
4326 * rindex() is at least self-consistent in
4327 * that it implicitly clamps positions greater
4328 * than the string length to be the string
4329 * length. Where Perl completely loses
4330 * coherence, however, is when the specified
4331 * substring is the empty string (""). In
4332 * this case, even if the position is
4333 * negative, rindex() returns 0 -- and even if
4334 * the position is greater than the length,
4335 * index() returns the string length. These
4336 * semantics violate the notion that index()
4337 * should never return a value less than the
4338 * specified position and that rindex() should
4339 * never return a value greater than the
4340 * specified position. (One assumes that
4341 * these semantics are artifacts of Perl's
4342 * implementation and not the results of
4343 * deliberate design -- it beggars belief that
4344 * even Larry Wall could desire such oddness.)
4345 * While in the abstract one would wish for
4346 * consistent position semantics across
4347 * substr(), index() and rindex() -- or at the
4348 * very least self-consistent position
4349 * semantics for index() and rindex() -- we
4350 * instead opt to keep with the extant Perl
4351 * semantics, in all their broken glory. (Do
4352 * we have more desire to maintain Perl's
4353 * semantics than Perl does? Probably.)
4354 */
4355 if (subr == DIF_SUBR_RINDEX) {
4356 if (pos < 0) {
4357 if (sublen == 0)
4358 regs[rd] = 0;
4359 break;
4360 }
4361
4362 if (pos > len)
4363 pos = len;
4364 } else {
4365 if (pos < 0)
4366 pos = 0;
4367
4368 if (pos >= len) {
4369 if (sublen == 0)
4370 regs[rd] = len;
4371 break;
4372 }
4373 }
4374
4375 addr = orig + pos;
4376 }
4377 }
4378
4379 for (regs[rd] = notfound; addr != limit; addr += inc) {
4380 if (dtrace_strncmp(addr, substr, sublen) == 0) {
4381 if (subr != DIF_SUBR_STRSTR) {
4382 /*
4383 * As D index() and rindex() are
4384 * modeled on Perl (and not on awk),
4385 * we return a zero-based (and not a
4386 * one-based) index. (For you Perl
4387 * weenies: no, we're not going to add
4388 * $[ -- and shouldn't you be at a con
4389 * or something?)
4390 */
4391 regs[rd] = (uintptr_t)(addr - orig);
4392 break;
4393 }
4394
4395 ASSERT(subr == DIF_SUBR_STRSTR);
4396 regs[rd] = (uintptr_t)addr;
4397 break;
4398 }
4399 }
4400
4401 break;
4402 }
4403
4404 case DIF_SUBR_STRTOK: {
4405 uintptr_t addr = tupregs[0].dttk_value;
4406 uintptr_t tokaddr = tupregs[1].dttk_value;
4407 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4408 uintptr_t limit, toklimit = tokaddr + size;
4409 uint8_t c, tokmap[32]; /* 256 / 8 */
4410 char *dest = (char *)mstate->dtms_scratch_ptr;
4411 int i;
4412
4413 /*
4414 * Check both the token buffer and (later) the input buffer,
4415 * since both could be non-scratch addresses.
4416 */
4417 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
4418 regs[rd] = NULL;
4419 break;
4420 }
4421
4422 if (!DTRACE_INSCRATCH(mstate, size)) {
4423 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4424 regs[rd] = NULL;
4425 break;
4426 }
4427
4428 if (addr == NULL) {
4429 /*
4430 * If the address specified is NULL, we use our saved
4431 * strtok pointer from the mstate. Note that this
4432 * means that the saved strtok pointer is _only_
4433 * valid within multiple enablings of the same probe --
4434 * it behaves like an implicit clause-local variable.
4435 */
4436 addr = mstate->dtms_strtok;
4437 } else {
4438 /*
4439 * If the user-specified address is non-NULL we must
4440 * access check it. This is the only time we have
4441 * a chance to do so, since this address may reside
4442 * in the string table of this clause-- future calls
4443 * (when we fetch addr from mstate->dtms_strtok)
4444 * would fail this access check.
4445 */
4446 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
4447 regs[rd] = NULL;
4448 break;
4449 }
4450 }
4451
4452 /*
4453 * First, zero the token map, and then process the token
4454 * string -- setting a bit in the map for every character
4455 * found in the token string.
4456 */
4457 for (i = 0; i < sizeof (tokmap); i++)
4458 tokmap[i] = 0;
4459
4460 for (; tokaddr < toklimit; tokaddr++) {
4461 if ((c = dtrace_load8(tokaddr)) == '\0')
4462 break;
4463
4464 ASSERT((c >> 3) < sizeof (tokmap));
4465 tokmap[c >> 3] |= (1 << (c & 0x7));
4466 }
4467
4468 for (limit = addr + size; addr < limit; addr++) {
4469 /*
4470 * We're looking for a character that is _not_ contained
4471 * in the token string.
4472 */
4473 if ((c = dtrace_load8(addr)) == '\0')
4474 break;
4475
4476 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4477 break;
4478 }
4479
4480 if (c == '\0') {
4481 /*
4482 * We reached the end of the string without finding
4483 * any character that was not in the token string.
4484 * We return NULL in this case, and we set the saved
4485 * address to NULL as well.
4486 */
4487 regs[rd] = NULL;
4488 mstate->dtms_strtok = NULL;
4489 break;
4490 }
4491
4492 /*
4493 * From here on, we're copying into the destination string.
4494 */
4495 for (i = 0; addr < limit && i < size - 1; addr++) {
4496 if ((c = dtrace_load8(addr)) == '\0')
4497 break;
4498
4499 if (tokmap[c >> 3] & (1 << (c & 0x7)))
4500 break;
4501
4502 ASSERT(i < size);
4503 dest[i++] = c;
4504 }
4505
4506 ASSERT(i < size);
4507 dest[i] = '\0';
4508 regs[rd] = (uintptr_t)dest;
4509 mstate->dtms_scratch_ptr += size;
4510 mstate->dtms_strtok = addr;
4511 break;
4512 }
4513
4514 case DIF_SUBR_SUBSTR: {
4515 uintptr_t s = tupregs[0].dttk_value;
4516 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4517 char *d = (char *)mstate->dtms_scratch_ptr;
4518 int64_t index = (int64_t)tupregs[1].dttk_value;
4519 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4520 size_t len = dtrace_strlen((char *)s, size);
4521 int64_t i;
4522
4523 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4524 regs[rd] = NULL;
4525 break;
4526 }
4527
4528 if (!DTRACE_INSCRATCH(mstate, size)) {
4529 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4530 regs[rd] = NULL;
4531 break;
4532 }
4533
4534 if (nargs <= 2)
4535 remaining = (int64_t)size;
4536
4537 if (index < 0) {
4538 index += len;
4539
4540 if (index < 0 && index + remaining > 0) {
4541 remaining += index;
4542 index = 0;
4543 }
4544 }
4545
4546 if (index >= len || index < 0) {
4547 remaining = 0;
4548 } else if (remaining < 0) {
4549 remaining += len - index;
4550 } else if (index + remaining > size) {
4551 remaining = size - index;
4552 }
4553
4554 for (i = 0; i < remaining; i++) {
4555 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4556 break;
4557 }
4558
4559 d[i] = '\0';
4560
4561 mstate->dtms_scratch_ptr += size;
4562 regs[rd] = (uintptr_t)d;
4563 break;
4564 }
4565
4566 case DIF_SUBR_JSON: {
4567 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4568 uintptr_t json = tupregs[0].dttk_value;
4569 size_t jsonlen = dtrace_strlen((char *)json, size);
4570 uintptr_t elem = tupregs[1].dttk_value;
4571 size_t elemlen = dtrace_strlen((char *)elem, size);
4572
4573 char *dest = (char *)mstate->dtms_scratch_ptr;
4574 char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
4575 char *ee = elemlist;
4576 int nelems = 1;
4577 uintptr_t cur;
4578
4579 if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
4580 !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
4581 regs[rd] = NULL;
4582 break;
4583 }
4584
4585 if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
4586 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4587 regs[rd] = NULL;
4588 break;
4589 }
4590
4591 /*
4592 * Read the element selector and split it up into a packed list
4593 * of strings.
4594 */
4595 for (cur = elem; cur < elem + elemlen; cur++) {
4596 char cc = dtrace_load8(cur);
4597
4598 if (cur == elem && cc == '[') {
4599 /*
4600 * If the first element selector key is
4601 * actually an array index then ignore the
4602 * bracket.
4603 */
4604 continue;
4605 }
4606
4607 if (cc == ']')
4608 continue;
4609
4610 if (cc == '.' || cc == '[') {
4611 nelems++;
4612 cc = '\0';
4613 }
4614
4615 *ee++ = cc;
4616 }
4617 *ee++ = '\0';
4618
4619 if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
4620 nelems, dest)) != NULL)
4621 mstate->dtms_scratch_ptr += jsonlen + 1;
4622 break;
4623 }
4624
4625 case DIF_SUBR_TOUPPER:
4626 case DIF_SUBR_TOLOWER: {
4627 uintptr_t s = tupregs[0].dttk_value;
4628 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4629 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4630 size_t len = dtrace_strlen((char *)s, size);
4631 char lower, upper, convert;
4632 int64_t i;
4633
4634 if (subr == DIF_SUBR_TOUPPER) {
4635 lower = 'a';
4636 upper = 'z';
4637 convert = 'A';
4638 } else {
4639 lower = 'A';
4640 upper = 'Z';
4641 convert = 'a';
4642 }
4643
4644 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4645 regs[rd] = NULL;
4646 break;
4647 }
4648
4649 if (!DTRACE_INSCRATCH(mstate, size)) {
4650 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4651 regs[rd] = NULL;
4652 break;
4653 }
4654
4655 for (i = 0; i < size - 1; i++) {
4656 if ((c = dtrace_load8(s + i)) == '\0')
4657 break;
4658
4659 if (c >= lower && c <= upper)
4660 c = convert + (c - lower);
4661
4662 dest[i] = c;
4663 }
4664
4665 ASSERT(i < size);
4666 dest[i] = '\0';
4667 regs[rd] = (uintptr_t)dest;
4668 mstate->dtms_scratch_ptr += size;
4669 break;
4670 }
4671
4672 case DIF_SUBR_GETMAJOR:
4673 #ifdef _LP64
4674 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4675 #else
4676 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4677 #endif
4678 break;
4679
4680 case DIF_SUBR_GETMINOR:
4681 #ifdef _LP64
4682 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4683 #else
4684 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4685 #endif
4686 break;
4687
4688 case DIF_SUBR_DDI_PATHNAME: {
4689 /*
4690 * This one is a galactic mess. We are going to roughly
4691 * emulate ddi_pathname(), but it's made more complicated
4692 * by the fact that we (a) want to include the minor name and
4693 * (b) must proceed iteratively instead of recursively.
4694 */
4695 uintptr_t dest = mstate->dtms_scratch_ptr;
4696 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4697 char *start = (char *)dest, *end = start + size - 1;
4698 uintptr_t daddr = tupregs[0].dttk_value;
4699 int64_t minor = (int64_t)tupregs[1].dttk_value;
4700 char *s;
4701 int i, len, depth = 0;
4702
4703 /*
4704 * Due to all the pointer jumping we do and context we must
4705 * rely upon, we just mandate that the user must have kernel
4706 * read privileges to use this routine.
4707 */
4708 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4709 *flags |= CPU_DTRACE_KPRIV;
4710 *illval = daddr;
4711 regs[rd] = NULL;
4712 }
4713
4714 if (!DTRACE_INSCRATCH(mstate, size)) {
4715 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4716 regs[rd] = NULL;
4717 break;
4718 }
4719
4720 *end = '\0';
4721
4722 /*
4723 * We want to have a name for the minor. In order to do this,
4724 * we need to walk the minor list from the devinfo. We want
4725 * to be sure that we don't infinitely walk a circular list,
4726 * so we check for circularity by sending a scout pointer
4727 * ahead two elements for every element that we iterate over;
4728 * if the list is circular, these will ultimately point to the
4729 * same element. You may recognize this little trick as the
4730 * answer to a stupid interview question -- one that always
4731 * seems to be asked by those who had to have it laboriously
4732 * explained to them, and who can't even concisely describe
4733 * the conditions under which one would be forced to resort to
4734 * this technique. Needless to say, those conditions are
4735 * found here -- and probably only here. Is this the only use
4736 * of this infamous trick in shipping, production code? If it
4737 * isn't, it probably should be...
4738 */
4739 if (minor != -1) {
4740 uintptr_t maddr = dtrace_loadptr(daddr +
4741 offsetof(struct dev_info, devi_minor));
4742
4743 uintptr_t next = offsetof(struct ddi_minor_data, next);
4744 uintptr_t name = offsetof(struct ddi_minor_data,
4745 d_minor) + offsetof(struct ddi_minor, name);
4746 uintptr_t dev = offsetof(struct ddi_minor_data,
4747 d_minor) + offsetof(struct ddi_minor, dev);
4748 uintptr_t scout;
4749
4750 if (maddr != NULL)
4751 scout = dtrace_loadptr(maddr + next);
4752
4753 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4754 uint64_t m;
4755 #ifdef _LP64
4756 m = dtrace_load64(maddr + dev) & MAXMIN64;
4757 #else
4758 m = dtrace_load32(maddr + dev) & MAXMIN;
4759 #endif
4760 if (m != minor) {
4761 maddr = dtrace_loadptr(maddr + next);
4762
4763 if (scout == NULL)
4764 continue;
4765
4766 scout = dtrace_loadptr(scout + next);
4767
4768 if (scout == NULL)
4769 continue;
4770
4771 scout = dtrace_loadptr(scout + next);
4772
4773 if (scout == NULL)
4774 continue;
4775
4776 if (scout == maddr) {
4777 *flags |= CPU_DTRACE_ILLOP;
4778 break;
4779 }
4780
4781 continue;
4782 }
4783
4784 /*
4785 * We have the minor data. Now we need to
4786 * copy the minor's name into the end of the
4787 * pathname.
4788 */
4789 s = (char *)dtrace_loadptr(maddr + name);
4790 len = dtrace_strlen(s, size);
4791
4792 if (*flags & CPU_DTRACE_FAULT)
4793 break;
4794
4795 if (len != 0) {
4796 if ((end -= (len + 1)) < start)
4797 break;
4798
4799 *end = ':';
4800 }
4801
4802 for (i = 1; i <= len; i++)
4803 end[i] = dtrace_load8((uintptr_t)s++);
4804 break;
4805 }
4806 }
4807
4808 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4809 ddi_node_state_t devi_state;
4810
4811 devi_state = dtrace_load32(daddr +
4812 offsetof(struct dev_info, devi_node_state));
4813
4814 if (*flags & CPU_DTRACE_FAULT)
4815 break;
4816
4817 if (devi_state >= DS_INITIALIZED) {
4818 s = (char *)dtrace_loadptr(daddr +
4819 offsetof(struct dev_info, devi_addr));
4820 len = dtrace_strlen(s, size);
4821
4822 if (*flags & CPU_DTRACE_FAULT)
4823 break;
4824
4825 if (len != 0) {
4826 if ((end -= (len + 1)) < start)
4827 break;
4828
4829 *end = '@';
4830 }
4831
4832 for (i = 1; i <= len; i++)
4833 end[i] = dtrace_load8((uintptr_t)s++);
4834 }
4835
4836 /*
4837 * Now for the node name...
4838 */
4839 s = (char *)dtrace_loadptr(daddr +
4840 offsetof(struct dev_info, devi_node_name));
4841
4842 daddr = dtrace_loadptr(daddr +
4843 offsetof(struct dev_info, devi_parent));
4844
4845 /*
4846 * If our parent is NULL (that is, if we're the root
4847 * node), we're going to use the special path
4848 * "devices".
4849 */
4850 if (daddr == NULL)
4851 s = "devices";
4852
4853 len = dtrace_strlen(s, size);
4854 if (*flags & CPU_DTRACE_FAULT)
4855 break;
4856
4857 if ((end -= (len + 1)) < start)
4858 break;
4859
4860 for (i = 1; i <= len; i++)
4861 end[i] = dtrace_load8((uintptr_t)s++);
4862 *end = '/';
4863
4864 if (depth++ > dtrace_devdepth_max) {
4865 *flags |= CPU_DTRACE_ILLOP;
4866 break;
4867 }
4868 }
4869
4870 if (end < start)
4871 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4872
4873 if (daddr == NULL) {
4874 regs[rd] = (uintptr_t)end;
4875 mstate->dtms_scratch_ptr += size;
4876 }
4877
4878 break;
4879 }
4880
4881 case DIF_SUBR_STRJOIN: {
4882 char *d = (char *)mstate->dtms_scratch_ptr;
4883 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4884 uintptr_t s1 = tupregs[0].dttk_value;
4885 uintptr_t s2 = tupregs[1].dttk_value;
4886 int i = 0;
4887
4888 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4889 !dtrace_strcanload(s2, size, mstate, vstate)) {
4890 regs[rd] = NULL;
4891 break;
4892 }
4893
4894 if (!DTRACE_INSCRATCH(mstate, size)) {
4895 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4896 regs[rd] = NULL;
4897 break;
4898 }
4899
4900 for (;;) {
4901 if (i >= size) {
4902 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4903 regs[rd] = NULL;
4904 break;
4905 }
4906
4907 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4908 i--;
4909 break;
4910 }
4911 }
4912
4913 for (;;) {
4914 if (i >= size) {
4915 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4916 regs[rd] = NULL;
4917 break;
4918 }
4919
4920 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4921 break;
4922 }
4923
4924 if (i < size) {
4925 mstate->dtms_scratch_ptr += i;
4926 regs[rd] = (uintptr_t)d;
4927 }
4928
4929 break;
4930 }
4931
4932 case DIF_SUBR_STRTOLL: {
4933 uintptr_t s = tupregs[0].dttk_value;
4934 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4935 int base = 10;
4936
4937 if (nargs > 1) {
4938 if ((base = tupregs[1].dttk_value) <= 1 ||
4939 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4940 *flags |= CPU_DTRACE_ILLOP;
4941 break;
4942 }
4943 }
4944
4945 if (!dtrace_strcanload(s, size, mstate, vstate)) {
4946 regs[rd] = INT64_MIN;
4947 break;
4948 }
4949
4950 regs[rd] = dtrace_strtoll((char *)s, base, size);
4951 break;
4952 }
4953
4954 case DIF_SUBR_LLTOSTR: {
4955 int64_t i = (int64_t)tupregs[0].dttk_value;
4956 uint64_t val, digit;
4957 uint64_t size = 65; /* enough room for 2^64 in binary */
4958 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4959 int base = 10;
4960
4961 if (nargs > 1) {
4962 if ((base = tupregs[1].dttk_value) <= 1 ||
4963 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4964 *flags |= CPU_DTRACE_ILLOP;
4965 break;
4966 }
4967 }
4968
4969 val = (base == 10 && i < 0) ? i * -1 : i;
4970
4971 if (!DTRACE_INSCRATCH(mstate, size)) {
4972 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4973 regs[rd] = NULL;
4974 break;
4975 }
4976
4977 for (*end-- = '\0'; val; val /= base) {
4978 if ((digit = val % base) <= '9' - '0') {
4979 *end-- = '0' + digit;
4980 } else {
4981 *end-- = 'a' + (digit - ('9' - '0') - 1);
4982 }
4983 }
4984
4985 if (i == 0 && base == 16)
4986 *end-- = '0';
4987
4988 if (base == 16)
4989 *end-- = 'x';
4990
4991 if (i == 0 || base == 8 || base == 16)
4992 *end-- = '0';
4993
4994 if (i < 0 && base == 10)
4995 *end-- = '-';
4996
4997 regs[rd] = (uintptr_t)end + 1;
4998 mstate->dtms_scratch_ptr += size;
4999 break;
5000 }
5001
5002 case DIF_SUBR_HTONS:
5003 case DIF_SUBR_NTOHS:
5004 #ifdef _BIG_ENDIAN
5005 regs[rd] = (uint16_t)tupregs[0].dttk_value;
5006 #else
5007 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
5008 #endif
5009 break;
5010
5011
5012 case DIF_SUBR_HTONL:
5013 case DIF_SUBR_NTOHL:
5014 #ifdef _BIG_ENDIAN
5015 regs[rd] = (uint32_t)tupregs[0].dttk_value;
5016 #else
5017 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5018 #endif
5019 break;
5020
5021
5022 case DIF_SUBR_HTONLL:
5023 case DIF_SUBR_NTOHLL:
5024 #ifdef _BIG_ENDIAN
5025 regs[rd] = (uint64_t)tupregs[0].dttk_value;
5026 #else
5027 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5028 #endif
5029 break;
5030
5031
5032 case DIF_SUBR_DIRNAME:
5033 case DIF_SUBR_BASENAME: {
5034 char *dest = (char *)mstate->dtms_scratch_ptr;
5035 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5036 uintptr_t src = tupregs[0].dttk_value;
5037 int i, j, len = dtrace_strlen((char *)src, size);
5038 int lastbase = -1, firstbase = -1, lastdir = -1;
5039 int start, end;
5040
5041 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5042 regs[rd] = NULL;
5043 break;
5044 }
5045
5046 if (!DTRACE_INSCRATCH(mstate, size)) {
5047 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5048 regs[rd] = NULL;
5049 break;
5050 }
5051
5052 /*
5053 * The basename and dirname for a zero-length string is
5054 * defined to be "."
5055 */
5056 if (len == 0) {
5057 len = 1;
5058 src = (uintptr_t)".";
5059 }
5060
5061 /*
5062 * Start from the back of the string, moving back toward the
5063 * front until we see a character that isn't a slash. That
5064 * character is the last character in the basename.
5065 */
5066 for (i = len - 1; i >= 0; i--) {
5067 if (dtrace_load8(src + i) != '/')
5068 break;
5069 }
5070
5071 if (i >= 0)
5072 lastbase = i;
5073
5074 /*
5075 * Starting from the last character in the basename, move
5076 * towards the front until we find a slash. The character
5077 * that we processed immediately before that is the first
5078 * character in the basename.
5079 */
5080 for (; i >= 0; i--) {
5081 if (dtrace_load8(src + i) == '/')
5082 break;
5083 }
5084
5085 if (i >= 0)
5086 firstbase = i + 1;
5087
5088 /*
5089 * Now keep going until we find a non-slash character. That
5090 * character is the last character in the dirname.
5091 */
5092 for (; i >= 0; i--) {
5093 if (dtrace_load8(src + i) != '/')
5094 break;
5095 }
5096
5097 if (i >= 0)
5098 lastdir = i;
5099
5100 ASSERT(!(lastbase == -1 && firstbase != -1));
5101 ASSERT(!(firstbase == -1 && lastdir != -1));
5102
5103 if (lastbase == -1) {
5104 /*
5105 * We didn't find a non-slash character. We know that
5106 * the length is non-zero, so the whole string must be
5107 * slashes. In either the dirname or the basename
5108 * case, we return '/'.
5109 */
5110 ASSERT(firstbase == -1);
5111 firstbase = lastbase = lastdir = 0;
5112 }
5113
5114 if (firstbase == -1) {
5115 /*
5116 * The entire string consists only of a basename
5117 * component. If we're looking for dirname, we need
5118 * to change our string to be just "."; if we're
5119 * looking for a basename, we'll just set the first
5120 * character of the basename to be 0.
5121 */
5122 if (subr == DIF_SUBR_DIRNAME) {
5123 ASSERT(lastdir == -1);
5124 src = (uintptr_t)".";
5125 lastdir = 0;
5126 } else {
5127 firstbase = 0;
5128 }
5129 }
5130
5131 if (subr == DIF_SUBR_DIRNAME) {
5132 if (lastdir == -1) {
5133 /*
5134 * We know that we have a slash in the name --
5135 * or lastdir would be set to 0, above. And
5136 * because lastdir is -1, we know that this
5137 * slash must be the first character. (That
5138 * is, the full string must be of the form
5139 * "/basename".) In this case, the last
5140 * character of the directory name is 0.
5141 */
5142 lastdir = 0;
5143 }
5144
5145 start = 0;
5146 end = lastdir;
5147 } else {
5148 ASSERT(subr == DIF_SUBR_BASENAME);
5149 ASSERT(firstbase != -1 && lastbase != -1);
5150 start = firstbase;
5151 end = lastbase;
5152 }
5153
5154 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5155 dest[j] = dtrace_load8(src + i);
5156
5157 dest[j] = '\0';
5158 regs[rd] = (uintptr_t)dest;
5159 mstate->dtms_scratch_ptr += size;
5160 break;
5161 }
5162
5163 case DIF_SUBR_GETF: {
5164 uintptr_t fd = tupregs[0].dttk_value;
5165 uf_info_t *finfo = &curthread->t_procp->p_user.u_finfo;
5166 file_t *fp;
5167
5168 if (!dtrace_priv_proc(state, mstate)) {
5169 regs[rd] = NULL;
5170 break;
5171 }
5172
5173 /*
5174 * This is safe because fi_nfiles only increases, and the
5175 * fi_list array is not freed when the array size doubles.
5176 * (See the comment in flist_grow() for details on the
5177 * management of the u_finfo structure.)
5178 */
5179 fp = fd < finfo->fi_nfiles ? finfo->fi_list[fd].uf_file : NULL;
5180
5181 mstate->dtms_getf = fp;
5182 regs[rd] = (uintptr_t)fp;
5183 break;
5184 }
5185
5186 case DIF_SUBR_CLEANPATH: {
5187 char *dest = (char *)mstate->dtms_scratch_ptr, c;
5188 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5189 uintptr_t src = tupregs[0].dttk_value;
5190 int i = 0, j = 0;
5191 zone_t *z;
5192
5193 if (!dtrace_strcanload(src, size, mstate, vstate)) {
5194 regs[rd] = NULL;
5195 break;
5196 }
5197
5198 if (!DTRACE_INSCRATCH(mstate, size)) {
5199 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5200 regs[rd] = NULL;
5201 break;
5202 }
5203
5204 /*
5205 * Move forward, loading each character.
5206 */
5207 do {
5208 c = dtrace_load8(src + i++);
5209 next:
5210 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
5211 break;
5212
5213 if (c != '/') {
5214 dest[j++] = c;
5215 continue;
5216 }
5217
5218 c = dtrace_load8(src + i++);
5219
5220 if (c == '/') {
5221 /*
5222 * We have two slashes -- we can just advance
5223 * to the next character.
5224 */
5225 goto next;
5226 }
5227
5228 if (c != '.') {
5229 /*
5230 * This is not "." and it's not ".." -- we can
5231 * just store the "/" and this character and
5232 * drive on.
5233 */
5234 dest[j++] = '/';
5235 dest[j++] = c;
5236 continue;
5237 }
5238
5239 c = dtrace_load8(src + i++);
5240
5241 if (c == '/') {
5242 /*
5243 * This is a "/./" component. We're not going
5244 * to store anything in the destination buffer;
5245 * we're just going to go to the next component.
5246 */
5247 goto next;
5248 }
5249
5250 if (c != '.') {
5251 /*
5252 * This is not ".." -- we can just store the
5253 * "/." and this character and continue
5254 * processing.
5255 */
5256 dest[j++] = '/';
5257 dest[j++] = '.';
5258 dest[j++] = c;
5259 continue;
5260 }
5261
5262 c = dtrace_load8(src + i++);
5263
5264 if (c != '/' && c != '\0') {
5265 /*
5266 * This is not ".." -- it's "..[mumble]".
5267 * We'll store the "/.." and this character
5268 * and continue processing.
5269 */
5270 dest[j++] = '/';
5271 dest[j++] = '.';
5272 dest[j++] = '.';
5273 dest[j++] = c;
5274 continue;
5275 }
5276
5277 /*
5278 * This is "/../" or "/..\0". We need to back up
5279 * our destination pointer until we find a "/".
5280 */
5281 i--;
5282 while (j != 0 && dest[--j] != '/')
5283 continue;
5284
5285 if (c == '\0')
5286 dest[++j] = '/';
5287 } while (c != '\0');
5288
5289 dest[j] = '\0';
5290
5291 if (mstate->dtms_getf != NULL &&
5292 !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5293 (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5294 /*
5295 * If we've done a getf() as a part of this ECB and we
5296 * don't have kernel access (and we're not in the global
5297 * zone), check if the path we cleaned up begins with
5298 * the zone's root path, and trim it off if so. Note
5299 * that this is an output cleanliness issue, not a
5300 * security issue: knowing one's zone root path does
5301 * not enable privilege escalation.
5302 */
5303 if (strstr(dest, z->zone_rootpath) == dest)
5304 dest += strlen(z->zone_rootpath) - 1;
5305 }
5306
5307 regs[rd] = (uintptr_t)dest;
5308 mstate->dtms_scratch_ptr += size;
5309 break;
5310 }
5311
5312 case DIF_SUBR_INET_NTOA:
5313 case DIF_SUBR_INET_NTOA6:
5314 case DIF_SUBR_INET_NTOP: {
5315 size_t size;
5316 int af, argi, i;
5317 char *base, *end;
5318
5319 if (subr == DIF_SUBR_INET_NTOP) {
5320 af = (int)tupregs[0].dttk_value;
5321 argi = 1;
5322 } else {
5323 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5324 argi = 0;
5325 }
5326
5327 if (af == AF_INET) {
5328 ipaddr_t ip4;
5329 uint8_t *ptr8, val;
5330
5331 /*
5332 * Safely load the IPv4 address.
5333 */
5334 ip4 = dtrace_load32(tupregs[argi].dttk_value);
5335
5336 /*
5337 * Check an IPv4 string will fit in scratch.
5338 */
5339 size = INET_ADDRSTRLEN;
5340 if (!DTRACE_INSCRATCH(mstate, size)) {
5341 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5342 regs[rd] = NULL;
5343 break;
5344 }
5345 base = (char *)mstate->dtms_scratch_ptr;
5346 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5347
5348 /*
5349 * Stringify as a dotted decimal quad.
5350 */
5351 *end-- = '\0';
5352 ptr8 = (uint8_t *)&ip4;
5353 for (i = 3; i >= 0; i--) {
5354 val = ptr8[i];
5355
5356 if (val == 0) {
5357 *end-- = '0';
5358 } else {
5359 for (; val; val /= 10) {
5360 *end-- = '0' + (val % 10);
5361 }
5362 }
5363
5364 if (i > 0)
5365 *end-- = '.';
5366 }
5367 ASSERT(end + 1 >= base);
5368
5369 } else if (af == AF_INET6) {
5370 struct in6_addr ip6;
5371 int firstzero, tryzero, numzero, v6end;
5372 uint16_t val;
5373 const char digits[] = "0123456789abcdef";
5374
5375 /*
5376 * Stringify using RFC 1884 convention 2 - 16 bit
5377 * hexadecimal values with a zero-run compression.
5378 * Lower case hexadecimal digits are used.
5379 * eg, fe80::214:4fff:fe0b:76c8.
5380 * The IPv4 embedded form is returned for inet_ntop,
5381 * just the IPv4 string is returned for inet_ntoa6.
5382 */
5383
5384 /*
5385 * Safely load the IPv6 address.
5386 */
5387 dtrace_bcopy(
5388 (void *)(uintptr_t)tupregs[argi].dttk_value,
5389 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5390
5391 /*
5392 * Check an IPv6 string will fit in scratch.
5393 */
5394 size = INET6_ADDRSTRLEN;
5395 if (!DTRACE_INSCRATCH(mstate, size)) {
5396 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5397 regs[rd] = NULL;
5398 break;
5399 }
5400 base = (char *)mstate->dtms_scratch_ptr;
5401 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5402 *end-- = '\0';
5403
5404 /*
5405 * Find the longest run of 16 bit zero values
5406 * for the single allowed zero compression - "::".
5407 */
5408 firstzero = -1;
5409 tryzero = -1;
5410 numzero = 1;
5411 for (i = 0; i < sizeof (struct in6_addr); i++) {
5412 if (ip6._S6_un._S6_u8[i] == 0 &&
5413 tryzero == -1 && i % 2 == 0) {
5414 tryzero = i;
5415 continue;
5416 }
5417
5418 if (tryzero != -1 &&
5419 (ip6._S6_un._S6_u8[i] != 0 ||
5420 i == sizeof (struct in6_addr) - 1)) {
5421
5422 if (i - tryzero <= numzero) {
5423 tryzero = -1;
5424 continue;
5425 }
5426
5427 firstzero = tryzero;
5428 numzero = i - i % 2 - tryzero;
5429 tryzero = -1;
5430
5431 if (ip6._S6_un._S6_u8[i] == 0 &&
5432 i == sizeof (struct in6_addr) - 1)
5433 numzero += 2;
5434 }
5435 }
5436 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5437
5438 /*
5439 * Check for an IPv4 embedded address.
5440 */
5441 v6end = sizeof (struct in6_addr) - 2;
5442 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5443 IN6_IS_ADDR_V4COMPAT(&ip6)) {
5444 for (i = sizeof (struct in6_addr) - 1;
5445 i >= DTRACE_V4MAPPED_OFFSET; i--) {
5446 ASSERT(end >= base);
5447
5448 val = ip6._S6_un._S6_u8[i];
5449
5450 if (val == 0) {
5451 *end-- = '0';
5452 } else {
5453 for (; val; val /= 10) {
5454 *end-- = '0' + val % 10;
5455 }
5456 }
5457
5458 if (i > DTRACE_V4MAPPED_OFFSET)
5459 *end-- = '.';
5460 }
5461
5462 if (subr == DIF_SUBR_INET_NTOA6)
5463 goto inetout;
5464
5465 /*
5466 * Set v6end to skip the IPv4 address that
5467 * we have already stringified.
5468 */
5469 v6end = 10;
5470 }
5471
5472 /*
5473 * Build the IPv6 string by working through the
5474 * address in reverse.
5475 */
5476 for (i = v6end; i >= 0; i -= 2) {
5477 ASSERT(end >= base);
5478
5479 if (i == firstzero + numzero - 2) {
5480 *end-- = ':';
5481 *end-- = ':';
5482 i -= numzero - 2;
5483 continue;
5484 }
5485
5486 if (i < 14 && i != firstzero - 2)
5487 *end-- = ':';
5488
5489 val = (ip6._S6_un._S6_u8[i] << 8) +
5490 ip6._S6_un._S6_u8[i + 1];
5491
5492 if (val == 0) {
5493 *end-- = '0';
5494 } else {
5495 for (; val; val /= 16) {
5496 *end-- = digits[val % 16];
5497 }
5498 }
5499 }
5500 ASSERT(end + 1 >= base);
5501
5502 } else {
5503 /*
5504 * The user didn't use AH_INET or AH_INET6.
5505 */
5506 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5507 regs[rd] = NULL;
5508 break;
5509 }
5510
5511 inetout: regs[rd] = (uintptr_t)end + 1;
5512 mstate->dtms_scratch_ptr += size;
5513 break;
5514 }
5515
5516 }
5517 }
5518
5519 /*
5520 * Emulate the execution of DTrace IR instructions specified by the given
5521 * DIF object. This function is deliberately void of assertions as all of
5522 * the necessary checks are handled by a call to dtrace_difo_validate().
5523 */
5524 static uint64_t
5525 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
5526 dtrace_vstate_t *vstate, dtrace_state_t *state)
5527 {
5528 const dif_instr_t *text = difo->dtdo_buf;
5529 const uint_t textlen = difo->dtdo_len;
5530 const char *strtab = difo->dtdo_strtab;
5531 const uint64_t *inttab = difo->dtdo_inttab;
5532
5533 uint64_t rval = 0;
5534 dtrace_statvar_t *svar;
5535 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5536 dtrace_difv_t *v;
5537 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5538 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
5539
5540 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5541 uint64_t regs[DIF_DIR_NREGS];
5542 uint64_t *tmp;
5543
5544 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5545 int64_t cc_r;
5546 uint_t pc = 0, id, opc;
5547 uint8_t ttop = 0;
5548 dif_instr_t instr;
5549 uint_t r1, r2, rd;
5550
5551 /*
5552 * We stash the current DIF object into the machine state: we need it
5553 * for subsequent access checking.
5554 */
5555 mstate->dtms_difo = difo;
5556
5557 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
5558
5559 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5560 opc = pc;
5561
5562 instr = text[pc++];
5563 r1 = DIF_INSTR_R1(instr);
5564 r2 = DIF_INSTR_R2(instr);
5565 rd = DIF_INSTR_RD(instr);
5566
5567 switch (DIF_INSTR_OP(instr)) {
5568 case DIF_OP_OR:
5569 regs[rd] = regs[r1] | regs[r2];
5570 break;
5571 case DIF_OP_XOR:
5572 regs[rd] = regs[r1] ^ regs[r2];
5573 break;
5574 case DIF_OP_AND:
5575 regs[rd] = regs[r1] & regs[r2];
5576 break;
5577 case DIF_OP_SLL:
5578 regs[rd] = regs[r1] << regs[r2];
5579 break;
5580 case DIF_OP_SRL:
5581 regs[rd] = regs[r1] >> regs[r2];
5582 break;
5583 case DIF_OP_SUB:
5584 regs[rd] = regs[r1] - regs[r2];
5585 break;
5586 case DIF_OP_ADD:
5587 regs[rd] = regs[r1] + regs[r2];
5588 break;
5589 case DIF_OP_MUL:
5590 regs[rd] = regs[r1] * regs[r2];
5591 break;
5592 case DIF_OP_SDIV:
5593 if (regs[r2] == 0) {
5594 regs[rd] = 0;
5595 *flags |= CPU_DTRACE_DIVZERO;
5596 } else {
5597 regs[rd] = (int64_t)regs[r1] /
5598 (int64_t)regs[r2];
5599 }
5600 break;
5601
5602 case DIF_OP_UDIV:
5603 if (regs[r2] == 0) {
5604 regs[rd] = 0;
5605 *flags |= CPU_DTRACE_DIVZERO;
5606 } else {
5607 regs[rd] = regs[r1] / regs[r2];
5608 }
5609 break;
5610
5611 case DIF_OP_SREM:
5612 if (regs[r2] == 0) {
5613 regs[rd] = 0;
5614 *flags |= CPU_DTRACE_DIVZERO;
5615 } else {
5616 regs[rd] = (int64_t)regs[r1] %
5617 (int64_t)regs[r2];
5618 }
5619 break;
5620
5621 case DIF_OP_UREM:
5622 if (regs[r2] == 0) {
5623 regs[rd] = 0;
5624 *flags |= CPU_DTRACE_DIVZERO;
5625 } else {
5626 regs[rd] = regs[r1] % regs[r2];
5627 }
5628 break;
5629
5630 case DIF_OP_NOT:
5631 regs[rd] = ~regs[r1];
5632 break;
5633 case DIF_OP_MOV:
5634 regs[rd] = regs[r1];
5635 break;
5636 case DIF_OP_CMP:
5637 cc_r = regs[r1] - regs[r2];
5638 cc_n = cc_r < 0;
5639 cc_z = cc_r == 0;
5640 cc_v = 0;
5641 cc_c = regs[r1] < regs[r2];
5642 break;
5643 case DIF_OP_TST:
5644 cc_n = cc_v = cc_c = 0;
5645 cc_z = regs[r1] == 0;
5646 break;
5647 case DIF_OP_BA:
5648 pc = DIF_INSTR_LABEL(instr);
5649 break;
5650 case DIF_OP_BE:
5651 if (cc_z)
5652 pc = DIF_INSTR_LABEL(instr);
5653 break;
5654 case DIF_OP_BNE:
5655 if (cc_z == 0)
5656 pc = DIF_INSTR_LABEL(instr);
5657 break;
5658 case DIF_OP_BG:
5659 if ((cc_z | (cc_n ^ cc_v)) == 0)
5660 pc = DIF_INSTR_LABEL(instr);
5661 break;
5662 case DIF_OP_BGU:
5663 if ((cc_c | cc_z) == 0)
5664 pc = DIF_INSTR_LABEL(instr);
5665 break;
5666 case DIF_OP_BGE:
5667 if ((cc_n ^ cc_v) == 0)
5668 pc = DIF_INSTR_LABEL(instr);
5669 break;
5670 case DIF_OP_BGEU:
5671 if (cc_c == 0)
5672 pc = DIF_INSTR_LABEL(instr);
5673 break;
5674 case DIF_OP_BL:
5675 if (cc_n ^ cc_v)
5676 pc = DIF_INSTR_LABEL(instr);
5677 break;
5678 case DIF_OP_BLU:
5679 if (cc_c)
5680 pc = DIF_INSTR_LABEL(instr);
5681 break;
5682 case DIF_OP_BLE:
5683 if (cc_z | (cc_n ^ cc_v))
5684 pc = DIF_INSTR_LABEL(instr);
5685 break;
5686 case DIF_OP_BLEU:
5687 if (cc_c | cc_z)
5688 pc = DIF_INSTR_LABEL(instr);
5689 break;
5690 case DIF_OP_RLDSB:
5691 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5692 break;
5693 /*FALLTHROUGH*/
5694 case DIF_OP_LDSB:
5695 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5696 break;
5697 case DIF_OP_RLDSH:
5698 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5699 break;
5700 /*FALLTHROUGH*/
5701 case DIF_OP_LDSH:
5702 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5703 break;
5704 case DIF_OP_RLDSW:
5705 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5706 break;
5707 /*FALLTHROUGH*/
5708 case DIF_OP_LDSW:
5709 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5710 break;
5711 case DIF_OP_RLDUB:
5712 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5713 break;
5714 /*FALLTHROUGH*/
5715 case DIF_OP_LDUB:
5716 regs[rd] = dtrace_load8(regs[r1]);
5717 break;
5718 case DIF_OP_RLDUH:
5719 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5720 break;
5721 /*FALLTHROUGH*/
5722 case DIF_OP_LDUH:
5723 regs[rd] = dtrace_load16(regs[r1]);
5724 break;
5725 case DIF_OP_RLDUW:
5726 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5727 break;
5728 /*FALLTHROUGH*/
5729 case DIF_OP_LDUW:
5730 regs[rd] = dtrace_load32(regs[r1]);
5731 break;
5732 case DIF_OP_RLDX:
5733 if (!dtrace_canload(regs[r1], 8, mstate, vstate))
5734 break;
5735 /*FALLTHROUGH*/
5736 case DIF_OP_LDX:
5737 regs[rd] = dtrace_load64(regs[r1]);
5738 break;
5739 case DIF_OP_ULDSB:
5740 regs[rd] = (int8_t)
5741 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5742 break;
5743 case DIF_OP_ULDSH:
5744 regs[rd] = (int16_t)
5745 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5746 break;
5747 case DIF_OP_ULDSW:
5748 regs[rd] = (int32_t)
5749 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5750 break;
5751 case DIF_OP_ULDUB:
5752 regs[rd] =
5753 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5754 break;
5755 case DIF_OP_ULDUH:
5756 regs[rd] =
5757 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5758 break;
5759 case DIF_OP_ULDUW:
5760 regs[rd] =
5761 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5762 break;
5763 case DIF_OP_ULDX:
5764 regs[rd] =
5765 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5766 break;
5767 case DIF_OP_RET:
5768 rval = regs[rd];
5769 pc = textlen;
5770 break;
5771 case DIF_OP_NOP:
5772 break;
5773 case DIF_OP_SETX:
5774 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5775 break;
5776 case DIF_OP_SETS:
5777 regs[rd] = (uint64_t)(uintptr_t)
5778 (strtab + DIF_INSTR_STRING(instr));
5779 break;
5780 case DIF_OP_SCMP: {
5781 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5782 uintptr_t s1 = regs[r1];
5783 uintptr_t s2 = regs[r2];
5784
5785 if (s1 != NULL &&
5786 !dtrace_strcanload(s1, sz, mstate, vstate))
5787 break;
5788 if (s2 != NULL &&
5789 !dtrace_strcanload(s2, sz, mstate, vstate))
5790 break;
5791
5792 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5793
5794 cc_n = cc_r < 0;
5795 cc_z = cc_r == 0;
5796 cc_v = cc_c = 0;
5797 break;
5798 }
5799 case DIF_OP_LDGA:
5800 regs[rd] = dtrace_dif_variable(mstate, state,
5801 r1, regs[r2]);
5802 break;
5803 case DIF_OP_LDGS:
5804 id = DIF_INSTR_VAR(instr);
5805
5806 if (id >= DIF_VAR_OTHER_UBASE) {
5807 uintptr_t a;
5808
5809 id -= DIF_VAR_OTHER_UBASE;
5810 svar = vstate->dtvs_globals[id];
5811 ASSERT(svar != NULL);
5812 v = &svar->dtsv_var;
5813
5814 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5815 regs[rd] = svar->dtsv_data;
5816 break;
5817 }
5818
5819 a = (uintptr_t)svar->dtsv_data;
5820
5821 if (*(uint8_t *)a == UINT8_MAX) {
5822 /*
5823 * If the 0th byte is set to UINT8_MAX
5824 * then this is to be treated as a
5825 * reference to a NULL variable.
5826 */
5827 regs[rd] = NULL;
5828 } else {
5829 regs[rd] = a + sizeof (uint64_t);
5830 }
5831
5832 break;
5833 }
5834
5835 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5836 break;
5837
5838 case DIF_OP_STGS:
5839 id = DIF_INSTR_VAR(instr);
5840
5841 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5842 id -= DIF_VAR_OTHER_UBASE;
5843
5844 svar = vstate->dtvs_globals[id];
5845 ASSERT(svar != NULL);
5846 v = &svar->dtsv_var;
5847
5848 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5849 uintptr_t a = (uintptr_t)svar->dtsv_data;
5850
5851 ASSERT(a != NULL);
5852 ASSERT(svar->dtsv_size != 0);
5853
5854 if (regs[rd] == NULL) {
5855 *(uint8_t *)a = UINT8_MAX;
5856 break;
5857 } else {
5858 *(uint8_t *)a = 0;
5859 a += sizeof (uint64_t);
5860 }
5861 if (!dtrace_vcanload(
5862 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5863 mstate, vstate))
5864 break;
5865
5866 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5867 (void *)a, &v->dtdv_type);
5868 break;
5869 }
5870
5871 svar->dtsv_data = regs[rd];
5872 break;
5873
5874 case DIF_OP_LDTA:
5875 /*
5876 * There are no DTrace built-in thread-local arrays at
5877 * present. This opcode is saved for future work.
5878 */
5879 *flags |= CPU_DTRACE_ILLOP;
5880 regs[rd] = 0;
5881 break;
5882
5883 case DIF_OP_LDLS:
5884 id = DIF_INSTR_VAR(instr);
5885
5886 if (id < DIF_VAR_OTHER_UBASE) {
5887 /*
5888 * For now, this has no meaning.
5889 */
5890 regs[rd] = 0;
5891 break;
5892 }
5893
5894 id -= DIF_VAR_OTHER_UBASE;
5895
5896 ASSERT(id < vstate->dtvs_nlocals);
5897 ASSERT(vstate->dtvs_locals != NULL);
5898
5899 svar = vstate->dtvs_locals[id];
5900 ASSERT(svar != NULL);
5901 v = &svar->dtsv_var;
5902
5903 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5904 uintptr_t a = (uintptr_t)svar->dtsv_data;
5905 size_t sz = v->dtdv_type.dtdt_size;
5906
5907 sz += sizeof (uint64_t);
5908 ASSERT(svar->dtsv_size == NCPU * sz);
5909 a += CPU->cpu_id * sz;
5910
5911 if (*(uint8_t *)a == UINT8_MAX) {
5912 /*
5913 * If the 0th byte is set to UINT8_MAX
5914 * then this is to be treated as a
5915 * reference to a NULL variable.
5916 */
5917 regs[rd] = NULL;
5918 } else {
5919 regs[rd] = a + sizeof (uint64_t);
5920 }
5921
5922 break;
5923 }
5924
5925 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5926 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5927 regs[rd] = tmp[CPU->cpu_id];
5928 break;
5929
5930 case DIF_OP_STLS:
5931 id = DIF_INSTR_VAR(instr);
5932
5933 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5934 id -= DIF_VAR_OTHER_UBASE;
5935 ASSERT(id < vstate->dtvs_nlocals);
5936
5937 ASSERT(vstate->dtvs_locals != NULL);
5938 svar = vstate->dtvs_locals[id];
5939 ASSERT(svar != NULL);
5940 v = &svar->dtsv_var;
5941
5942 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5943 uintptr_t a = (uintptr_t)svar->dtsv_data;
5944 size_t sz = v->dtdv_type.dtdt_size;
5945
5946 sz += sizeof (uint64_t);
5947 ASSERT(svar->dtsv_size == NCPU * sz);
5948 a += CPU->cpu_id * sz;
5949
5950 if (regs[rd] == NULL) {
5951 *(uint8_t *)a = UINT8_MAX;
5952 break;
5953 } else {
5954 *(uint8_t *)a = 0;
5955 a += sizeof (uint64_t);
5956 }
5957
5958 if (!dtrace_vcanload(
5959 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5960 mstate, vstate))
5961 break;
5962
5963 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5964 (void *)a, &v->dtdv_type);
5965 break;
5966 }
5967
5968 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5969 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5970 tmp[CPU->cpu_id] = regs[rd];
5971 break;
5972
5973 case DIF_OP_LDTS: {
5974 dtrace_dynvar_t *dvar;
5975 dtrace_key_t *key;
5976
5977 id = DIF_INSTR_VAR(instr);
5978 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5979 id -= DIF_VAR_OTHER_UBASE;
5980 v = &vstate->dtvs_tlocals[id];
5981
5982 key = &tupregs[DIF_DTR_NREGS];
5983 key[0].dttk_value = (uint64_t)id;
5984 key[0].dttk_size = 0;
5985 DTRACE_TLS_THRKEY(key[1].dttk_value);
5986 key[1].dttk_size = 0;
5987
5988 dvar = dtrace_dynvar(dstate, 2, key,
5989 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5990 mstate, vstate);
5991
5992 if (dvar == NULL) {
5993 regs[rd] = 0;
5994 break;
5995 }
5996
5997 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5998 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5999 } else {
6000 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6001 }
6002
6003 break;
6004 }
6005
6006 case DIF_OP_STTS: {
6007 dtrace_dynvar_t *dvar;
6008 dtrace_key_t *key;
6009
6010 id = DIF_INSTR_VAR(instr);
6011 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6012 id -= DIF_VAR_OTHER_UBASE;
6013
6014 key = &tupregs[DIF_DTR_NREGS];
6015 key[0].dttk_value = (uint64_t)id;
6016 key[0].dttk_size = 0;
6017 DTRACE_TLS_THRKEY(key[1].dttk_value);
6018 key[1].dttk_size = 0;
6019 v = &vstate->dtvs_tlocals[id];
6020
6021 dvar = dtrace_dynvar(dstate, 2, key,
6022 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6023 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6024 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6025 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6026
6027 /*
6028 * Given that we're storing to thread-local data,
6029 * we need to flush our predicate cache.
6030 */
6031 curthread->t_predcache = NULL;
6032
6033 if (dvar == NULL)
6034 break;
6035
6036 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6037 if (!dtrace_vcanload(
6038 (void *)(uintptr_t)regs[rd],
6039 &v->dtdv_type, mstate, vstate))
6040 break;
6041
6042 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6043 dvar->dtdv_data, &v->dtdv_type);
6044 } else {
6045 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6046 }
6047
6048 break;
6049 }
6050
6051 case DIF_OP_SRA:
6052 regs[rd] = (int64_t)regs[r1] >> regs[r2];
6053 break;
6054
6055 case DIF_OP_CALL:
6056 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6057 regs, tupregs, ttop, mstate, state);
6058 break;
6059
6060 case DIF_OP_PUSHTR:
6061 if (ttop == DIF_DTR_NREGS) {
6062 *flags |= CPU_DTRACE_TUPOFLOW;
6063 break;
6064 }
6065
6066 if (r1 == DIF_TYPE_STRING) {
6067 /*
6068 * If this is a string type and the size is 0,
6069 * we'll use the system-wide default string
6070 * size. Note that we are _not_ looking at
6071 * the value of the DTRACEOPT_STRSIZE option;
6072 * had this been set, we would expect to have
6073 * a non-zero size value in the "pushtr".
6074 */
6075 tupregs[ttop].dttk_size =
6076 dtrace_strlen((char *)(uintptr_t)regs[rd],
6077 regs[r2] ? regs[r2] :
6078 dtrace_strsize_default) + 1;
6079 } else {
6080 tupregs[ttop].dttk_size = regs[r2];
6081 }
6082
6083 tupregs[ttop++].dttk_value = regs[rd];
6084 break;
6085
6086 case DIF_OP_PUSHTV:
6087 if (ttop == DIF_DTR_NREGS) {
6088 *flags |= CPU_DTRACE_TUPOFLOW;
6089 break;
6090 }
6091
6092 tupregs[ttop].dttk_value = regs[rd];
6093 tupregs[ttop++].dttk_size = 0;
6094 break;
6095
6096 case DIF_OP_POPTS:
6097 if (ttop != 0)
6098 ttop--;
6099 break;
6100
6101 case DIF_OP_FLUSHTS:
6102 ttop = 0;
6103 break;
6104
6105 case DIF_OP_LDGAA:
6106 case DIF_OP_LDTAA: {
6107 dtrace_dynvar_t *dvar;
6108 dtrace_key_t *key = tupregs;
6109 uint_t nkeys = ttop;
6110
6111 id = DIF_INSTR_VAR(instr);
6112 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6113 id -= DIF_VAR_OTHER_UBASE;
6114
6115 key[nkeys].dttk_value = (uint64_t)id;
6116 key[nkeys++].dttk_size = 0;
6117
6118 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6119 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6120 key[nkeys++].dttk_size = 0;
6121 v = &vstate->dtvs_tlocals[id];
6122 } else {
6123 v = &vstate->dtvs_globals[id]->dtsv_var;
6124 }
6125
6126 dvar = dtrace_dynvar(dstate, nkeys, key,
6127 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6128 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6129 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6130
6131 if (dvar == NULL) {
6132 regs[rd] = 0;
6133 break;
6134 }
6135
6136 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6137 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6138 } else {
6139 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6140 }
6141
6142 break;
6143 }
6144
6145 case DIF_OP_STGAA:
6146 case DIF_OP_STTAA: {
6147 dtrace_dynvar_t *dvar;
6148 dtrace_key_t *key = tupregs;
6149 uint_t nkeys = ttop;
6150
6151 id = DIF_INSTR_VAR(instr);
6152 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6153 id -= DIF_VAR_OTHER_UBASE;
6154
6155 key[nkeys].dttk_value = (uint64_t)id;
6156 key[nkeys++].dttk_size = 0;
6157
6158 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6159 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6160 key[nkeys++].dttk_size = 0;
6161 v = &vstate->dtvs_tlocals[id];
6162 } else {
6163 v = &vstate->dtvs_globals[id]->dtsv_var;
6164 }
6165
6166 dvar = dtrace_dynvar(dstate, nkeys, key,
6167 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6168 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6169 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6170 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6171
6172 if (dvar == NULL)
6173 break;
6174
6175 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6176 if (!dtrace_vcanload(
6177 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6178 mstate, vstate))
6179 break;
6180
6181 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6182 dvar->dtdv_data, &v->dtdv_type);
6183 } else {
6184 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6185 }
6186
6187 break;
6188 }
6189
6190 case DIF_OP_ALLOCS: {
6191 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6192 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6193
6194 /*
6195 * Rounding up the user allocation size could have
6196 * overflowed large, bogus allocations (like -1ULL) to
6197 * 0.
6198 */
6199 if (size < regs[r1] ||
6200 !DTRACE_INSCRATCH(mstate, size)) {
6201 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6202 regs[rd] = NULL;
6203 break;
6204 }
6205
6206 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6207 mstate->dtms_scratch_ptr += size;
6208 regs[rd] = ptr;
6209 break;
6210 }
6211
6212 case DIF_OP_COPYS:
6213 if (!dtrace_canstore(regs[rd], regs[r2],
6214 mstate, vstate)) {
6215 *flags |= CPU_DTRACE_BADADDR;
6216 *illval = regs[rd];
6217 break;
6218 }
6219
6220 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6221 break;
6222
6223 dtrace_bcopy((void *)(uintptr_t)regs[r1],
6224 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6225 break;
6226
6227 case DIF_OP_STB:
6228 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6229 *flags |= CPU_DTRACE_BADADDR;
6230 *illval = regs[rd];
6231 break;
6232 }
6233 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6234 break;
6235
6236 case DIF_OP_STH:
6237 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6238 *flags |= CPU_DTRACE_BADADDR;
6239 *illval = regs[rd];
6240 break;
6241 }
6242 if (regs[rd] & 1) {
6243 *flags |= CPU_DTRACE_BADALIGN;
6244 *illval = regs[rd];
6245 break;
6246 }
6247 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6248 break;
6249
6250 case DIF_OP_STW:
6251 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6252 *flags |= CPU_DTRACE_BADADDR;
6253 *illval = regs[rd];
6254 break;
6255 }
6256 if (regs[rd] & 3) {
6257 *flags |= CPU_DTRACE_BADALIGN;
6258 *illval = regs[rd];
6259 break;
6260 }
6261 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6262 break;
6263
6264 case DIF_OP_STX:
6265 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6266 *flags |= CPU_DTRACE_BADADDR;
6267 *illval = regs[rd];
6268 break;
6269 }
6270 if (regs[rd] & 7) {
6271 *flags |= CPU_DTRACE_BADALIGN;
6272 *illval = regs[rd];
6273 break;
6274 }
6275 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6276 break;
6277 }
6278 }
6279
6280 if (!(*flags & CPU_DTRACE_FAULT))
6281 return (rval);
6282
6283 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6284 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6285
6286 return (0);
6287 }
6288
6289 static void
6290 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6291 {
6292 dtrace_probe_t *probe = ecb->dte_probe;
6293 dtrace_provider_t *prov = probe->dtpr_provider;
6294 char c[DTRACE_FULLNAMELEN + 80], *str;
6295 char *msg = "dtrace: breakpoint action at probe ";
6296 char *ecbmsg = " (ecb ";
6297 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6298 uintptr_t val = (uintptr_t)ecb;
6299 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6300
6301 if (dtrace_destructive_disallow)
6302 return;
6303
6304 /*
6305 * It's impossible to be taking action on the NULL probe.
6306 */
6307 ASSERT(probe != NULL);
6308
6309 /*
6310 * This is a poor man's (destitute man's?) sprintf(): we want to
6311 * print the provider name, module name, function name and name of
6312 * the probe, along with the hex address of the ECB with the breakpoint
6313 * action -- all of which we must place in the character buffer by
6314 * hand.
6315 */
6316 while (*msg != '\0')
6317 c[i++] = *msg++;
6318
6319 for (str = prov->dtpv_name; *str != '\0'; str++)
6320 c[i++] = *str;
6321 c[i++] = ':';
6322
6323 for (str = probe->dtpr_mod; *str != '\0'; str++)
6324 c[i++] = *str;
6325 c[i++] = ':';
6326
6327 for (str = probe->dtpr_func; *str != '\0'; str++)
6328 c[i++] = *str;
6329 c[i++] = ':';
6330
6331 for (str = probe->dtpr_name; *str != '\0'; str++)
6332 c[i++] = *str;
6333
6334 while (*ecbmsg != '\0')
6335 c[i++] = *ecbmsg++;
6336
6337 while (shift >= 0) {
6338 mask = (uintptr_t)0xf << shift;
6339
6340 if (val >= ((uintptr_t)1 << shift))
6341 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6342 shift -= 4;
6343 }
6344
6345 c[i++] = ')';
6346 c[i] = '\0';
6347
6348 debug_enter(c);
6349 }
6350
6351 static void
6352 dtrace_action_panic(dtrace_ecb_t *ecb)
6353 {
6354 dtrace_probe_t *probe = ecb->dte_probe;
6355
6356 /*
6357 * It's impossible to be taking action on the NULL probe.
6358 */
6359 ASSERT(probe != NULL);
6360
6361 if (dtrace_destructive_disallow)
6362 return;
6363
6364 if (dtrace_panicked != NULL)
6365 return;
6366
6367 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6368 return;
6369
6370 /*
6371 * We won the right to panic. (We want to be sure that only one
6372 * thread calls panic() from dtrace_probe(), and that panic() is
6373 * called exactly once.)
6374 */
6375 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6376 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6377 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6378 }
6379
6380 static void
6381 dtrace_action_raise(uint64_t sig)
6382 {
6383 if (dtrace_destructive_disallow)
6384 return;
6385
6386 if (sig >= NSIG) {
6387 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6388 return;
6389 }
6390
6391 /*
6392 * raise() has a queue depth of 1 -- we ignore all subsequent
6393 * invocations of the raise() action.
6394 */
6395 if (curthread->t_dtrace_sig == 0)
6396 curthread->t_dtrace_sig = (uint8_t)sig;
6397
6398 curthread->t_sig_check = 1;
6399 aston(curthread);
6400 }
6401
6402 static void
6403 dtrace_action_stop(void)
6404 {
6405 if (dtrace_destructive_disallow)
6406 return;
6407
6408 if (!curthread->t_dtrace_stop) {
6409 curthread->t_dtrace_stop = 1;
6410 curthread->t_sig_check = 1;
6411 aston(curthread);
6412 }
6413 }
6414
6415 static void
6416 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
6417 {
6418 hrtime_t now;
6419 volatile uint16_t *flags;
6420 cpu_t *cpu = CPU;
6421
6422 if (dtrace_destructive_disallow)
6423 return;
6424
6425 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
6426
6427 now = dtrace_gethrtime();
6428
6429 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
6430 /*
6431 * We need to advance the mark to the current time.
6432 */
6433 cpu->cpu_dtrace_chillmark = now;
6434 cpu->cpu_dtrace_chilled = 0;
6435 }
6436
6437 /*
6438 * Now check to see if the requested chill time would take us over
6439 * the maximum amount of time allowed in the chill interval. (Or
6440 * worse, if the calculation itself induces overflow.)
6441 */
6442 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
6443 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
6444 *flags |= CPU_DTRACE_ILLOP;
6445 return;
6446 }
6447
6448 while (dtrace_gethrtime() - now < val)
6449 continue;
6450
6451 /*
6452 * Normally, we assure that the value of the variable "timestamp" does
6453 * not change within an ECB. The presence of chill() represents an
6454 * exception to this rule, however.
6455 */
6456 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
6457 cpu->cpu_dtrace_chilled += val;
6458 }
6459
6460 static void
6461 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
6462 uint64_t *buf, uint64_t arg)
6463 {
6464 int nframes = DTRACE_USTACK_NFRAMES(arg);
6465 int strsize = DTRACE_USTACK_STRSIZE(arg);
6466 uint64_t *pcs = &buf[1], *fps;
6467 char *str = (char *)&pcs[nframes];
6468 int size, offs = 0, i, j;
6469 uintptr_t old = mstate->dtms_scratch_ptr, saved;
6470 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6471 char *sym;
6472
6473 /*
6474 * Should be taking a faster path if string space has not been
6475 * allocated.
6476 */
6477 ASSERT(strsize != 0);
6478
6479 /*
6480 * We will first allocate some temporary space for the frame pointers.
6481 */
6482 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6483 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
6484 (nframes * sizeof (uint64_t));
6485
6486 if (!DTRACE_INSCRATCH(mstate, size)) {
6487 /*
6488 * Not enough room for our frame pointers -- need to indicate
6489 * that we ran out of scratch space.
6490 */
6491 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6492 return;
6493 }
6494
6495 mstate->dtms_scratch_ptr += size;
6496 saved = mstate->dtms_scratch_ptr;
6497
6498 /*
6499 * Now get a stack with both program counters and frame pointers.
6500 */
6501 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6502 dtrace_getufpstack(buf, fps, nframes + 1);
6503 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6504
6505 /*
6506 * If that faulted, we're cooked.
6507 */
6508 if (*flags & CPU_DTRACE_FAULT)
6509 goto out;
6510
6511 /*
6512 * Now we want to walk up the stack, calling the USTACK helper. For
6513 * each iteration, we restore the scratch pointer.
6514 */
6515 for (i = 0; i < nframes; i++) {
6516 mstate->dtms_scratch_ptr = saved;
6517
6518 if (offs >= strsize)
6519 break;
6520
6521 sym = (char *)(uintptr_t)dtrace_helper(
6522 DTRACE_HELPER_ACTION_USTACK,
6523 mstate, state, pcs[i], fps[i]);
6524
6525 /*
6526 * If we faulted while running the helper, we're going to
6527 * clear the fault and null out the corresponding string.
6528 */
6529 if (*flags & CPU_DTRACE_FAULT) {
6530 *flags &= ~CPU_DTRACE_FAULT;
6531 str[offs++] = '\0';
6532 continue;
6533 }
6534
6535 if (sym == NULL) {
6536 str[offs++] = '\0';
6537 continue;
6538 }
6539
6540 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6541
6542 /*
6543 * Now copy in the string that the helper returned to us.
6544 */
6545 for (j = 0; offs + j < strsize; j++) {
6546 if ((str[offs + j] = sym[j]) == '\0')
6547 break;
6548 }
6549
6550 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6551
6552 offs += j + 1;
6553 }
6554
6555 if (offs >= strsize) {
6556 /*
6557 * If we didn't have room for all of the strings, we don't
6558 * abort processing -- this needn't be a fatal error -- but we
6559 * still want to increment a counter (dts_stkstroverflows) to
6560 * allow this condition to be warned about. (If this is from
6561 * a jstack() action, it is easily tuned via jstackstrsize.)
6562 */
6563 dtrace_error(&state->dts_stkstroverflows);
6564 }
6565
6566 while (offs < strsize)
6567 str[offs++] = '\0';
6568
6569 out:
6570 mstate->dtms_scratch_ptr = old;
6571 }
6572
6573 /*
6574 * If you're looking for the epicenter of DTrace, you just found it. This
6575 * is the function called by the provider to fire a probe -- from which all
6576 * subsequent probe-context DTrace activity emanates.
6577 */
6578 void
6579 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6580 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6581 {
6582 processorid_t cpuid;
6583 dtrace_icookie_t cookie;
6584 dtrace_probe_t *probe;
6585 dtrace_mstate_t mstate;
6586 dtrace_ecb_t *ecb;
6587 dtrace_action_t *act;
6588 intptr_t offs;
6589 size_t size;
6590 int vtime, onintr;
6591 volatile uint16_t *flags;
6592 hrtime_t now;
6593
6594 /*
6595 * Kick out immediately if this CPU is still being born (in which case
6596 * curthread will be set to -1) or the current thread can't allow
6597 * probes in its current context.
6598 */
6599 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6600 return;
6601
6602 cookie = dtrace_interrupt_disable();
6603 probe = dtrace_probes[id - 1];
6604 cpuid = CPU->cpu_id;
6605 onintr = CPU_ON_INTR(CPU);
6606
6607 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6608 probe->dtpr_predcache == curthread->t_predcache) {
6609 /*
6610 * We have hit in the predicate cache; we know that
6611 * this predicate would evaluate to be false.
6612 */
6613 dtrace_interrupt_enable(cookie);
6614 return;
6615 }
6616
6617 if (panic_quiesce) {
6618 /*
6619 * We don't trace anything if we're panicking.
6620 */
6621 dtrace_interrupt_enable(cookie);
6622 return;
6623 }
6624
6625 now = dtrace_gethrtime();
6626 vtime = dtrace_vtime_references != 0;
6627
6628 if (vtime && curthread->t_dtrace_start)
6629 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6630
6631 mstate.dtms_difo = NULL;
6632 mstate.dtms_probe = probe;
6633 mstate.dtms_strtok = NULL;
6634 mstate.dtms_arg[0] = arg0;
6635 mstate.dtms_arg[1] = arg1;
6636 mstate.dtms_arg[2] = arg2;
6637 mstate.dtms_arg[3] = arg3;
6638 mstate.dtms_arg[4] = arg4;
6639
6640 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6641
6642 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6643 dtrace_predicate_t *pred = ecb->dte_predicate;
6644 dtrace_state_t *state = ecb->dte_state;
6645 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6646 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6647 dtrace_vstate_t *vstate = &state->dts_vstate;
6648 dtrace_provider_t *prov = probe->dtpr_provider;
6649 uint64_t tracememsize = 0;
6650 int committed = 0;
6651 caddr_t tomax;
6652
6653 /*
6654 * A little subtlety with the following (seemingly innocuous)
6655 * declaration of the automatic 'val': by looking at the
6656 * code, you might think that it could be declared in the
6657 * action processing loop, below. (That is, it's only used in
6658 * the action processing loop.) However, it must be declared
6659 * out of that scope because in the case of DIF expression
6660 * arguments to aggregating actions, one iteration of the
6661 * action loop will use the last iteration's value.
6662 */
6663 #ifdef lint
6664 uint64_t val = 0;
6665 #else
6666 uint64_t val;
6667 #endif
6668
6669 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6670 mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
6671 mstate.dtms_getf = NULL;
6672
6673 *flags &= ~CPU_DTRACE_ERROR;
6674
6675 if (prov == dtrace_provider) {
6676 /*
6677 * If dtrace itself is the provider of this probe,
6678 * we're only going to continue processing the ECB if
6679 * arg0 (the dtrace_state_t) is equal to the ECB's
6680 * creating state. (This prevents disjoint consumers
6681 * from seeing one another's metaprobes.)
6682 */
6683 if (arg0 != (uint64_t)(uintptr_t)state)
6684 continue;
6685 }
6686
6687 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6688 /*
6689 * We're not currently active. If our provider isn't
6690 * the dtrace pseudo provider, we're not interested.
6691 */
6692 if (prov != dtrace_provider)
6693 continue;
6694
6695 /*
6696 * Now we must further check if we are in the BEGIN
6697 * probe. If we are, we will only continue processing
6698 * if we're still in WARMUP -- if one BEGIN enabling
6699 * has invoked the exit() action, we don't want to
6700 * evaluate subsequent BEGIN enablings.
6701 */
6702 if (probe->dtpr_id == dtrace_probeid_begin &&
6703 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6704 ASSERT(state->dts_activity ==
6705 DTRACE_ACTIVITY_DRAINING);
6706 continue;
6707 }
6708 }
6709
6710 if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
6711 continue;
6712
6713 if (now - state->dts_alive > dtrace_deadman_timeout) {
6714 /*
6715 * We seem to be dead. Unless we (a) have kernel
6716 * destructive permissions (b) have explicitly enabled
6717 * destructive actions and (c) destructive actions have
6718 * not been disabled, we're going to transition into
6719 * the KILLED state, from which no further processing
6720 * on this state will be performed.
6721 */
6722 if (!dtrace_priv_kernel_destructive(state) ||
6723 !state->dts_cred.dcr_destructive ||
6724 dtrace_destructive_disallow) {
6725 void *activity = &state->dts_activity;
6726 dtrace_activity_t current;
6727
6728 do {
6729 current = state->dts_activity;
6730 } while (dtrace_cas32(activity, current,
6731 DTRACE_ACTIVITY_KILLED) != current);
6732
6733 continue;
6734 }
6735 }
6736
6737 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6738 ecb->dte_alignment, state, &mstate)) < 0)
6739 continue;
6740
6741 tomax = buf->dtb_tomax;
6742 ASSERT(tomax != NULL);
6743
6744 if (ecb->dte_size != 0) {
6745 dtrace_rechdr_t dtrh;
6746 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6747 mstate.dtms_timestamp = dtrace_gethrtime();
6748 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6749 }
6750 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6751 dtrh.dtrh_epid = ecb->dte_epid;
6752 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6753 mstate.dtms_timestamp);
6754 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6755 }
6756
6757 mstate.dtms_epid = ecb->dte_epid;
6758 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6759
6760 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6761 mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
6762
6763 if (pred != NULL) {
6764 dtrace_difo_t *dp = pred->dtp_difo;
6765 int rval;
6766
6767 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6768
6769 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6770 dtrace_cacheid_t cid = probe->dtpr_predcache;
6771
6772 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6773 /*
6774 * Update the predicate cache...
6775 */
6776 ASSERT(cid == pred->dtp_cacheid);
6777 curthread->t_predcache = cid;
6778 }
6779
6780 continue;
6781 }
6782 }
6783
6784 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6785 act != NULL; act = act->dta_next) {
6786 size_t valoffs;
6787 dtrace_difo_t *dp;
6788 dtrace_recdesc_t *rec = &act->dta_rec;
6789
6790 size = rec->dtrd_size;
6791 valoffs = offs + rec->dtrd_offset;
6792
6793 if (DTRACEACT_ISAGG(act->dta_kind)) {
6794 uint64_t v = 0xbad;
6795 dtrace_aggregation_t *agg;
6796
6797 agg = (dtrace_aggregation_t *)act;
6798
6799 if ((dp = act->dta_difo) != NULL)
6800 v = dtrace_dif_emulate(dp,
6801 &mstate, vstate, state);
6802
6803 if (*flags & CPU_DTRACE_ERROR)
6804 continue;
6805
6806 /*
6807 * Note that we always pass the expression
6808 * value from the previous iteration of the
6809 * action loop. This value will only be used
6810 * if there is an expression argument to the
6811 * aggregating action, denoted by the
6812 * dtag_hasarg field.
6813 */
6814 dtrace_aggregate(agg, buf,
6815 offs, aggbuf, v, val);
6816 continue;
6817 }
6818
6819 switch (act->dta_kind) {
6820 case DTRACEACT_STOP:
6821 if (dtrace_priv_proc_destructive(state,
6822 &mstate))
6823 dtrace_action_stop();
6824 continue;
6825
6826 case DTRACEACT_BREAKPOINT:
6827 if (dtrace_priv_kernel_destructive(state))
6828 dtrace_action_breakpoint(ecb);
6829 continue;
6830
6831 case DTRACEACT_PANIC:
6832 if (dtrace_priv_kernel_destructive(state))
6833 dtrace_action_panic(ecb);
6834 continue;
6835
6836 case DTRACEACT_STACK:
6837 if (!dtrace_priv_kernel(state))
6838 continue;
6839
6840 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6841 size / sizeof (pc_t), probe->dtpr_aframes,
6842 DTRACE_ANCHORED(probe) ? NULL :
6843 (uint32_t *)arg0);
6844
6845 continue;
6846
6847 case DTRACEACT_JSTACK:
6848 case DTRACEACT_USTACK:
6849 if (!dtrace_priv_proc(state, &mstate))
6850 continue;
6851
6852 /*
6853 * See comment in DIF_VAR_PID.
6854 */
6855 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6856 CPU_ON_INTR(CPU)) {
6857 int depth = DTRACE_USTACK_NFRAMES(
6858 rec->dtrd_arg) + 1;
6859
6860 dtrace_bzero((void *)(tomax + valoffs),
6861 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6862 + depth * sizeof (uint64_t));
6863
6864 continue;
6865 }
6866
6867 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6868 curproc->p_dtrace_helpers != NULL) {
6869 /*
6870 * This is the slow path -- we have
6871 * allocated string space, and we're
6872 * getting the stack of a process that
6873 * has helpers. Call into a separate
6874 * routine to perform this processing.
6875 */
6876 dtrace_action_ustack(&mstate, state,
6877 (uint64_t *)(tomax + valoffs),
6878 rec->dtrd_arg);
6879 continue;
6880 }
6881
6882 /*
6883 * Clear the string space, since there's no
6884 * helper to do it for us.
6885 */
6886 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
6887 int depth = DTRACE_USTACK_NFRAMES(
6888 rec->dtrd_arg);
6889 size_t strsize = DTRACE_USTACK_STRSIZE(
6890 rec->dtrd_arg);
6891 uint64_t *buf = (uint64_t *)(tomax +
6892 valoffs);
6893 void *strspace = &buf[depth + 1];
6894
6895 dtrace_bzero(strspace,
6896 MIN(depth, strsize));
6897 }
6898
6899 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6900 dtrace_getupcstack((uint64_t *)
6901 (tomax + valoffs),
6902 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6903 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6904 continue;
6905
6906 default:
6907 break;
6908 }
6909
6910 dp = act->dta_difo;
6911 ASSERT(dp != NULL);
6912
6913 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6914
6915 if (*flags & CPU_DTRACE_ERROR)
6916 continue;
6917
6918 switch (act->dta_kind) {
6919 case DTRACEACT_SPECULATE: {
6920 dtrace_rechdr_t *dtrh;
6921
6922 ASSERT(buf == &state->dts_buffer[cpuid]);
6923 buf = dtrace_speculation_buffer(state,
6924 cpuid, val);
6925
6926 if (buf == NULL) {
6927 *flags |= CPU_DTRACE_DROP;
6928 continue;
6929 }
6930
6931 offs = dtrace_buffer_reserve(buf,
6932 ecb->dte_needed, ecb->dte_alignment,
6933 state, NULL);
6934
6935 if (offs < 0) {
6936 *flags |= CPU_DTRACE_DROP;
6937 continue;
6938 }
6939
6940 tomax = buf->dtb_tomax;
6941 ASSERT(tomax != NULL);
6942
6943 if (ecb->dte_size == 0)
6944 continue;
6945
6946 ASSERT3U(ecb->dte_size, >=,
6947 sizeof (dtrace_rechdr_t));
6948 dtrh = ((void *)(tomax + offs));
6949 dtrh->dtrh_epid = ecb->dte_epid;
6950 /*
6951 * When the speculation is committed, all of
6952 * the records in the speculative buffer will
6953 * have their timestamps set to the commit
6954 * time. Until then, it is set to a sentinel
6955 * value, for debugability.
6956 */
6957 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6958 continue;
6959 }
6960
6961 case DTRACEACT_CHILL:
6962 if (dtrace_priv_kernel_destructive(state))
6963 dtrace_action_chill(&mstate, val);
6964 continue;
6965
6966 case DTRACEACT_RAISE:
6967 if (dtrace_priv_proc_destructive(state,
6968 &mstate))
6969 dtrace_action_raise(val);
6970 continue;
6971
6972 case DTRACEACT_COMMIT:
6973 ASSERT(!committed);
6974
6975 /*
6976 * We need to commit our buffer state.
6977 */
6978 if (ecb->dte_size)
6979 buf->dtb_offset = offs + ecb->dte_size;
6980 buf = &state->dts_buffer[cpuid];
6981 dtrace_speculation_commit(state, cpuid, val);
6982 committed = 1;
6983 continue;
6984
6985 case DTRACEACT_DISCARD:
6986 dtrace_speculation_discard(state, cpuid, val);
6987 continue;
6988
6989 case DTRACEACT_DIFEXPR:
6990 case DTRACEACT_LIBACT:
6991 case DTRACEACT_PRINTF:
6992 case DTRACEACT_PRINTA:
6993 case DTRACEACT_SYSTEM:
6994 case DTRACEACT_FREOPEN:
6995 case DTRACEACT_TRACEMEM:
6996 break;
6997
6998 case DTRACEACT_TRACEMEM_DYNSIZE:
6999 tracememsize = val;
7000 break;
7001
7002 case DTRACEACT_SYM:
7003 case DTRACEACT_MOD:
7004 if (!dtrace_priv_kernel(state))
7005 continue;
7006 break;
7007
7008 case DTRACEACT_USYM:
7009 case DTRACEACT_UMOD:
7010 case DTRACEACT_UADDR: {
7011 struct pid *pid = curthread->t_procp->p_pidp;
7012
7013 if (!dtrace_priv_proc(state, &mstate))
7014 continue;
7015
7016 DTRACE_STORE(uint64_t, tomax,
7017 valoffs, (uint64_t)pid->pid_id);
7018 DTRACE_STORE(uint64_t, tomax,
7019 valoffs + sizeof (uint64_t), val);
7020
7021 continue;
7022 }
7023
7024 case DTRACEACT_EXIT: {
7025 /*
7026 * For the exit action, we are going to attempt
7027 * to atomically set our activity to be
7028 * draining. If this fails (either because
7029 * another CPU has beat us to the exit action,
7030 * or because our current activity is something
7031 * other than ACTIVE or WARMUP), we will
7032 * continue. This assures that the exit action
7033 * can be successfully recorded at most once
7034 * when we're in the ACTIVE state. If we're
7035 * encountering the exit() action while in
7036 * COOLDOWN, however, we want to honor the new
7037 * status code. (We know that we're the only
7038 * thread in COOLDOWN, so there is no race.)
7039 */
7040 void *activity = &state->dts_activity;
7041 dtrace_activity_t current = state->dts_activity;
7042
7043 if (current == DTRACE_ACTIVITY_COOLDOWN)
7044 break;
7045
7046 if (current != DTRACE_ACTIVITY_WARMUP)
7047 current = DTRACE_ACTIVITY_ACTIVE;
7048
7049 if (dtrace_cas32(activity, current,
7050 DTRACE_ACTIVITY_DRAINING) != current) {
7051 *flags |= CPU_DTRACE_DROP;
7052 continue;
7053 }
7054
7055 break;
7056 }
7057
7058 default:
7059 ASSERT(0);
7060 }
7061
7062 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
7063 uintptr_t end = valoffs + size;
7064
7065 if (tracememsize != 0 &&
7066 valoffs + tracememsize < end) {
7067 end = valoffs + tracememsize;
7068 tracememsize = 0;
7069 }
7070
7071 if (!dtrace_vcanload((void *)(uintptr_t)val,
7072 &dp->dtdo_rtype, &mstate, vstate))
7073 continue;
7074
7075 /*
7076 * If this is a string, we're going to only
7077 * load until we find the zero byte -- after
7078 * which we'll store zero bytes.
7079 */
7080 if (dp->dtdo_rtype.dtdt_kind ==
7081 DIF_TYPE_STRING) {
7082 char c = '\0' + 1;
7083 int intuple = act->dta_intuple;
7084 size_t s;
7085
7086 for (s = 0; s < size; s++) {
7087 if (c != '\0')
7088 c = dtrace_load8(val++);
7089
7090 DTRACE_STORE(uint8_t, tomax,
7091 valoffs++, c);
7092
7093 if (c == '\0' && intuple)
7094 break;
7095 }
7096
7097 continue;
7098 }
7099
7100 while (valoffs < end) {
7101 DTRACE_STORE(uint8_t, tomax, valoffs++,
7102 dtrace_load8(val++));
7103 }
7104
7105 continue;
7106 }
7107
7108 switch (size) {
7109 case 0:
7110 break;
7111
7112 case sizeof (uint8_t):
7113 DTRACE_STORE(uint8_t, tomax, valoffs, val);
7114 break;
7115 case sizeof (uint16_t):
7116 DTRACE_STORE(uint16_t, tomax, valoffs, val);
7117 break;
7118 case sizeof (uint32_t):
7119 DTRACE_STORE(uint32_t, tomax, valoffs, val);
7120 break;
7121 case sizeof (uint64_t):
7122 DTRACE_STORE(uint64_t, tomax, valoffs, val);
7123 break;
7124 default:
7125 /*
7126 * Any other size should have been returned by
7127 * reference, not by value.
7128 */
7129 ASSERT(0);
7130 break;
7131 }
7132 }
7133
7134 if (*flags & CPU_DTRACE_DROP)
7135 continue;
7136
7137 if (*flags & CPU_DTRACE_FAULT) {
7138 int ndx;
7139 dtrace_action_t *err;
7140
7141 buf->dtb_errors++;
7142
7143 if (probe->dtpr_id == dtrace_probeid_error) {
7144 /*
7145 * There's nothing we can do -- we had an
7146 * error on the error probe. We bump an
7147 * error counter to at least indicate that
7148 * this condition happened.
7149 */
7150 dtrace_error(&state->dts_dblerrors);
7151 continue;
7152 }
7153
7154 if (vtime) {
7155 /*
7156 * Before recursing on dtrace_probe(), we
7157 * need to explicitly clear out our start
7158 * time to prevent it from being accumulated
7159 * into t_dtrace_vtime.
7160 */
7161 curthread->t_dtrace_start = 0;
7162 }
7163
7164 /*
7165 * Iterate over the actions to figure out which action
7166 * we were processing when we experienced the error.
7167 * Note that act points _past_ the faulting action; if
7168 * act is ecb->dte_action, the fault was in the
7169 * predicate, if it's ecb->dte_action->dta_next it's
7170 * in action #1, and so on.
7171 */
7172 for (err = ecb->dte_action, ndx = 0;
7173 err != act; err = err->dta_next, ndx++)
7174 continue;
7175
7176 dtrace_probe_error(state, ecb->dte_epid, ndx,
7177 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7178 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7179 cpu_core[cpuid].cpuc_dtrace_illval);
7180
7181 continue;
7182 }
7183
7184 if (!committed)
7185 buf->dtb_offset = offs + ecb->dte_size;
7186 }
7187
7188 if (vtime)
7189 curthread->t_dtrace_start = dtrace_gethrtime();
7190
7191 dtrace_interrupt_enable(cookie);
7192 }
7193
7194 /*
7195 * DTrace Probe Hashing Functions
7196 *
7197 * The functions in this section (and indeed, the functions in remaining
7198 * sections) are not _called_ from probe context. (Any exceptions to this are
7199 * marked with a "Note:".) Rather, they are called from elsewhere in the
7200 * DTrace framework to look-up probes in, add probes to and remove probes from
7201 * the DTrace probe hashes. (Each probe is hashed by each element of the
7202 * probe tuple -- allowing for fast lookups, regardless of what was
7203 * specified.)
7204 */
7205 static uint_t
7206 dtrace_hash_str(char *p)
7207 {
7208 unsigned int g;
7209 uint_t hval = 0;
7210
7211 while (*p) {
7212 hval = (hval << 4) + *p++;
7213 if ((g = (hval & 0xf0000000)) != 0)
7214 hval ^= g >> 24;
7215 hval &= ~g;
7216 }
7217 return (hval);
7218 }
7219
7220 static dtrace_hash_t *
7221 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
7222 {
7223 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
7224
7225 hash->dth_stroffs = stroffs;
7226 hash->dth_nextoffs = nextoffs;
7227 hash->dth_prevoffs = prevoffs;
7228
7229 hash->dth_size = 1;
7230 hash->dth_mask = hash->dth_size - 1;
7231
7232 hash->dth_tab = kmem_zalloc(hash->dth_size *
7233 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
7234
7235 return (hash);
7236 }
7237
7238 static void
7239 dtrace_hash_destroy(dtrace_hash_t *hash)
7240 {
7241 #ifdef DEBUG
7242 int i;
7243
7244 for (i = 0; i < hash->dth_size; i++)
7245 ASSERT(hash->dth_tab[i] == NULL);
7246 #endif
7247
7248 kmem_free(hash->dth_tab,
7249 hash->dth_size * sizeof (dtrace_hashbucket_t *));
7250 kmem_free(hash, sizeof (dtrace_hash_t));
7251 }
7252
7253 static void
7254 dtrace_hash_resize(dtrace_hash_t *hash)
7255 {
7256 int size = hash->dth_size, i, ndx;
7257 int new_size = hash->dth_size << 1;
7258 int new_mask = new_size - 1;
7259 dtrace_hashbucket_t **new_tab, *bucket, *next;
7260
7261 ASSERT((new_size & new_mask) == 0);
7262
7263 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
7264
7265 for (i = 0; i < size; i++) {
7266 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
7267 dtrace_probe_t *probe = bucket->dthb_chain;
7268
7269 ASSERT(probe != NULL);
7270 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
7271
7272 next = bucket->dthb_next;
7273 bucket->dthb_next = new_tab[ndx];
7274 new_tab[ndx] = bucket;
7275 }
7276 }
7277
7278 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
7279 hash->dth_tab = new_tab;
7280 hash->dth_size = new_size;
7281 hash->dth_mask = new_mask;
7282 }
7283
7284 static void
7285 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
7286 {
7287 int hashval = DTRACE_HASHSTR(hash, new);
7288 int ndx = hashval & hash->dth_mask;
7289 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7290 dtrace_probe_t **nextp, **prevp;
7291
7292 for (; bucket != NULL; bucket = bucket->dthb_next) {
7293 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
7294 goto add;
7295 }
7296
7297 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
7298 dtrace_hash_resize(hash);
7299 dtrace_hash_add(hash, new);
7300 return;
7301 }
7302
7303 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
7304 bucket->dthb_next = hash->dth_tab[ndx];
7305 hash->dth_tab[ndx] = bucket;
7306 hash->dth_nbuckets++;
7307
7308 add:
7309 nextp = DTRACE_HASHNEXT(hash, new);
7310 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
7311 *nextp = bucket->dthb_chain;
7312
7313 if (bucket->dthb_chain != NULL) {
7314 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
7315 ASSERT(*prevp == NULL);
7316 *prevp = new;
7317 }
7318
7319 bucket->dthb_chain = new;
7320 bucket->dthb_len++;
7321 }
7322
7323 static dtrace_probe_t *
7324 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
7325 {
7326 int hashval = DTRACE_HASHSTR(hash, template);
7327 int ndx = hashval & hash->dth_mask;
7328 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7329
7330 for (; bucket != NULL; bucket = bucket->dthb_next) {
7331 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7332 return (bucket->dthb_chain);
7333 }
7334
7335 return (NULL);
7336 }
7337
7338 static int
7339 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
7340 {
7341 int hashval = DTRACE_HASHSTR(hash, template);
7342 int ndx = hashval & hash->dth_mask;
7343 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7344
7345 for (; bucket != NULL; bucket = bucket->dthb_next) {
7346 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7347 return (bucket->dthb_len);
7348 }
7349
7350 return (NULL);
7351 }
7352
7353 static void
7354 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7355 {
7356 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7357 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7358
7359 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7360 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7361
7362 /*
7363 * Find the bucket that we're removing this probe from.
7364 */
7365 for (; bucket != NULL; bucket = bucket->dthb_next) {
7366 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7367 break;
7368 }
7369
7370 ASSERT(bucket != NULL);
7371
7372 if (*prevp == NULL) {
7373 if (*nextp == NULL) {
7374 /*
7375 * The removed probe was the only probe on this
7376 * bucket; we need to remove the bucket.
7377 */
7378 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7379
7380 ASSERT(bucket->dthb_chain == probe);
7381 ASSERT(b != NULL);
7382
7383 if (b == bucket) {
7384 hash->dth_tab[ndx] = bucket->dthb_next;
7385 } else {
7386 while (b->dthb_next != bucket)
7387 b = b->dthb_next;
7388 b->dthb_next = bucket->dthb_next;
7389 }
7390
7391 ASSERT(hash->dth_nbuckets > 0);
7392 hash->dth_nbuckets--;
7393 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7394 return;
7395 }
7396
7397 bucket->dthb_chain = *nextp;
7398 } else {
7399 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7400 }
7401
7402 if (*nextp != NULL)
7403 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7404 }
7405
7406 /*
7407 * DTrace Utility Functions
7408 *
7409 * These are random utility functions that are _not_ called from probe context.
7410 */
7411 static int
7412 dtrace_badattr(const dtrace_attribute_t *a)
7413 {
7414 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7415 a->dtat_data > DTRACE_STABILITY_MAX ||
7416 a->dtat_class > DTRACE_CLASS_MAX);
7417 }
7418
7419 /*
7420 * Return a duplicate copy of a string. If the specified string is NULL,
7421 * this function returns a zero-length string.
7422 */
7423 static char *
7424 dtrace_strdup(const char *str)
7425 {
7426 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7427
7428 if (str != NULL)
7429 (void) strcpy(new, str);
7430
7431 return (new);
7432 }
7433
7434 #define DTRACE_ISALPHA(c) \
7435 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7436
7437 static int
7438 dtrace_badname(const char *s)
7439 {
7440 char c;
7441
7442 if (s == NULL || (c = *s++) == '\0')
7443 return (0);
7444
7445 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7446 return (1);
7447
7448 while ((c = *s++) != '\0') {
7449 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7450 c != '-' && c != '_' && c != '.' && c != '`')
7451 return (1);
7452 }
7453
7454 return (0);
7455 }
7456
7457 static void
7458 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7459 {
7460 uint32_t priv;
7461
7462 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7463 /*
7464 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7465 */
7466 priv = DTRACE_PRIV_ALL;
7467 } else {
7468 *uidp = crgetuid(cr);
7469 *zoneidp = crgetzonedid(cr);
7470
7471 priv = 0;
7472 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7473 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7474 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7475 priv |= DTRACE_PRIV_USER;
7476 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7477 priv |= DTRACE_PRIV_PROC;
7478 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7479 priv |= DTRACE_PRIV_OWNER;
7480 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7481 priv |= DTRACE_PRIV_ZONEOWNER;
7482 }
7483
7484 *privp = priv;
7485 }
7486
7487 #ifdef DTRACE_ERRDEBUG
7488 static void
7489 dtrace_errdebug(const char *str)
7490 {
7491 int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
7492 int occupied = 0;
7493
7494 mutex_enter(&dtrace_errlock);
7495 dtrace_errlast = str;
7496 dtrace_errthread = curthread;
7497
7498 while (occupied++ < DTRACE_ERRHASHSZ) {
7499 if (dtrace_errhash[hval].dter_msg == str) {
7500 dtrace_errhash[hval].dter_count++;
7501 goto out;
7502 }
7503
7504 if (dtrace_errhash[hval].dter_msg != NULL) {
7505 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7506 continue;
7507 }
7508
7509 dtrace_errhash[hval].dter_msg = str;
7510 dtrace_errhash[hval].dter_count = 1;
7511 goto out;
7512 }
7513
7514 panic("dtrace: undersized error hash");
7515 out:
7516 mutex_exit(&dtrace_errlock);
7517 }
7518 #endif
7519
7520 /*
7521 * DTrace Matching Functions
7522 *
7523 * These functions are used to match groups of probes, given some elements of
7524 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7525 */
7526 static int
7527 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7528 zoneid_t zoneid)
7529 {
7530 if (priv != DTRACE_PRIV_ALL) {
7531 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7532 uint32_t match = priv & ppriv;
7533
7534 /*
7535 * No PRIV_DTRACE_* privileges...
7536 */
7537 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7538 DTRACE_PRIV_KERNEL)) == 0)
7539 return (0);
7540
7541 /*
7542 * No matching bits, but there were bits to match...
7543 */
7544 if (match == 0 && ppriv != 0)
7545 return (0);
7546
7547 /*
7548 * Need to have permissions to the process, but don't...
7549 */
7550 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7551 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7552 return (0);
7553 }
7554
7555 /*
7556 * Need to be in the same zone unless we possess the
7557 * privilege to examine all zones.
7558 */
7559 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7560 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7561 return (0);
7562 }
7563 }
7564
7565 return (1);
7566 }
7567
7568 /*
7569 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7570 * consists of input pattern strings and an ops-vector to evaluate them.
7571 * This function returns >0 for match, 0 for no match, and <0 for error.
7572 */
7573 static int
7574 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7575 uint32_t priv, uid_t uid, zoneid_t zoneid)
7576 {
7577 dtrace_provider_t *pvp = prp->dtpr_provider;
7578 int rv;
7579
7580 if (pvp->dtpv_defunct)
7581 return (0);
7582
7583 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7584 return (rv);
7585
7586 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7587 return (rv);
7588
7589 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7590 return (rv);
7591
7592 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7593 return (rv);
7594
7595 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7596 return (0);
7597
7598 return (rv);
7599 }
7600
7601 /*
7602 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7603 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7604 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7605 * In addition, all of the recursion cases except for '*' matching have been
7606 * unwound. For '*', we still implement recursive evaluation, but a depth
7607 * counter is maintained and matching is aborted if we recurse too deep.
7608 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7609 */
7610 static int
7611 dtrace_match_glob(const char *s, const char *p, int depth)
7612 {
7613 const char *olds;
7614 char s1, c;
7615 int gs;
7616
7617 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7618 return (-1);
7619
7620 if (s == NULL)
7621 s = ""; /* treat NULL as empty string */
7622
7623 top:
7624 olds = s;
7625 s1 = *s++;
7626
7627 if (p == NULL)
7628 return (0);
7629
7630 if ((c = *p++) == '\0')
7631 return (s1 == '\0');
7632
7633 switch (c) {
7634 case '[': {
7635 int ok = 0, notflag = 0;
7636 char lc = '\0';
7637
7638 if (s1 == '\0')
7639 return (0);
7640
7641 if (*p == '!') {
7642 notflag = 1;
7643 p++;
7644 }
7645
7646 if ((c = *p++) == '\0')
7647 return (0);
7648
7649 do {
7650 if (c == '-' && lc != '\0' && *p != ']') {
7651 if ((c = *p++) == '\0')
7652 return (0);
7653 if (c == '\\' && (c = *p++) == '\0')
7654 return (0);
7655
7656 if (notflag) {
7657 if (s1 < lc || s1 > c)
7658 ok++;
7659 else
7660 return (0);
7661 } else if (lc <= s1 && s1 <= c)
7662 ok++;
7663
7664 } else if (c == '\\' && (c = *p++) == '\0')
7665 return (0);
7666
7667 lc = c; /* save left-hand 'c' for next iteration */
7668
7669 if (notflag) {
7670 if (s1 != c)
7671 ok++;
7672 else
7673 return (0);
7674 } else if (s1 == c)
7675 ok++;
7676
7677 if ((c = *p++) == '\0')
7678 return (0);
7679
7680 } while (c != ']');
7681
7682 if (ok)
7683 goto top;
7684
7685 return (0);
7686 }
7687
7688 case '\\':
7689 if ((c = *p++) == '\0')
7690 return (0);
7691 /*FALLTHRU*/
7692
7693 default:
7694 if (c != s1)
7695 return (0);
7696 /*FALLTHRU*/
7697
7698 case '?':
7699 if (s1 != '\0')
7700 goto top;
7701 return (0);
7702
7703 case '*':
7704 while (*p == '*')
7705 p++; /* consecutive *'s are identical to a single one */
7706
7707 if (*p == '\0')
7708 return (1);
7709
7710 for (s = olds; *s != '\0'; s++) {
7711 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7712 return (gs);
7713 }
7714
7715 return (0);
7716 }
7717 }
7718
7719 /*ARGSUSED*/
7720 static int
7721 dtrace_match_string(const char *s, const char *p, int depth)
7722 {
7723 return (s != NULL && strcmp(s, p) == 0);
7724 }
7725
7726 /*ARGSUSED*/
7727 static int
7728 dtrace_match_nul(const char *s, const char *p, int depth)
7729 {
7730 return (1); /* always match the empty pattern */
7731 }
7732
7733 /*ARGSUSED*/
7734 static int
7735 dtrace_match_nonzero(const char *s, const char *p, int depth)
7736 {
7737 return (s != NULL && s[0] != '\0');
7738 }
7739
7740 static int
7741 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7742 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7743 {
7744 dtrace_probe_t template, *probe;
7745 dtrace_hash_t *hash = NULL;
7746 int len, rc, best = INT_MAX, nmatched = 0;
7747 dtrace_id_t i;
7748
7749 ASSERT(MUTEX_HELD(&dtrace_lock));
7750
7751 /*
7752 * If the probe ID is specified in the key, just lookup by ID and
7753 * invoke the match callback once if a matching probe is found.
7754 */
7755 if (pkp->dtpk_id != DTRACE_IDNONE) {
7756 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7757 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7758 if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
7759 return (DTRACE_MATCH_FAIL);
7760 nmatched++;
7761 }
7762 return (nmatched);
7763 }
7764
7765 template.dtpr_mod = (char *)pkp->dtpk_mod;
7766 template.dtpr_func = (char *)pkp->dtpk_func;
7767 template.dtpr_name = (char *)pkp->dtpk_name;
7768
7769 /*
7770 * We want to find the most distinct of the module name, function
7771 * name, and name. So for each one that is not a glob pattern or
7772 * empty string, we perform a lookup in the corresponding hash and
7773 * use the hash table with the fewest collisions to do our search.
7774 */
7775 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7776 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7777 best = len;
7778 hash = dtrace_bymod;
7779 }
7780
7781 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7782 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7783 best = len;
7784 hash = dtrace_byfunc;
7785 }
7786
7787 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7788 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7789 best = len;
7790 hash = dtrace_byname;
7791 }
7792
7793 /*
7794 * If we did not select a hash table, iterate over every probe and
7795 * invoke our callback for each one that matches our input probe key.
7796 */
7797 if (hash == NULL) {
7798 for (i = 0; i < dtrace_nprobes; i++) {
7799 if ((probe = dtrace_probes[i]) == NULL ||
7800 dtrace_match_probe(probe, pkp, priv, uid,
7801 zoneid) <= 0)
7802 continue;
7803
7804 nmatched++;
7805
7806 if ((rc = (*matched)(probe, arg)) !=
7807 DTRACE_MATCH_NEXT) {
7808 if (rc == DTRACE_MATCH_FAIL)
7809 return (DTRACE_MATCH_FAIL);
7810 break;
7811 }
7812 }
7813
7814 return (nmatched);
7815 }
7816
7817 /*
7818 * If we selected a hash table, iterate over each probe of the same key
7819 * name and invoke the callback for every probe that matches the other
7820 * attributes of our input probe key.
7821 */
7822 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7823 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7824
7825 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7826 continue;
7827
7828 nmatched++;
7829
7830 if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
7831 if (rc == DTRACE_MATCH_FAIL)
7832 return (DTRACE_MATCH_FAIL);
7833 break;
7834 }
7835 }
7836
7837 return (nmatched);
7838 }
7839
7840 /*
7841 * Return the function pointer dtrace_probecmp() should use to compare the
7842 * specified pattern with a string. For NULL or empty patterns, we select
7843 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7844 * For non-empty non-glob strings, we use dtrace_match_string().
7845 */
7846 static dtrace_probekey_f *
7847 dtrace_probekey_func(const char *p)
7848 {
7849 char c;
7850
7851 if (p == NULL || *p == '\0')
7852 return (&dtrace_match_nul);
7853
7854 while ((c = *p++) != '\0') {
7855 if (c == '[' || c == '?' || c == '*' || c == '\\')
7856 return (&dtrace_match_glob);
7857 }
7858
7859 return (&dtrace_match_string);
7860 }
7861
7862 /*
7863 * Build a probe comparison key for use with dtrace_match_probe() from the
7864 * given probe description. By convention, a null key only matches anchored
7865 * probes: if each field is the empty string, reset dtpk_fmatch to
7866 * dtrace_match_nonzero().
7867 */
7868 static void
7869 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7870 {
7871 pkp->dtpk_prov = pdp->dtpd_provider;
7872 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7873
7874 pkp->dtpk_mod = pdp->dtpd_mod;
7875 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7876
7877 pkp->dtpk_func = pdp->dtpd_func;
7878 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7879
7880 pkp->dtpk_name = pdp->dtpd_name;
7881 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7882
7883 pkp->dtpk_id = pdp->dtpd_id;
7884
7885 if (pkp->dtpk_id == DTRACE_IDNONE &&
7886 pkp->dtpk_pmatch == &dtrace_match_nul &&
7887 pkp->dtpk_mmatch == &dtrace_match_nul &&
7888 pkp->dtpk_fmatch == &dtrace_match_nul &&
7889 pkp->dtpk_nmatch == &dtrace_match_nul)
7890 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7891 }
7892
7893 /*
7894 * DTrace Provider-to-Framework API Functions
7895 *
7896 * These functions implement much of the Provider-to-Framework API, as
7897 * described in <sys/dtrace.h>. The parts of the API not in this section are
7898 * the functions in the API for probe management (found below), and
7899 * dtrace_probe() itself (found above).
7900 */
7901
7902 /*
7903 * Register the calling provider with the DTrace framework. This should
7904 * generally be called by DTrace providers in their attach(9E) entry point.
7905 */
7906 int
7907 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7908 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7909 {
7910 dtrace_provider_t *provider;
7911
7912 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7913 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7914 "arguments", name ? name : "<NULL>");
7915 return (EINVAL);
7916 }
7917
7918 if (name[0] == '\0' || dtrace_badname(name)) {
7919 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7920 "provider name", name);
7921 return (EINVAL);
7922 }
7923
7924 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7925 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7926 pops->dtps_destroy == NULL ||
7927 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7928 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7929 "provider ops", name);
7930 return (EINVAL);
7931 }
7932
7933 if (dtrace_badattr(&pap->dtpa_provider) ||
7934 dtrace_badattr(&pap->dtpa_mod) ||
7935 dtrace_badattr(&pap->dtpa_func) ||
7936 dtrace_badattr(&pap->dtpa_name) ||
7937 dtrace_badattr(&pap->dtpa_args)) {
7938 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7939 "provider attributes", name);
7940 return (EINVAL);
7941 }
7942
7943 if (priv & ~DTRACE_PRIV_ALL) {
7944 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7945 "privilege attributes", name);
7946 return (EINVAL);
7947 }
7948
7949 if ((priv & DTRACE_PRIV_KERNEL) &&
7950 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7951 pops->dtps_mode == NULL) {
7952 cmn_err(CE_WARN, "failed to register provider '%s': need "
7953 "dtps_mode() op for given privilege attributes", name);
7954 return (EINVAL);
7955 }
7956
7957 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7958 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7959 (void) strcpy(provider->dtpv_name, name);
7960
7961 provider->dtpv_attr = *pap;
7962 provider->dtpv_priv.dtpp_flags = priv;
7963 if (cr != NULL) {
7964 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7965 provider->dtpv_priv.dtpp_zoneid = crgetzonedid(cr);
7966 }
7967 provider->dtpv_pops = *pops;
7968
7969 if (pops->dtps_provide == NULL) {
7970 ASSERT(pops->dtps_provide_module != NULL);
7971 provider->dtpv_pops.dtps_provide =
7972 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
7973 }
7974
7975 if (pops->dtps_provide_module == NULL) {
7976 ASSERT(pops->dtps_provide != NULL);
7977 provider->dtpv_pops.dtps_provide_module =
7978 (void (*)(void *, struct modctl *))dtrace_nullop;
7979 }
7980
7981 if (pops->dtps_suspend == NULL) {
7982 ASSERT(pops->dtps_resume == NULL);
7983 provider->dtpv_pops.dtps_suspend =
7984 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7985 provider->dtpv_pops.dtps_resume =
7986 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7987 }
7988
7989 provider->dtpv_arg = arg;
7990 *idp = (dtrace_provider_id_t)provider;
7991
7992 if (pops == &dtrace_provider_ops) {
7993 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7994 ASSERT(MUTEX_HELD(&dtrace_lock));
7995 ASSERT(dtrace_anon.dta_enabling == NULL);
7996
7997 /*
7998 * We make sure that the DTrace provider is at the head of
7999 * the provider chain.
8000 */
8001 provider->dtpv_next = dtrace_provider;
8002 dtrace_provider = provider;
8003 return (0);
8004 }
8005
8006 mutex_enter(&dtrace_provider_lock);
8007 mutex_enter(&dtrace_lock);
8008
8009 /*
8010 * If there is at least one provider registered, we'll add this
8011 * provider after the first provider.
8012 */
8013 if (dtrace_provider != NULL) {
8014 provider->dtpv_next = dtrace_provider->dtpv_next;
8015 dtrace_provider->dtpv_next = provider;
8016 } else {
8017 dtrace_provider = provider;
8018 }
8019
8020 if (dtrace_retained != NULL) {
8021 dtrace_enabling_provide(provider);
8022
8023 /*
8024 * Now we need to call dtrace_enabling_matchall() -- which
8025 * will acquire cpu_lock and dtrace_lock. We therefore need
8026 * to drop all of our locks before calling into it...
8027 */
8028 mutex_exit(&dtrace_lock);
8029 mutex_exit(&dtrace_provider_lock);
8030 dtrace_enabling_matchall();
8031
8032 return (0);
8033 }
8034
8035 mutex_exit(&dtrace_lock);
8036 mutex_exit(&dtrace_provider_lock);
8037
8038 return (0);
8039 }
8040
8041 /*
8042 * Unregister the specified provider from the DTrace framework. This should
8043 * generally be called by DTrace providers in their detach(9E) entry point.
8044 */
8045 int
8046 dtrace_unregister(dtrace_provider_id_t id)
8047 {
8048 dtrace_provider_t *old = (dtrace_provider_t *)id;
8049 dtrace_provider_t *prev = NULL;
8050 int i, self = 0, noreap = 0;
8051 dtrace_probe_t *probe, *first = NULL;
8052
8053 if (old->dtpv_pops.dtps_enable ==
8054 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
8055 /*
8056 * If DTrace itself is the provider, we're called with locks
8057 * already held.
8058 */
8059 ASSERT(old == dtrace_provider);
8060 ASSERT(dtrace_devi != NULL);
8061 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8062 ASSERT(MUTEX_HELD(&dtrace_lock));
8063 self = 1;
8064
8065 if (dtrace_provider->dtpv_next != NULL) {
8066 /*
8067 * There's another provider here; return failure.
8068 */
8069 return (EBUSY);
8070 }
8071 } else {
8072 mutex_enter(&dtrace_provider_lock);
8073 mutex_enter(&mod_lock);
8074 mutex_enter(&dtrace_lock);
8075 }
8076
8077 /*
8078 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8079 * probes, we refuse to let providers slither away, unless this
8080 * provider has already been explicitly invalidated.
8081 */
8082 if (!old->dtpv_defunct &&
8083 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8084 dtrace_anon.dta_state->dts_necbs > 0))) {
8085 if (!self) {
8086 mutex_exit(&dtrace_lock);
8087 mutex_exit(&mod_lock);
8088 mutex_exit(&dtrace_provider_lock);
8089 }
8090 return (EBUSY);
8091 }
8092
8093 /*
8094 * Attempt to destroy the probes associated with this provider.
8095 */
8096 for (i = 0; i < dtrace_nprobes; i++) {
8097 if ((probe = dtrace_probes[i]) == NULL)
8098 continue;
8099
8100 if (probe->dtpr_provider != old)
8101 continue;
8102
8103 if (probe->dtpr_ecb == NULL)
8104 continue;
8105
8106 /*
8107 * If we are trying to unregister a defunct provider, and the
8108 * provider was made defunct within the interval dictated by
8109 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8110 * attempt to reap our enablings. To denote that the provider
8111 * should reattempt to unregister itself at some point in the
8112 * future, we will return a differentiable error code (EAGAIN
8113 * instead of EBUSY) in this case.
8114 */
8115 if (dtrace_gethrtime() - old->dtpv_defunct >
8116 dtrace_unregister_defunct_reap)
8117 noreap = 1;
8118
8119 if (!self) {
8120 mutex_exit(&dtrace_lock);
8121 mutex_exit(&mod_lock);
8122 mutex_exit(&dtrace_provider_lock);
8123 }
8124
8125 if (noreap)
8126 return (EBUSY);
8127
8128 (void) taskq_dispatch(dtrace_taskq,
8129 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8130
8131 return (EAGAIN);
8132 }
8133
8134 /*
8135 * All of the probes for this provider are disabled; we can safely
8136 * remove all of them from their hash chains and from the probe array.
8137 */
8138 for (i = 0; i < dtrace_nprobes; i++) {
8139 if ((probe = dtrace_probes[i]) == NULL)
8140 continue;
8141
8142 if (probe->dtpr_provider != old)
8143 continue;
8144
8145 dtrace_probes[i] = NULL;
8146
8147 dtrace_hash_remove(dtrace_bymod, probe);
8148 dtrace_hash_remove(dtrace_byfunc, probe);
8149 dtrace_hash_remove(dtrace_byname, probe);
8150
8151 if (first == NULL) {
8152 first = probe;
8153 probe->dtpr_nextmod = NULL;
8154 } else {
8155 probe->dtpr_nextmod = first;
8156 first = probe;
8157 }
8158 }
8159
8160 /*
8161 * The provider's probes have been removed from the hash chains and
8162 * from the probe array. Now issue a dtrace_sync() to be sure that
8163 * everyone has cleared out from any probe array processing.
8164 */
8165 dtrace_sync();
8166
8167 for (probe = first; probe != NULL; probe = first) {
8168 first = probe->dtpr_nextmod;
8169
8170 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8171 probe->dtpr_arg);
8172 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8173 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8174 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8175 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8176 kmem_free(probe, sizeof (dtrace_probe_t));
8177 }
8178
8179 if ((prev = dtrace_provider) == old) {
8180 ASSERT(self || dtrace_devi == NULL);
8181 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8182 dtrace_provider = old->dtpv_next;
8183 } else {
8184 while (prev != NULL && prev->dtpv_next != old)
8185 prev = prev->dtpv_next;
8186
8187 if (prev == NULL) {
8188 panic("attempt to unregister non-existent "
8189 "dtrace provider %p\n", (void *)id);
8190 }
8191
8192 prev->dtpv_next = old->dtpv_next;
8193 }
8194
8195 if (!self) {
8196 mutex_exit(&dtrace_lock);
8197 mutex_exit(&mod_lock);
8198 mutex_exit(&dtrace_provider_lock);
8199 }
8200
8201 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
8202 kmem_free(old, sizeof (dtrace_provider_t));
8203
8204 return (0);
8205 }
8206
8207 /*
8208 * Invalidate the specified provider. All subsequent probe lookups for the
8209 * specified provider will fail, but its probes will not be removed.
8210 */
8211 void
8212 dtrace_invalidate(dtrace_provider_id_t id)
8213 {
8214 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
8215
8216 ASSERT(pvp->dtpv_pops.dtps_enable !=
8217 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8218
8219 mutex_enter(&dtrace_provider_lock);
8220 mutex_enter(&dtrace_lock);
8221
8222 pvp->dtpv_defunct = dtrace_gethrtime();
8223
8224 mutex_exit(&dtrace_lock);
8225 mutex_exit(&dtrace_provider_lock);
8226 }
8227
8228 /*
8229 * Indicate whether or not DTrace has attached.
8230 */
8231 int
8232 dtrace_attached(void)
8233 {
8234 /*
8235 * dtrace_provider will be non-NULL iff the DTrace driver has
8236 * attached. (It's non-NULL because DTrace is always itself a
8237 * provider.)
8238 */
8239 return (dtrace_provider != NULL);
8240 }
8241
8242 /*
8243 * Remove all the unenabled probes for the given provider. This function is
8244 * not unlike dtrace_unregister(), except that it doesn't remove the provider
8245 * -- just as many of its associated probes as it can.
8246 */
8247 int
8248 dtrace_condense(dtrace_provider_id_t id)
8249 {
8250 dtrace_provider_t *prov = (dtrace_provider_t *)id;
8251 int i;
8252 dtrace_probe_t *probe;
8253
8254 /*
8255 * Make sure this isn't the dtrace provider itself.
8256 */
8257 ASSERT(prov->dtpv_pops.dtps_enable !=
8258 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8259
8260 mutex_enter(&dtrace_provider_lock);
8261 mutex_enter(&dtrace_lock);
8262
8263 /*
8264 * Attempt to destroy the probes associated with this provider.
8265 */
8266 for (i = 0; i < dtrace_nprobes; i++) {
8267 if ((probe = dtrace_probes[i]) == NULL)
8268 continue;
8269
8270 if (probe->dtpr_provider != prov)
8271 continue;
8272
8273 if (probe->dtpr_ecb != NULL)
8274 continue;
8275
8276 dtrace_probes[i] = NULL;
8277
8278 dtrace_hash_remove(dtrace_bymod, probe);
8279 dtrace_hash_remove(dtrace_byfunc, probe);
8280 dtrace_hash_remove(dtrace_byname, probe);
8281
8282 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
8283 probe->dtpr_arg);
8284 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8285 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8286 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8287 kmem_free(probe, sizeof (dtrace_probe_t));
8288 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
8289 }
8290
8291 mutex_exit(&dtrace_lock);
8292 mutex_exit(&dtrace_provider_lock);
8293
8294 return (0);
8295 }
8296
8297 /*
8298 * DTrace Probe Management Functions
8299 *
8300 * The functions in this section perform the DTrace probe management,
8301 * including functions to create probes, look-up probes, and call into the
8302 * providers to request that probes be provided. Some of these functions are
8303 * in the Provider-to-Framework API; these functions can be identified by the
8304 * fact that they are not declared "static".
8305 */
8306
8307 /*
8308 * Create a probe with the specified module name, function name, and name.
8309 */
8310 dtrace_id_t
8311 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
8312 const char *func, const char *name, int aframes, void *arg)
8313 {
8314 dtrace_probe_t *probe, **probes;
8315 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
8316 dtrace_id_t id;
8317
8318 if (provider == dtrace_provider) {
8319 ASSERT(MUTEX_HELD(&dtrace_lock));
8320 } else {
8321 mutex_enter(&dtrace_lock);
8322 }
8323
8324 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
8325 VM_BESTFIT | VM_SLEEP);
8326 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8327
8328 probe->dtpr_id = id;
8329 probe->dtpr_gen = dtrace_probegen++;
8330 probe->dtpr_mod = dtrace_strdup(mod);
8331 probe->dtpr_func = dtrace_strdup(func);
8332 probe->dtpr_name = dtrace_strdup(name);
8333 probe->dtpr_arg = arg;
8334 probe->dtpr_aframes = aframes;
8335 probe->dtpr_provider = provider;
8336
8337 dtrace_hash_add(dtrace_bymod, probe);
8338 dtrace_hash_add(dtrace_byfunc, probe);
8339 dtrace_hash_add(dtrace_byname, probe);
8340
8341 if (id - 1 >= dtrace_nprobes) {
8342 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8343 size_t nsize = osize << 1;
8344
8345 if (nsize == 0) {
8346 ASSERT(osize == 0);
8347 ASSERT(dtrace_probes == NULL);
8348 nsize = sizeof (dtrace_probe_t *);
8349 }
8350
8351 probes = kmem_zalloc(nsize, KM_SLEEP);
8352
8353 if (dtrace_probes == NULL) {
8354 ASSERT(osize == 0);
8355 dtrace_probes = probes;
8356 dtrace_nprobes = 1;
8357 } else {
8358 dtrace_probe_t **oprobes = dtrace_probes;
8359
8360 bcopy(oprobes, probes, osize);
8361 dtrace_membar_producer();
8362 dtrace_probes = probes;
8363
8364 dtrace_sync();
8365
8366 /*
8367 * All CPUs are now seeing the new probes array; we can
8368 * safely free the old array.
8369 */
8370 kmem_free(oprobes, osize);
8371 dtrace_nprobes <<= 1;
8372 }
8373
8374 ASSERT(id - 1 < dtrace_nprobes);
8375 }
8376
8377 ASSERT(dtrace_probes[id - 1] == NULL);
8378 dtrace_probes[id - 1] = probe;
8379
8380 if (provider != dtrace_provider)
8381 mutex_exit(&dtrace_lock);
8382
8383 return (id);
8384 }
8385
8386 static dtrace_probe_t *
8387 dtrace_probe_lookup_id(dtrace_id_t id)
8388 {
8389 ASSERT(MUTEX_HELD(&dtrace_lock));
8390
8391 if (id == 0 || id > dtrace_nprobes)
8392 return (NULL);
8393
8394 return (dtrace_probes[id - 1]);
8395 }
8396
8397 static int
8398 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8399 {
8400 *((dtrace_id_t *)arg) = probe->dtpr_id;
8401
8402 return (DTRACE_MATCH_DONE);
8403 }
8404
8405 /*
8406 * Look up a probe based on provider and one or more of module name, function
8407 * name and probe name.
8408 */
8409 dtrace_id_t
8410 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
8411 const char *func, const char *name)
8412 {
8413 dtrace_probekey_t pkey;
8414 dtrace_id_t id;
8415 int match;
8416
8417 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8418 pkey.dtpk_pmatch = &dtrace_match_string;
8419 pkey.dtpk_mod = mod;
8420 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8421 pkey.dtpk_func = func;
8422 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8423 pkey.dtpk_name = name;
8424 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8425 pkey.dtpk_id = DTRACE_IDNONE;
8426
8427 mutex_enter(&dtrace_lock);
8428 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8429 dtrace_probe_lookup_match, &id);
8430 mutex_exit(&dtrace_lock);
8431
8432 ASSERT(match == 1 || match == 0);
8433 return (match ? id : 0);
8434 }
8435
8436 /*
8437 * Returns the probe argument associated with the specified probe.
8438 */
8439 void *
8440 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8441 {
8442 dtrace_probe_t *probe;
8443 void *rval = NULL;
8444
8445 mutex_enter(&dtrace_lock);
8446
8447 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8448 probe->dtpr_provider == (dtrace_provider_t *)id)
8449 rval = probe->dtpr_arg;
8450
8451 mutex_exit(&dtrace_lock);
8452
8453 return (rval);
8454 }
8455
8456 /*
8457 * Copy a probe into a probe description.
8458 */
8459 static void
8460 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8461 {
8462 bzero(pdp, sizeof (dtrace_probedesc_t));
8463 pdp->dtpd_id = prp->dtpr_id;
8464
8465 (void) strncpy(pdp->dtpd_provider,
8466 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8467
8468 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8469 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8470 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8471 }
8472
8473 /*
8474 * Called to indicate that a probe -- or probes -- should be provided by a
8475 * specfied provider. If the specified description is NULL, the provider will
8476 * be told to provide all of its probes. (This is done whenever a new
8477 * consumer comes along, or whenever a retained enabling is to be matched.) If
8478 * the specified description is non-NULL, the provider is given the
8479 * opportunity to dynamically provide the specified probe, allowing providers
8480 * to support the creation of probes on-the-fly. (So-called _autocreated_
8481 * probes.) If the provider is NULL, the operations will be applied to all
8482 * providers; if the provider is non-NULL the operations will only be applied
8483 * to the specified provider. The dtrace_provider_lock must be held, and the
8484 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8485 * will need to grab the dtrace_lock when it reenters the framework through
8486 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8487 */
8488 static void
8489 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8490 {
8491 struct modctl *ctl;
8492 int all = 0;
8493
8494 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8495
8496 if (prv == NULL) {
8497 all = 1;
8498 prv = dtrace_provider;
8499 }
8500
8501 do {
8502 /*
8503 * First, call the blanket provide operation.
8504 */
8505 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8506
8507 /*
8508 * Now call the per-module provide operation. We will grab
8509 * mod_lock to prevent the list from being modified. Note
8510 * that this also prevents the mod_busy bits from changing.
8511 * (mod_busy can only be changed with mod_lock held.)
8512 */
8513 mutex_enter(&mod_lock);
8514
8515 ctl = &modules;
8516 do {
8517 if (ctl->mod_busy || ctl->mod_mp == NULL)
8518 continue;
8519
8520 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8521
8522 } while ((ctl = ctl->mod_next) != &modules);
8523
8524 mutex_exit(&mod_lock);
8525 } while (all && (prv = prv->dtpv_next) != NULL);
8526 }
8527
8528 /*
8529 * Iterate over each probe, and call the Framework-to-Provider API function
8530 * denoted by offs.
8531 */
8532 static void
8533 dtrace_probe_foreach(uintptr_t offs)
8534 {
8535 dtrace_provider_t *prov;
8536 void (*func)(void *, dtrace_id_t, void *);
8537 dtrace_probe_t *probe;
8538 dtrace_icookie_t cookie;
8539 int i;
8540
8541 /*
8542 * We disable interrupts to walk through the probe array. This is
8543 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8544 * won't see stale data.
8545 */
8546 cookie = dtrace_interrupt_disable();
8547
8548 for (i = 0; i < dtrace_nprobes; i++) {
8549 if ((probe = dtrace_probes[i]) == NULL)
8550 continue;
8551
8552 if (probe->dtpr_ecb == NULL) {
8553 /*
8554 * This probe isn't enabled -- don't call the function.
8555 */
8556 continue;
8557 }
8558
8559 prov = probe->dtpr_provider;
8560 func = *((void(**)(void *, dtrace_id_t, void *))
8561 ((uintptr_t)&prov->dtpv_pops + offs));
8562
8563 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8564 }
8565
8566 dtrace_interrupt_enable(cookie);
8567 }
8568
8569 static int
8570 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8571 {
8572 dtrace_probekey_t pkey;
8573 uint32_t priv;
8574 uid_t uid;
8575 zoneid_t zoneid;
8576 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
8577
8578 ASSERT(MUTEX_HELD(&dtrace_lock));
8579 dtrace_ecb_create_cache = NULL;
8580
8581 if (desc == NULL) {
8582 /*
8583 * If we're passed a NULL description, we're being asked to
8584 * create an ECB with a NULL probe.
8585 */
8586 (void) dtrace_ecb_create_enable(NULL, enab);
8587 return (0);
8588 }
8589
8590 dtrace_probekey(desc, &pkey);
8591 dtrace_cred2priv(state->dts_cred.dcr_cred, &priv, &uid, &zoneid);
8592
8593 if ((priv & DTRACE_PRIV_ZONEOWNER) &&
8594 state->dts_options[DTRACEOPT_ZONE] != DTRACEOPT_UNSET) {
8595 /*
8596 * If we have the privilege of instrumenting all zones but we
8597 * have been told to instrument but one, we will spoof this up
8598 * depriving ourselves of DTRACE_PRIV_ZONEOWNER for purposes
8599 * of dtrace_match(). (Note that DTRACEOPT_ZONE is not for
8600 * security but rather for performance: it allows the global
8601 * zone to instrument USDT probes in a local zone without
8602 * requiring all zones to be instrumented.)
8603 */
8604 priv &= ~DTRACE_PRIV_ZONEOWNER;
8605 zoneid = state->dts_options[DTRACEOPT_ZONE];
8606 }
8607
8608 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8609 enab));
8610 }
8611
8612 /*
8613 * DTrace Helper Provider Functions
8614 */
8615 static void
8616 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8617 {
8618 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8619 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8620 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8621 }
8622
8623 static void
8624 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8625 const dof_provider_t *dofprov, char *strtab)
8626 {
8627 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8628 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8629 dofprov->dofpv_provattr);
8630 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8631 dofprov->dofpv_modattr);
8632 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8633 dofprov->dofpv_funcattr);
8634 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8635 dofprov->dofpv_nameattr);
8636 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8637 dofprov->dofpv_argsattr);
8638 }
8639
8640 static void
8641 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8642 {
8643 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8644 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8645 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8646 dof_provider_t *provider;
8647 dof_probe_t *probe;
8648 uint32_t *off, *enoff;
8649 uint8_t *arg;
8650 char *strtab;
8651 uint_t i, nprobes;
8652 dtrace_helper_provdesc_t dhpv;
8653 dtrace_helper_probedesc_t dhpb;
8654 dtrace_meta_t *meta = dtrace_meta_pid;
8655 dtrace_mops_t *mops = &meta->dtm_mops;
8656 void *parg;
8657
8658 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8659 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8660 provider->dofpv_strtab * dof->dofh_secsize);
8661 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8662 provider->dofpv_probes * dof->dofh_secsize);
8663 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8664 provider->dofpv_prargs * dof->dofh_secsize);
8665 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8666 provider->dofpv_proffs * dof->dofh_secsize);
8667
8668 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8669 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8670 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8671 enoff = NULL;
8672
8673 /*
8674 * See dtrace_helper_provider_validate().
8675 */
8676 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8677 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8678 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8679 provider->dofpv_prenoffs * dof->dofh_secsize);
8680 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8681 }
8682
8683 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8684
8685 /*
8686 * Create the provider.
8687 */
8688 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8689
8690 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8691 return;
8692
8693 meta->dtm_count++;
8694
8695 /*
8696 * Create the probes.
8697 */
8698 for (i = 0; i < nprobes; i++) {
8699 probe = (dof_probe_t *)(uintptr_t)(daddr +
8700 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8701
8702 dhpb.dthpb_mod = dhp->dofhp_mod;
8703 dhpb.dthpb_func = strtab + probe->dofpr_func;
8704 dhpb.dthpb_name = strtab + probe->dofpr_name;
8705 dhpb.dthpb_base = probe->dofpr_addr;
8706 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8707 dhpb.dthpb_noffs = probe->dofpr_noffs;
8708 if (enoff != NULL) {
8709 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8710 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8711 } else {
8712 dhpb.dthpb_enoffs = NULL;
8713 dhpb.dthpb_nenoffs = 0;
8714 }
8715 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8716 dhpb.dthpb_nargc = probe->dofpr_nargc;
8717 dhpb.dthpb_xargc = probe->dofpr_xargc;
8718 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8719 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8720
8721 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8722 }
8723 }
8724
8725 static void
8726 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8727 {
8728 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8729 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8730 int i;
8731
8732 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8733
8734 for (i = 0; i < dof->dofh_secnum; i++) {
8735 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8736 dof->dofh_secoff + i * dof->dofh_secsize);
8737
8738 if (sec->dofs_type != DOF_SECT_PROVIDER)
8739 continue;
8740
8741 dtrace_helper_provide_one(dhp, sec, pid);
8742 }
8743
8744 /*
8745 * We may have just created probes, so we must now rematch against
8746 * any retained enablings. Note that this call will acquire both
8747 * cpu_lock and dtrace_lock; the fact that we are holding
8748 * dtrace_meta_lock now is what defines the ordering with respect to
8749 * these three locks.
8750 */
8751 dtrace_enabling_matchall();
8752 }
8753
8754 static void
8755 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8756 {
8757 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8758 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8759 dof_sec_t *str_sec;
8760 dof_provider_t *provider;
8761 char *strtab;
8762 dtrace_helper_provdesc_t dhpv;
8763 dtrace_meta_t *meta = dtrace_meta_pid;
8764 dtrace_mops_t *mops = &meta->dtm_mops;
8765
8766 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8767 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8768 provider->dofpv_strtab * dof->dofh_secsize);
8769
8770 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8771
8772 /*
8773 * Create the provider.
8774 */
8775 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8776
8777 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8778
8779 meta->dtm_count--;
8780 }
8781
8782 static void
8783 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8784 {
8785 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8786 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8787 int i;
8788
8789 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8790
8791 for (i = 0; i < dof->dofh_secnum; i++) {
8792 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8793 dof->dofh_secoff + i * dof->dofh_secsize);
8794
8795 if (sec->dofs_type != DOF_SECT_PROVIDER)
8796 continue;
8797
8798 dtrace_helper_provider_remove_one(dhp, sec, pid);
8799 }
8800 }
8801
8802 /*
8803 * DTrace Meta Provider-to-Framework API Functions
8804 *
8805 * These functions implement the Meta Provider-to-Framework API, as described
8806 * in <sys/dtrace.h>.
8807 */
8808 int
8809 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8810 dtrace_meta_provider_id_t *idp)
8811 {
8812 dtrace_meta_t *meta;
8813 dtrace_helpers_t *help, *next;
8814 int i;
8815
8816 *idp = DTRACE_METAPROVNONE;
8817
8818 /*
8819 * We strictly don't need the name, but we hold onto it for
8820 * debuggability. All hail error queues!
8821 */
8822 if (name == NULL) {
8823 cmn_err(CE_WARN, "failed to register meta-provider: "
8824 "invalid name");
8825 return (EINVAL);
8826 }
8827
8828 if (mops == NULL ||
8829 mops->dtms_create_probe == NULL ||
8830 mops->dtms_provide_pid == NULL ||
8831 mops->dtms_remove_pid == NULL) {
8832 cmn_err(CE_WARN, "failed to register meta-register %s: "
8833 "invalid ops", name);
8834 return (EINVAL);
8835 }
8836
8837 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8838 meta->dtm_mops = *mops;
8839 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8840 (void) strcpy(meta->dtm_name, name);
8841 meta->dtm_arg = arg;
8842
8843 mutex_enter(&dtrace_meta_lock);
8844 mutex_enter(&dtrace_lock);
8845
8846 if (dtrace_meta_pid != NULL) {
8847 mutex_exit(&dtrace_lock);
8848 mutex_exit(&dtrace_meta_lock);
8849 cmn_err(CE_WARN, "failed to register meta-register %s: "
8850 "user-land meta-provider exists", name);
8851 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8852 kmem_free(meta, sizeof (dtrace_meta_t));
8853 return (EINVAL);
8854 }
8855
8856 dtrace_meta_pid = meta;
8857 *idp = (dtrace_meta_provider_id_t)meta;
8858
8859 /*
8860 * If there are providers and probes ready to go, pass them
8861 * off to the new meta provider now.
8862 */
8863
8864 help = dtrace_deferred_pid;
8865 dtrace_deferred_pid = NULL;
8866
8867 mutex_exit(&dtrace_lock);
8868
8869 while (help != NULL) {
8870 for (i = 0; i < help->dthps_nprovs; i++) {
8871 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8872 help->dthps_pid);
8873 }
8874
8875 next = help->dthps_next;
8876 help->dthps_next = NULL;
8877 help->dthps_prev = NULL;
8878 help->dthps_deferred = 0;
8879 help = next;
8880 }
8881
8882 mutex_exit(&dtrace_meta_lock);
8883
8884 return (0);
8885 }
8886
8887 int
8888 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8889 {
8890 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8891
8892 mutex_enter(&dtrace_meta_lock);
8893 mutex_enter(&dtrace_lock);
8894
8895 if (old == dtrace_meta_pid) {
8896 pp = &dtrace_meta_pid;
8897 } else {
8898 panic("attempt to unregister non-existent "
8899 "dtrace meta-provider %p\n", (void *)old);
8900 }
8901
8902 if (old->dtm_count != 0) {
8903 mutex_exit(&dtrace_lock);
8904 mutex_exit(&dtrace_meta_lock);
8905 return (EBUSY);
8906 }
8907
8908 *pp = NULL;
8909
8910 mutex_exit(&dtrace_lock);
8911 mutex_exit(&dtrace_meta_lock);
8912
8913 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8914 kmem_free(old, sizeof (dtrace_meta_t));
8915
8916 return (0);
8917 }
8918
8919
8920 /*
8921 * DTrace DIF Object Functions
8922 */
8923 static int
8924 dtrace_difo_err(uint_t pc, const char *format, ...)
8925 {
8926 if (dtrace_err_verbose) {
8927 va_list alist;
8928
8929 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8930 va_start(alist, format);
8931 (void) vuprintf(format, alist);
8932 va_end(alist);
8933 }
8934
8935 #ifdef DTRACE_ERRDEBUG
8936 dtrace_errdebug(format);
8937 #endif
8938 return (1);
8939 }
8940
8941 /*
8942 * Validate a DTrace DIF object by checking the IR instructions. The following
8943 * rules are currently enforced by dtrace_difo_validate():
8944 *
8945 * 1. Each instruction must have a valid opcode
8946 * 2. Each register, string, variable, or subroutine reference must be valid
8947 * 3. No instruction can modify register %r0 (must be zero)
8948 * 4. All instruction reserved bits must be set to zero
8949 * 5. The last instruction must be a "ret" instruction
8950 * 6. All branch targets must reference a valid instruction _after_ the branch
8951 */
8952 static int
8953 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8954 cred_t *cr)
8955 {
8956 int err = 0, i;
8957 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8958 int kcheckload;
8959 uint_t pc;
8960
8961 kcheckload = cr == NULL ||
8962 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8963
8964 dp->dtdo_destructive = 0;
8965
8966 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8967 dif_instr_t instr = dp->dtdo_buf[pc];
8968
8969 uint_t r1 = DIF_INSTR_R1(instr);
8970 uint_t r2 = DIF_INSTR_R2(instr);
8971 uint_t rd = DIF_INSTR_RD(instr);
8972 uint_t rs = DIF_INSTR_RS(instr);
8973 uint_t label = DIF_INSTR_LABEL(instr);
8974 uint_t v = DIF_INSTR_VAR(instr);
8975 uint_t subr = DIF_INSTR_SUBR(instr);
8976 uint_t type = DIF_INSTR_TYPE(instr);
8977 uint_t op = DIF_INSTR_OP(instr);
8978
8979 switch (op) {
8980 case DIF_OP_OR:
8981 case DIF_OP_XOR:
8982 case DIF_OP_AND:
8983 case DIF_OP_SLL:
8984 case DIF_OP_SRL:
8985 case DIF_OP_SRA:
8986 case DIF_OP_SUB:
8987 case DIF_OP_ADD:
8988 case DIF_OP_MUL:
8989 case DIF_OP_SDIV:
8990 case DIF_OP_UDIV:
8991 case DIF_OP_SREM:
8992 case DIF_OP_UREM:
8993 case DIF_OP_COPYS:
8994 if (r1 >= nregs)
8995 err += efunc(pc, "invalid register %u\n", r1);
8996 if (r2 >= nregs)
8997 err += efunc(pc, "invalid register %u\n", r2);
8998 if (rd >= nregs)
8999 err += efunc(pc, "invalid register %u\n", rd);
9000 if (rd == 0)
9001 err += efunc(pc, "cannot write to %r0\n");
9002 break;
9003 case DIF_OP_NOT:
9004 case DIF_OP_MOV:
9005 case DIF_OP_ALLOCS:
9006 if (r1 >= nregs)
9007 err += efunc(pc, "invalid register %u\n", r1);
9008 if (r2 != 0)
9009 err += efunc(pc, "non-zero reserved bits\n");
9010 if (rd >= nregs)
9011 err += efunc(pc, "invalid register %u\n", rd);
9012 if (rd == 0)
9013 err += efunc(pc, "cannot write to %r0\n");
9014 break;
9015 case DIF_OP_LDSB:
9016 case DIF_OP_LDSH:
9017 case DIF_OP_LDSW:
9018 case DIF_OP_LDUB:
9019 case DIF_OP_LDUH:
9020 case DIF_OP_LDUW:
9021 case DIF_OP_LDX:
9022 if (r1 >= nregs)
9023 err += efunc(pc, "invalid register %u\n", r1);
9024 if (r2 != 0)
9025 err += efunc(pc, "non-zero reserved bits\n");
9026 if (rd >= nregs)
9027 err += efunc(pc, "invalid register %u\n", rd);
9028 if (rd == 0)
9029 err += efunc(pc, "cannot write to %r0\n");
9030 if (kcheckload)
9031 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9032 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9033 break;
9034 case DIF_OP_RLDSB:
9035 case DIF_OP_RLDSH:
9036 case DIF_OP_RLDSW:
9037 case DIF_OP_RLDUB:
9038 case DIF_OP_RLDUH:
9039 case DIF_OP_RLDUW:
9040 case DIF_OP_RLDX:
9041 if (r1 >= nregs)
9042 err += efunc(pc, "invalid register %u\n", r1);
9043 if (r2 != 0)
9044 err += efunc(pc, "non-zero reserved bits\n");
9045 if (rd >= nregs)
9046 err += efunc(pc, "invalid register %u\n", rd);
9047 if (rd == 0)
9048 err += efunc(pc, "cannot write to %r0\n");
9049 break;
9050 case DIF_OP_ULDSB:
9051 case DIF_OP_ULDSH:
9052 case DIF_OP_ULDSW:
9053 case DIF_OP_ULDUB:
9054 case DIF_OP_ULDUH:
9055 case DIF_OP_ULDUW:
9056 case DIF_OP_ULDX:
9057 if (r1 >= nregs)
9058 err += efunc(pc, "invalid register %u\n", r1);
9059 if (r2 != 0)
9060 err += efunc(pc, "non-zero reserved bits\n");
9061 if (rd >= nregs)
9062 err += efunc(pc, "invalid register %u\n", rd);
9063 if (rd == 0)
9064 err += efunc(pc, "cannot write to %r0\n");
9065 break;
9066 case DIF_OP_STB:
9067 case DIF_OP_STH:
9068 case DIF_OP_STW:
9069 case DIF_OP_STX:
9070 if (r1 >= nregs)
9071 err += efunc(pc, "invalid register %u\n", r1);
9072 if (r2 != 0)
9073 err += efunc(pc, "non-zero reserved bits\n");
9074 if (rd >= nregs)
9075 err += efunc(pc, "invalid register %u\n", rd);
9076 if (rd == 0)
9077 err += efunc(pc, "cannot write to 0 address\n");
9078 break;
9079 case DIF_OP_CMP:
9080 case DIF_OP_SCMP:
9081 if (r1 >= nregs)
9082 err += efunc(pc, "invalid register %u\n", r1);
9083 if (r2 >= nregs)
9084 err += efunc(pc, "invalid register %u\n", r2);
9085 if (rd != 0)
9086 err += efunc(pc, "non-zero reserved bits\n");
9087 break;
9088 case DIF_OP_TST:
9089 if (r1 >= nregs)
9090 err += efunc(pc, "invalid register %u\n", r1);
9091 if (r2 != 0 || rd != 0)
9092 err += efunc(pc, "non-zero reserved bits\n");
9093 break;
9094 case DIF_OP_BA:
9095 case DIF_OP_BE:
9096 case DIF_OP_BNE:
9097 case DIF_OP_BG:
9098 case DIF_OP_BGU:
9099 case DIF_OP_BGE:
9100 case DIF_OP_BGEU:
9101 case DIF_OP_BL:
9102 case DIF_OP_BLU:
9103 case DIF_OP_BLE:
9104 case DIF_OP_BLEU:
9105 if (label >= dp->dtdo_len) {
9106 err += efunc(pc, "invalid branch target %u\n",
9107 label);
9108 }
9109 if (label <= pc) {
9110 err += efunc(pc, "backward branch to %u\n",
9111 label);
9112 }
9113 break;
9114 case DIF_OP_RET:
9115 if (r1 != 0 || r2 != 0)
9116 err += efunc(pc, "non-zero reserved bits\n");
9117 if (rd >= nregs)
9118 err += efunc(pc, "invalid register %u\n", rd);
9119 break;
9120 case DIF_OP_NOP:
9121 case DIF_OP_POPTS:
9122 case DIF_OP_FLUSHTS:
9123 if (r1 != 0 || r2 != 0 || rd != 0)
9124 err += efunc(pc, "non-zero reserved bits\n");
9125 break;
9126 case DIF_OP_SETX:
9127 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9128 err += efunc(pc, "invalid integer ref %u\n",
9129 DIF_INSTR_INTEGER(instr));
9130 }
9131 if (rd >= nregs)
9132 err += efunc(pc, "invalid register %u\n", rd);
9133 if (rd == 0)
9134 err += efunc(pc, "cannot write to %r0\n");
9135 break;
9136 case DIF_OP_SETS:
9137 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9138 err += efunc(pc, "invalid string ref %u\n",
9139 DIF_INSTR_STRING(instr));
9140 }
9141 if (rd >= nregs)
9142 err += efunc(pc, "invalid register %u\n", rd);
9143 if (rd == 0)
9144 err += efunc(pc, "cannot write to %r0\n");
9145 break;
9146 case DIF_OP_LDGA:
9147 case DIF_OP_LDTA:
9148 if (r1 > DIF_VAR_ARRAY_MAX)
9149 err += efunc(pc, "invalid array %u\n", r1);
9150 if (r2 >= nregs)
9151 err += efunc(pc, "invalid register %u\n", r2);
9152 if (rd >= nregs)
9153 err += efunc(pc, "invalid register %u\n", rd);
9154 if (rd == 0)
9155 err += efunc(pc, "cannot write to %r0\n");
9156 break;
9157 case DIF_OP_LDGS:
9158 case DIF_OP_LDTS:
9159 case DIF_OP_LDLS:
9160 case DIF_OP_LDGAA:
9161 case DIF_OP_LDTAA:
9162 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9163 err += efunc(pc, "invalid variable %u\n", v);
9164 if (rd >= nregs)
9165 err += efunc(pc, "invalid register %u\n", rd);
9166 if (rd == 0)
9167 err += efunc(pc, "cannot write to %r0\n");
9168 break;
9169 case DIF_OP_STGS:
9170 case DIF_OP_STTS:
9171 case DIF_OP_STLS:
9172 case DIF_OP_STGAA:
9173 case DIF_OP_STTAA:
9174 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9175 err += efunc(pc, "invalid variable %u\n", v);
9176 if (rs >= nregs)
9177 err += efunc(pc, "invalid register %u\n", rd);
9178 break;
9179 case DIF_OP_CALL:
9180 if (subr > DIF_SUBR_MAX)
9181 err += efunc(pc, "invalid subr %u\n", subr);
9182 if (rd >= nregs)
9183 err += efunc(pc, "invalid register %u\n", rd);
9184 if (rd == 0)
9185 err += efunc(pc, "cannot write to %r0\n");
9186
9187 if (subr == DIF_SUBR_COPYOUT ||
9188 subr == DIF_SUBR_COPYOUTSTR) {
9189 dp->dtdo_destructive = 1;
9190 }
9191
9192 if (subr == DIF_SUBR_GETF) {
9193 /*
9194 * If we have a getf() we need to record that
9195 * in our state. Note that our state can be
9196 * NULL if this is a helper -- but in that
9197 * case, the call to getf() is itself illegal,
9198 * and will be caught (slightly later) when
9199 * the helper is validated.
9200 */
9201 if (vstate->dtvs_state != NULL)
9202 vstate->dtvs_state->dts_getf++;
9203 }
9204
9205 break;
9206 case DIF_OP_PUSHTR:
9207 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
9208 err += efunc(pc, "invalid ref type %u\n", type);
9209 if (r2 >= nregs)
9210 err += efunc(pc, "invalid register %u\n", r2);
9211 if (rs >= nregs)
9212 err += efunc(pc, "invalid register %u\n", rs);
9213 break;
9214 case DIF_OP_PUSHTV:
9215 if (type != DIF_TYPE_CTF)
9216 err += efunc(pc, "invalid val type %u\n", type);
9217 if (r2 >= nregs)
9218 err += efunc(pc, "invalid register %u\n", r2);
9219 if (rs >= nregs)
9220 err += efunc(pc, "invalid register %u\n", rs);
9221 break;
9222 default:
9223 err += efunc(pc, "invalid opcode %u\n",
9224 DIF_INSTR_OP(instr));
9225 }
9226 }
9227
9228 if (dp->dtdo_len != 0 &&
9229 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
9230 err += efunc(dp->dtdo_len - 1,
9231 "expected 'ret' as last DIF instruction\n");
9232 }
9233
9234 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
9235 /*
9236 * If we're not returning by reference, the size must be either
9237 * 0 or the size of one of the base types.
9238 */
9239 switch (dp->dtdo_rtype.dtdt_size) {
9240 case 0:
9241 case sizeof (uint8_t):
9242 case sizeof (uint16_t):
9243 case sizeof (uint32_t):
9244 case sizeof (uint64_t):
9245 break;
9246
9247 default:
9248 err += efunc(dp->dtdo_len - 1, "bad return size\n");
9249 }
9250 }
9251
9252 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
9253 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
9254 dtrace_diftype_t *vt, *et;
9255 uint_t id, ndx;
9256
9257 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
9258 v->dtdv_scope != DIFV_SCOPE_THREAD &&
9259 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
9260 err += efunc(i, "unrecognized variable scope %d\n",
9261 v->dtdv_scope);
9262 break;
9263 }
9264
9265 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
9266 v->dtdv_kind != DIFV_KIND_SCALAR) {
9267 err += efunc(i, "unrecognized variable type %d\n",
9268 v->dtdv_kind);
9269 break;
9270 }
9271
9272 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
9273 err += efunc(i, "%d exceeds variable id limit\n", id);
9274 break;
9275 }
9276
9277 if (id < DIF_VAR_OTHER_UBASE)
9278 continue;
9279
9280 /*
9281 * For user-defined variables, we need to check that this
9282 * definition is identical to any previous definition that we
9283 * encountered.
9284 */
9285 ndx = id - DIF_VAR_OTHER_UBASE;
9286
9287 switch (v->dtdv_scope) {
9288 case DIFV_SCOPE_GLOBAL:
9289 if (ndx < vstate->dtvs_nglobals) {
9290 dtrace_statvar_t *svar;
9291
9292 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
9293 existing = &svar->dtsv_var;
9294 }
9295
9296 break;
9297
9298 case DIFV_SCOPE_THREAD:
9299 if (ndx < vstate->dtvs_ntlocals)
9300 existing = &vstate->dtvs_tlocals[ndx];
9301 break;
9302
9303 case DIFV_SCOPE_LOCAL:
9304 if (ndx < vstate->dtvs_nlocals) {
9305 dtrace_statvar_t *svar;
9306
9307 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
9308 existing = &svar->dtsv_var;
9309 }
9310
9311 break;
9312 }
9313
9314 vt = &v->dtdv_type;
9315
9316 if (vt->dtdt_flags & DIF_TF_BYREF) {
9317 if (vt->dtdt_size == 0) {
9318 err += efunc(i, "zero-sized variable\n");
9319 break;
9320 }
9321
9322 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
9323 vt->dtdt_size > dtrace_global_maxsize) {
9324 err += efunc(i, "oversized by-ref global\n");
9325 break;
9326 }
9327 }
9328
9329 if (existing == NULL || existing->dtdv_id == 0)
9330 continue;
9331
9332 ASSERT(existing->dtdv_id == v->dtdv_id);
9333 ASSERT(existing->dtdv_scope == v->dtdv_scope);
9334
9335 if (existing->dtdv_kind != v->dtdv_kind)
9336 err += efunc(i, "%d changed variable kind\n", id);
9337
9338 et = &existing->dtdv_type;
9339
9340 if (vt->dtdt_flags != et->dtdt_flags) {
9341 err += efunc(i, "%d changed variable type flags\n", id);
9342 break;
9343 }
9344
9345 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
9346 err += efunc(i, "%d changed variable type size\n", id);
9347 break;
9348 }
9349 }
9350
9351 return (err);
9352 }
9353
9354 /*
9355 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
9356 * are much more constrained than normal DIFOs. Specifically, they may
9357 * not:
9358 *
9359 * 1. Make calls to subroutines other than copyin(), copyinstr() or
9360 * miscellaneous string routines
9361 * 2. Access DTrace variables other than the args[] array, and the
9362 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9363 * 3. Have thread-local variables.
9364 * 4. Have dynamic variables.
9365 */
9366 static int
9367 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9368 {
9369 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9370 int err = 0;
9371 uint_t pc;
9372
9373 for (pc = 0; pc < dp->dtdo_len; pc++) {
9374 dif_instr_t instr = dp->dtdo_buf[pc];
9375
9376 uint_t v = DIF_INSTR_VAR(instr);
9377 uint_t subr = DIF_INSTR_SUBR(instr);
9378 uint_t op = DIF_INSTR_OP(instr);
9379
9380 switch (op) {
9381 case DIF_OP_OR:
9382 case DIF_OP_XOR:
9383 case DIF_OP_AND:
9384 case DIF_OP_SLL:
9385 case DIF_OP_SRL:
9386 case DIF_OP_SRA:
9387 case DIF_OP_SUB:
9388 case DIF_OP_ADD:
9389 case DIF_OP_MUL:
9390 case DIF_OP_SDIV:
9391 case DIF_OP_UDIV:
9392 case DIF_OP_SREM:
9393 case DIF_OP_UREM:
9394 case DIF_OP_COPYS:
9395 case DIF_OP_NOT:
9396 case DIF_OP_MOV:
9397 case DIF_OP_RLDSB:
9398 case DIF_OP_RLDSH:
9399 case DIF_OP_RLDSW:
9400 case DIF_OP_RLDUB:
9401 case DIF_OP_RLDUH:
9402 case DIF_OP_RLDUW:
9403 case DIF_OP_RLDX:
9404 case DIF_OP_ULDSB:
9405 case DIF_OP_ULDSH:
9406 case DIF_OP_ULDSW:
9407 case DIF_OP_ULDUB:
9408 case DIF_OP_ULDUH:
9409 case DIF_OP_ULDUW:
9410 case DIF_OP_ULDX:
9411 case DIF_OP_STB:
9412 case DIF_OP_STH:
9413 case DIF_OP_STW:
9414 case DIF_OP_STX:
9415 case DIF_OP_ALLOCS:
9416 case DIF_OP_CMP:
9417 case DIF_OP_SCMP:
9418 case DIF_OP_TST:
9419 case DIF_OP_BA:
9420 case DIF_OP_BE:
9421 case DIF_OP_BNE:
9422 case DIF_OP_BG:
9423 case DIF_OP_BGU:
9424 case DIF_OP_BGE:
9425 case DIF_OP_BGEU:
9426 case DIF_OP_BL:
9427 case DIF_OP_BLU:
9428 case DIF_OP_BLE:
9429 case DIF_OP_BLEU:
9430 case DIF_OP_RET:
9431 case DIF_OP_NOP:
9432 case DIF_OP_POPTS:
9433 case DIF_OP_FLUSHTS:
9434 case DIF_OP_SETX:
9435 case DIF_OP_SETS:
9436 case DIF_OP_LDGA:
9437 case DIF_OP_LDLS:
9438 case DIF_OP_STGS:
9439 case DIF_OP_STLS:
9440 case DIF_OP_PUSHTR:
9441 case DIF_OP_PUSHTV:
9442 break;
9443
9444 case DIF_OP_LDGS:
9445 if (v >= DIF_VAR_OTHER_UBASE)
9446 break;
9447
9448 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9449 break;
9450
9451 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9452 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9453 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9454 v == DIF_VAR_UID || v == DIF_VAR_GID)
9455 break;
9456
9457 err += efunc(pc, "illegal variable %u\n", v);
9458 break;
9459
9460 case DIF_OP_LDTA:
9461 case DIF_OP_LDTS:
9462 case DIF_OP_LDGAA:
9463 case DIF_OP_LDTAA:
9464 err += efunc(pc, "illegal dynamic variable load\n");
9465 break;
9466
9467 case DIF_OP_STTS:
9468 case DIF_OP_STGAA:
9469 case DIF_OP_STTAA:
9470 err += efunc(pc, "illegal dynamic variable store\n");
9471 break;
9472
9473 case DIF_OP_CALL:
9474 if (subr == DIF_SUBR_ALLOCA ||
9475 subr == DIF_SUBR_BCOPY ||
9476 subr == DIF_SUBR_COPYIN ||
9477 subr == DIF_SUBR_COPYINTO ||
9478 subr == DIF_SUBR_COPYINSTR ||
9479 subr == DIF_SUBR_INDEX ||
9480 subr == DIF_SUBR_INET_NTOA ||
9481 subr == DIF_SUBR_INET_NTOA6 ||
9482 subr == DIF_SUBR_INET_NTOP ||
9483 subr == DIF_SUBR_JSON ||
9484 subr == DIF_SUBR_LLTOSTR ||
9485 subr == DIF_SUBR_STRTOLL ||
9486 subr == DIF_SUBR_RINDEX ||
9487 subr == DIF_SUBR_STRCHR ||
9488 subr == DIF_SUBR_STRJOIN ||
9489 subr == DIF_SUBR_STRRCHR ||
9490 subr == DIF_SUBR_STRSTR ||
9491 subr == DIF_SUBR_HTONS ||
9492 subr == DIF_SUBR_HTONL ||
9493 subr == DIF_SUBR_HTONLL ||
9494 subr == DIF_SUBR_NTOHS ||
9495 subr == DIF_SUBR_NTOHL ||
9496 subr == DIF_SUBR_NTOHLL)
9497 break;
9498
9499 err += efunc(pc, "invalid subr %u\n", subr);
9500 break;
9501
9502 default:
9503 err += efunc(pc, "invalid opcode %u\n",
9504 DIF_INSTR_OP(instr));
9505 }
9506 }
9507
9508 return (err);
9509 }
9510
9511 /*
9512 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9513 * basis; 0 if not.
9514 */
9515 static int
9516 dtrace_difo_cacheable(dtrace_difo_t *dp)
9517 {
9518 int i;
9519
9520 if (dp == NULL)
9521 return (0);
9522
9523 for (i = 0; i < dp->dtdo_varlen; i++) {
9524 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9525
9526 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9527 continue;
9528
9529 switch (v->dtdv_id) {
9530 case DIF_VAR_CURTHREAD:
9531 case DIF_VAR_PID:
9532 case DIF_VAR_TID:
9533 case DIF_VAR_EXECNAME:
9534 case DIF_VAR_ZONENAME:
9535 break;
9536
9537 default:
9538 return (0);
9539 }
9540 }
9541
9542 /*
9543 * This DIF object may be cacheable. Now we need to look for any
9544 * array loading instructions, any memory loading instructions, or
9545 * any stores to thread-local variables.
9546 */
9547 for (i = 0; i < dp->dtdo_len; i++) {
9548 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9549
9550 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9551 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9552 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9553 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9554 return (0);
9555 }
9556
9557 return (1);
9558 }
9559
9560 static void
9561 dtrace_difo_hold(dtrace_difo_t *dp)
9562 {
9563 int i;
9564
9565 ASSERT(MUTEX_HELD(&dtrace_lock));
9566
9567 dp->dtdo_refcnt++;
9568 ASSERT(dp->dtdo_refcnt != 0);
9569
9570 /*
9571 * We need to check this DIF object for references to the variable
9572 * DIF_VAR_VTIMESTAMP.
9573 */
9574 for (i = 0; i < dp->dtdo_varlen; i++) {
9575 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9576
9577 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9578 continue;
9579
9580 if (dtrace_vtime_references++ == 0)
9581 dtrace_vtime_enable();
9582 }
9583 }
9584
9585 /*
9586 * This routine calculates the dynamic variable chunksize for a given DIF
9587 * object. The calculation is not fool-proof, and can probably be tricked by
9588 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9589 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9590 * if a dynamic variable size exceeds the chunksize.
9591 */
9592 static void
9593 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9594 {
9595 uint64_t sval;
9596 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9597 const dif_instr_t *text = dp->dtdo_buf;
9598 uint_t pc, srd = 0;
9599 uint_t ttop = 0;
9600 size_t size, ksize;
9601 uint_t id, i;
9602
9603 for (pc = 0; pc < dp->dtdo_len; pc++) {
9604 dif_instr_t instr = text[pc];
9605 uint_t op = DIF_INSTR_OP(instr);
9606 uint_t rd = DIF_INSTR_RD(instr);
9607 uint_t r1 = DIF_INSTR_R1(instr);
9608 uint_t nkeys = 0;
9609 uchar_t scope;
9610
9611 dtrace_key_t *key = tupregs;
9612
9613 switch (op) {
9614 case DIF_OP_SETX:
9615 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9616 srd = rd;
9617 continue;
9618
9619 case DIF_OP_STTS:
9620 key = &tupregs[DIF_DTR_NREGS];
9621 key[0].dttk_size = 0;
9622 key[1].dttk_size = 0;
9623 nkeys = 2;
9624 scope = DIFV_SCOPE_THREAD;
9625 break;
9626
9627 case DIF_OP_STGAA:
9628 case DIF_OP_STTAA:
9629 nkeys = ttop;
9630
9631 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9632 key[nkeys++].dttk_size = 0;
9633
9634 key[nkeys++].dttk_size = 0;
9635
9636 if (op == DIF_OP_STTAA) {
9637 scope = DIFV_SCOPE_THREAD;
9638 } else {
9639 scope = DIFV_SCOPE_GLOBAL;
9640 }
9641
9642 break;
9643
9644 case DIF_OP_PUSHTR:
9645 if (ttop == DIF_DTR_NREGS)
9646 return;
9647
9648 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9649 /*
9650 * If the register for the size of the "pushtr"
9651 * is %r0 (or the value is 0) and the type is
9652 * a string, we'll use the system-wide default
9653 * string size.
9654 */
9655 tupregs[ttop++].dttk_size =
9656 dtrace_strsize_default;
9657 } else {
9658 if (srd == 0)
9659 return;
9660
9661 tupregs[ttop++].dttk_size = sval;
9662 }
9663
9664 break;
9665
9666 case DIF_OP_PUSHTV:
9667 if (ttop == DIF_DTR_NREGS)
9668 return;
9669
9670 tupregs[ttop++].dttk_size = 0;
9671 break;
9672
9673 case DIF_OP_FLUSHTS:
9674 ttop = 0;
9675 break;
9676
9677 case DIF_OP_POPTS:
9678 if (ttop != 0)
9679 ttop--;
9680 break;
9681 }
9682
9683 sval = 0;
9684 srd = 0;
9685
9686 if (nkeys == 0)
9687 continue;
9688
9689 /*
9690 * We have a dynamic variable allocation; calculate its size.
9691 */
9692 for (ksize = 0, i = 0; i < nkeys; i++)
9693 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9694
9695 size = sizeof (dtrace_dynvar_t);
9696 size += sizeof (dtrace_key_t) * (nkeys - 1);
9697 size += ksize;
9698
9699 /*
9700 * Now we need to determine the size of the stored data.
9701 */
9702 id = DIF_INSTR_VAR(instr);
9703
9704 for (i = 0; i < dp->dtdo_varlen; i++) {
9705 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9706
9707 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9708 size += v->dtdv_type.dtdt_size;
9709 break;
9710 }
9711 }
9712
9713 if (i == dp->dtdo_varlen)
9714 return;
9715
9716 /*
9717 * We have the size. If this is larger than the chunk size
9718 * for our dynamic variable state, reset the chunk size.
9719 */
9720 size = P2ROUNDUP(size, sizeof (uint64_t));
9721
9722 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9723 vstate->dtvs_dynvars.dtds_chunksize = size;
9724 }
9725 }
9726
9727 static void
9728 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9729 {
9730 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9731 uint_t id;
9732
9733 ASSERT(MUTEX_HELD(&dtrace_lock));
9734 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9735
9736 for (i = 0; i < dp->dtdo_varlen; i++) {
9737 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9738 dtrace_statvar_t *svar, ***svarp;
9739 size_t dsize = 0;
9740 uint8_t scope = v->dtdv_scope;
9741 int *np;
9742
9743 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9744 continue;
9745
9746 id -= DIF_VAR_OTHER_UBASE;
9747
9748 switch (scope) {
9749 case DIFV_SCOPE_THREAD:
9750 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9751 dtrace_difv_t *tlocals;
9752
9753 if ((ntlocals = (otlocals << 1)) == 0)
9754 ntlocals = 1;
9755
9756 osz = otlocals * sizeof (dtrace_difv_t);
9757 nsz = ntlocals * sizeof (dtrace_difv_t);
9758
9759 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9760
9761 if (osz != 0) {
9762 bcopy(vstate->dtvs_tlocals,
9763 tlocals, osz);
9764 kmem_free(vstate->dtvs_tlocals, osz);
9765 }
9766
9767 vstate->dtvs_tlocals = tlocals;
9768 vstate->dtvs_ntlocals = ntlocals;
9769 }
9770
9771 vstate->dtvs_tlocals[id] = *v;
9772 continue;
9773
9774 case DIFV_SCOPE_LOCAL:
9775 np = &vstate->dtvs_nlocals;
9776 svarp = &vstate->dtvs_locals;
9777
9778 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9779 dsize = NCPU * (v->dtdv_type.dtdt_size +
9780 sizeof (uint64_t));
9781 else
9782 dsize = NCPU * sizeof (uint64_t);
9783
9784 break;
9785
9786 case DIFV_SCOPE_GLOBAL:
9787 np = &vstate->dtvs_nglobals;
9788 svarp = &vstate->dtvs_globals;
9789
9790 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9791 dsize = v->dtdv_type.dtdt_size +
9792 sizeof (uint64_t);
9793
9794 break;
9795
9796 default:
9797 ASSERT(0);
9798 }
9799
9800 while (id >= (oldsvars = *np)) {
9801 dtrace_statvar_t **statics;
9802 int newsvars, oldsize, newsize;
9803
9804 if ((newsvars = (oldsvars << 1)) == 0)
9805 newsvars = 1;
9806
9807 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9808 newsize = newsvars * sizeof (dtrace_statvar_t *);
9809
9810 statics = kmem_zalloc(newsize, KM_SLEEP);
9811
9812 if (oldsize != 0) {
9813 bcopy(*svarp, statics, oldsize);
9814 kmem_free(*svarp, oldsize);
9815 }
9816
9817 *svarp = statics;
9818 *np = newsvars;
9819 }
9820
9821 if ((svar = (*svarp)[id]) == NULL) {
9822 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9823 svar->dtsv_var = *v;
9824
9825 if ((svar->dtsv_size = dsize) != 0) {
9826 svar->dtsv_data = (uint64_t)(uintptr_t)
9827 kmem_zalloc(dsize, KM_SLEEP);
9828 }
9829
9830 (*svarp)[id] = svar;
9831 }
9832
9833 svar->dtsv_refcnt++;
9834 }
9835
9836 dtrace_difo_chunksize(dp, vstate);
9837 dtrace_difo_hold(dp);
9838 }
9839
9840 static dtrace_difo_t *
9841 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9842 {
9843 dtrace_difo_t *new;
9844 size_t sz;
9845
9846 ASSERT(dp->dtdo_buf != NULL);
9847 ASSERT(dp->dtdo_refcnt != 0);
9848
9849 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9850
9851 ASSERT(dp->dtdo_buf != NULL);
9852 sz = dp->dtdo_len * sizeof (dif_instr_t);
9853 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9854 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9855 new->dtdo_len = dp->dtdo_len;
9856
9857 if (dp->dtdo_strtab != NULL) {
9858 ASSERT(dp->dtdo_strlen != 0);
9859 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9860 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9861 new->dtdo_strlen = dp->dtdo_strlen;
9862 }
9863
9864 if (dp->dtdo_inttab != NULL) {
9865 ASSERT(dp->dtdo_intlen != 0);
9866 sz = dp->dtdo_intlen * sizeof (uint64_t);
9867 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9868 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9869 new->dtdo_intlen = dp->dtdo_intlen;
9870 }
9871
9872 if (dp->dtdo_vartab != NULL) {
9873 ASSERT(dp->dtdo_varlen != 0);
9874 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9875 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9876 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9877 new->dtdo_varlen = dp->dtdo_varlen;
9878 }
9879
9880 dtrace_difo_init(new, vstate);
9881 return (new);
9882 }
9883
9884 static void
9885 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9886 {
9887 int i;
9888
9889 ASSERT(dp->dtdo_refcnt == 0);
9890
9891 for (i = 0; i < dp->dtdo_varlen; i++) {
9892 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9893 dtrace_statvar_t *svar, **svarp;
9894 uint_t id;
9895 uint8_t scope = v->dtdv_scope;
9896 int *np;
9897
9898 switch (scope) {
9899 case DIFV_SCOPE_THREAD:
9900 continue;
9901
9902 case DIFV_SCOPE_LOCAL:
9903 np = &vstate->dtvs_nlocals;
9904 svarp = vstate->dtvs_locals;
9905 break;
9906
9907 case DIFV_SCOPE_GLOBAL:
9908 np = &vstate->dtvs_nglobals;
9909 svarp = vstate->dtvs_globals;
9910 break;
9911
9912 default:
9913 ASSERT(0);
9914 }
9915
9916 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9917 continue;
9918
9919 id -= DIF_VAR_OTHER_UBASE;
9920 ASSERT(id < *np);
9921
9922 svar = svarp[id];
9923 ASSERT(svar != NULL);
9924 ASSERT(svar->dtsv_refcnt > 0);
9925
9926 if (--svar->dtsv_refcnt > 0)
9927 continue;
9928
9929 if (svar->dtsv_size != 0) {
9930 ASSERT(svar->dtsv_data != NULL);
9931 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9932 svar->dtsv_size);
9933 }
9934
9935 kmem_free(svar, sizeof (dtrace_statvar_t));
9936 svarp[id] = NULL;
9937 }
9938
9939 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9940 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9941 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9942 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9943
9944 kmem_free(dp, sizeof (dtrace_difo_t));
9945 }
9946
9947 static void
9948 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9949 {
9950 int i;
9951
9952 ASSERT(MUTEX_HELD(&dtrace_lock));
9953 ASSERT(dp->dtdo_refcnt != 0);
9954
9955 for (i = 0; i < dp->dtdo_varlen; i++) {
9956 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9957
9958 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9959 continue;
9960
9961 ASSERT(dtrace_vtime_references > 0);
9962 if (--dtrace_vtime_references == 0)
9963 dtrace_vtime_disable();
9964 }
9965
9966 if (--dp->dtdo_refcnt == 0)
9967 dtrace_difo_destroy(dp, vstate);
9968 }
9969
9970 /*
9971 * DTrace Format Functions
9972 */
9973 static uint16_t
9974 dtrace_format_add(dtrace_state_t *state, char *str)
9975 {
9976 char *fmt, **new;
9977 uint16_t ndx, len = strlen(str) + 1;
9978
9979 fmt = kmem_zalloc(len, KM_SLEEP);
9980 bcopy(str, fmt, len);
9981
9982 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9983 if (state->dts_formats[ndx] == NULL) {
9984 state->dts_formats[ndx] = fmt;
9985 return (ndx + 1);
9986 }
9987 }
9988
9989 if (state->dts_nformats == USHRT_MAX) {
9990 /*
9991 * This is only likely if a denial-of-service attack is being
9992 * attempted. As such, it's okay to fail silently here.
9993 */
9994 kmem_free(fmt, len);
9995 return (0);
9996 }
9997
9998 /*
9999 * For simplicity, we always resize the formats array to be exactly the
10000 * number of formats.
10001 */
10002 ndx = state->dts_nformats++;
10003 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10004
10005 if (state->dts_formats != NULL) {
10006 ASSERT(ndx != 0);
10007 bcopy(state->dts_formats, new, ndx * sizeof (char *));
10008 kmem_free(state->dts_formats, ndx * sizeof (char *));
10009 }
10010
10011 state->dts_formats = new;
10012 state->dts_formats[ndx] = fmt;
10013
10014 return (ndx + 1);
10015 }
10016
10017 static void
10018 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10019 {
10020 char *fmt;
10021
10022 ASSERT(state->dts_formats != NULL);
10023 ASSERT(format <= state->dts_nformats);
10024 ASSERT(state->dts_formats[format - 1] != NULL);
10025
10026 fmt = state->dts_formats[format - 1];
10027 kmem_free(fmt, strlen(fmt) + 1);
10028 state->dts_formats[format - 1] = NULL;
10029 }
10030
10031 static void
10032 dtrace_format_destroy(dtrace_state_t *state)
10033 {
10034 int i;
10035
10036 if (state->dts_nformats == 0) {
10037 ASSERT(state->dts_formats == NULL);
10038 return;
10039 }
10040
10041 ASSERT(state->dts_formats != NULL);
10042
10043 for (i = 0; i < state->dts_nformats; i++) {
10044 char *fmt = state->dts_formats[i];
10045
10046 if (fmt == NULL)
10047 continue;
10048
10049 kmem_free(fmt, strlen(fmt) + 1);
10050 }
10051
10052 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10053 state->dts_nformats = 0;
10054 state->dts_formats = NULL;
10055 }
10056
10057 /*
10058 * DTrace Predicate Functions
10059 */
10060 static dtrace_predicate_t *
10061 dtrace_predicate_create(dtrace_difo_t *dp)
10062 {
10063 dtrace_predicate_t *pred;
10064
10065 ASSERT(MUTEX_HELD(&dtrace_lock));
10066 ASSERT(dp->dtdo_refcnt != 0);
10067
10068 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10069 pred->dtp_difo = dp;
10070 pred->dtp_refcnt = 1;
10071
10072 if (!dtrace_difo_cacheable(dp))
10073 return (pred);
10074
10075 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10076 /*
10077 * This is only theoretically possible -- we have had 2^32
10078 * cacheable predicates on this machine. We cannot allow any
10079 * more predicates to become cacheable: as unlikely as it is,
10080 * there may be a thread caching a (now stale) predicate cache
10081 * ID. (N.B.: the temptation is being successfully resisted to
10082 * have this cmn_err() "Holy shit -- we executed this code!")
10083 */
10084 return (pred);
10085 }
10086
10087 pred->dtp_cacheid = dtrace_predcache_id++;
10088
10089 return (pred);
10090 }
10091
10092 static void
10093 dtrace_predicate_hold(dtrace_predicate_t *pred)
10094 {
10095 ASSERT(MUTEX_HELD(&dtrace_lock));
10096 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10097 ASSERT(pred->dtp_refcnt > 0);
10098
10099 pred->dtp_refcnt++;
10100 }
10101
10102 static void
10103 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10104 {
10105 dtrace_difo_t *dp = pred->dtp_difo;
10106
10107 ASSERT(MUTEX_HELD(&dtrace_lock));
10108 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
10109 ASSERT(pred->dtp_refcnt > 0);
10110
10111 if (--pred->dtp_refcnt == 0) {
10112 dtrace_difo_release(pred->dtp_difo, vstate);
10113 kmem_free(pred, sizeof (dtrace_predicate_t));
10114 }
10115 }
10116
10117 /*
10118 * DTrace Action Description Functions
10119 */
10120 static dtrace_actdesc_t *
10121 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
10122 uint64_t uarg, uint64_t arg)
10123 {
10124 dtrace_actdesc_t *act;
10125
10126 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
10127 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
10128
10129 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
10130 act->dtad_kind = kind;
10131 act->dtad_ntuple = ntuple;
10132 act->dtad_uarg = uarg;
10133 act->dtad_arg = arg;
10134 act->dtad_refcnt = 1;
10135
10136 return (act);
10137 }
10138
10139 static void
10140 dtrace_actdesc_hold(dtrace_actdesc_t *act)
10141 {
10142 ASSERT(act->dtad_refcnt >= 1);
10143 act->dtad_refcnt++;
10144 }
10145
10146 static void
10147 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
10148 {
10149 dtrace_actkind_t kind = act->dtad_kind;
10150 dtrace_difo_t *dp;
10151
10152 ASSERT(act->dtad_refcnt >= 1);
10153
10154 if (--act->dtad_refcnt != 0)
10155 return;
10156
10157 if ((dp = act->dtad_difo) != NULL)
10158 dtrace_difo_release(dp, vstate);
10159
10160 if (DTRACEACT_ISPRINTFLIKE(kind)) {
10161 char *str = (char *)(uintptr_t)act->dtad_arg;
10162
10163 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
10164 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
10165
10166 if (str != NULL)
10167 kmem_free(str, strlen(str) + 1);
10168 }
10169
10170 kmem_free(act, sizeof (dtrace_actdesc_t));
10171 }
10172
10173 /*
10174 * DTrace ECB Functions
10175 */
10176 static dtrace_ecb_t *
10177 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
10178 {
10179 dtrace_ecb_t *ecb;
10180 dtrace_epid_t epid;
10181
10182 ASSERT(MUTEX_HELD(&dtrace_lock));
10183
10184 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
10185 ecb->dte_predicate = NULL;
10186 ecb->dte_probe = probe;
10187
10188 /*
10189 * The default size is the size of the default action: recording
10190 * the header.
10191 */
10192 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
10193 ecb->dte_alignment = sizeof (dtrace_epid_t);
10194
10195 epid = state->dts_epid++;
10196
10197 if (epid - 1 >= state->dts_necbs) {
10198 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
10199 int necbs = state->dts_necbs << 1;
10200
10201 ASSERT(epid == state->dts_necbs + 1);
10202
10203 if (necbs == 0) {
10204 ASSERT(oecbs == NULL);
10205 necbs = 1;
10206 }
10207
10208 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
10209
10210 if (oecbs != NULL)
10211 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
10212
10213 dtrace_membar_producer();
10214 state->dts_ecbs = ecbs;
10215
10216 if (oecbs != NULL) {
10217 /*
10218 * If this state is active, we must dtrace_sync()
10219 * before we can free the old dts_ecbs array: we're
10220 * coming in hot, and there may be active ring
10221 * buffer processing (which indexes into the dts_ecbs
10222 * array) on another CPU.
10223 */
10224 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
10225 dtrace_sync();
10226
10227 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
10228 }
10229
10230 dtrace_membar_producer();
10231 state->dts_necbs = necbs;
10232 }
10233
10234 ecb->dte_state = state;
10235
10236 ASSERT(state->dts_ecbs[epid - 1] == NULL);
10237 dtrace_membar_producer();
10238 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
10239
10240 return (ecb);
10241 }
10242
10243 static int
10244 dtrace_ecb_enable(dtrace_ecb_t *ecb)
10245 {
10246 dtrace_probe_t *probe = ecb->dte_probe;
10247
10248 ASSERT(MUTEX_HELD(&cpu_lock));
10249 ASSERT(MUTEX_HELD(&dtrace_lock));
10250 ASSERT(ecb->dte_next == NULL);
10251
10252 if (probe == NULL) {
10253 /*
10254 * This is the NULL probe -- there's nothing to do.
10255 */
10256 return (0);
10257 }
10258
10259 if (probe->dtpr_ecb == NULL) {
10260 dtrace_provider_t *prov = probe->dtpr_provider;
10261
10262 /*
10263 * We're the first ECB on this probe.
10264 */
10265 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
10266
10267 if (ecb->dte_predicate != NULL)
10268 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
10269
10270 return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
10271 probe->dtpr_id, probe->dtpr_arg));
10272 } else {
10273 /*
10274 * This probe is already active. Swing the last pointer to
10275 * point to the new ECB, and issue a dtrace_sync() to assure
10276 * that all CPUs have seen the change.
10277 */
10278 ASSERT(probe->dtpr_ecb_last != NULL);
10279 probe->dtpr_ecb_last->dte_next = ecb;
10280 probe->dtpr_ecb_last = ecb;
10281 probe->dtpr_predcache = 0;
10282
10283 dtrace_sync();
10284 return (0);
10285 }
10286 }
10287
10288 static void
10289 dtrace_ecb_resize(dtrace_ecb_t *ecb)
10290 {
10291 dtrace_action_t *act;
10292 uint32_t curneeded = UINT32_MAX;
10293 uint32_t aggbase = UINT32_MAX;
10294
10295 /*
10296 * If we record anything, we always record the dtrace_rechdr_t. (And
10297 * we always record it first.)
10298 */
10299 ecb->dte_size = sizeof (dtrace_rechdr_t);
10300 ecb->dte_alignment = sizeof (dtrace_epid_t);
10301
10302 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10303 dtrace_recdesc_t *rec = &act->dta_rec;
10304 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
10305
10306 ecb->dte_alignment = MAX(ecb->dte_alignment,
10307 rec->dtrd_alignment);
10308
10309 if (DTRACEACT_ISAGG(act->dta_kind)) {
10310 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10311
10312 ASSERT(rec->dtrd_size != 0);
10313 ASSERT(agg->dtag_first != NULL);
10314 ASSERT(act->dta_prev->dta_intuple);
10315 ASSERT(aggbase != UINT32_MAX);
10316 ASSERT(curneeded != UINT32_MAX);
10317
10318 agg->dtag_base = aggbase;
10319
10320 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10321 rec->dtrd_offset = curneeded;
10322 curneeded += rec->dtrd_size;
10323 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
10324
10325 aggbase = UINT32_MAX;
10326 curneeded = UINT32_MAX;
10327 } else if (act->dta_intuple) {
10328 if (curneeded == UINT32_MAX) {
10329 /*
10330 * This is the first record in a tuple. Align
10331 * curneeded to be at offset 4 in an 8-byte
10332 * aligned block.
10333 */
10334 ASSERT(act->dta_prev == NULL ||
10335 !act->dta_prev->dta_intuple);
10336 ASSERT3U(aggbase, ==, UINT32_MAX);
10337 curneeded = P2PHASEUP(ecb->dte_size,
10338 sizeof (uint64_t), sizeof (dtrace_aggid_t));
10339
10340 aggbase = curneeded - sizeof (dtrace_aggid_t);
10341 ASSERT(IS_P2ALIGNED(aggbase,
10342 sizeof (uint64_t)));
10343 }
10344 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10345 rec->dtrd_offset = curneeded;
10346 curneeded += rec->dtrd_size;
10347 } else {
10348 /* tuples must be followed by an aggregation */
10349 ASSERT(act->dta_prev == NULL ||
10350 !act->dta_prev->dta_intuple);
10351
10352 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10353 rec->dtrd_alignment);
10354 rec->dtrd_offset = ecb->dte_size;
10355 ecb->dte_size += rec->dtrd_size;
10356 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10357 }
10358 }
10359
10360 if ((act = ecb->dte_action) != NULL &&
10361 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10362 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10363 /*
10364 * If the size is still sizeof (dtrace_rechdr_t), then all
10365 * actions store no data; set the size to 0.
10366 */
10367 ecb->dte_size = 0;
10368 }
10369
10370 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10371 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10372 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10373 ecb->dte_needed);
10374 }
10375
10376 static dtrace_action_t *
10377 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10378 {
10379 dtrace_aggregation_t *agg;
10380 size_t size = sizeof (uint64_t);
10381 int ntuple = desc->dtad_ntuple;
10382 dtrace_action_t *act;
10383 dtrace_recdesc_t *frec;
10384 dtrace_aggid_t aggid;
10385 dtrace_state_t *state = ecb->dte_state;
10386
10387 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10388 agg->dtag_ecb = ecb;
10389
10390 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10391
10392 switch (desc->dtad_kind) {
10393 case DTRACEAGG_MIN:
10394 agg->dtag_initial = INT64_MAX;
10395 agg->dtag_aggregate = dtrace_aggregate_min;
10396 break;
10397
10398 case DTRACEAGG_MAX:
10399 agg->dtag_initial = INT64_MIN;
10400 agg->dtag_aggregate = dtrace_aggregate_max;
10401 break;
10402
10403 case DTRACEAGG_COUNT:
10404 agg->dtag_aggregate = dtrace_aggregate_count;
10405 break;
10406
10407 case DTRACEAGG_QUANTIZE:
10408 agg->dtag_aggregate = dtrace_aggregate_quantize;
10409 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10410 sizeof (uint64_t);
10411 break;
10412
10413 case DTRACEAGG_LQUANTIZE: {
10414 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10415 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10416
10417 agg->dtag_initial = desc->dtad_arg;
10418 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10419
10420 if (step == 0 || levels == 0)
10421 goto err;
10422
10423 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10424 break;
10425 }
10426
10427 case DTRACEAGG_LLQUANTIZE: {
10428 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10429 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10430 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10431 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10432 int64_t v;
10433
10434 agg->dtag_initial = desc->dtad_arg;
10435 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10436
10437 if (factor < 2 || low >= high || nsteps < factor)
10438 goto err;
10439
10440 /*
10441 * Now check that the number of steps evenly divides a power
10442 * of the factor. (This assures both integer bucket size and
10443 * linearity within each magnitude.)
10444 */
10445 for (v = factor; v < nsteps; v *= factor)
10446 continue;
10447
10448 if ((v % nsteps) || (nsteps % factor))
10449 goto err;
10450
10451 size = (dtrace_aggregate_llquantize_bucket(factor,
10452 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10453 break;
10454 }
10455
10456 case DTRACEAGG_AVG:
10457 agg->dtag_aggregate = dtrace_aggregate_avg;
10458 size = sizeof (uint64_t) * 2;
10459 break;
10460
10461 case DTRACEAGG_STDDEV:
10462 agg->dtag_aggregate = dtrace_aggregate_stddev;
10463 size = sizeof (uint64_t) * 4;
10464 break;
10465
10466 case DTRACEAGG_SUM:
10467 agg->dtag_aggregate = dtrace_aggregate_sum;
10468 break;
10469
10470 default:
10471 goto err;
10472 }
10473
10474 agg->dtag_action.dta_rec.dtrd_size = size;
10475
10476 if (ntuple == 0)
10477 goto err;
10478
10479 /*
10480 * We must make sure that we have enough actions for the n-tuple.
10481 */
10482 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10483 if (DTRACEACT_ISAGG(act->dta_kind))
10484 break;
10485
10486 if (--ntuple == 0) {
10487 /*
10488 * This is the action with which our n-tuple begins.
10489 */
10490 agg->dtag_first = act;
10491 goto success;
10492 }
10493 }
10494
10495 /*
10496 * This n-tuple is short by ntuple elements. Return failure.
10497 */
10498 ASSERT(ntuple != 0);
10499 err:
10500 kmem_free(agg, sizeof (dtrace_aggregation_t));
10501 return (NULL);
10502
10503 success:
10504 /*
10505 * If the last action in the tuple has a size of zero, it's actually
10506 * an expression argument for the aggregating action.
10507 */
10508 ASSERT(ecb->dte_action_last != NULL);
10509 act = ecb->dte_action_last;
10510
10511 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10512 ASSERT(act->dta_difo != NULL);
10513
10514 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10515 agg->dtag_hasarg = 1;
10516 }
10517
10518 /*
10519 * We need to allocate an id for this aggregation.
10520 */
10521 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10522 VM_BESTFIT | VM_SLEEP);
10523
10524 if (aggid - 1 >= state->dts_naggregations) {
10525 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10526 dtrace_aggregation_t **aggs;
10527 int naggs = state->dts_naggregations << 1;
10528 int onaggs = state->dts_naggregations;
10529
10530 ASSERT(aggid == state->dts_naggregations + 1);
10531
10532 if (naggs == 0) {
10533 ASSERT(oaggs == NULL);
10534 naggs = 1;
10535 }
10536
10537 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10538
10539 if (oaggs != NULL) {
10540 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10541 kmem_free(oaggs, onaggs * sizeof (*aggs));
10542 }
10543
10544 state->dts_aggregations = aggs;
10545 state->dts_naggregations = naggs;
10546 }
10547
10548 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10549 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10550
10551 frec = &agg->dtag_first->dta_rec;
10552 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10553 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10554
10555 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10556 ASSERT(!act->dta_intuple);
10557 act->dta_intuple = 1;
10558 }
10559
10560 return (&agg->dtag_action);
10561 }
10562
10563 static void
10564 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10565 {
10566 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10567 dtrace_state_t *state = ecb->dte_state;
10568 dtrace_aggid_t aggid = agg->dtag_id;
10569
10570 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10571 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10572
10573 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10574 state->dts_aggregations[aggid - 1] = NULL;
10575
10576 kmem_free(agg, sizeof (dtrace_aggregation_t));
10577 }
10578
10579 static int
10580 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10581 {
10582 dtrace_action_t *action, *last;
10583 dtrace_difo_t *dp = desc->dtad_difo;
10584 uint32_t size = 0, align = sizeof (uint8_t), mask;
10585 uint16_t format = 0;
10586 dtrace_recdesc_t *rec;
10587 dtrace_state_t *state = ecb->dte_state;
10588 dtrace_optval_t *opt = state->dts_options, nframes, strsize;
10589 uint64_t arg = desc->dtad_arg;
10590
10591 ASSERT(MUTEX_HELD(&dtrace_lock));
10592 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10593
10594 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10595 /*
10596 * If this is an aggregating action, there must be neither
10597 * a speculate nor a commit on the action chain.
10598 */
10599 dtrace_action_t *act;
10600
10601 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10602 if (act->dta_kind == DTRACEACT_COMMIT)
10603 return (EINVAL);
10604
10605 if (act->dta_kind == DTRACEACT_SPECULATE)
10606 return (EINVAL);
10607 }
10608
10609 action = dtrace_ecb_aggregation_create(ecb, desc);
10610
10611 if (action == NULL)
10612 return (EINVAL);
10613 } else {
10614 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10615 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10616 dp != NULL && dp->dtdo_destructive)) {
10617 state->dts_destructive = 1;
10618 }
10619
10620 switch (desc->dtad_kind) {
10621 case DTRACEACT_PRINTF:
10622 case DTRACEACT_PRINTA:
10623 case DTRACEACT_SYSTEM:
10624 case DTRACEACT_FREOPEN:
10625 case DTRACEACT_DIFEXPR:
10626 /*
10627 * We know that our arg is a string -- turn it into a
10628 * format.
10629 */
10630 if (arg == NULL) {
10631 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10632 desc->dtad_kind == DTRACEACT_DIFEXPR);
10633 format = 0;
10634 } else {
10635 ASSERT(arg != NULL);
10636 ASSERT(arg > KERNELBASE);
10637 format = dtrace_format_add(state,
10638 (char *)(uintptr_t)arg);
10639 }
10640
10641 /*FALLTHROUGH*/
10642 case DTRACEACT_LIBACT:
10643 case DTRACEACT_TRACEMEM:
10644 case DTRACEACT_TRACEMEM_DYNSIZE:
10645 if (dp == NULL)
10646 return (EINVAL);
10647
10648 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10649 break;
10650
10651 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10652 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10653 return (EINVAL);
10654
10655 size = opt[DTRACEOPT_STRSIZE];
10656 }
10657
10658 break;
10659
10660 case DTRACEACT_STACK:
10661 if ((nframes = arg) == 0) {
10662 nframes = opt[DTRACEOPT_STACKFRAMES];
10663 ASSERT(nframes > 0);
10664 arg = nframes;
10665 }
10666
10667 size = nframes * sizeof (pc_t);
10668 break;
10669
10670 case DTRACEACT_JSTACK:
10671 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10672 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10673
10674 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10675 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10676
10677 arg = DTRACE_USTACK_ARG(nframes, strsize);
10678
10679 /*FALLTHROUGH*/
10680 case DTRACEACT_USTACK:
10681 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10682 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10683 strsize = DTRACE_USTACK_STRSIZE(arg);
10684 nframes = opt[DTRACEOPT_USTACKFRAMES];
10685 ASSERT(nframes > 0);
10686 arg = DTRACE_USTACK_ARG(nframes, strsize);
10687 }
10688
10689 /*
10690 * Save a slot for the pid.
10691 */
10692 size = (nframes + 1) * sizeof (uint64_t);
10693 size += DTRACE_USTACK_STRSIZE(arg);
10694 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10695
10696 break;
10697
10698 case DTRACEACT_SYM:
10699 case DTRACEACT_MOD:
10700 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10701 sizeof (uint64_t)) ||
10702 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10703 return (EINVAL);
10704 break;
10705
10706 case DTRACEACT_USYM:
10707 case DTRACEACT_UMOD:
10708 case DTRACEACT_UADDR:
10709 if (dp == NULL ||
10710 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10711 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10712 return (EINVAL);
10713
10714 /*
10715 * We have a slot for the pid, plus a slot for the
10716 * argument. To keep things simple (aligned with
10717 * bitness-neutral sizing), we store each as a 64-bit
10718 * quantity.
10719 */
10720 size = 2 * sizeof (uint64_t);
10721 break;
10722
10723 case DTRACEACT_STOP:
10724 case DTRACEACT_BREAKPOINT:
10725 case DTRACEACT_PANIC:
10726 break;
10727
10728 case DTRACEACT_CHILL:
10729 case DTRACEACT_DISCARD:
10730 case DTRACEACT_RAISE:
10731 if (dp == NULL)
10732 return (EINVAL);
10733 break;
10734
10735 case DTRACEACT_EXIT:
10736 if (dp == NULL ||
10737 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10738 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10739 return (EINVAL);
10740 break;
10741
10742 case DTRACEACT_SPECULATE:
10743 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10744 return (EINVAL);
10745
10746 if (dp == NULL)
10747 return (EINVAL);
10748
10749 state->dts_speculates = 1;
10750 break;
10751
10752 case DTRACEACT_COMMIT: {
10753 dtrace_action_t *act = ecb->dte_action;
10754
10755 for (; act != NULL; act = act->dta_next) {
10756 if (act->dta_kind == DTRACEACT_COMMIT)
10757 return (EINVAL);
10758 }
10759
10760 if (dp == NULL)
10761 return (EINVAL);
10762 break;
10763 }
10764
10765 default:
10766 return (EINVAL);
10767 }
10768
10769 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10770 /*
10771 * If this is a data-storing action or a speculate,
10772 * we must be sure that there isn't a commit on the
10773 * action chain.
10774 */
10775 dtrace_action_t *act = ecb->dte_action;
10776
10777 for (; act != NULL; act = act->dta_next) {
10778 if (act->dta_kind == DTRACEACT_COMMIT)
10779 return (EINVAL);
10780 }
10781 }
10782
10783 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10784 action->dta_rec.dtrd_size = size;
10785 }
10786
10787 action->dta_refcnt = 1;
10788 rec = &action->dta_rec;
10789 size = rec->dtrd_size;
10790
10791 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10792 if (!(size & mask)) {
10793 align = mask + 1;
10794 break;
10795 }
10796 }
10797
10798 action->dta_kind = desc->dtad_kind;
10799
10800 if ((action->dta_difo = dp) != NULL)
10801 dtrace_difo_hold(dp);
10802
10803 rec->dtrd_action = action->dta_kind;
10804 rec->dtrd_arg = arg;
10805 rec->dtrd_uarg = desc->dtad_uarg;
10806 rec->dtrd_alignment = (uint16_t)align;
10807 rec->dtrd_format = format;
10808
10809 if ((last = ecb->dte_action_last) != NULL) {
10810 ASSERT(ecb->dte_action != NULL);
10811 action->dta_prev = last;
10812 last->dta_next = action;
10813 } else {
10814 ASSERT(ecb->dte_action == NULL);
10815 ecb->dte_action = action;
10816 }
10817
10818 ecb->dte_action_last = action;
10819
10820 return (0);
10821 }
10822
10823 static void
10824 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10825 {
10826 dtrace_action_t *act = ecb->dte_action, *next;
10827 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10828 dtrace_difo_t *dp;
10829 uint16_t format;
10830
10831 if (act != NULL && act->dta_refcnt > 1) {
10832 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10833 act->dta_refcnt--;
10834 } else {
10835 for (; act != NULL; act = next) {
10836 next = act->dta_next;
10837 ASSERT(next != NULL || act == ecb->dte_action_last);
10838 ASSERT(act->dta_refcnt == 1);
10839
10840 if ((format = act->dta_rec.dtrd_format) != 0)
10841 dtrace_format_remove(ecb->dte_state, format);
10842
10843 if ((dp = act->dta_difo) != NULL)
10844 dtrace_difo_release(dp, vstate);
10845
10846 if (DTRACEACT_ISAGG(act->dta_kind)) {
10847 dtrace_ecb_aggregation_destroy(ecb, act);
10848 } else {
10849 kmem_free(act, sizeof (dtrace_action_t));
10850 }
10851 }
10852 }
10853
10854 ecb->dte_action = NULL;
10855 ecb->dte_action_last = NULL;
10856 ecb->dte_size = 0;
10857 }
10858
10859 static void
10860 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10861 {
10862 /*
10863 * We disable the ECB by removing it from its probe.
10864 */
10865 dtrace_ecb_t *pecb, *prev = NULL;
10866 dtrace_probe_t *probe = ecb->dte_probe;
10867
10868 ASSERT(MUTEX_HELD(&dtrace_lock));
10869
10870 if (probe == NULL) {
10871 /*
10872 * This is the NULL probe; there is nothing to disable.
10873 */
10874 return;
10875 }
10876
10877 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10878 if (pecb == ecb)
10879 break;
10880 prev = pecb;
10881 }
10882
10883 ASSERT(pecb != NULL);
10884
10885 if (prev == NULL) {
10886 probe->dtpr_ecb = ecb->dte_next;
10887 } else {
10888 prev->dte_next = ecb->dte_next;
10889 }
10890
10891 if (ecb == probe->dtpr_ecb_last) {
10892 ASSERT(ecb->dte_next == NULL);
10893 probe->dtpr_ecb_last = prev;
10894 }
10895
10896 /*
10897 * The ECB has been disconnected from the probe; now sync to assure
10898 * that all CPUs have seen the change before returning.
10899 */
10900 dtrace_sync();
10901
10902 if (probe->dtpr_ecb == NULL) {
10903 /*
10904 * That was the last ECB on the probe; clear the predicate
10905 * cache ID for the probe, disable it and sync one more time
10906 * to assure that we'll never hit it again.
10907 */
10908 dtrace_provider_t *prov = probe->dtpr_provider;
10909
10910 ASSERT(ecb->dte_next == NULL);
10911 ASSERT(probe->dtpr_ecb_last == NULL);
10912 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10913 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10914 probe->dtpr_id, probe->dtpr_arg);
10915 dtrace_sync();
10916 } else {
10917 /*
10918 * There is at least one ECB remaining on the probe. If there
10919 * is _exactly_ one, set the probe's predicate cache ID to be
10920 * the predicate cache ID of the remaining ECB.
10921 */
10922 ASSERT(probe->dtpr_ecb_last != NULL);
10923 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10924
10925 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10926 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10927
10928 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10929
10930 if (p != NULL)
10931 probe->dtpr_predcache = p->dtp_cacheid;
10932 }
10933
10934 ecb->dte_next = NULL;
10935 }
10936 }
10937
10938 static void
10939 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10940 {
10941 dtrace_state_t *state = ecb->dte_state;
10942 dtrace_vstate_t *vstate = &state->dts_vstate;
10943 dtrace_predicate_t *pred;
10944 dtrace_epid_t epid = ecb->dte_epid;
10945
10946 ASSERT(MUTEX_HELD(&dtrace_lock));
10947 ASSERT(ecb->dte_next == NULL);
10948 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10949
10950 if ((pred = ecb->dte_predicate) != NULL)
10951 dtrace_predicate_release(pred, vstate);
10952
10953 dtrace_ecb_action_remove(ecb);
10954
10955 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10956 state->dts_ecbs[epid - 1] = NULL;
10957
10958 kmem_free(ecb, sizeof (dtrace_ecb_t));
10959 }
10960
10961 static dtrace_ecb_t *
10962 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10963 dtrace_enabling_t *enab)
10964 {
10965 dtrace_ecb_t *ecb;
10966 dtrace_predicate_t *pred;
10967 dtrace_actdesc_t *act;
10968 dtrace_provider_t *prov;
10969 dtrace_ecbdesc_t *desc = enab->dten_current;
10970
10971 ASSERT(MUTEX_HELD(&dtrace_lock));
10972 ASSERT(state != NULL);
10973
10974 ecb = dtrace_ecb_add(state, probe);
10975 ecb->dte_uarg = desc->dted_uarg;
10976
10977 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10978 dtrace_predicate_hold(pred);
10979 ecb->dte_predicate = pred;
10980 }
10981
10982 if (probe != NULL) {
10983 /*
10984 * If the provider shows more leg than the consumer is old
10985 * enough to see, we need to enable the appropriate implicit
10986 * predicate bits to prevent the ecb from activating at
10987 * revealing times.
10988 *
10989 * Providers specifying DTRACE_PRIV_USER at register time
10990 * are stating that they need the /proc-style privilege
10991 * model to be enforced, and this is what DTRACE_COND_OWNER
10992 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10993 */
10994 prov = probe->dtpr_provider;
10995 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10996 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10997 ecb->dte_cond |= DTRACE_COND_OWNER;
10998
10999 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11000 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11001 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11002
11003 /*
11004 * If the provider shows us kernel innards and the user
11005 * is lacking sufficient privilege, enable the
11006 * DTRACE_COND_USERMODE implicit predicate.
11007 */
11008 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11009 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11010 ecb->dte_cond |= DTRACE_COND_USERMODE;
11011 }
11012
11013 if (dtrace_ecb_create_cache != NULL) {
11014 /*
11015 * If we have a cached ecb, we'll use its action list instead
11016 * of creating our own (saving both time and space).
11017 */
11018 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11019 dtrace_action_t *act = cached->dte_action;
11020
11021 if (act != NULL) {
11022 ASSERT(act->dta_refcnt > 0);
11023 act->dta_refcnt++;
11024 ecb->dte_action = act;
11025 ecb->dte_action_last = cached->dte_action_last;
11026 ecb->dte_needed = cached->dte_needed;
11027 ecb->dte_size = cached->dte_size;
11028 ecb->dte_alignment = cached->dte_alignment;
11029 }
11030
11031 return (ecb);
11032 }
11033
11034 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11035 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11036 dtrace_ecb_destroy(ecb);
11037 return (NULL);
11038 }
11039 }
11040
11041 dtrace_ecb_resize(ecb);
11042
11043 return (dtrace_ecb_create_cache = ecb);
11044 }
11045
11046 static int
11047 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11048 {
11049 dtrace_ecb_t *ecb;
11050 dtrace_enabling_t *enab = arg;
11051 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11052
11053 ASSERT(state != NULL);
11054
11055 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11056 /*
11057 * This probe was created in a generation for which this
11058 * enabling has previously created ECBs; we don't want to
11059 * enable it again, so just kick out.
11060 */
11061 return (DTRACE_MATCH_NEXT);
11062 }
11063
11064 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11065 return (DTRACE_MATCH_DONE);
11066
11067 if (dtrace_ecb_enable(ecb) < 0)
11068 return (DTRACE_MATCH_FAIL);
11069
11070 return (DTRACE_MATCH_NEXT);
11071 }
11072
11073 static dtrace_ecb_t *
11074 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11075 {
11076 dtrace_ecb_t *ecb;
11077
11078 ASSERT(MUTEX_HELD(&dtrace_lock));
11079
11080 if (id == 0 || id > state->dts_necbs)
11081 return (NULL);
11082
11083 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
11084 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
11085
11086 return (state->dts_ecbs[id - 1]);
11087 }
11088
11089 static dtrace_aggregation_t *
11090 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
11091 {
11092 dtrace_aggregation_t *agg;
11093
11094 ASSERT(MUTEX_HELD(&dtrace_lock));
11095
11096 if (id == 0 || id > state->dts_naggregations)
11097 return (NULL);
11098
11099 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
11100 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
11101 agg->dtag_id == id);
11102
11103 return (state->dts_aggregations[id - 1]);
11104 }
11105
11106 /*
11107 * DTrace Buffer Functions
11108 *
11109 * The following functions manipulate DTrace buffers. Most of these functions
11110 * are called in the context of establishing or processing consumer state;
11111 * exceptions are explicitly noted.
11112 */
11113
11114 /*
11115 * Note: called from cross call context. This function switches the two
11116 * buffers on a given CPU. The atomicity of this operation is assured by
11117 * disabling interrupts while the actual switch takes place; the disabling of
11118 * interrupts serializes the execution with any execution of dtrace_probe() on
11119 * the same CPU.
11120 */
11121 static void
11122 dtrace_buffer_switch(dtrace_buffer_t *buf)
11123 {
11124 caddr_t tomax = buf->dtb_tomax;
11125 caddr_t xamot = buf->dtb_xamot;
11126 dtrace_icookie_t cookie;
11127 hrtime_t now;
11128
11129 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11130 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
11131
11132 cookie = dtrace_interrupt_disable();
11133 now = dtrace_gethrtime();
11134 buf->dtb_tomax = xamot;
11135 buf->dtb_xamot = tomax;
11136 buf->dtb_xamot_drops = buf->dtb_drops;
11137 buf->dtb_xamot_offset = buf->dtb_offset;
11138 buf->dtb_xamot_errors = buf->dtb_errors;
11139 buf->dtb_xamot_flags = buf->dtb_flags;
11140 buf->dtb_offset = 0;
11141 buf->dtb_drops = 0;
11142 buf->dtb_errors = 0;
11143 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
11144 buf->dtb_interval = now - buf->dtb_switched;
11145 buf->dtb_switched = now;
11146 dtrace_interrupt_enable(cookie);
11147 }
11148
11149 /*
11150 * Note: called from cross call context. This function activates a buffer
11151 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
11152 * is guaranteed by the disabling of interrupts.
11153 */
11154 static void
11155 dtrace_buffer_activate(dtrace_state_t *state)
11156 {
11157 dtrace_buffer_t *buf;
11158 dtrace_icookie_t cookie = dtrace_interrupt_disable();
11159
11160 buf = &state->dts_buffer[CPU->cpu_id];
11161
11162 if (buf->dtb_tomax != NULL) {
11163 /*
11164 * We might like to assert that the buffer is marked inactive,
11165 * but this isn't necessarily true: the buffer for the CPU
11166 * that processes the BEGIN probe has its buffer activated
11167 * manually. In this case, we take the (harmless) action
11168 * re-clearing the bit INACTIVE bit.
11169 */
11170 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
11171 }
11172
11173 dtrace_interrupt_enable(cookie);
11174 }
11175
11176 static int
11177 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
11178 processorid_t cpu, int *factor)
11179 {
11180 cpu_t *cp;
11181 dtrace_buffer_t *buf;
11182 int allocated = 0, desired = 0;
11183
11184 ASSERT(MUTEX_HELD(&cpu_lock));
11185 ASSERT(MUTEX_HELD(&dtrace_lock));
11186
11187 *factor = 1;
11188
11189 if (size > dtrace_nonroot_maxsize &&
11190 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
11191 return (EFBIG);
11192
11193 cp = cpu_list;
11194
11195 do {
11196 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11197 continue;
11198
11199 buf = &bufs[cp->cpu_id];
11200
11201 /*
11202 * If there is already a buffer allocated for this CPU, it
11203 * is only possible that this is a DR event. In this case,
11204 * the buffer size must match our specified size.
11205 */
11206 if (buf->dtb_tomax != NULL) {
11207 ASSERT(buf->dtb_size == size);
11208 continue;
11209 }
11210
11211 ASSERT(buf->dtb_xamot == NULL);
11212
11213 if ((buf->dtb_tomax = kmem_zalloc(size,
11214 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11215 goto err;
11216
11217 buf->dtb_size = size;
11218 buf->dtb_flags = flags;
11219 buf->dtb_offset = 0;
11220 buf->dtb_drops = 0;
11221
11222 if (flags & DTRACEBUF_NOSWITCH)
11223 continue;
11224
11225 if ((buf->dtb_xamot = kmem_zalloc(size,
11226 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11227 goto err;
11228 } while ((cp = cp->cpu_next) != cpu_list);
11229
11230 return (0);
11231
11232 err:
11233 cp = cpu_list;
11234
11235 do {
11236 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11237 continue;
11238
11239 buf = &bufs[cp->cpu_id];
11240 desired += 2;
11241
11242 if (buf->dtb_xamot != NULL) {
11243 ASSERT(buf->dtb_tomax != NULL);
11244 ASSERT(buf->dtb_size == size);
11245 kmem_free(buf->dtb_xamot, size);
11246 allocated++;
11247 }
11248
11249 if (buf->dtb_tomax != NULL) {
11250 ASSERT(buf->dtb_size == size);
11251 kmem_free(buf->dtb_tomax, size);
11252 allocated++;
11253 }
11254
11255 buf->dtb_tomax = NULL;
11256 buf->dtb_xamot = NULL;
11257 buf->dtb_size = 0;
11258 } while ((cp = cp->cpu_next) != cpu_list);
11259
11260 *factor = desired / (allocated > 0 ? allocated : 1);
11261
11262 return (ENOMEM);
11263 }
11264
11265 /*
11266 * Note: called from probe context. This function just increments the drop
11267 * count on a buffer. It has been made a function to allow for the
11268 * possibility of understanding the source of mysterious drop counts. (A
11269 * problem for which one may be particularly disappointed that DTrace cannot
11270 * be used to understand DTrace.)
11271 */
11272 static void
11273 dtrace_buffer_drop(dtrace_buffer_t *buf)
11274 {
11275 buf->dtb_drops++;
11276 }
11277
11278 /*
11279 * Note: called from probe context. This function is called to reserve space
11280 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
11281 * mstate. Returns the new offset in the buffer, or a negative value if an
11282 * error has occurred.
11283 */
11284 static intptr_t
11285 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11286 dtrace_state_t *state, dtrace_mstate_t *mstate)
11287 {
11288 intptr_t offs = buf->dtb_offset, soffs;
11289 intptr_t woffs;
11290 caddr_t tomax;
11291 size_t total;
11292
11293 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11294 return (-1);
11295
11296 if ((tomax = buf->dtb_tomax) == NULL) {
11297 dtrace_buffer_drop(buf);
11298 return (-1);
11299 }
11300
11301 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11302 while (offs & (align - 1)) {
11303 /*
11304 * Assert that our alignment is off by a number which
11305 * is itself sizeof (uint32_t) aligned.
11306 */
11307 ASSERT(!((align - (offs & (align - 1))) &
11308 (sizeof (uint32_t) - 1)));
11309 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11310 offs += sizeof (uint32_t);
11311 }
11312
11313 if ((soffs = offs + needed) > buf->dtb_size) {
11314 dtrace_buffer_drop(buf);
11315 return (-1);
11316 }
11317
11318 if (mstate == NULL)
11319 return (offs);
11320
11321 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11322 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11323 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11324
11325 return (offs);
11326 }
11327
11328 if (buf->dtb_flags & DTRACEBUF_FILL) {
11329 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11330 (buf->dtb_flags & DTRACEBUF_FULL))
11331 return (-1);
11332 goto out;
11333 }
11334
11335 total = needed + (offs & (align - 1));
11336
11337 /*
11338 * For a ring buffer, life is quite a bit more complicated. Before
11339 * we can store any padding, we need to adjust our wrapping offset.
11340 * (If we've never before wrapped or we're not about to, no adjustment
11341 * is required.)
11342 */
11343 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11344 offs + total > buf->dtb_size) {
11345 woffs = buf->dtb_xamot_offset;
11346
11347 if (offs + total > buf->dtb_size) {
11348 /*
11349 * We can't fit in the end of the buffer. First, a
11350 * sanity check that we can fit in the buffer at all.
11351 */
11352 if (total > buf->dtb_size) {
11353 dtrace_buffer_drop(buf);
11354 return (-1);
11355 }
11356
11357 /*
11358 * We're going to be storing at the top of the buffer,
11359 * so now we need to deal with the wrapped offset. We
11360 * only reset our wrapped offset to 0 if it is
11361 * currently greater than the current offset. If it
11362 * is less than the current offset, it is because a
11363 * previous allocation induced a wrap -- but the
11364 * allocation didn't subsequently take the space due
11365 * to an error or false predicate evaluation. In this
11366 * case, we'll just leave the wrapped offset alone: if
11367 * the wrapped offset hasn't been advanced far enough
11368 * for this allocation, it will be adjusted in the
11369 * lower loop.
11370 */
11371 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11372 if (woffs >= offs)
11373 woffs = 0;
11374 } else {
11375 woffs = 0;
11376 }
11377
11378 /*
11379 * Now we know that we're going to be storing to the
11380 * top of the buffer and that there is room for us
11381 * there. We need to clear the buffer from the current
11382 * offset to the end (there may be old gunk there).
11383 */
11384 while (offs < buf->dtb_size)
11385 tomax[offs++] = 0;
11386
11387 /*
11388 * We need to set our offset to zero. And because we
11389 * are wrapping, we need to set the bit indicating as
11390 * much. We can also adjust our needed space back
11391 * down to the space required by the ECB -- we know
11392 * that the top of the buffer is aligned.
11393 */
11394 offs = 0;
11395 total = needed;
11396 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11397 } else {
11398 /*
11399 * There is room for us in the buffer, so we simply
11400 * need to check the wrapped offset.
11401 */
11402 if (woffs < offs) {
11403 /*
11404 * The wrapped offset is less than the offset.
11405 * This can happen if we allocated buffer space
11406 * that induced a wrap, but then we didn't
11407 * subsequently take the space due to an error
11408 * or false predicate evaluation. This is
11409 * okay; we know that _this_ allocation isn't
11410 * going to induce a wrap. We still can't
11411 * reset the wrapped offset to be zero,
11412 * however: the space may have been trashed in
11413 * the previous failed probe attempt. But at
11414 * least the wrapped offset doesn't need to
11415 * be adjusted at all...
11416 */
11417 goto out;
11418 }
11419 }
11420
11421 while (offs + total > woffs) {
11422 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11423 size_t size;
11424
11425 if (epid == DTRACE_EPIDNONE) {
11426 size = sizeof (uint32_t);
11427 } else {
11428 ASSERT3U(epid, <=, state->dts_necbs);
11429 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11430
11431 size = state->dts_ecbs[epid - 1]->dte_size;
11432 }
11433
11434 ASSERT(woffs + size <= buf->dtb_size);
11435 ASSERT(size != 0);
11436
11437 if (woffs + size == buf->dtb_size) {
11438 /*
11439 * We've reached the end of the buffer; we want
11440 * to set the wrapped offset to 0 and break
11441 * out. However, if the offs is 0, then we're
11442 * in a strange edge-condition: the amount of
11443 * space that we want to reserve plus the size
11444 * of the record that we're overwriting is
11445 * greater than the size of the buffer. This
11446 * is problematic because if we reserve the
11447 * space but subsequently don't consume it (due
11448 * to a failed predicate or error) the wrapped
11449 * offset will be 0 -- yet the EPID at offset 0
11450 * will not be committed. This situation is
11451 * relatively easy to deal with: if we're in
11452 * this case, the buffer is indistinguishable
11453 * from one that hasn't wrapped; we need only
11454 * finish the job by clearing the wrapped bit,
11455 * explicitly setting the offset to be 0, and
11456 * zero'ing out the old data in the buffer.
11457 */
11458 if (offs == 0) {
11459 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11460 buf->dtb_offset = 0;
11461 woffs = total;
11462
11463 while (woffs < buf->dtb_size)
11464 tomax[woffs++] = 0;
11465 }
11466
11467 woffs = 0;
11468 break;
11469 }
11470
11471 woffs += size;
11472 }
11473
11474 /*
11475 * We have a wrapped offset. It may be that the wrapped offset
11476 * has become zero -- that's okay.
11477 */
11478 buf->dtb_xamot_offset = woffs;
11479 }
11480
11481 out:
11482 /*
11483 * Now we can plow the buffer with any necessary padding.
11484 */
11485 while (offs & (align - 1)) {
11486 /*
11487 * Assert that our alignment is off by a number which
11488 * is itself sizeof (uint32_t) aligned.
11489 */
11490 ASSERT(!((align - (offs & (align - 1))) &
11491 (sizeof (uint32_t) - 1)));
11492 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11493 offs += sizeof (uint32_t);
11494 }
11495
11496 if (buf->dtb_flags & DTRACEBUF_FILL) {
11497 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11498 buf->dtb_flags |= DTRACEBUF_FULL;
11499 return (-1);
11500 }
11501 }
11502
11503 if (mstate == NULL)
11504 return (offs);
11505
11506 /*
11507 * For ring buffers and fill buffers, the scratch space is always
11508 * the inactive buffer.
11509 */
11510 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11511 mstate->dtms_scratch_size = buf->dtb_size;
11512 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11513
11514 return (offs);
11515 }
11516
11517 static void
11518 dtrace_buffer_polish(dtrace_buffer_t *buf)
11519 {
11520 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11521 ASSERT(MUTEX_HELD(&dtrace_lock));
11522
11523 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11524 return;
11525
11526 /*
11527 * We need to polish the ring buffer. There are three cases:
11528 *
11529 * - The first (and presumably most common) is that there is no gap
11530 * between the buffer offset and the wrapped offset. In this case,
11531 * there is nothing in the buffer that isn't valid data; we can
11532 * mark the buffer as polished and return.
11533 *
11534 * - The second (less common than the first but still more common
11535 * than the third) is that there is a gap between the buffer offset
11536 * and the wrapped offset, and the wrapped offset is larger than the
11537 * buffer offset. This can happen because of an alignment issue, or
11538 * can happen because of a call to dtrace_buffer_reserve() that
11539 * didn't subsequently consume the buffer space. In this case,
11540 * we need to zero the data from the buffer offset to the wrapped
11541 * offset.
11542 *
11543 * - The third (and least common) is that there is a gap between the
11544 * buffer offset and the wrapped offset, but the wrapped offset is
11545 * _less_ than the buffer offset. This can only happen because a
11546 * call to dtrace_buffer_reserve() induced a wrap, but the space
11547 * was not subsequently consumed. In this case, we need to zero the
11548 * space from the offset to the end of the buffer _and_ from the
11549 * top of the buffer to the wrapped offset.
11550 */
11551 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11552 bzero(buf->dtb_tomax + buf->dtb_offset,
11553 buf->dtb_xamot_offset - buf->dtb_offset);
11554 }
11555
11556 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11557 bzero(buf->dtb_tomax + buf->dtb_offset,
11558 buf->dtb_size - buf->dtb_offset);
11559 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11560 }
11561 }
11562
11563 /*
11564 * This routine determines if data generated at the specified time has likely
11565 * been entirely consumed at user-level. This routine is called to determine
11566 * if an ECB on a defunct probe (but for an active enabling) can be safely
11567 * disabled and destroyed.
11568 */
11569 static int
11570 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11571 {
11572 int i;
11573
11574 for (i = 0; i < NCPU; i++) {
11575 dtrace_buffer_t *buf = &bufs[i];
11576
11577 if (buf->dtb_size == 0)
11578 continue;
11579
11580 if (buf->dtb_flags & DTRACEBUF_RING)
11581 return (0);
11582
11583 if (!buf->dtb_switched && buf->dtb_offset != 0)
11584 return (0);
11585
11586 if (buf->dtb_switched - buf->dtb_interval < when)
11587 return (0);
11588 }
11589
11590 return (1);
11591 }
11592
11593 static void
11594 dtrace_buffer_free(dtrace_buffer_t *bufs)
11595 {
11596 int i;
11597
11598 for (i = 0; i < NCPU; i++) {
11599 dtrace_buffer_t *buf = &bufs[i];
11600
11601 if (buf->dtb_tomax == NULL) {
11602 ASSERT(buf->dtb_xamot == NULL);
11603 ASSERT(buf->dtb_size == 0);
11604 continue;
11605 }
11606
11607 if (buf->dtb_xamot != NULL) {
11608 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11609 kmem_free(buf->dtb_xamot, buf->dtb_size);
11610 }
11611
11612 kmem_free(buf->dtb_tomax, buf->dtb_size);
11613 buf->dtb_size = 0;
11614 buf->dtb_tomax = NULL;
11615 buf->dtb_xamot = NULL;
11616 }
11617 }
11618
11619 /*
11620 * DTrace Enabling Functions
11621 */
11622 static dtrace_enabling_t *
11623 dtrace_enabling_create(dtrace_vstate_t *vstate)
11624 {
11625 dtrace_enabling_t *enab;
11626
11627 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11628 enab->dten_vstate = vstate;
11629
11630 return (enab);
11631 }
11632
11633 static void
11634 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11635 {
11636 dtrace_ecbdesc_t **ndesc;
11637 size_t osize, nsize;
11638
11639 /*
11640 * We can't add to enablings after we've enabled them, or after we've
11641 * retained them.
11642 */
11643 ASSERT(enab->dten_probegen == 0);
11644 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11645
11646 if (enab->dten_ndesc < enab->dten_maxdesc) {
11647 enab->dten_desc[enab->dten_ndesc++] = ecb;
11648 return;
11649 }
11650
11651 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11652
11653 if (enab->dten_maxdesc == 0) {
11654 enab->dten_maxdesc = 1;
11655 } else {
11656 enab->dten_maxdesc <<= 1;
11657 }
11658
11659 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11660
11661 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11662 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11663 bcopy(enab->dten_desc, ndesc, osize);
11664 kmem_free(enab->dten_desc, osize);
11665
11666 enab->dten_desc = ndesc;
11667 enab->dten_desc[enab->dten_ndesc++] = ecb;
11668 }
11669
11670 static void
11671 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11672 dtrace_probedesc_t *pd)
11673 {
11674 dtrace_ecbdesc_t *new;
11675 dtrace_predicate_t *pred;
11676 dtrace_actdesc_t *act;
11677
11678 /*
11679 * We're going to create a new ECB description that matches the
11680 * specified ECB in every way, but has the specified probe description.
11681 */
11682 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11683
11684 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11685 dtrace_predicate_hold(pred);
11686
11687 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11688 dtrace_actdesc_hold(act);
11689
11690 new->dted_action = ecb->dted_action;
11691 new->dted_pred = ecb->dted_pred;
11692 new->dted_probe = *pd;
11693 new->dted_uarg = ecb->dted_uarg;
11694
11695 dtrace_enabling_add(enab, new);
11696 }
11697
11698 static void
11699 dtrace_enabling_dump(dtrace_enabling_t *enab)
11700 {
11701 int i;
11702
11703 for (i = 0; i < enab->dten_ndesc; i++) {
11704 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11705
11706 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11707 desc->dtpd_provider, desc->dtpd_mod,
11708 desc->dtpd_func, desc->dtpd_name);
11709 }
11710 }
11711
11712 static void
11713 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11714 {
11715 int i;
11716 dtrace_ecbdesc_t *ep;
11717 dtrace_vstate_t *vstate = enab->dten_vstate;
11718
11719 ASSERT(MUTEX_HELD(&dtrace_lock));
11720
11721 for (i = 0; i < enab->dten_ndesc; i++) {
11722 dtrace_actdesc_t *act, *next;
11723 dtrace_predicate_t *pred;
11724
11725 ep = enab->dten_desc[i];
11726
11727 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11728 dtrace_predicate_release(pred, vstate);
11729
11730 for (act = ep->dted_action; act != NULL; act = next) {
11731 next = act->dtad_next;
11732 dtrace_actdesc_release(act, vstate);
11733 }
11734
11735 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11736 }
11737
11738 kmem_free(enab->dten_desc,
11739 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11740
11741 /*
11742 * If this was a retained enabling, decrement the dts_nretained count
11743 * and take it off of the dtrace_retained list.
11744 */
11745 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11746 dtrace_retained == enab) {
11747 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11748 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11749 enab->dten_vstate->dtvs_state->dts_nretained--;
11750 dtrace_retained_gen++;
11751 }
11752
11753 if (enab->dten_prev == NULL) {
11754 if (dtrace_retained == enab) {
11755 dtrace_retained = enab->dten_next;
11756
11757 if (dtrace_retained != NULL)
11758 dtrace_retained->dten_prev = NULL;
11759 }
11760 } else {
11761 ASSERT(enab != dtrace_retained);
11762 ASSERT(dtrace_retained != NULL);
11763 enab->dten_prev->dten_next = enab->dten_next;
11764 }
11765
11766 if (enab->dten_next != NULL) {
11767 ASSERT(dtrace_retained != NULL);
11768 enab->dten_next->dten_prev = enab->dten_prev;
11769 }
11770
11771 kmem_free(enab, sizeof (dtrace_enabling_t));
11772 }
11773
11774 static int
11775 dtrace_enabling_retain(dtrace_enabling_t *enab)
11776 {
11777 dtrace_state_t *state;
11778
11779 ASSERT(MUTEX_HELD(&dtrace_lock));
11780 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11781 ASSERT(enab->dten_vstate != NULL);
11782
11783 state = enab->dten_vstate->dtvs_state;
11784 ASSERT(state != NULL);
11785
11786 /*
11787 * We only allow each state to retain dtrace_retain_max enablings.
11788 */
11789 if (state->dts_nretained >= dtrace_retain_max)
11790 return (ENOSPC);
11791
11792 state->dts_nretained++;
11793 dtrace_retained_gen++;
11794
11795 if (dtrace_retained == NULL) {
11796 dtrace_retained = enab;
11797 return (0);
11798 }
11799
11800 enab->dten_next = dtrace_retained;
11801 dtrace_retained->dten_prev = enab;
11802 dtrace_retained = enab;
11803
11804 return (0);
11805 }
11806
11807 static int
11808 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11809 dtrace_probedesc_t *create)
11810 {
11811 dtrace_enabling_t *new, *enab;
11812 int found = 0, err = ENOENT;
11813
11814 ASSERT(MUTEX_HELD(&dtrace_lock));
11815 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11816 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11817 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11818 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11819
11820 new = dtrace_enabling_create(&state->dts_vstate);
11821
11822 /*
11823 * Iterate over all retained enablings, looking for enablings that
11824 * match the specified state.
11825 */
11826 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11827 int i;
11828
11829 /*
11830 * dtvs_state can only be NULL for helper enablings -- and
11831 * helper enablings can't be retained.
11832 */
11833 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11834
11835 if (enab->dten_vstate->dtvs_state != state)
11836 continue;
11837
11838 /*
11839 * Now iterate over each probe description; we're looking for
11840 * an exact match to the specified probe description.
11841 */
11842 for (i = 0; i < enab->dten_ndesc; i++) {
11843 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11844 dtrace_probedesc_t *pd = &ep->dted_probe;
11845
11846 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11847 continue;
11848
11849 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11850 continue;
11851
11852 if (strcmp(pd->dtpd_func, match->dtpd_func))
11853 continue;
11854
11855 if (strcmp(pd->dtpd_name, match->dtpd_name))
11856 continue;
11857
11858 /*
11859 * We have a winning probe! Add it to our growing
11860 * enabling.
11861 */
11862 found = 1;
11863 dtrace_enabling_addlike(new, ep, create);
11864 }
11865 }
11866
11867 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11868 dtrace_enabling_destroy(new);
11869 return (err);
11870 }
11871
11872 return (0);
11873 }
11874
11875 static void
11876 dtrace_enabling_retract(dtrace_state_t *state)
11877 {
11878 dtrace_enabling_t *enab, *next;
11879
11880 ASSERT(MUTEX_HELD(&dtrace_lock));
11881
11882 /*
11883 * Iterate over all retained enablings, destroy the enablings retained
11884 * for the specified state.
11885 */
11886 for (enab = dtrace_retained; enab != NULL; enab = next) {
11887 next = enab->dten_next;
11888
11889 /*
11890 * dtvs_state can only be NULL for helper enablings -- and
11891 * helper enablings can't be retained.
11892 */
11893 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11894
11895 if (enab->dten_vstate->dtvs_state == state) {
11896 ASSERT(state->dts_nretained > 0);
11897 dtrace_enabling_destroy(enab);
11898 }
11899 }
11900
11901 ASSERT(state->dts_nretained == 0);
11902 }
11903
11904 static int
11905 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11906 {
11907 int i = 0;
11908 int total_matched = 0, matched = 0;
11909
11910 ASSERT(MUTEX_HELD(&cpu_lock));
11911 ASSERT(MUTEX_HELD(&dtrace_lock));
11912
11913 for (i = 0; i < enab->dten_ndesc; i++) {
11914 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11915
11916 enab->dten_current = ep;
11917 enab->dten_error = 0;
11918
11919 /*
11920 * If a provider failed to enable a probe then get out and
11921 * let the consumer know we failed.
11922 */
11923 if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
11924 return (EBUSY);
11925
11926 total_matched += matched;
11927
11928 if (enab->dten_error != 0) {
11929 /*
11930 * If we get an error half-way through enabling the
11931 * probes, we kick out -- perhaps with some number of
11932 * them enabled. Leaving enabled probes enabled may
11933 * be slightly confusing for user-level, but we expect
11934 * that no one will attempt to actually drive on in
11935 * the face of such errors. If this is an anonymous
11936 * enabling (indicated with a NULL nmatched pointer),
11937 * we cmn_err() a message. We aren't expecting to
11938 * get such an error -- such as it can exist at all,
11939 * it would be a result of corrupted DOF in the driver
11940 * properties.
11941 */
11942 if (nmatched == NULL) {
11943 cmn_err(CE_WARN, "dtrace_enabling_match() "
11944 "error on %p: %d", (void *)ep,
11945 enab->dten_error);
11946 }
11947
11948 return (enab->dten_error);
11949 }
11950 }
11951
11952 enab->dten_probegen = dtrace_probegen;
11953 if (nmatched != NULL)
11954 *nmatched = total_matched;
11955
11956 return (0);
11957 }
11958
11959 static void
11960 dtrace_enabling_matchall(void)
11961 {
11962 dtrace_enabling_t *enab;
11963
11964 mutex_enter(&cpu_lock);
11965 mutex_enter(&dtrace_lock);
11966
11967 /*
11968 * Iterate over all retained enablings to see if any probes match
11969 * against them. We only perform this operation on enablings for which
11970 * we have sufficient permissions by virtue of being in the global zone
11971 * or in the same zone as the DTrace client. Because we can be called
11972 * after dtrace_detach() has been called, we cannot assert that there
11973 * are retained enablings. We can safely load from dtrace_retained,
11974 * however: the taskq_destroy() at the end of dtrace_detach() will
11975 * block pending our completion.
11976 */
11977 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11978 dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
11979 cred_t *cr = dcr->dcr_cred;
11980 zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
11981
11982 if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
11983 (zone == GLOBAL_ZONEID || getzoneid() == zone)))
11984 (void) dtrace_enabling_match(enab, NULL);
11985 }
11986
11987 mutex_exit(&dtrace_lock);
11988 mutex_exit(&cpu_lock);
11989 }
11990
11991 /*
11992 * If an enabling is to be enabled without having matched probes (that is, if
11993 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11994 * enabling must be _primed_ by creating an ECB for every ECB description.
11995 * This must be done to assure that we know the number of speculations, the
11996 * number of aggregations, the minimum buffer size needed, etc. before we
11997 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11998 * enabling any probes, we create ECBs for every ECB decription, but with a
11999 * NULL probe -- which is exactly what this function does.
12000 */
12001 static void
12002 dtrace_enabling_prime(dtrace_state_t *state)
12003 {
12004 dtrace_enabling_t *enab;
12005 int i;
12006
12007 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12008 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12009
12010 if (enab->dten_vstate->dtvs_state != state)
12011 continue;
12012
12013 /*
12014 * We don't want to prime an enabling more than once, lest
12015 * we allow a malicious user to induce resource exhaustion.
12016 * (The ECBs that result from priming an enabling aren't
12017 * leaked -- but they also aren't deallocated until the
12018 * consumer state is destroyed.)
12019 */
12020 if (enab->dten_primed)
12021 continue;
12022
12023 for (i = 0; i < enab->dten_ndesc; i++) {
12024 enab->dten_current = enab->dten_desc[i];
12025 (void) dtrace_probe_enable(NULL, enab);
12026 }
12027
12028 enab->dten_primed = 1;
12029 }
12030 }
12031
12032 /*
12033 * Called to indicate that probes should be provided due to retained
12034 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
12035 * must take an initial lap through the enabling calling the dtps_provide()
12036 * entry point explicitly to allow for autocreated probes.
12037 */
12038 static void
12039 dtrace_enabling_provide(dtrace_provider_t *prv)
12040 {
12041 int i, all = 0;
12042 dtrace_probedesc_t desc;
12043 dtrace_genid_t gen;
12044
12045 ASSERT(MUTEX_HELD(&dtrace_lock));
12046 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
12047
12048 if (prv == NULL) {
12049 all = 1;
12050 prv = dtrace_provider;
12051 }
12052
12053 do {
12054 dtrace_enabling_t *enab;
12055 void *parg = prv->dtpv_arg;
12056
12057 retry:
12058 gen = dtrace_retained_gen;
12059 for (enab = dtrace_retained; enab != NULL;
12060 enab = enab->dten_next) {
12061 for (i = 0; i < enab->dten_ndesc; i++) {
12062 desc = enab->dten_desc[i]->dted_probe;
12063 mutex_exit(&dtrace_lock);
12064 prv->dtpv_pops.dtps_provide(parg, &desc);
12065 mutex_enter(&dtrace_lock);
12066 /*
12067 * Process the retained enablings again if
12068 * they have changed while we weren't holding
12069 * dtrace_lock.
12070 */
12071 if (gen != dtrace_retained_gen)
12072 goto retry;
12073 }
12074 }
12075 } while (all && (prv = prv->dtpv_next) != NULL);
12076
12077 mutex_exit(&dtrace_lock);
12078 dtrace_probe_provide(NULL, all ? NULL : prv);
12079 mutex_enter(&dtrace_lock);
12080 }
12081
12082 /*
12083 * Called to reap ECBs that are attached to probes from defunct providers.
12084 */
12085 static void
12086 dtrace_enabling_reap(void)
12087 {
12088 dtrace_provider_t *prov;
12089 dtrace_probe_t *probe;
12090 dtrace_ecb_t *ecb;
12091 hrtime_t when;
12092 int i;
12093
12094 mutex_enter(&cpu_lock);
12095 mutex_enter(&dtrace_lock);
12096
12097 for (i = 0; i < dtrace_nprobes; i++) {
12098 if ((probe = dtrace_probes[i]) == NULL)
12099 continue;
12100
12101 if (probe->dtpr_ecb == NULL)
12102 continue;
12103
12104 prov = probe->dtpr_provider;
12105
12106 if ((when = prov->dtpv_defunct) == 0)
12107 continue;
12108
12109 /*
12110 * We have ECBs on a defunct provider: we want to reap these
12111 * ECBs to allow the provider to unregister. The destruction
12112 * of these ECBs must be done carefully: if we destroy the ECB
12113 * and the consumer later wishes to consume an EPID that
12114 * corresponds to the destroyed ECB (and if the EPID metadata
12115 * has not been previously consumed), the consumer will abort
12116 * processing on the unknown EPID. To reduce (but not, sadly,
12117 * eliminate) the possibility of this, we will only destroy an
12118 * ECB for a defunct provider if, for the state that
12119 * corresponds to the ECB:
12120 *
12121 * (a) There is no speculative tracing (which can effectively
12122 * cache an EPID for an arbitrary amount of time).
12123 *
12124 * (b) The principal buffers have been switched twice since the
12125 * provider became defunct.
12126 *
12127 * (c) The aggregation buffers are of zero size or have been
12128 * switched twice since the provider became defunct.
12129 *
12130 * We use dts_speculates to determine (a) and call a function
12131 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
12132 * that as soon as we've been unable to destroy one of the ECBs
12133 * associated with the probe, we quit trying -- reaping is only
12134 * fruitful in as much as we can destroy all ECBs associated
12135 * with the defunct provider's probes.
12136 */
12137 while ((ecb = probe->dtpr_ecb) != NULL) {
12138 dtrace_state_t *state = ecb->dte_state;
12139 dtrace_buffer_t *buf = state->dts_buffer;
12140 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
12141
12142 if (state->dts_speculates)
12143 break;
12144
12145 if (!dtrace_buffer_consumed(buf, when))
12146 break;
12147
12148 if (!dtrace_buffer_consumed(aggbuf, when))
12149 break;
12150
12151 dtrace_ecb_disable(ecb);
12152 ASSERT(probe->dtpr_ecb != ecb);
12153 dtrace_ecb_destroy(ecb);
12154 }
12155 }
12156
12157 mutex_exit(&dtrace_lock);
12158 mutex_exit(&cpu_lock);
12159 }
12160
12161 /*
12162 * DTrace DOF Functions
12163 */
12164 /*ARGSUSED*/
12165 static void
12166 dtrace_dof_error(dof_hdr_t *dof, const char *str)
12167 {
12168 if (dtrace_err_verbose)
12169 cmn_err(CE_WARN, "failed to process DOF: %s", str);
12170
12171 #ifdef DTRACE_ERRDEBUG
12172 dtrace_errdebug(str);
12173 #endif
12174 }
12175
12176 /*
12177 * Create DOF out of a currently enabled state. Right now, we only create
12178 * DOF containing the run-time options -- but this could be expanded to create
12179 * complete DOF representing the enabled state.
12180 */
12181 static dof_hdr_t *
12182 dtrace_dof_create(dtrace_state_t *state)
12183 {
12184 dof_hdr_t *dof;
12185 dof_sec_t *sec;
12186 dof_optdesc_t *opt;
12187 int i, len = sizeof (dof_hdr_t) +
12188 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
12189 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12190
12191 ASSERT(MUTEX_HELD(&dtrace_lock));
12192
12193 dof = kmem_zalloc(len, KM_SLEEP);
12194 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
12195 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
12196 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
12197 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
12198
12199 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
12200 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
12201 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
12202 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
12203 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
12204 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
12205
12206 dof->dofh_flags = 0;
12207 dof->dofh_hdrsize = sizeof (dof_hdr_t);
12208 dof->dofh_secsize = sizeof (dof_sec_t);
12209 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
12210 dof->dofh_secoff = sizeof (dof_hdr_t);
12211 dof->dofh_loadsz = len;
12212 dof->dofh_filesz = len;
12213 dof->dofh_pad = 0;
12214
12215 /*
12216 * Fill in the option section header...
12217 */
12218 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
12219 sec->dofs_type = DOF_SECT_OPTDESC;
12220 sec->dofs_align = sizeof (uint64_t);
12221 sec->dofs_flags = DOF_SECF_LOAD;
12222 sec->dofs_entsize = sizeof (dof_optdesc_t);
12223
12224 opt = (dof_optdesc_t *)((uintptr_t)sec +
12225 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
12226
12227 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
12228 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12229
12230 for (i = 0; i < DTRACEOPT_MAX; i++) {
12231 opt[i].dofo_option = i;
12232 opt[i].dofo_strtab = DOF_SECIDX_NONE;
12233 opt[i].dofo_value = state->dts_options[i];
12234 }
12235
12236 return (dof);
12237 }
12238
12239 static dof_hdr_t *
12240 dtrace_dof_copyin(uintptr_t uarg, int *errp)
12241 {
12242 dof_hdr_t hdr, *dof;
12243
12244 ASSERT(!MUTEX_HELD(&dtrace_lock));
12245
12246 /*
12247 * First, we're going to copyin() the sizeof (dof_hdr_t).
12248 */
12249 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
12250 dtrace_dof_error(NULL, "failed to copyin DOF header");
12251 *errp = EFAULT;
12252 return (NULL);
12253 }
12254
12255 /*
12256 * Now we'll allocate the entire DOF and copy it in -- provided
12257 * that the length isn't outrageous.
12258 */
12259 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
12260 dtrace_dof_error(&hdr, "load size exceeds maximum");
12261 *errp = E2BIG;
12262 return (NULL);
12263 }
12264
12265 if (hdr.dofh_loadsz < sizeof (hdr)) {
12266 dtrace_dof_error(&hdr, "invalid load size");
12267 *errp = EINVAL;
12268 return (NULL);
12269 }
12270
12271 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12272
12273 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
12274 dof->dofh_loadsz != hdr.dofh_loadsz) {
12275 kmem_free(dof, hdr.dofh_loadsz);
12276 *errp = EFAULT;
12277 return (NULL);
12278 }
12279
12280 return (dof);
12281 }
12282
12283 static dof_hdr_t *
12284 dtrace_dof_property(const char *name)
12285 {
12286 uchar_t *buf;
12287 uint64_t loadsz;
12288 unsigned int len, i;
12289 dof_hdr_t *dof;
12290
12291 /*
12292 * Unfortunately, array of values in .conf files are always (and
12293 * only) interpreted to be integer arrays. We must read our DOF
12294 * as an integer array, and then squeeze it into a byte array.
12295 */
12296 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12297 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12298 return (NULL);
12299
12300 for (i = 0; i < len; i++)
12301 buf[i] = (uchar_t)(((int *)buf)[i]);
12302
12303 if (len < sizeof (dof_hdr_t)) {
12304 ddi_prop_free(buf);
12305 dtrace_dof_error(NULL, "truncated header");
12306 return (NULL);
12307 }
12308
12309 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12310 ddi_prop_free(buf);
12311 dtrace_dof_error(NULL, "truncated DOF");
12312 return (NULL);
12313 }
12314
12315 if (loadsz >= dtrace_dof_maxsize) {
12316 ddi_prop_free(buf);
12317 dtrace_dof_error(NULL, "oversized DOF");
12318 return (NULL);
12319 }
12320
12321 dof = kmem_alloc(loadsz, KM_SLEEP);
12322 bcopy(buf, dof, loadsz);
12323 ddi_prop_free(buf);
12324
12325 return (dof);
12326 }
12327
12328 static void
12329 dtrace_dof_destroy(dof_hdr_t *dof)
12330 {
12331 kmem_free(dof, dof->dofh_loadsz);
12332 }
12333
12334 /*
12335 * Return the dof_sec_t pointer corresponding to a given section index. If the
12336 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12337 * a type other than DOF_SECT_NONE is specified, the header is checked against
12338 * this type and NULL is returned if the types do not match.
12339 */
12340 static dof_sec_t *
12341 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12342 {
12343 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12344 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12345
12346 if (i >= dof->dofh_secnum) {
12347 dtrace_dof_error(dof, "referenced section index is invalid");
12348 return (NULL);
12349 }
12350
12351 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12352 dtrace_dof_error(dof, "referenced section is not loadable");
12353 return (NULL);
12354 }
12355
12356 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12357 dtrace_dof_error(dof, "referenced section is the wrong type");
12358 return (NULL);
12359 }
12360
12361 return (sec);
12362 }
12363
12364 static dtrace_probedesc_t *
12365 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12366 {
12367 dof_probedesc_t *probe;
12368 dof_sec_t *strtab;
12369 uintptr_t daddr = (uintptr_t)dof;
12370 uintptr_t str;
12371 size_t size;
12372
12373 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12374 dtrace_dof_error(dof, "invalid probe section");
12375 return (NULL);
12376 }
12377
12378 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12379 dtrace_dof_error(dof, "bad alignment in probe description");
12380 return (NULL);
12381 }
12382
12383 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12384 dtrace_dof_error(dof, "truncated probe description");
12385 return (NULL);
12386 }
12387
12388 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12389 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12390
12391 if (strtab == NULL)
12392 return (NULL);
12393
12394 str = daddr + strtab->dofs_offset;
12395 size = strtab->dofs_size;
12396
12397 if (probe->dofp_provider >= strtab->dofs_size) {
12398 dtrace_dof_error(dof, "corrupt probe provider");
12399 return (NULL);
12400 }
12401
12402 (void) strncpy(desc->dtpd_provider,
12403 (char *)(str + probe->dofp_provider),
12404 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12405
12406 if (probe->dofp_mod >= strtab->dofs_size) {
12407 dtrace_dof_error(dof, "corrupt probe module");
12408 return (NULL);
12409 }
12410
12411 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12412 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12413
12414 if (probe->dofp_func >= strtab->dofs_size) {
12415 dtrace_dof_error(dof, "corrupt probe function");
12416 return (NULL);
12417 }
12418
12419 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12420 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12421
12422 if (probe->dofp_name >= strtab->dofs_size) {
12423 dtrace_dof_error(dof, "corrupt probe name");
12424 return (NULL);
12425 }
12426
12427 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12428 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12429
12430 return (desc);
12431 }
12432
12433 static dtrace_difo_t *
12434 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12435 cred_t *cr)
12436 {
12437 dtrace_difo_t *dp;
12438 size_t ttl = 0;
12439 dof_difohdr_t *dofd;
12440 uintptr_t daddr = (uintptr_t)dof;
12441 size_t max = dtrace_difo_maxsize;
12442 int i, l, n;
12443
12444 static const struct {
12445 int section;
12446 int bufoffs;
12447 int lenoffs;
12448 int entsize;
12449 int align;
12450 const char *msg;
12451 } difo[] = {
12452 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12453 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12454 sizeof (dif_instr_t), "multiple DIF sections" },
12455
12456 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12457 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12458 sizeof (uint64_t), "multiple integer tables" },
12459
12460 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12461 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12462 sizeof (char), "multiple string tables" },
12463
12464 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12465 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12466 sizeof (uint_t), "multiple variable tables" },
12467
12468 { DOF_SECT_NONE, 0, 0, 0, NULL }
12469 };
12470
12471 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12472 dtrace_dof_error(dof, "invalid DIFO header section");
12473 return (NULL);
12474 }
12475
12476 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12477 dtrace_dof_error(dof, "bad alignment in DIFO header");
12478 return (NULL);
12479 }
12480
12481 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12482 sec->dofs_size % sizeof (dof_secidx_t)) {
12483 dtrace_dof_error(dof, "bad size in DIFO header");
12484 return (NULL);
12485 }
12486
12487 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12488 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12489
12490 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12491 dp->dtdo_rtype = dofd->dofd_rtype;
12492
12493 for (l = 0; l < n; l++) {
12494 dof_sec_t *subsec;
12495 void **bufp;
12496 uint32_t *lenp;
12497
12498 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12499 dofd->dofd_links[l])) == NULL)
12500 goto err; /* invalid section link */
12501
12502 if (ttl + subsec->dofs_size > max) {
12503 dtrace_dof_error(dof, "exceeds maximum size");
12504 goto err;
12505 }
12506
12507 ttl += subsec->dofs_size;
12508
12509 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12510 if (subsec->dofs_type != difo[i].section)
12511 continue;
12512
12513 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12514 dtrace_dof_error(dof, "section not loaded");
12515 goto err;
12516 }
12517
12518 if (subsec->dofs_align != difo[i].align) {
12519 dtrace_dof_error(dof, "bad alignment");
12520 goto err;
12521 }
12522
12523 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12524 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12525
12526 if (*bufp != NULL) {
12527 dtrace_dof_error(dof, difo[i].msg);
12528 goto err;
12529 }
12530
12531 if (difo[i].entsize != subsec->dofs_entsize) {
12532 dtrace_dof_error(dof, "entry size mismatch");
12533 goto err;
12534 }
12535
12536 if (subsec->dofs_entsize != 0 &&
12537 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12538 dtrace_dof_error(dof, "corrupt entry size");
12539 goto err;
12540 }
12541
12542 *lenp = subsec->dofs_size;
12543 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12544 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12545 *bufp, subsec->dofs_size);
12546
12547 if (subsec->dofs_entsize != 0)
12548 *lenp /= subsec->dofs_entsize;
12549
12550 break;
12551 }
12552
12553 /*
12554 * If we encounter a loadable DIFO sub-section that is not
12555 * known to us, assume this is a broken program and fail.
12556 */
12557 if (difo[i].section == DOF_SECT_NONE &&
12558 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12559 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12560 goto err;
12561 }
12562 }
12563
12564 if (dp->dtdo_buf == NULL) {
12565 /*
12566 * We can't have a DIF object without DIF text.
12567 */
12568 dtrace_dof_error(dof, "missing DIF text");
12569 goto err;
12570 }
12571
12572 /*
12573 * Before we validate the DIF object, run through the variable table
12574 * looking for the strings -- if any of their size are under, we'll set
12575 * their size to be the system-wide default string size. Note that
12576 * this should _not_ happen if the "strsize" option has been set --
12577 * in this case, the compiler should have set the size to reflect the
12578 * setting of the option.
12579 */
12580 for (i = 0; i < dp->dtdo_varlen; i++) {
12581 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12582 dtrace_diftype_t *t = &v->dtdv_type;
12583
12584 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12585 continue;
12586
12587 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12588 t->dtdt_size = dtrace_strsize_default;
12589 }
12590
12591 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12592 goto err;
12593
12594 dtrace_difo_init(dp, vstate);
12595 return (dp);
12596
12597 err:
12598 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12599 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12600 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12601 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12602
12603 kmem_free(dp, sizeof (dtrace_difo_t));
12604 return (NULL);
12605 }
12606
12607 static dtrace_predicate_t *
12608 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12609 cred_t *cr)
12610 {
12611 dtrace_difo_t *dp;
12612
12613 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12614 return (NULL);
12615
12616 return (dtrace_predicate_create(dp));
12617 }
12618
12619 static dtrace_actdesc_t *
12620 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12621 cred_t *cr)
12622 {
12623 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12624 dof_actdesc_t *desc;
12625 dof_sec_t *difosec;
12626 size_t offs;
12627 uintptr_t daddr = (uintptr_t)dof;
12628 uint64_t arg;
12629 dtrace_actkind_t kind;
12630
12631 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12632 dtrace_dof_error(dof, "invalid action section");
12633 return (NULL);
12634 }
12635
12636 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12637 dtrace_dof_error(dof, "truncated action description");
12638 return (NULL);
12639 }
12640
12641 if (sec->dofs_align != sizeof (uint64_t)) {
12642 dtrace_dof_error(dof, "bad alignment in action description");
12643 return (NULL);
12644 }
12645
12646 if (sec->dofs_size < sec->dofs_entsize) {
12647 dtrace_dof_error(dof, "section entry size exceeds total size");
12648 return (NULL);
12649 }
12650
12651 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12652 dtrace_dof_error(dof, "bad entry size in action description");
12653 return (NULL);
12654 }
12655
12656 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12657 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12658 return (NULL);
12659 }
12660
12661 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12662 desc = (dof_actdesc_t *)(daddr +
12663 (uintptr_t)sec->dofs_offset + offs);
12664 kind = (dtrace_actkind_t)desc->dofa_kind;
12665
12666 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12667 (kind != DTRACEACT_PRINTA ||
12668 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12669 (kind == DTRACEACT_DIFEXPR &&
12670 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12671 dof_sec_t *strtab;
12672 char *str, *fmt;
12673 uint64_t i;
12674
12675 /*
12676 * The argument to these actions is an index into the
12677 * DOF string table. For printf()-like actions, this
12678 * is the format string. For print(), this is the
12679 * CTF type of the expression result.
12680 */
12681 if ((strtab = dtrace_dof_sect(dof,
12682 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12683 goto err;
12684
12685 str = (char *)((uintptr_t)dof +
12686 (uintptr_t)strtab->dofs_offset);
12687
12688 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12689 if (str[i] == '\0')
12690 break;
12691 }
12692
12693 if (i >= strtab->dofs_size) {
12694 dtrace_dof_error(dof, "bogus format string");
12695 goto err;
12696 }
12697
12698 if (i == desc->dofa_arg) {
12699 dtrace_dof_error(dof, "empty format string");
12700 goto err;
12701 }
12702
12703 i -= desc->dofa_arg;
12704 fmt = kmem_alloc(i + 1, KM_SLEEP);
12705 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12706 arg = (uint64_t)(uintptr_t)fmt;
12707 } else {
12708 if (kind == DTRACEACT_PRINTA) {
12709 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12710 arg = 0;
12711 } else {
12712 arg = desc->dofa_arg;
12713 }
12714 }
12715
12716 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12717 desc->dofa_uarg, arg);
12718
12719 if (last != NULL) {
12720 last->dtad_next = act;
12721 } else {
12722 first = act;
12723 }
12724
12725 last = act;
12726
12727 if (desc->dofa_difo == DOF_SECIDX_NONE)
12728 continue;
12729
12730 if ((difosec = dtrace_dof_sect(dof,
12731 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12732 goto err;
12733
12734 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12735
12736 if (act->dtad_difo == NULL)
12737 goto err;
12738 }
12739
12740 ASSERT(first != NULL);
12741 return (first);
12742
12743 err:
12744 for (act = first; act != NULL; act = next) {
12745 next = act->dtad_next;
12746 dtrace_actdesc_release(act, vstate);
12747 }
12748
12749 return (NULL);
12750 }
12751
12752 static dtrace_ecbdesc_t *
12753 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12754 cred_t *cr)
12755 {
12756 dtrace_ecbdesc_t *ep;
12757 dof_ecbdesc_t *ecb;
12758 dtrace_probedesc_t *desc;
12759 dtrace_predicate_t *pred = NULL;
12760
12761 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12762 dtrace_dof_error(dof, "truncated ECB description");
12763 return (NULL);
12764 }
12765
12766 if (sec->dofs_align != sizeof (uint64_t)) {
12767 dtrace_dof_error(dof, "bad alignment in ECB description");
12768 return (NULL);
12769 }
12770
12771 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12772 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12773
12774 if (sec == NULL)
12775 return (NULL);
12776
12777 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12778 ep->dted_uarg = ecb->dofe_uarg;
12779 desc = &ep->dted_probe;
12780
12781 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12782 goto err;
12783
12784 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12785 if ((sec = dtrace_dof_sect(dof,
12786 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12787 goto err;
12788
12789 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12790 goto err;
12791
12792 ep->dted_pred.dtpdd_predicate = pred;
12793 }
12794
12795 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12796 if ((sec = dtrace_dof_sect(dof,
12797 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12798 goto err;
12799
12800 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12801
12802 if (ep->dted_action == NULL)
12803 goto err;
12804 }
12805
12806 return (ep);
12807
12808 err:
12809 if (pred != NULL)
12810 dtrace_predicate_release(pred, vstate);
12811 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12812 return (NULL);
12813 }
12814
12815 /*
12816 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12817 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12818 * site of any user SETX relocations to account for load object base address.
12819 * In the future, if we need other relocations, this function can be extended.
12820 */
12821 static int
12822 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12823 {
12824 uintptr_t daddr = (uintptr_t)dof;
12825 dof_relohdr_t *dofr =
12826 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12827 dof_sec_t *ss, *rs, *ts;
12828 dof_relodesc_t *r;
12829 uint_t i, n;
12830
12831 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12832 sec->dofs_align != sizeof (dof_secidx_t)) {
12833 dtrace_dof_error(dof, "invalid relocation header");
12834 return (-1);
12835 }
12836
12837 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12838 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12839 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12840
12841 if (ss == NULL || rs == NULL || ts == NULL)
12842 return (-1); /* dtrace_dof_error() has been called already */
12843
12844 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12845 rs->dofs_align != sizeof (uint64_t)) {
12846 dtrace_dof_error(dof, "invalid relocation section");
12847 return (-1);
12848 }
12849
12850 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12851 n = rs->dofs_size / rs->dofs_entsize;
12852
12853 for (i = 0; i < n; i++) {
12854 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12855
12856 switch (r->dofr_type) {
12857 case DOF_RELO_NONE:
12858 break;
12859 case DOF_RELO_SETX:
12860 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12861 sizeof (uint64_t) > ts->dofs_size) {
12862 dtrace_dof_error(dof, "bad relocation offset");
12863 return (-1);
12864 }
12865
12866 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12867 dtrace_dof_error(dof, "misaligned setx relo");
12868 return (-1);
12869 }
12870
12871 *(uint64_t *)taddr += ubase;
12872 break;
12873 default:
12874 dtrace_dof_error(dof, "invalid relocation type");
12875 return (-1);
12876 }
12877
12878 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12879 }
12880
12881 return (0);
12882 }
12883
12884 /*
12885 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12886 * header: it should be at the front of a memory region that is at least
12887 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12888 * size. It need not be validated in any other way.
12889 */
12890 static int
12891 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12892 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12893 {
12894 uint64_t len = dof->dofh_loadsz, seclen;
12895 uintptr_t daddr = (uintptr_t)dof;
12896 dtrace_ecbdesc_t *ep;
12897 dtrace_enabling_t *enab;
12898 uint_t i;
12899
12900 ASSERT(MUTEX_HELD(&dtrace_lock));
12901 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12902
12903 /*
12904 * Check the DOF header identification bytes. In addition to checking
12905 * valid settings, we also verify that unused bits/bytes are zeroed so
12906 * we can use them later without fear of regressing existing binaries.
12907 */
12908 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12909 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12910 dtrace_dof_error(dof, "DOF magic string mismatch");
12911 return (-1);
12912 }
12913
12914 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12915 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12916 dtrace_dof_error(dof, "DOF has invalid data model");
12917 return (-1);
12918 }
12919
12920 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12921 dtrace_dof_error(dof, "DOF encoding mismatch");
12922 return (-1);
12923 }
12924
12925 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12926 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12927 dtrace_dof_error(dof, "DOF version mismatch");
12928 return (-1);
12929 }
12930
12931 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12932 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12933 return (-1);
12934 }
12935
12936 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12937 dtrace_dof_error(dof, "DOF uses too many integer registers");
12938 return (-1);
12939 }
12940
12941 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12942 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12943 return (-1);
12944 }
12945
12946 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12947 if (dof->dofh_ident[i] != 0) {
12948 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12949 return (-1);
12950 }
12951 }
12952
12953 if (dof->dofh_flags & ~DOF_FL_VALID) {
12954 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12955 return (-1);
12956 }
12957
12958 if (dof->dofh_secsize == 0) {
12959 dtrace_dof_error(dof, "zero section header size");
12960 return (-1);
12961 }
12962
12963 /*
12964 * Check that the section headers don't exceed the amount of DOF
12965 * data. Note that we cast the section size and number of sections
12966 * to uint64_t's to prevent possible overflow in the multiplication.
12967 */
12968 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12969
12970 if (dof->dofh_secoff > len || seclen > len ||
12971 dof->dofh_secoff + seclen > len) {
12972 dtrace_dof_error(dof, "truncated section headers");
12973 return (-1);
12974 }
12975
12976 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12977 dtrace_dof_error(dof, "misaligned section headers");
12978 return (-1);
12979 }
12980
12981 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12982 dtrace_dof_error(dof, "misaligned section size");
12983 return (-1);
12984 }
12985
12986 /*
12987 * Take an initial pass through the section headers to be sure that
12988 * the headers don't have stray offsets. If the 'noprobes' flag is
12989 * set, do not permit sections relating to providers, probes, or args.
12990 */
12991 for (i = 0; i < dof->dofh_secnum; i++) {
12992 dof_sec_t *sec = (dof_sec_t *)(daddr +
12993 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12994
12995 if (noprobes) {
12996 switch (sec->dofs_type) {
12997 case DOF_SECT_PROVIDER:
12998 case DOF_SECT_PROBES:
12999 case DOF_SECT_PRARGS:
13000 case DOF_SECT_PROFFS:
13001 dtrace_dof_error(dof, "illegal sections "
13002 "for enabling");
13003 return (-1);
13004 }
13005 }
13006
13007 if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
13008 !(sec->dofs_flags & DOF_SECF_LOAD)) {
13009 dtrace_dof_error(dof, "loadable section with load "
13010 "flag unset");
13011 return (-1);
13012 }
13013
13014 if (!(sec->dofs_flags & DOF_SECF_LOAD))
13015 continue; /* just ignore non-loadable sections */
13016
13017 if (sec->dofs_align & (sec->dofs_align - 1)) {
13018 dtrace_dof_error(dof, "bad section alignment");
13019 return (-1);
13020 }
13021
13022 if (sec->dofs_offset & (sec->dofs_align - 1)) {
13023 dtrace_dof_error(dof, "misaligned section");
13024 return (-1);
13025 }
13026
13027 if (sec->dofs_offset > len || sec->dofs_size > len ||
13028 sec->dofs_offset + sec->dofs_size > len) {
13029 dtrace_dof_error(dof, "corrupt section header");
13030 return (-1);
13031 }
13032
13033 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
13034 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
13035 dtrace_dof_error(dof, "non-terminating string table");
13036 return (-1);
13037 }
13038 }
13039
13040 /*
13041 * Take a second pass through the sections and locate and perform any
13042 * relocations that are present. We do this after the first pass to
13043 * be sure that all sections have had their headers validated.
13044 */
13045 for (i = 0; i < dof->dofh_secnum; i++) {
13046 dof_sec_t *sec = (dof_sec_t *)(daddr +
13047 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13048
13049 if (!(sec->dofs_flags & DOF_SECF_LOAD))
13050 continue; /* skip sections that are not loadable */
13051
13052 switch (sec->dofs_type) {
13053 case DOF_SECT_URELHDR:
13054 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
13055 return (-1);
13056 break;
13057 }
13058 }
13059
13060 if ((enab = *enabp) == NULL)
13061 enab = *enabp = dtrace_enabling_create(vstate);
13062
13063 for (i = 0; i < dof->dofh_secnum; i++) {
13064 dof_sec_t *sec = (dof_sec_t *)(daddr +
13065 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13066
13067 if (sec->dofs_type != DOF_SECT_ECBDESC)
13068 continue;
13069
13070 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
13071 dtrace_enabling_destroy(enab);
13072 *enabp = NULL;
13073 return (-1);
13074 }
13075
13076 dtrace_enabling_add(enab, ep);
13077 }
13078
13079 return (0);
13080 }
13081
13082 /*
13083 * Process DOF for any options. This routine assumes that the DOF has been
13084 * at least processed by dtrace_dof_slurp().
13085 */
13086 static int
13087 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
13088 {
13089 int i, rval;
13090 uint32_t entsize;
13091 size_t offs;
13092 dof_optdesc_t *desc;
13093
13094 for (i = 0; i < dof->dofh_secnum; i++) {
13095 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
13096 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13097
13098 if (sec->dofs_type != DOF_SECT_OPTDESC)
13099 continue;
13100
13101 if (sec->dofs_align != sizeof (uint64_t)) {
13102 dtrace_dof_error(dof, "bad alignment in "
13103 "option description");
13104 return (EINVAL);
13105 }
13106
13107 if ((entsize = sec->dofs_entsize) == 0) {
13108 dtrace_dof_error(dof, "zeroed option entry size");
13109 return (EINVAL);
13110 }
13111
13112 if (entsize < sizeof (dof_optdesc_t)) {
13113 dtrace_dof_error(dof, "bad option entry size");
13114 return (EINVAL);
13115 }
13116
13117 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
13118 desc = (dof_optdesc_t *)((uintptr_t)dof +
13119 (uintptr_t)sec->dofs_offset + offs);
13120
13121 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
13122 dtrace_dof_error(dof, "non-zero option string");
13123 return (EINVAL);
13124 }
13125
13126 if (desc->dofo_value == DTRACEOPT_UNSET) {
13127 dtrace_dof_error(dof, "unset option");
13128 return (EINVAL);
13129 }
13130
13131 if ((rval = dtrace_state_option(state,
13132 desc->dofo_option, desc->dofo_value)) != 0) {
13133 dtrace_dof_error(dof, "rejected option");
13134 return (rval);
13135 }
13136 }
13137 }
13138
13139 return (0);
13140 }
13141
13142 /*
13143 * DTrace Consumer State Functions
13144 */
13145 int
13146 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
13147 {
13148 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
13149 void *base;
13150 uintptr_t limit;
13151 dtrace_dynvar_t *dvar, *next, *start;
13152 int i;
13153
13154 ASSERT(MUTEX_HELD(&dtrace_lock));
13155 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
13156
13157 bzero(dstate, sizeof (dtrace_dstate_t));
13158
13159 if ((dstate->dtds_chunksize = chunksize) == 0)
13160 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
13161
13162 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
13163 size = min;
13164
13165 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
13166 return (ENOMEM);
13167
13168 dstate->dtds_size = size;
13169 dstate->dtds_base = base;
13170 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
13171 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
13172
13173 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
13174
13175 if (hashsize != 1 && (hashsize & 1))
13176 hashsize--;
13177
13178 dstate->dtds_hashsize = hashsize;
13179 dstate->dtds_hash = dstate->dtds_base;
13180
13181 /*
13182 * Set all of our hash buckets to point to the single sink, and (if
13183 * it hasn't already been set), set the sink's hash value to be the
13184 * sink sentinel value. The sink is needed for dynamic variable
13185 * lookups to know that they have iterated over an entire, valid hash
13186 * chain.
13187 */
13188 for (i = 0; i < hashsize; i++)
13189 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
13190
13191 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
13192 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
13193
13194 /*
13195 * Determine number of active CPUs. Divide free list evenly among
13196 * active CPUs.
13197 */
13198 start = (dtrace_dynvar_t *)
13199 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
13200 limit = (uintptr_t)base + size;
13201
13202 maxper = (limit - (uintptr_t)start) / NCPU;
13203 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13204
13205 for (i = 0; i < NCPU; i++) {
13206 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13207
13208 /*
13209 * If we don't even have enough chunks to make it once through
13210 * NCPUs, we're just going to allocate everything to the first
13211 * CPU. And if we're on the last CPU, we're going to allocate
13212 * whatever is left over. In either case, we set the limit to
13213 * be the limit of the dynamic variable space.
13214 */
13215 if (maxper == 0 || i == NCPU - 1) {
13216 limit = (uintptr_t)base + size;
13217 start = NULL;
13218 } else {
13219 limit = (uintptr_t)start + maxper;
13220 start = (dtrace_dynvar_t *)limit;
13221 }
13222
13223 ASSERT(limit <= (uintptr_t)base + size);
13224
13225 for (;;) {
13226 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13227 dstate->dtds_chunksize);
13228
13229 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13230 break;
13231
13232 dvar->dtdv_next = next;
13233 dvar = next;
13234 }
13235
13236 if (maxper == 0)
13237 break;
13238 }
13239
13240 return (0);
13241 }
13242
13243 void
13244 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13245 {
13246 ASSERT(MUTEX_HELD(&cpu_lock));
13247
13248 if (dstate->dtds_base == NULL)
13249 return;
13250
13251 kmem_free(dstate->dtds_base, dstate->dtds_size);
13252 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13253 }
13254
13255 static void
13256 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13257 {
13258 /*
13259 * Logical XOR, where are you?
13260 */
13261 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13262
13263 if (vstate->dtvs_nglobals > 0) {
13264 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13265 sizeof (dtrace_statvar_t *));
13266 }
13267
13268 if (vstate->dtvs_ntlocals > 0) {
13269 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13270 sizeof (dtrace_difv_t));
13271 }
13272
13273 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13274
13275 if (vstate->dtvs_nlocals > 0) {
13276 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13277 sizeof (dtrace_statvar_t *));
13278 }
13279 }
13280
13281 static void
13282 dtrace_state_clean(dtrace_state_t *state)
13283 {
13284 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13285 return;
13286
13287 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13288 dtrace_speculation_clean(state);
13289 }
13290
13291 static void
13292 dtrace_state_deadman(dtrace_state_t *state)
13293 {
13294 hrtime_t now;
13295
13296 dtrace_sync();
13297
13298 now = dtrace_gethrtime();
13299
13300 if (state != dtrace_anon.dta_state &&
13301 now - state->dts_laststatus >= dtrace_deadman_user)
13302 return;
13303
13304 /*
13305 * We must be sure that dts_alive never appears to be less than the
13306 * value upon entry to dtrace_state_deadman(), and because we lack a
13307 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13308 * store INT64_MAX to it, followed by a memory barrier, followed by
13309 * the new value. This assures that dts_alive never appears to be
13310 * less than its true value, regardless of the order in which the
13311 * stores to the underlying storage are issued.
13312 */
13313 state->dts_alive = INT64_MAX;
13314 dtrace_membar_producer();
13315 state->dts_alive = now;
13316 }
13317
13318 dtrace_state_t *
13319 dtrace_state_create(dev_t *devp, cred_t *cr)
13320 {
13321 minor_t minor;
13322 major_t major;
13323 char c[30];
13324 dtrace_state_t *state;
13325 dtrace_optval_t *opt;
13326 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13327
13328 ASSERT(MUTEX_HELD(&dtrace_lock));
13329 ASSERT(MUTEX_HELD(&cpu_lock));
13330
13331 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13332 VM_BESTFIT | VM_SLEEP);
13333
13334 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13335 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13336 return (NULL);
13337 }
13338
13339 state = ddi_get_soft_state(dtrace_softstate, minor);
13340 state->dts_epid = DTRACE_EPIDNONE + 1;
13341
13342 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
13343 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13344 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13345
13346 if (devp != NULL) {
13347 major = getemajor(*devp);
13348 } else {
13349 major = ddi_driver_major(dtrace_devi);
13350 }
13351
13352 state->dts_dev = makedevice(major, minor);
13353
13354 if (devp != NULL)
13355 *devp = state->dts_dev;
13356
13357 /*
13358 * We allocate NCPU buffers. On the one hand, this can be quite
13359 * a bit of memory per instance (nearly 36K on a Starcat). On the
13360 * other hand, it saves an additional memory reference in the probe
13361 * path.
13362 */
13363 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13364 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13365 state->dts_cleaner = CYCLIC_NONE;
13366 state->dts_deadman = CYCLIC_NONE;
13367 state->dts_vstate.dtvs_state = state;
13368
13369 for (i = 0; i < DTRACEOPT_MAX; i++)
13370 state->dts_options[i] = DTRACEOPT_UNSET;
13371
13372 /*
13373 * Set the default options.
13374 */
13375 opt = state->dts_options;
13376 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13377 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13378 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13379 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13380 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13381 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13382 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13383 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13384 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13385 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13386 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13387 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13388 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13389 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13390
13391 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13392
13393 /*
13394 * Depending on the user credentials, we set flag bits which alter probe
13395 * visibility or the amount of destructiveness allowed. In the case of
13396 * actual anonymous tracing, or the possession of all privileges, all of
13397 * the normal checks are bypassed.
13398 */
13399 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13400 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13401 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13402 } else {
13403 /*
13404 * Set up the credentials for this instantiation. We take a
13405 * hold on the credential to prevent it from disappearing on
13406 * us; this in turn prevents the zone_t referenced by this
13407 * credential from disappearing. This means that we can
13408 * examine the credential and the zone from probe context.
13409 */
13410 crhold(cr);
13411 state->dts_cred.dcr_cred = cr;
13412
13413 /*
13414 * CRA_PROC means "we have *some* privilege for dtrace" and
13415 * unlocks the use of variables like pid, zonename, etc.
13416 */
13417 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13418 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13419 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13420 }
13421
13422 /*
13423 * dtrace_user allows use of syscall and profile providers.
13424 * If the user also has proc_owner and/or proc_zone, we
13425 * extend the scope to include additional visibility and
13426 * destructive power.
13427 */
13428 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13429 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13430 state->dts_cred.dcr_visible |=
13431 DTRACE_CRV_ALLPROC;
13432
13433 state->dts_cred.dcr_action |=
13434 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13435 }
13436
13437 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13438 state->dts_cred.dcr_visible |=
13439 DTRACE_CRV_ALLZONE;
13440
13441 state->dts_cred.dcr_action |=
13442 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13443 }
13444
13445 /*
13446 * If we have all privs in whatever zone this is,
13447 * we can do destructive things to processes which
13448 * have altered credentials.
13449 */
13450 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13451 cr->cr_zone->zone_privset)) {
13452 state->dts_cred.dcr_action |=
13453 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13454 }
13455 }
13456
13457 /*
13458 * Holding the dtrace_kernel privilege also implies that
13459 * the user has the dtrace_user privilege from a visibility
13460 * perspective. But without further privileges, some
13461 * destructive actions are not available.
13462 */
13463 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13464 /*
13465 * Make all probes in all zones visible. However,
13466 * this doesn't mean that all actions become available
13467 * to all zones.
13468 */
13469 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13470 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13471
13472 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13473 DTRACE_CRA_PROC;
13474 /*
13475 * Holding proc_owner means that destructive actions
13476 * for *this* zone are allowed.
13477 */
13478 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13479 state->dts_cred.dcr_action |=
13480 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13481
13482 /*
13483 * Holding proc_zone means that destructive actions
13484 * for this user/group ID in all zones is allowed.
13485 */
13486 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13487 state->dts_cred.dcr_action |=
13488 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13489
13490 /*
13491 * If we have all privs in whatever zone this is,
13492 * we can do destructive things to processes which
13493 * have altered credentials.
13494 */
13495 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13496 cr->cr_zone->zone_privset)) {
13497 state->dts_cred.dcr_action |=
13498 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13499 }
13500 }
13501
13502 /*
13503 * Holding the dtrace_proc privilege gives control over fasttrap
13504 * and pid providers. We need to grant wider destructive
13505 * privileges in the event that the user has proc_owner and/or
13506 * proc_zone.
13507 */
13508 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13509 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13510 state->dts_cred.dcr_action |=
13511 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13512
13513 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13514 state->dts_cred.dcr_action |=
13515 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13516 }
13517 }
13518
13519 return (state);
13520 }
13521
13522 static int
13523 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13524 {
13525 dtrace_optval_t *opt = state->dts_options, size;
13526 processorid_t cpu;
13527 int flags = 0, rval, factor, divisor = 1;
13528
13529 ASSERT(MUTEX_HELD(&dtrace_lock));
13530 ASSERT(MUTEX_HELD(&cpu_lock));
13531 ASSERT(which < DTRACEOPT_MAX);
13532 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13533 (state == dtrace_anon.dta_state &&
13534 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13535
13536 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13537 return (0);
13538
13539 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13540 cpu = opt[DTRACEOPT_CPU];
13541
13542 if (which == DTRACEOPT_SPECSIZE)
13543 flags |= DTRACEBUF_NOSWITCH;
13544
13545 if (which == DTRACEOPT_BUFSIZE) {
13546 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13547 flags |= DTRACEBUF_RING;
13548
13549 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13550 flags |= DTRACEBUF_FILL;
13551
13552 if (state != dtrace_anon.dta_state ||
13553 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13554 flags |= DTRACEBUF_INACTIVE;
13555 }
13556
13557 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
13558 /*
13559 * The size must be 8-byte aligned. If the size is not 8-byte
13560 * aligned, drop it down by the difference.
13561 */
13562 if (size & (sizeof (uint64_t) - 1))
13563 size -= size & (sizeof (uint64_t) - 1);
13564
13565 if (size < state->dts_reserve) {
13566 /*
13567 * Buffers always must be large enough to accommodate
13568 * their prereserved space. We return E2BIG instead
13569 * of ENOMEM in this case to allow for user-level
13570 * software to differentiate the cases.
13571 */
13572 return (E2BIG);
13573 }
13574
13575 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
13576
13577 if (rval != ENOMEM) {
13578 opt[which] = size;
13579 return (rval);
13580 }
13581
13582 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13583 return (rval);
13584
13585 for (divisor = 2; divisor < factor; divisor <<= 1)
13586 continue;
13587 }
13588
13589 return (ENOMEM);
13590 }
13591
13592 static int
13593 dtrace_state_buffers(dtrace_state_t *state)
13594 {
13595 dtrace_speculation_t *spec = state->dts_speculations;
13596 int rval, i;
13597
13598 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13599 DTRACEOPT_BUFSIZE)) != 0)
13600 return (rval);
13601
13602 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13603 DTRACEOPT_AGGSIZE)) != 0)
13604 return (rval);
13605
13606 for (i = 0; i < state->dts_nspeculations; i++) {
13607 if ((rval = dtrace_state_buffer(state,
13608 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13609 return (rval);
13610 }
13611
13612 return (0);
13613 }
13614
13615 static void
13616 dtrace_state_prereserve(dtrace_state_t *state)
13617 {
13618 dtrace_ecb_t *ecb;
13619 dtrace_probe_t *probe;
13620
13621 state->dts_reserve = 0;
13622
13623 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13624 return;
13625
13626 /*
13627 * If our buffer policy is a "fill" buffer policy, we need to set the
13628 * prereserved space to be the space required by the END probes.
13629 */
13630 probe = dtrace_probes[dtrace_probeid_end - 1];
13631 ASSERT(probe != NULL);
13632
13633 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13634 if (ecb->dte_state != state)
13635 continue;
13636
13637 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13638 }
13639 }
13640
13641 static int
13642 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13643 {
13644 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13645 dtrace_speculation_t *spec;
13646 dtrace_buffer_t *buf;
13647 cyc_handler_t hdlr;
13648 cyc_time_t when;
13649 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13650 dtrace_icookie_t cookie;
13651
13652 mutex_enter(&cpu_lock);
13653 mutex_enter(&dtrace_lock);
13654
13655 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13656 rval = EBUSY;
13657 goto out;
13658 }
13659
13660 /*
13661 * Before we can perform any checks, we must prime all of the
13662 * retained enablings that correspond to this state.
13663 */
13664 dtrace_enabling_prime(state);
13665
13666 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13667 rval = EACCES;
13668 goto out;
13669 }
13670
13671 dtrace_state_prereserve(state);
13672
13673 /*
13674 * Now we want to do is try to allocate our speculations.
13675 * We do not automatically resize the number of speculations; if
13676 * this fails, we will fail the operation.
13677 */
13678 nspec = opt[DTRACEOPT_NSPEC];
13679 ASSERT(nspec != DTRACEOPT_UNSET);
13680
13681 if (nspec > INT_MAX) {
13682 rval = ENOMEM;
13683 goto out;
13684 }
13685
13686 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
13687 KM_NOSLEEP | KM_NORMALPRI);
13688
13689 if (spec == NULL) {
13690 rval = ENOMEM;
13691 goto out;
13692 }
13693
13694 state->dts_speculations = spec;
13695 state->dts_nspeculations = (int)nspec;
13696
13697 for (i = 0; i < nspec; i++) {
13698 if ((buf = kmem_zalloc(bufsize,
13699 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
13700 rval = ENOMEM;
13701 goto err;
13702 }
13703
13704 spec[i].dtsp_buffer = buf;
13705 }
13706
13707 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13708 if (dtrace_anon.dta_state == NULL) {
13709 rval = ENOENT;
13710 goto out;
13711 }
13712
13713 if (state->dts_necbs != 0) {
13714 rval = EALREADY;
13715 goto out;
13716 }
13717
13718 state->dts_anon = dtrace_anon_grab();
13719 ASSERT(state->dts_anon != NULL);
13720 state = state->dts_anon;
13721
13722 /*
13723 * We want "grabanon" to be set in the grabbed state, so we'll
13724 * copy that option value from the grabbing state into the
13725 * grabbed state.
13726 */
13727 state->dts_options[DTRACEOPT_GRABANON] =
13728 opt[DTRACEOPT_GRABANON];
13729
13730 *cpu = dtrace_anon.dta_beganon;
13731
13732 /*
13733 * If the anonymous state is active (as it almost certainly
13734 * is if the anonymous enabling ultimately matched anything),
13735 * we don't allow any further option processing -- but we
13736 * don't return failure.
13737 */
13738 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13739 goto out;
13740 }
13741
13742 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13743 opt[DTRACEOPT_AGGSIZE] != 0) {
13744 if (state->dts_aggregations == NULL) {
13745 /*
13746 * We're not going to create an aggregation buffer
13747 * because we don't have any ECBs that contain
13748 * aggregations -- set this option to 0.
13749 */
13750 opt[DTRACEOPT_AGGSIZE] = 0;
13751 } else {
13752 /*
13753 * If we have an aggregation buffer, we must also have
13754 * a buffer to use as scratch.
13755 */
13756 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13757 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13758 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13759 }
13760 }
13761 }
13762
13763 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13764 opt[DTRACEOPT_SPECSIZE] != 0) {
13765 if (!state->dts_speculates) {
13766 /*
13767 * We're not going to create speculation buffers
13768 * because we don't have any ECBs that actually
13769 * speculate -- set the speculation size to 0.
13770 */
13771 opt[DTRACEOPT_SPECSIZE] = 0;
13772 }
13773 }
13774
13775 /*
13776 * The bare minimum size for any buffer that we're actually going to
13777 * do anything to is sizeof (uint64_t).
13778 */
13779 sz = sizeof (uint64_t);
13780
13781 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13782 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13783 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13784 /*
13785 * A buffer size has been explicitly set to 0 (or to a size
13786 * that will be adjusted to 0) and we need the space -- we
13787 * need to return failure. We return ENOSPC to differentiate
13788 * it from failing to allocate a buffer due to failure to meet
13789 * the reserve (for which we return E2BIG).
13790 */
13791 rval = ENOSPC;
13792 goto out;
13793 }
13794
13795 if ((rval = dtrace_state_buffers(state)) != 0)
13796 goto err;
13797
13798 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13799 sz = dtrace_dstate_defsize;
13800
13801 do {
13802 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13803
13804 if (rval == 0)
13805 break;
13806
13807 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13808 goto err;
13809 } while (sz >>= 1);
13810
13811 opt[DTRACEOPT_DYNVARSIZE] = sz;
13812
13813 if (rval != 0)
13814 goto err;
13815
13816 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13817 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13818
13819 if (opt[DTRACEOPT_CLEANRATE] == 0)
13820 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13821
13822 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13823 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13824
13825 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13826 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13827
13828 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13829 hdlr.cyh_arg = state;
13830 hdlr.cyh_level = CY_LOW_LEVEL;
13831
13832 when.cyt_when = 0;
13833 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13834
13835 state->dts_cleaner = cyclic_add(&hdlr, &when);
13836
13837 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13838 hdlr.cyh_arg = state;
13839 hdlr.cyh_level = CY_LOW_LEVEL;
13840
13841 when.cyt_when = 0;
13842 when.cyt_interval = dtrace_deadman_interval;
13843
13844 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13845 state->dts_deadman = cyclic_add(&hdlr, &when);
13846
13847 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13848
13849 if (state->dts_getf != 0 &&
13850 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13851 /*
13852 * We don't have kernel privs but we have at least one call
13853 * to getf(); we need to bump our zone's count, and (if
13854 * this is the first enabling to have an unprivileged call
13855 * to getf()) we need to hook into closef().
13856 */
13857 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
13858
13859 if (dtrace_getf++ == 0) {
13860 ASSERT(dtrace_closef == NULL);
13861 dtrace_closef = dtrace_getf_barrier;
13862 }
13863 }
13864
13865 /*
13866 * Now it's time to actually fire the BEGIN probe. We need to disable
13867 * interrupts here both to record the CPU on which we fired the BEGIN
13868 * probe (the data from this CPU will be processed first at user
13869 * level) and to manually activate the buffer for this CPU.
13870 */
13871 cookie = dtrace_interrupt_disable();
13872 *cpu = CPU->cpu_id;
13873 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13874 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13875
13876 dtrace_probe(dtrace_probeid_begin,
13877 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13878 dtrace_interrupt_enable(cookie);
13879 /*
13880 * We may have had an exit action from a BEGIN probe; only change our
13881 * state to ACTIVE if we're still in WARMUP.
13882 */
13883 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13884 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13885
13886 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13887 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13888
13889 /*
13890 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13891 * want each CPU to transition its principal buffer out of the
13892 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13893 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13894 * atomically transition from processing none of a state's ECBs to
13895 * processing all of them.
13896 */
13897 dtrace_xcall(DTRACE_CPUALL,
13898 (dtrace_xcall_t)dtrace_buffer_activate, state);
13899 goto out;
13900
13901 err:
13902 dtrace_buffer_free(state->dts_buffer);
13903 dtrace_buffer_free(state->dts_aggbuffer);
13904
13905 if ((nspec = state->dts_nspeculations) == 0) {
13906 ASSERT(state->dts_speculations == NULL);
13907 goto out;
13908 }
13909
13910 spec = state->dts_speculations;
13911 ASSERT(spec != NULL);
13912
13913 for (i = 0; i < state->dts_nspeculations; i++) {
13914 if ((buf = spec[i].dtsp_buffer) == NULL)
13915 break;
13916
13917 dtrace_buffer_free(buf);
13918 kmem_free(buf, bufsize);
13919 }
13920
13921 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13922 state->dts_nspeculations = 0;
13923 state->dts_speculations = NULL;
13924
13925 out:
13926 mutex_exit(&dtrace_lock);
13927 mutex_exit(&cpu_lock);
13928
13929 return (rval);
13930 }
13931
13932 static int
13933 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13934 {
13935 dtrace_icookie_t cookie;
13936
13937 ASSERT(MUTEX_HELD(&dtrace_lock));
13938
13939 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13940 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13941 return (EINVAL);
13942
13943 /*
13944 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13945 * to be sure that every CPU has seen it. See below for the details
13946 * on why this is done.
13947 */
13948 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13949 dtrace_sync();
13950
13951 /*
13952 * By this point, it is impossible for any CPU to be still processing
13953 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13954 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13955 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13956 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13957 * iff we're in the END probe.
13958 */
13959 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13960 dtrace_sync();
13961 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13962
13963 /*
13964 * Finally, we can release the reserve and call the END probe. We
13965 * disable interrupts across calling the END probe to allow us to
13966 * return the CPU on which we actually called the END probe. This
13967 * allows user-land to be sure that this CPU's principal buffer is
13968 * processed last.
13969 */
13970 state->dts_reserve = 0;
13971
13972 cookie = dtrace_interrupt_disable();
13973 *cpu = CPU->cpu_id;
13974 dtrace_probe(dtrace_probeid_end,
13975 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13976 dtrace_interrupt_enable(cookie);
13977
13978 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13979 dtrace_sync();
13980
13981 if (state->dts_getf != 0 &&
13982 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13983 /*
13984 * We don't have kernel privs but we have at least one call
13985 * to getf(); we need to lower our zone's count, and (if
13986 * this is the last enabling to have an unprivileged call
13987 * to getf()) we need to clear the closef() hook.
13988 */
13989 ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
13990 ASSERT(dtrace_closef == dtrace_getf_barrier);
13991 ASSERT(dtrace_getf > 0);
13992
13993 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
13994
13995 if (--dtrace_getf == 0)
13996 dtrace_closef = NULL;
13997 }
13998
13999 return (0);
14000 }
14001
14002 static int
14003 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
14004 dtrace_optval_t val)
14005 {
14006 ASSERT(MUTEX_HELD(&dtrace_lock));
14007
14008 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14009 return (EBUSY);
14010
14011 if (option >= DTRACEOPT_MAX)
14012 return (EINVAL);
14013
14014 if (option != DTRACEOPT_CPU && val < 0)
14015 return (EINVAL);
14016
14017 switch (option) {
14018 case DTRACEOPT_DESTRUCTIVE:
14019 if (dtrace_destructive_disallow)
14020 return (EACCES);
14021
14022 state->dts_cred.dcr_destructive = 1;
14023 break;
14024
14025 case DTRACEOPT_BUFSIZE:
14026 case DTRACEOPT_DYNVARSIZE:
14027 case DTRACEOPT_AGGSIZE:
14028 case DTRACEOPT_SPECSIZE:
14029 case DTRACEOPT_STRSIZE:
14030 if (val < 0)
14031 return (EINVAL);
14032
14033 if (val >= LONG_MAX) {
14034 /*
14035 * If this is an otherwise negative value, set it to
14036 * the highest multiple of 128m less than LONG_MAX.
14037 * Technically, we're adjusting the size without
14038 * regard to the buffer resizing policy, but in fact,
14039 * this has no effect -- if we set the buffer size to
14040 * ~LONG_MAX and the buffer policy is ultimately set to
14041 * be "manual", the buffer allocation is guaranteed to
14042 * fail, if only because the allocation requires two
14043 * buffers. (We set the the size to the highest
14044 * multiple of 128m because it ensures that the size
14045 * will remain a multiple of a megabyte when
14046 * repeatedly halved -- all the way down to 15m.)
14047 */
14048 val = LONG_MAX - (1 << 27) + 1;
14049 }
14050 }
14051
14052 state->dts_options[option] = val;
14053
14054 return (0);
14055 }
14056
14057 static void
14058 dtrace_state_destroy(dtrace_state_t *state)
14059 {
14060 dtrace_ecb_t *ecb;
14061 dtrace_vstate_t *vstate = &state->dts_vstate;
14062 minor_t minor = getminor(state->dts_dev);
14063 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14064 dtrace_speculation_t *spec = state->dts_speculations;
14065 int nspec = state->dts_nspeculations;
14066 uint32_t match;
14067
14068 ASSERT(MUTEX_HELD(&dtrace_lock));
14069 ASSERT(MUTEX_HELD(&cpu_lock));
14070
14071 /*
14072 * First, retract any retained enablings for this state.
14073 */
14074 dtrace_enabling_retract(state);
14075 ASSERT(state->dts_nretained == 0);
14076
14077 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
14078 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
14079 /*
14080 * We have managed to come into dtrace_state_destroy() on a
14081 * hot enabling -- almost certainly because of a disorderly
14082 * shutdown of a consumer. (That is, a consumer that is
14083 * exiting without having called dtrace_stop().) In this case,
14084 * we're going to set our activity to be KILLED, and then
14085 * issue a sync to be sure that everyone is out of probe
14086 * context before we start blowing away ECBs.
14087 */
14088 state->dts_activity = DTRACE_ACTIVITY_KILLED;
14089 dtrace_sync();
14090 }
14091
14092 /*
14093 * Release the credential hold we took in dtrace_state_create().
14094 */
14095 if (state->dts_cred.dcr_cred != NULL)
14096 crfree(state->dts_cred.dcr_cred);
14097
14098 /*
14099 * Now we can safely disable and destroy any enabled probes. Because
14100 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
14101 * (especially if they're all enabled), we take two passes through the
14102 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
14103 * in the second we disable whatever is left over.
14104 */
14105 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
14106 for (i = 0; i < state->dts_necbs; i++) {
14107 if ((ecb = state->dts_ecbs[i]) == NULL)
14108 continue;
14109
14110 if (match && ecb->dte_probe != NULL) {
14111 dtrace_probe_t *probe = ecb->dte_probe;
14112 dtrace_provider_t *prov = probe->dtpr_provider;
14113
14114 if (!(prov->dtpv_priv.dtpp_flags & match))
14115 continue;
14116 }
14117
14118 dtrace_ecb_disable(ecb);
14119 dtrace_ecb_destroy(ecb);
14120 }
14121
14122 if (!match)
14123 break;
14124 }
14125
14126 /*
14127 * Before we free the buffers, perform one more sync to assure that
14128 * every CPU is out of probe context.
14129 */
14130 dtrace_sync();
14131
14132 dtrace_buffer_free(state->dts_buffer);
14133 dtrace_buffer_free(state->dts_aggbuffer);
14134
14135 for (i = 0; i < nspec; i++)
14136 dtrace_buffer_free(spec[i].dtsp_buffer);
14137
14138 if (state->dts_cleaner != CYCLIC_NONE)
14139 cyclic_remove(state->dts_cleaner);
14140
14141 if (state->dts_deadman != CYCLIC_NONE)
14142 cyclic_remove(state->dts_deadman);
14143
14144 dtrace_dstate_fini(&vstate->dtvs_dynvars);
14145 dtrace_vstate_fini(vstate);
14146 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14147
14148 if (state->dts_aggregations != NULL) {
14149 #ifdef DEBUG
14150 for (i = 0; i < state->dts_naggregations; i++)
14151 ASSERT(state->dts_aggregations[i] == NULL);
14152 #endif
14153 ASSERT(state->dts_naggregations > 0);
14154 kmem_free(state->dts_aggregations,
14155 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14156 }
14157
14158 kmem_free(state->dts_buffer, bufsize);
14159 kmem_free(state->dts_aggbuffer, bufsize);
14160
14161 for (i = 0; i < nspec; i++)
14162 kmem_free(spec[i].dtsp_buffer, bufsize);
14163
14164 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14165
14166 dtrace_format_destroy(state);
14167
14168 vmem_destroy(state->dts_aggid_arena);
14169 ddi_soft_state_free(dtrace_softstate, minor);
14170 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14171 }
14172
14173 /*
14174 * DTrace Anonymous Enabling Functions
14175 */
14176 static dtrace_state_t *
14177 dtrace_anon_grab(void)
14178 {
14179 dtrace_state_t *state;
14180
14181 ASSERT(MUTEX_HELD(&dtrace_lock));
14182
14183 if ((state = dtrace_anon.dta_state) == NULL) {
14184 ASSERT(dtrace_anon.dta_enabling == NULL);
14185 return (NULL);
14186 }
14187
14188 ASSERT(dtrace_anon.dta_enabling != NULL);
14189 ASSERT(dtrace_retained != NULL);
14190
14191 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14192 dtrace_anon.dta_enabling = NULL;
14193 dtrace_anon.dta_state = NULL;
14194
14195 return (state);
14196 }
14197
14198 static void
14199 dtrace_anon_property(void)
14200 {
14201 int i, rv;
14202 dtrace_state_t *state;
14203 dof_hdr_t *dof;
14204 char c[32]; /* enough for "dof-data-" + digits */
14205
14206 ASSERT(MUTEX_HELD(&dtrace_lock));
14207 ASSERT(MUTEX_HELD(&cpu_lock));
14208
14209 for (i = 0; ; i++) {
14210 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14211
14212 dtrace_err_verbose = 1;
14213
14214 if ((dof = dtrace_dof_property(c)) == NULL) {
14215 dtrace_err_verbose = 0;
14216 break;
14217 }
14218
14219 /*
14220 * We want to create anonymous state, so we need to transition
14221 * the kernel debugger to indicate that DTrace is active. If
14222 * this fails (e.g. because the debugger has modified text in
14223 * some way), we won't continue with the processing.
14224 */
14225 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14226 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14227 "enabling ignored.");
14228 dtrace_dof_destroy(dof);
14229 break;
14230 }
14231
14232 /*
14233 * If we haven't allocated an anonymous state, we'll do so now.
14234 */
14235 if ((state = dtrace_anon.dta_state) == NULL) {
14236 state = dtrace_state_create(NULL, NULL);
14237 dtrace_anon.dta_state = state;
14238
14239 if (state == NULL) {
14240 /*
14241 * This basically shouldn't happen: the only
14242 * failure mode from dtrace_state_create() is a
14243 * failure of ddi_soft_state_zalloc() that
14244 * itself should never happen. Still, the
14245 * interface allows for a failure mode, and
14246 * we want to fail as gracefully as possible:
14247 * we'll emit an error message and cease
14248 * processing anonymous state in this case.
14249 */
14250 cmn_err(CE_WARN, "failed to create "
14251 "anonymous state");
14252 dtrace_dof_destroy(dof);
14253 break;
14254 }
14255 }
14256
14257 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14258 &dtrace_anon.dta_enabling, 0, B_TRUE);
14259
14260 if (rv == 0)
14261 rv = dtrace_dof_options(dof, state);
14262
14263 dtrace_err_verbose = 0;
14264 dtrace_dof_destroy(dof);
14265
14266 if (rv != 0) {
14267 /*
14268 * This is malformed DOF; chuck any anonymous state
14269 * that we created.
14270 */
14271 ASSERT(dtrace_anon.dta_enabling == NULL);
14272 dtrace_state_destroy(state);
14273 dtrace_anon.dta_state = NULL;
14274 break;
14275 }
14276
14277 ASSERT(dtrace_anon.dta_enabling != NULL);
14278 }
14279
14280 if (dtrace_anon.dta_enabling != NULL) {
14281 int rval;
14282
14283 /*
14284 * dtrace_enabling_retain() can only fail because we are
14285 * trying to retain more enablings than are allowed -- but
14286 * we only have one anonymous enabling, and we are guaranteed
14287 * to be allowed at least one retained enabling; we assert
14288 * that dtrace_enabling_retain() returns success.
14289 */
14290 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14291 ASSERT(rval == 0);
14292
14293 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14294 }
14295 }
14296
14297 /*
14298 * DTrace Helper Functions
14299 */
14300 static void
14301 dtrace_helper_trace(dtrace_helper_action_t *helper,
14302 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14303 {
14304 uint32_t size, next, nnext, i;
14305 dtrace_helptrace_t *ent, *buffer;
14306 uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14307
14308 if ((buffer = dtrace_helptrace_buffer) == NULL)
14309 return;
14310
14311 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14312
14313 /*
14314 * What would a tracing framework be without its own tracing
14315 * framework? (Well, a hell of a lot simpler, for starters...)
14316 */
14317 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14318 sizeof (uint64_t) - sizeof (uint64_t);
14319
14320 /*
14321 * Iterate until we can allocate a slot in the trace buffer.
14322 */
14323 do {
14324 next = dtrace_helptrace_next;
14325
14326 if (next + size < dtrace_helptrace_bufsize) {
14327 nnext = next + size;
14328 } else {
14329 nnext = size;
14330 }
14331 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14332
14333 /*
14334 * We have our slot; fill it in.
14335 */
14336 if (nnext == size) {
14337 dtrace_helptrace_wrapped++;
14338 next = 0;
14339 }
14340
14341 ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
14342 ent->dtht_helper = helper;
14343 ent->dtht_where = where;
14344 ent->dtht_nlocals = vstate->dtvs_nlocals;
14345
14346 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14347 mstate->dtms_fltoffs : -1;
14348 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14349 ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
14350
14351 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14352 dtrace_statvar_t *svar;
14353
14354 if ((svar = vstate->dtvs_locals[i]) == NULL)
14355 continue;
14356
14357 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14358 ent->dtht_locals[i] =
14359 ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
14360 }
14361 }
14362
14363 static uint64_t
14364 dtrace_helper(int which, dtrace_mstate_t *mstate,
14365 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14366 {
14367 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14368 uint64_t sarg0 = mstate->dtms_arg[0];
14369 uint64_t sarg1 = mstate->dtms_arg[1];
14370 uint64_t rval;
14371 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14372 dtrace_helper_action_t *helper;
14373 dtrace_vstate_t *vstate;
14374 dtrace_difo_t *pred;
14375 int i, trace = dtrace_helptrace_buffer != NULL;
14376
14377 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14378
14379 if (helpers == NULL)
14380 return (0);
14381
14382 if ((helper = helpers->dthps_actions[which]) == NULL)
14383 return (0);
14384
14385 vstate = &helpers->dthps_vstate;
14386 mstate->dtms_arg[0] = arg0;
14387 mstate->dtms_arg[1] = arg1;
14388
14389 /*
14390 * Now iterate over each helper. If its predicate evaluates to 'true',
14391 * we'll call the corresponding actions. Note that the below calls
14392 * to dtrace_dif_emulate() may set faults in machine state. This is
14393 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14394 * the stored DIF offset with its own (which is the desired behavior).
14395 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14396 * from machine state; this is okay, too.
14397 */
14398 for (; helper != NULL; helper = helper->dtha_next) {
14399 if ((pred = helper->dtha_predicate) != NULL) {
14400 if (trace)
14401 dtrace_helper_trace(helper, mstate, vstate, 0);
14402
14403 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14404 goto next;
14405
14406 if (*flags & CPU_DTRACE_FAULT)
14407 goto err;
14408 }
14409
14410 for (i = 0; i < helper->dtha_nactions; i++) {
14411 if (trace)
14412 dtrace_helper_trace(helper,
14413 mstate, vstate, i + 1);
14414
14415 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14416 mstate, vstate, state);
14417
14418 if (*flags & CPU_DTRACE_FAULT)
14419 goto err;
14420 }
14421
14422 next:
14423 if (trace)
14424 dtrace_helper_trace(helper, mstate, vstate,
14425 DTRACE_HELPTRACE_NEXT);
14426 }
14427
14428 if (trace)
14429 dtrace_helper_trace(helper, mstate, vstate,
14430 DTRACE_HELPTRACE_DONE);
14431
14432 /*
14433 * Restore the arg0 that we saved upon entry.
14434 */
14435 mstate->dtms_arg[0] = sarg0;
14436 mstate->dtms_arg[1] = sarg1;
14437
14438 return (rval);
14439
14440 err:
14441 if (trace)
14442 dtrace_helper_trace(helper, mstate, vstate,
14443 DTRACE_HELPTRACE_ERR);
14444
14445 /*
14446 * Restore the arg0 that we saved upon entry.
14447 */
14448 mstate->dtms_arg[0] = sarg0;
14449 mstate->dtms_arg[1] = sarg1;
14450
14451 return (NULL);
14452 }
14453
14454 static void
14455 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14456 dtrace_vstate_t *vstate)
14457 {
14458 int i;
14459
14460 if (helper->dtha_predicate != NULL)
14461 dtrace_difo_release(helper->dtha_predicate, vstate);
14462
14463 for (i = 0; i < helper->dtha_nactions; i++) {
14464 ASSERT(helper->dtha_actions[i] != NULL);
14465 dtrace_difo_release(helper->dtha_actions[i], vstate);
14466 }
14467
14468 kmem_free(helper->dtha_actions,
14469 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14470 kmem_free(helper, sizeof (dtrace_helper_action_t));
14471 }
14472
14473 static int
14474 dtrace_helper_destroygen(int gen)
14475 {
14476 proc_t *p = curproc;
14477 dtrace_helpers_t *help = p->p_dtrace_helpers;
14478 dtrace_vstate_t *vstate;
14479 int i;
14480
14481 ASSERT(MUTEX_HELD(&dtrace_lock));
14482
14483 if (help == NULL || gen > help->dthps_generation)
14484 return (EINVAL);
14485
14486 vstate = &help->dthps_vstate;
14487
14488 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14489 dtrace_helper_action_t *last = NULL, *h, *next;
14490
14491 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14492 next = h->dtha_next;
14493
14494 if (h->dtha_generation == gen) {
14495 if (last != NULL) {
14496 last->dtha_next = next;
14497 } else {
14498 help->dthps_actions[i] = next;
14499 }
14500
14501 dtrace_helper_action_destroy(h, vstate);
14502 } else {
14503 last = h;
14504 }
14505 }
14506 }
14507
14508 /*
14509 * Interate until we've cleared out all helper providers with the
14510 * given generation number.
14511 */
14512 for (;;) {
14513 dtrace_helper_provider_t *prov;
14514
14515 /*
14516 * Look for a helper provider with the right generation. We
14517 * have to start back at the beginning of the list each time
14518 * because we drop dtrace_lock. It's unlikely that we'll make
14519 * more than two passes.
14520 */
14521 for (i = 0; i < help->dthps_nprovs; i++) {
14522 prov = help->dthps_provs[i];
14523
14524 if (prov->dthp_generation == gen)
14525 break;
14526 }
14527
14528 /*
14529 * If there were no matches, we're done.
14530 */
14531 if (i == help->dthps_nprovs)
14532 break;
14533
14534 /*
14535 * Move the last helper provider into this slot.
14536 */
14537 help->dthps_nprovs--;
14538 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14539 help->dthps_provs[help->dthps_nprovs] = NULL;
14540
14541 mutex_exit(&dtrace_lock);
14542
14543 /*
14544 * If we have a meta provider, remove this helper provider.
14545 */
14546 mutex_enter(&dtrace_meta_lock);
14547 if (dtrace_meta_pid != NULL) {
14548 ASSERT(dtrace_deferred_pid == NULL);
14549 dtrace_helper_provider_remove(&prov->dthp_prov,
14550 p->p_pid);
14551 }
14552 mutex_exit(&dtrace_meta_lock);
14553
14554 dtrace_helper_provider_destroy(prov);
14555
14556 mutex_enter(&dtrace_lock);
14557 }
14558
14559 return (0);
14560 }
14561
14562 static int
14563 dtrace_helper_validate(dtrace_helper_action_t *helper)
14564 {
14565 int err = 0, i;
14566 dtrace_difo_t *dp;
14567
14568 if ((dp = helper->dtha_predicate) != NULL)
14569 err += dtrace_difo_validate_helper(dp);
14570
14571 for (i = 0; i < helper->dtha_nactions; i++)
14572 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14573
14574 return (err == 0);
14575 }
14576
14577 static int
14578 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14579 {
14580 dtrace_helpers_t *help;
14581 dtrace_helper_action_t *helper, *last;
14582 dtrace_actdesc_t *act;
14583 dtrace_vstate_t *vstate;
14584 dtrace_predicate_t *pred;
14585 int count = 0, nactions = 0, i;
14586
14587 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14588 return (EINVAL);
14589
14590 help = curproc->p_dtrace_helpers;
14591 last = help->dthps_actions[which];
14592 vstate = &help->dthps_vstate;
14593
14594 for (count = 0; last != NULL; last = last->dtha_next) {
14595 count++;
14596 if (last->dtha_next == NULL)
14597 break;
14598 }
14599
14600 /*
14601 * If we already have dtrace_helper_actions_max helper actions for this
14602 * helper action type, we'll refuse to add a new one.
14603 */
14604 if (count >= dtrace_helper_actions_max)
14605 return (ENOSPC);
14606
14607 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14608 helper->dtha_generation = help->dthps_generation;
14609
14610 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14611 ASSERT(pred->dtp_difo != NULL);
14612 dtrace_difo_hold(pred->dtp_difo);
14613 helper->dtha_predicate = pred->dtp_difo;
14614 }
14615
14616 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14617 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14618 goto err;
14619
14620 if (act->dtad_difo == NULL)
14621 goto err;
14622
14623 nactions++;
14624 }
14625
14626 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14627 (helper->dtha_nactions = nactions), KM_SLEEP);
14628
14629 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14630 dtrace_difo_hold(act->dtad_difo);
14631 helper->dtha_actions[i++] = act->dtad_difo;
14632 }
14633
14634 if (!dtrace_helper_validate(helper))
14635 goto err;
14636
14637 if (last == NULL) {
14638 help->dthps_actions[which] = helper;
14639 } else {
14640 last->dtha_next = helper;
14641 }
14642
14643 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14644 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14645 dtrace_helptrace_next = 0;
14646 }
14647
14648 return (0);
14649 err:
14650 dtrace_helper_action_destroy(helper, vstate);
14651 return (EINVAL);
14652 }
14653
14654 static void
14655 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14656 dof_helper_t *dofhp)
14657 {
14658 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14659
14660 mutex_enter(&dtrace_meta_lock);
14661 mutex_enter(&dtrace_lock);
14662
14663 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14664 /*
14665 * If the dtrace module is loaded but not attached, or if
14666 * there aren't isn't a meta provider registered to deal with
14667 * these provider descriptions, we need to postpone creating
14668 * the actual providers until later.
14669 */
14670
14671 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14672 dtrace_deferred_pid != help) {
14673 help->dthps_deferred = 1;
14674 help->dthps_pid = p->p_pid;
14675 help->dthps_next = dtrace_deferred_pid;
14676 help->dthps_prev = NULL;
14677 if (dtrace_deferred_pid != NULL)
14678 dtrace_deferred_pid->dthps_prev = help;
14679 dtrace_deferred_pid = help;
14680 }
14681
14682 mutex_exit(&dtrace_lock);
14683
14684 } else if (dofhp != NULL) {
14685 /*
14686 * If the dtrace module is loaded and we have a particular
14687 * helper provider description, pass that off to the
14688 * meta provider.
14689 */
14690
14691 mutex_exit(&dtrace_lock);
14692
14693 dtrace_helper_provide(dofhp, p->p_pid);
14694
14695 } else {
14696 /*
14697 * Otherwise, just pass all the helper provider descriptions
14698 * off to the meta provider.
14699 */
14700
14701 int i;
14702 mutex_exit(&dtrace_lock);
14703
14704 for (i = 0; i < help->dthps_nprovs; i++) {
14705 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14706 p->p_pid);
14707 }
14708 }
14709
14710 mutex_exit(&dtrace_meta_lock);
14711 }
14712
14713 static int
14714 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14715 {
14716 dtrace_helpers_t *help;
14717 dtrace_helper_provider_t *hprov, **tmp_provs;
14718 uint_t tmp_maxprovs, i;
14719
14720 ASSERT(MUTEX_HELD(&dtrace_lock));
14721
14722 help = curproc->p_dtrace_helpers;
14723 ASSERT(help != NULL);
14724
14725 /*
14726 * If we already have dtrace_helper_providers_max helper providers,
14727 * we're refuse to add a new one.
14728 */
14729 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14730 return (ENOSPC);
14731
14732 /*
14733 * Check to make sure this isn't a duplicate.
14734 */
14735 for (i = 0; i < help->dthps_nprovs; i++) {
14736 if (dofhp->dofhp_addr ==
14737 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14738 return (EALREADY);
14739 }
14740
14741 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14742 hprov->dthp_prov = *dofhp;
14743 hprov->dthp_ref = 1;
14744 hprov->dthp_generation = gen;
14745
14746 /*
14747 * Allocate a bigger table for helper providers if it's already full.
14748 */
14749 if (help->dthps_maxprovs == help->dthps_nprovs) {
14750 tmp_maxprovs = help->dthps_maxprovs;
14751 tmp_provs = help->dthps_provs;
14752
14753 if (help->dthps_maxprovs == 0)
14754 help->dthps_maxprovs = 2;
14755 else
14756 help->dthps_maxprovs *= 2;
14757 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14758 help->dthps_maxprovs = dtrace_helper_providers_max;
14759
14760 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14761
14762 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14763 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14764
14765 if (tmp_provs != NULL) {
14766 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14767 sizeof (dtrace_helper_provider_t *));
14768 kmem_free(tmp_provs, tmp_maxprovs *
14769 sizeof (dtrace_helper_provider_t *));
14770 }
14771 }
14772
14773 help->dthps_provs[help->dthps_nprovs] = hprov;
14774 help->dthps_nprovs++;
14775
14776 return (0);
14777 }
14778
14779 static void
14780 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14781 {
14782 mutex_enter(&dtrace_lock);
14783
14784 if (--hprov->dthp_ref == 0) {
14785 dof_hdr_t *dof;
14786 mutex_exit(&dtrace_lock);
14787 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14788 dtrace_dof_destroy(dof);
14789 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14790 } else {
14791 mutex_exit(&dtrace_lock);
14792 }
14793 }
14794
14795 static int
14796 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14797 {
14798 uintptr_t daddr = (uintptr_t)dof;
14799 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14800 dof_provider_t *provider;
14801 dof_probe_t *probe;
14802 uint8_t *arg;
14803 char *strtab, *typestr;
14804 dof_stridx_t typeidx;
14805 size_t typesz;
14806 uint_t nprobes, j, k;
14807
14808 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14809
14810 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14811 dtrace_dof_error(dof, "misaligned section offset");
14812 return (-1);
14813 }
14814
14815 /*
14816 * The section needs to be large enough to contain the DOF provider
14817 * structure appropriate for the given version.
14818 */
14819 if (sec->dofs_size <
14820 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14821 offsetof(dof_provider_t, dofpv_prenoffs) :
14822 sizeof (dof_provider_t))) {
14823 dtrace_dof_error(dof, "provider section too small");
14824 return (-1);
14825 }
14826
14827 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14828 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14829 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14830 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14831 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14832
14833 if (str_sec == NULL || prb_sec == NULL ||
14834 arg_sec == NULL || off_sec == NULL)
14835 return (-1);
14836
14837 enoff_sec = NULL;
14838
14839 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14840 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14841 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14842 provider->dofpv_prenoffs)) == NULL)
14843 return (-1);
14844
14845 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14846
14847 if (provider->dofpv_name >= str_sec->dofs_size ||
14848 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14849 dtrace_dof_error(dof, "invalid provider name");
14850 return (-1);
14851 }
14852
14853 if (prb_sec->dofs_entsize == 0 ||
14854 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14855 dtrace_dof_error(dof, "invalid entry size");
14856 return (-1);
14857 }
14858
14859 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14860 dtrace_dof_error(dof, "misaligned entry size");
14861 return (-1);
14862 }
14863
14864 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14865 dtrace_dof_error(dof, "invalid entry size");
14866 return (-1);
14867 }
14868
14869 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14870 dtrace_dof_error(dof, "misaligned section offset");
14871 return (-1);
14872 }
14873
14874 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14875 dtrace_dof_error(dof, "invalid entry size");
14876 return (-1);
14877 }
14878
14879 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14880
14881 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14882
14883 /*
14884 * Take a pass through the probes to check for errors.
14885 */
14886 for (j = 0; j < nprobes; j++) {
14887 probe = (dof_probe_t *)(uintptr_t)(daddr +
14888 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14889
14890 if (probe->dofpr_func >= str_sec->dofs_size) {
14891 dtrace_dof_error(dof, "invalid function name");
14892 return (-1);
14893 }
14894
14895 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14896 dtrace_dof_error(dof, "function name too long");
14897 return (-1);
14898 }
14899
14900 if (probe->dofpr_name >= str_sec->dofs_size ||
14901 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14902 dtrace_dof_error(dof, "invalid probe name");
14903 return (-1);
14904 }
14905
14906 /*
14907 * The offset count must not wrap the index, and the offsets
14908 * must also not overflow the section's data.
14909 */
14910 if (probe->dofpr_offidx + probe->dofpr_noffs <
14911 probe->dofpr_offidx ||
14912 (probe->dofpr_offidx + probe->dofpr_noffs) *
14913 off_sec->dofs_entsize > off_sec->dofs_size) {
14914 dtrace_dof_error(dof, "invalid probe offset");
14915 return (-1);
14916 }
14917
14918 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14919 /*
14920 * If there's no is-enabled offset section, make sure
14921 * there aren't any is-enabled offsets. Otherwise
14922 * perform the same checks as for probe offsets
14923 * (immediately above).
14924 */
14925 if (enoff_sec == NULL) {
14926 if (probe->dofpr_enoffidx != 0 ||
14927 probe->dofpr_nenoffs != 0) {
14928 dtrace_dof_error(dof, "is-enabled "
14929 "offsets with null section");
14930 return (-1);
14931 }
14932 } else if (probe->dofpr_enoffidx +
14933 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14934 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14935 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14936 dtrace_dof_error(dof, "invalid is-enabled "
14937 "offset");
14938 return (-1);
14939 }
14940
14941 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14942 dtrace_dof_error(dof, "zero probe and "
14943 "is-enabled offsets");
14944 return (-1);
14945 }
14946 } else if (probe->dofpr_noffs == 0) {
14947 dtrace_dof_error(dof, "zero probe offsets");
14948 return (-1);
14949 }
14950
14951 if (probe->dofpr_argidx + probe->dofpr_xargc <
14952 probe->dofpr_argidx ||
14953 (probe->dofpr_argidx + probe->dofpr_xargc) *
14954 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14955 dtrace_dof_error(dof, "invalid args");
14956 return (-1);
14957 }
14958
14959 typeidx = probe->dofpr_nargv;
14960 typestr = strtab + probe->dofpr_nargv;
14961 for (k = 0; k < probe->dofpr_nargc; k++) {
14962 if (typeidx >= str_sec->dofs_size) {
14963 dtrace_dof_error(dof, "bad "
14964 "native argument type");
14965 return (-1);
14966 }
14967
14968 typesz = strlen(typestr) + 1;
14969 if (typesz > DTRACE_ARGTYPELEN) {
14970 dtrace_dof_error(dof, "native "
14971 "argument type too long");
14972 return (-1);
14973 }
14974 typeidx += typesz;
14975 typestr += typesz;
14976 }
14977
14978 typeidx = probe->dofpr_xargv;
14979 typestr = strtab + probe->dofpr_xargv;
14980 for (k = 0; k < probe->dofpr_xargc; k++) {
14981 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14982 dtrace_dof_error(dof, "bad "
14983 "native argument index");
14984 return (-1);
14985 }
14986
14987 if (typeidx >= str_sec->dofs_size) {
14988 dtrace_dof_error(dof, "bad "
14989 "translated argument type");
14990 return (-1);
14991 }
14992
14993 typesz = strlen(typestr) + 1;
14994 if (typesz > DTRACE_ARGTYPELEN) {
14995 dtrace_dof_error(dof, "translated argument "
14996 "type too long");
14997 return (-1);
14998 }
14999
15000 typeidx += typesz;
15001 typestr += typesz;
15002 }
15003 }
15004
15005 return (0);
15006 }
15007
15008 static int
15009 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
15010 {
15011 dtrace_helpers_t *help;
15012 dtrace_vstate_t *vstate;
15013 dtrace_enabling_t *enab = NULL;
15014 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
15015 uintptr_t daddr = (uintptr_t)dof;
15016
15017 ASSERT(MUTEX_HELD(&dtrace_lock));
15018
15019 if ((help = curproc->p_dtrace_helpers) == NULL)
15020 help = dtrace_helpers_create(curproc);
15021
15022 vstate = &help->dthps_vstate;
15023
15024 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
15025 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
15026 dtrace_dof_destroy(dof);
15027 return (rv);
15028 }
15029
15030 /*
15031 * Look for helper providers and validate their descriptions.
15032 */
15033 if (dhp != NULL) {
15034 for (i = 0; i < dof->dofh_secnum; i++) {
15035 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
15036 dof->dofh_secoff + i * dof->dofh_secsize);
15037
15038 if (sec->dofs_type != DOF_SECT_PROVIDER)
15039 continue;
15040
15041 if (dtrace_helper_provider_validate(dof, sec) != 0) {
15042 dtrace_enabling_destroy(enab);
15043 dtrace_dof_destroy(dof);
15044 return (-1);
15045 }
15046
15047 nprovs++;
15048 }
15049 }
15050
15051 /*
15052 * Now we need to walk through the ECB descriptions in the enabling.
15053 */
15054 for (i = 0; i < enab->dten_ndesc; i++) {
15055 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
15056 dtrace_probedesc_t *desc = &ep->dted_probe;
15057
15058 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
15059 continue;
15060
15061 if (strcmp(desc->dtpd_mod, "helper") != 0)
15062 continue;
15063
15064 if (strcmp(desc->dtpd_func, "ustack") != 0)
15065 continue;
15066
15067 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
15068 ep)) != 0) {
15069 /*
15070 * Adding this helper action failed -- we are now going
15071 * to rip out the entire generation and return failure.
15072 */
15073 (void) dtrace_helper_destroygen(help->dthps_generation);
15074 dtrace_enabling_destroy(enab);
15075 dtrace_dof_destroy(dof);
15076 return (-1);
15077 }
15078
15079 nhelpers++;
15080 }
15081
15082 if (nhelpers < enab->dten_ndesc)
15083 dtrace_dof_error(dof, "unmatched helpers");
15084
15085 gen = help->dthps_generation++;
15086 dtrace_enabling_destroy(enab);
15087
15088 if (dhp != NULL && nprovs > 0) {
15089 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
15090 if (dtrace_helper_provider_add(dhp, gen) == 0) {
15091 mutex_exit(&dtrace_lock);
15092 dtrace_helper_provider_register(curproc, help, dhp);
15093 mutex_enter(&dtrace_lock);
15094
15095 destroy = 0;
15096 }
15097 }
15098
15099 if (destroy)
15100 dtrace_dof_destroy(dof);
15101
15102 return (gen);
15103 }
15104
15105 static dtrace_helpers_t *
15106 dtrace_helpers_create(proc_t *p)
15107 {
15108 dtrace_helpers_t *help;
15109
15110 ASSERT(MUTEX_HELD(&dtrace_lock));
15111 ASSERT(p->p_dtrace_helpers == NULL);
15112
15113 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
15114 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
15115 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
15116
15117 p->p_dtrace_helpers = help;
15118 dtrace_helpers++;
15119
15120 return (help);
15121 }
15122
15123 static void
15124 dtrace_helpers_destroy(void)
15125 {
15126 dtrace_helpers_t *help;
15127 dtrace_vstate_t *vstate;
15128 proc_t *p = curproc;
15129 int i;
15130
15131 mutex_enter(&dtrace_lock);
15132
15133 ASSERT(p->p_dtrace_helpers != NULL);
15134 ASSERT(dtrace_helpers > 0);
15135
15136 help = p->p_dtrace_helpers;
15137 vstate = &help->dthps_vstate;
15138
15139 /*
15140 * We're now going to lose the help from this process.
15141 */
15142 p->p_dtrace_helpers = NULL;
15143 dtrace_sync();
15144
15145 /*
15146 * Destory the helper actions.
15147 */
15148 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15149 dtrace_helper_action_t *h, *next;
15150
15151 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15152 next = h->dtha_next;
15153 dtrace_helper_action_destroy(h, vstate);
15154 h = next;
15155 }
15156 }
15157
15158 mutex_exit(&dtrace_lock);
15159
15160 /*
15161 * Destroy the helper providers.
15162 */
15163 if (help->dthps_maxprovs > 0) {
15164 mutex_enter(&dtrace_meta_lock);
15165 if (dtrace_meta_pid != NULL) {
15166 ASSERT(dtrace_deferred_pid == NULL);
15167
15168 for (i = 0; i < help->dthps_nprovs; i++) {
15169 dtrace_helper_provider_remove(
15170 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15171 }
15172 } else {
15173 mutex_enter(&dtrace_lock);
15174 ASSERT(help->dthps_deferred == 0 ||
15175 help->dthps_next != NULL ||
15176 help->dthps_prev != NULL ||
15177 help == dtrace_deferred_pid);
15178
15179 /*
15180 * Remove the helper from the deferred list.
15181 */
15182 if (help->dthps_next != NULL)
15183 help->dthps_next->dthps_prev = help->dthps_prev;
15184 if (help->dthps_prev != NULL)
15185 help->dthps_prev->dthps_next = help->dthps_next;
15186 if (dtrace_deferred_pid == help) {
15187 dtrace_deferred_pid = help->dthps_next;
15188 ASSERT(help->dthps_prev == NULL);
15189 }
15190
15191 mutex_exit(&dtrace_lock);
15192 }
15193
15194 mutex_exit(&dtrace_meta_lock);
15195
15196 for (i = 0; i < help->dthps_nprovs; i++) {
15197 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15198 }
15199
15200 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15201 sizeof (dtrace_helper_provider_t *));
15202 }
15203
15204 mutex_enter(&dtrace_lock);
15205
15206 dtrace_vstate_fini(&help->dthps_vstate);
15207 kmem_free(help->dthps_actions,
15208 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15209 kmem_free(help, sizeof (dtrace_helpers_t));
15210
15211 --dtrace_helpers;
15212 mutex_exit(&dtrace_lock);
15213 }
15214
15215 static void
15216 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15217 {
15218 dtrace_helpers_t *help, *newhelp;
15219 dtrace_helper_action_t *helper, *new, *last;
15220 dtrace_difo_t *dp;
15221 dtrace_vstate_t *vstate;
15222 int i, j, sz, hasprovs = 0;
15223
15224 mutex_enter(&dtrace_lock);
15225 ASSERT(from->p_dtrace_helpers != NULL);
15226 ASSERT(dtrace_helpers > 0);
15227
15228 help = from->p_dtrace_helpers;
15229 newhelp = dtrace_helpers_create(to);
15230 ASSERT(to->p_dtrace_helpers != NULL);
15231
15232 newhelp->dthps_generation = help->dthps_generation;
15233 vstate = &newhelp->dthps_vstate;
15234
15235 /*
15236 * Duplicate the helper actions.
15237 */
15238 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15239 if ((helper = help->dthps_actions[i]) == NULL)
15240 continue;
15241
15242 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15243 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15244 KM_SLEEP);
15245 new->dtha_generation = helper->dtha_generation;
15246
15247 if ((dp = helper->dtha_predicate) != NULL) {
15248 dp = dtrace_difo_duplicate(dp, vstate);
15249 new->dtha_predicate = dp;
15250 }
15251
15252 new->dtha_nactions = helper->dtha_nactions;
15253 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15254 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15255
15256 for (j = 0; j < new->dtha_nactions; j++) {
15257 dtrace_difo_t *dp = helper->dtha_actions[j];
15258
15259 ASSERT(dp != NULL);
15260 dp = dtrace_difo_duplicate(dp, vstate);
15261 new->dtha_actions[j] = dp;
15262 }
15263
15264 if (last != NULL) {
15265 last->dtha_next = new;
15266 } else {
15267 newhelp->dthps_actions[i] = new;
15268 }
15269
15270 last = new;
15271 }
15272 }
15273
15274 /*
15275 * Duplicate the helper providers and register them with the
15276 * DTrace framework.
15277 */
15278 if (help->dthps_nprovs > 0) {
15279 newhelp->dthps_nprovs = help->dthps_nprovs;
15280 newhelp->dthps_maxprovs = help->dthps_nprovs;
15281 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15282 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15283 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15284 newhelp->dthps_provs[i] = help->dthps_provs[i];
15285 newhelp->dthps_provs[i]->dthp_ref++;
15286 }
15287
15288 hasprovs = 1;
15289 }
15290
15291 mutex_exit(&dtrace_lock);
15292
15293 if (hasprovs)
15294 dtrace_helper_provider_register(to, newhelp, NULL);
15295 }
15296
15297 /*
15298 * DTrace Hook Functions
15299 */
15300 static void
15301 dtrace_module_loaded(struct modctl *ctl)
15302 {
15303 dtrace_provider_t *prv;
15304
15305 mutex_enter(&dtrace_provider_lock);
15306 mutex_enter(&mod_lock);
15307
15308 ASSERT(ctl->mod_busy);
15309
15310 /*
15311 * We're going to call each providers per-module provide operation
15312 * specifying only this module.
15313 */
15314 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15315 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15316
15317 mutex_exit(&mod_lock);
15318 mutex_exit(&dtrace_provider_lock);
15319
15320 /*
15321 * If we have any retained enablings, we need to match against them.
15322 * Enabling probes requires that cpu_lock be held, and we cannot hold
15323 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15324 * module. (In particular, this happens when loading scheduling
15325 * classes.) So if we have any retained enablings, we need to dispatch
15326 * our task queue to do the match for us.
15327 */
15328 mutex_enter(&dtrace_lock);
15329
15330 if (dtrace_retained == NULL) {
15331 mutex_exit(&dtrace_lock);
15332 return;
15333 }
15334
15335 (void) taskq_dispatch(dtrace_taskq,
15336 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15337
15338 mutex_exit(&dtrace_lock);
15339
15340 /*
15341 * And now, for a little heuristic sleaze: in general, we want to
15342 * match modules as soon as they load. However, we cannot guarantee
15343 * this, because it would lead us to the lock ordering violation
15344 * outlined above. The common case, of course, is that cpu_lock is
15345 * _not_ held -- so we delay here for a clock tick, hoping that that's
15346 * long enough for the task queue to do its work. If it's not, it's
15347 * not a serious problem -- it just means that the module that we
15348 * just loaded may not be immediately instrumentable.
15349 */
15350 delay(1);
15351 }
15352
15353 static void
15354 dtrace_module_unloaded(struct modctl *ctl)
15355 {
15356 dtrace_probe_t template, *probe, *first, *next;
15357 dtrace_provider_t *prov;
15358
15359 template.dtpr_mod = ctl->mod_modname;
15360
15361 mutex_enter(&dtrace_provider_lock);
15362 mutex_enter(&mod_lock);
15363 mutex_enter(&dtrace_lock);
15364
15365 if (dtrace_bymod == NULL) {
15366 /*
15367 * The DTrace module is loaded (obviously) but not attached;
15368 * we don't have any work to do.
15369 */
15370 mutex_exit(&dtrace_provider_lock);
15371 mutex_exit(&mod_lock);
15372 mutex_exit(&dtrace_lock);
15373 return;
15374 }
15375
15376 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15377 probe != NULL; probe = probe->dtpr_nextmod) {
15378 if (probe->dtpr_ecb != NULL) {
15379 mutex_exit(&dtrace_provider_lock);
15380 mutex_exit(&mod_lock);
15381 mutex_exit(&dtrace_lock);
15382
15383 /*
15384 * This shouldn't _actually_ be possible -- we're
15385 * unloading a module that has an enabled probe in it.
15386 * (It's normally up to the provider to make sure that
15387 * this can't happen.) However, because dtps_enable()
15388 * doesn't have a failure mode, there can be an
15389 * enable/unload race. Upshot: we don't want to
15390 * assert, but we're not going to disable the
15391 * probe, either.
15392 */
15393 if (dtrace_err_verbose) {
15394 cmn_err(CE_WARN, "unloaded module '%s' had "
15395 "enabled probes", ctl->mod_modname);
15396 }
15397
15398 return;
15399 }
15400 }
15401
15402 probe = first;
15403
15404 for (first = NULL; probe != NULL; probe = next) {
15405 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15406
15407 dtrace_probes[probe->dtpr_id - 1] = NULL;
15408
15409 next = probe->dtpr_nextmod;
15410 dtrace_hash_remove(dtrace_bymod, probe);
15411 dtrace_hash_remove(dtrace_byfunc, probe);
15412 dtrace_hash_remove(dtrace_byname, probe);
15413
15414 if (first == NULL) {
15415 first = probe;
15416 probe->dtpr_nextmod = NULL;
15417 } else {
15418 probe->dtpr_nextmod = first;
15419 first = probe;
15420 }
15421 }
15422
15423 /*
15424 * We've removed all of the module's probes from the hash chains and
15425 * from the probe array. Now issue a dtrace_sync() to be sure that
15426 * everyone has cleared out from any probe array processing.
15427 */
15428 dtrace_sync();
15429
15430 for (probe = first; probe != NULL; probe = first) {
15431 first = probe->dtpr_nextmod;
15432 prov = probe->dtpr_provider;
15433 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15434 probe->dtpr_arg);
15435 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15436 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15437 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15438 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15439 kmem_free(probe, sizeof (dtrace_probe_t));
15440 }
15441
15442 mutex_exit(&dtrace_lock);
15443 mutex_exit(&mod_lock);
15444 mutex_exit(&dtrace_provider_lock);
15445 }
15446
15447 void
15448 dtrace_suspend(void)
15449 {
15450 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15451 }
15452
15453 void
15454 dtrace_resume(void)
15455 {
15456 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15457 }
15458
15459 static int
15460 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15461 {
15462 ASSERT(MUTEX_HELD(&cpu_lock));
15463 mutex_enter(&dtrace_lock);
15464
15465 switch (what) {
15466 case CPU_CONFIG: {
15467 dtrace_state_t *state;
15468 dtrace_optval_t *opt, rs, c;
15469
15470 /*
15471 * For now, we only allocate a new buffer for anonymous state.
15472 */
15473 if ((state = dtrace_anon.dta_state) == NULL)
15474 break;
15475
15476 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15477 break;
15478
15479 opt = state->dts_options;
15480 c = opt[DTRACEOPT_CPU];
15481
15482 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15483 break;
15484
15485 /*
15486 * Regardless of what the actual policy is, we're going to
15487 * temporarily set our resize policy to be manual. We're
15488 * also going to temporarily set our CPU option to denote
15489 * the newly configured CPU.
15490 */
15491 rs = opt[DTRACEOPT_BUFRESIZE];
15492 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15493 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15494
15495 (void) dtrace_state_buffers(state);
15496
15497 opt[DTRACEOPT_BUFRESIZE] = rs;
15498 opt[DTRACEOPT_CPU] = c;
15499
15500 break;
15501 }
15502
15503 case CPU_UNCONFIG:
15504 /*
15505 * We don't free the buffer in the CPU_UNCONFIG case. (The
15506 * buffer will be freed when the consumer exits.)
15507 */
15508 break;
15509
15510 default:
15511 break;
15512 }
15513
15514 mutex_exit(&dtrace_lock);
15515 return (0);
15516 }
15517
15518 static void
15519 dtrace_cpu_setup_initial(processorid_t cpu)
15520 {
15521 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15522 }
15523
15524 static void
15525 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15526 {
15527 if (dtrace_toxranges >= dtrace_toxranges_max) {
15528 int osize, nsize;
15529 dtrace_toxrange_t *range;
15530
15531 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15532
15533 if (osize == 0) {
15534 ASSERT(dtrace_toxrange == NULL);
15535 ASSERT(dtrace_toxranges_max == 0);
15536 dtrace_toxranges_max = 1;
15537 } else {
15538 dtrace_toxranges_max <<= 1;
15539 }
15540
15541 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15542 range = kmem_zalloc(nsize, KM_SLEEP);
15543
15544 if (dtrace_toxrange != NULL) {
15545 ASSERT(osize != 0);
15546 bcopy(dtrace_toxrange, range, osize);
15547 kmem_free(dtrace_toxrange, osize);
15548 }
15549
15550 dtrace_toxrange = range;
15551 }
15552
15553 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
15554 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
15555
15556 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15557 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15558 dtrace_toxranges++;
15559 }
15560
15561 static void
15562 dtrace_getf_barrier()
15563 {
15564 /*
15565 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
15566 * that contain calls to getf(), this routine will be called on every
15567 * closef() before either the underlying vnode is released or the
15568 * file_t itself is freed. By the time we are here, it is essential
15569 * that the file_t can no longer be accessed from a call to getf()
15570 * in probe context -- that assures that a dtrace_sync() can be used
15571 * to clear out any enablings referring to the old structures.
15572 */
15573 if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
15574 kcred->cr_zone->zone_dtrace_getf != 0)
15575 dtrace_sync();
15576 }
15577
15578 /*
15579 * DTrace Driver Cookbook Functions
15580 */
15581 /*ARGSUSED*/
15582 static int
15583 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15584 {
15585 dtrace_provider_id_t id;
15586 dtrace_state_t *state = NULL;
15587 dtrace_enabling_t *enab;
15588
15589 mutex_enter(&cpu_lock);
15590 mutex_enter(&dtrace_provider_lock);
15591 mutex_enter(&dtrace_lock);
15592
15593 if (ddi_soft_state_init(&dtrace_softstate,
15594 sizeof (dtrace_state_t), 0) != 0) {
15595 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15596 mutex_exit(&cpu_lock);
15597 mutex_exit(&dtrace_provider_lock);
15598 mutex_exit(&dtrace_lock);
15599 return (DDI_FAILURE);
15600 }
15601
15602 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15603 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15604 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15605 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15606 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15607 ddi_remove_minor_node(devi, NULL);
15608 ddi_soft_state_fini(&dtrace_softstate);
15609 mutex_exit(&cpu_lock);
15610 mutex_exit(&dtrace_provider_lock);
15611 mutex_exit(&dtrace_lock);
15612 return (DDI_FAILURE);
15613 }
15614
15615 ddi_report_dev(devi);
15616 dtrace_devi = devi;
15617
15618 dtrace_modload = dtrace_module_loaded;
15619 dtrace_modunload = dtrace_module_unloaded;
15620 dtrace_cpu_init = dtrace_cpu_setup_initial;
15621 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15622 dtrace_helpers_fork = dtrace_helpers_duplicate;
15623 dtrace_cpustart_init = dtrace_suspend;
15624 dtrace_cpustart_fini = dtrace_resume;
15625 dtrace_debugger_init = dtrace_suspend;
15626 dtrace_debugger_fini = dtrace_resume;
15627
15628 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15629
15630 ASSERT(MUTEX_HELD(&cpu_lock));
15631
15632 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15633 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15634 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15635 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15636 VM_SLEEP | VMC_IDENTIFIER);
15637 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15638 1, INT_MAX, 0);
15639
15640 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15641 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15642 NULL, NULL, NULL, NULL, NULL, 0);
15643
15644 ASSERT(MUTEX_HELD(&cpu_lock));
15645 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15646 offsetof(dtrace_probe_t, dtpr_nextmod),
15647 offsetof(dtrace_probe_t, dtpr_prevmod));
15648
15649 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15650 offsetof(dtrace_probe_t, dtpr_nextfunc),
15651 offsetof(dtrace_probe_t, dtpr_prevfunc));
15652
15653 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15654 offsetof(dtrace_probe_t, dtpr_nextname),
15655 offsetof(dtrace_probe_t, dtpr_prevname));
15656
15657 if (dtrace_retain_max < 1) {
15658 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15659 "setting to 1", dtrace_retain_max);
15660 dtrace_retain_max = 1;
15661 }
15662
15663 /*
15664 * Now discover our toxic ranges.
15665 */
15666 dtrace_toxic_ranges(dtrace_toxrange_add);
15667
15668 /*
15669 * Before we register ourselves as a provider to our own framework,
15670 * we would like to assert that dtrace_provider is NULL -- but that's
15671 * not true if we were loaded as a dependency of a DTrace provider.
15672 * Once we've registered, we can assert that dtrace_provider is our
15673 * pseudo provider.
15674 */
15675 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15676 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15677
15678 ASSERT(dtrace_provider != NULL);
15679 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15680
15681 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15682 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15683 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15684 dtrace_provider, NULL, NULL, "END", 0, NULL);
15685 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15686 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15687
15688 dtrace_anon_property();
15689 mutex_exit(&cpu_lock);
15690
15691 /*
15692 * If there are already providers, we must ask them to provide their
15693 * probes, and then match any anonymous enabling against them. Note
15694 * that there should be no other retained enablings at this time:
15695 * the only retained enablings at this time should be the anonymous
15696 * enabling.
15697 */
15698 if (dtrace_anon.dta_enabling != NULL) {
15699 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15700
15701 dtrace_enabling_provide(NULL);
15702 state = dtrace_anon.dta_state;
15703
15704 /*
15705 * We couldn't hold cpu_lock across the above call to
15706 * dtrace_enabling_provide(), but we must hold it to actually
15707 * enable the probes. We have to drop all of our locks, pick
15708 * up cpu_lock, and regain our locks before matching the
15709 * retained anonymous enabling.
15710 */
15711 mutex_exit(&dtrace_lock);
15712 mutex_exit(&dtrace_provider_lock);
15713
15714 mutex_enter(&cpu_lock);
15715 mutex_enter(&dtrace_provider_lock);
15716 mutex_enter(&dtrace_lock);
15717
15718 if ((enab = dtrace_anon.dta_enabling) != NULL)
15719 (void) dtrace_enabling_match(enab, NULL);
15720
15721 mutex_exit(&cpu_lock);
15722 }
15723
15724 mutex_exit(&dtrace_lock);
15725 mutex_exit(&dtrace_provider_lock);
15726
15727 if (state != NULL) {
15728 /*
15729 * If we created any anonymous state, set it going now.
15730 */
15731 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15732 }
15733
15734 return (DDI_SUCCESS);
15735 }
15736
15737 /*ARGSUSED*/
15738 static int
15739 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15740 {
15741 dtrace_state_t *state;
15742 uint32_t priv;
15743 uid_t uid;
15744 zoneid_t zoneid;
15745
15746 if (getminor(*devp) == DTRACEMNRN_HELPER)
15747 return (0);
15748
15749 /*
15750 * If this wasn't an open with the "helper" minor, then it must be
15751 * the "dtrace" minor.
15752 */
15753 if (getminor(*devp) != DTRACEMNRN_DTRACE)
15754 return (ENXIO);
15755
15756 /*
15757 * If no DTRACE_PRIV_* bits are set in the credential, then the
15758 * caller lacks sufficient permission to do anything with DTrace.
15759 */
15760 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15761 if (priv == DTRACE_PRIV_NONE)
15762 return (EACCES);
15763
15764 /*
15765 * Ask all providers to provide all their probes.
15766 */
15767 mutex_enter(&dtrace_provider_lock);
15768 dtrace_probe_provide(NULL, NULL);
15769 mutex_exit(&dtrace_provider_lock);
15770
15771 mutex_enter(&cpu_lock);
15772 mutex_enter(&dtrace_lock);
15773 dtrace_opens++;
15774 dtrace_membar_producer();
15775
15776 /*
15777 * If the kernel debugger is active (that is, if the kernel debugger
15778 * modified text in some way), we won't allow the open.
15779 */
15780 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15781 dtrace_opens--;
15782 mutex_exit(&cpu_lock);
15783 mutex_exit(&dtrace_lock);
15784 return (EBUSY);
15785 }
15786
15787 if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
15788 /*
15789 * If DTrace helper tracing is enabled, we need to allocate the
15790 * trace buffer and initialize the values.
15791 */
15792 dtrace_helptrace_buffer =
15793 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15794 dtrace_helptrace_next = 0;
15795 dtrace_helptrace_wrapped = 0;
15796 dtrace_helptrace_enable = 0;
15797 }
15798
15799 state = dtrace_state_create(devp, cred_p);
15800 mutex_exit(&cpu_lock);
15801
15802 if (state == NULL) {
15803 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15804 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15805 mutex_exit(&dtrace_lock);
15806 return (EAGAIN);
15807 }
15808
15809 mutex_exit(&dtrace_lock);
15810
15811 return (0);
15812 }
15813
15814 /*ARGSUSED*/
15815 static int
15816 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15817 {
15818 minor_t minor = getminor(dev);
15819 dtrace_state_t *state;
15820 dtrace_helptrace_t *buf = NULL;
15821
15822 if (minor == DTRACEMNRN_HELPER)
15823 return (0);
15824
15825 state = ddi_get_soft_state(dtrace_softstate, minor);
15826
15827 mutex_enter(&cpu_lock);
15828 mutex_enter(&dtrace_lock);
15829
15830 if (state->dts_anon) {
15831 /*
15832 * There is anonymous state. Destroy that first.
15833 */
15834 ASSERT(dtrace_anon.dta_state == NULL);
15835 dtrace_state_destroy(state->dts_anon);
15836 }
15837
15838 if (dtrace_helptrace_disable) {
15839 /*
15840 * If we have been told to disable helper tracing, set the
15841 * buffer to NULL before calling into dtrace_state_destroy();
15842 * we take advantage of its dtrace_sync() to know that no
15843 * CPU is in probe context with enabled helper tracing
15844 * after it returns.
15845 */
15846 buf = dtrace_helptrace_buffer;
15847 dtrace_helptrace_buffer = NULL;
15848 }
15849
15850 dtrace_state_destroy(state);
15851 ASSERT(dtrace_opens > 0);
15852
15853 /*
15854 * Only relinquish control of the kernel debugger interface when there
15855 * are no consumers and no anonymous enablings.
15856 */
15857 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15858 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15859
15860 if (buf != NULL) {
15861 kmem_free(buf, dtrace_helptrace_bufsize);
15862 dtrace_helptrace_disable = 0;
15863 }
15864
15865 mutex_exit(&dtrace_lock);
15866 mutex_exit(&cpu_lock);
15867
15868 return (0);
15869 }
15870
15871 /*ARGSUSED*/
15872 static int
15873 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15874 {
15875 int rval;
15876 dof_helper_t help, *dhp = NULL;
15877
15878 switch (cmd) {
15879 case DTRACEHIOC_ADDDOF:
15880 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15881 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15882 return (EFAULT);
15883 }
15884
15885 dhp = &help;
15886 arg = (intptr_t)help.dofhp_dof;
15887 /*FALLTHROUGH*/
15888
15889 case DTRACEHIOC_ADD: {
15890 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15891
15892 if (dof == NULL)
15893 return (rval);
15894
15895 mutex_enter(&dtrace_lock);
15896
15897 /*
15898 * dtrace_helper_slurp() takes responsibility for the dof --
15899 * it may free it now or it may save it and free it later.
15900 */
15901 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15902 *rv = rval;
15903 rval = 0;
15904 } else {
15905 rval = EINVAL;
15906 }
15907
15908 mutex_exit(&dtrace_lock);
15909 return (rval);
15910 }
15911
15912 case DTRACEHIOC_REMOVE: {
15913 mutex_enter(&dtrace_lock);
15914 rval = dtrace_helper_destroygen(arg);
15915 mutex_exit(&dtrace_lock);
15916
15917 return (rval);
15918 }
15919
15920 default:
15921 break;
15922 }
15923
15924 return (ENOTTY);
15925 }
15926
15927 /*ARGSUSED*/
15928 static int
15929 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15930 {
15931 minor_t minor = getminor(dev);
15932 dtrace_state_t *state;
15933 int rval;
15934
15935 if (minor == DTRACEMNRN_HELPER)
15936 return (dtrace_ioctl_helper(cmd, arg, rv));
15937
15938 state = ddi_get_soft_state(dtrace_softstate, minor);
15939
15940 if (state->dts_anon) {
15941 ASSERT(dtrace_anon.dta_state == NULL);
15942 state = state->dts_anon;
15943 }
15944
15945 switch (cmd) {
15946 case DTRACEIOC_PROVIDER: {
15947 dtrace_providerdesc_t pvd;
15948 dtrace_provider_t *pvp;
15949
15950 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15951 return (EFAULT);
15952
15953 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15954 mutex_enter(&dtrace_provider_lock);
15955
15956 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15957 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15958 break;
15959 }
15960
15961 mutex_exit(&dtrace_provider_lock);
15962
15963 if (pvp == NULL)
15964 return (ESRCH);
15965
15966 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15967 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15968 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15969 return (EFAULT);
15970
15971 return (0);
15972 }
15973
15974 case DTRACEIOC_EPROBE: {
15975 dtrace_eprobedesc_t epdesc;
15976 dtrace_ecb_t *ecb;
15977 dtrace_action_t *act;
15978 void *buf;
15979 size_t size;
15980 uintptr_t dest;
15981 int nrecs;
15982
15983 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15984 return (EFAULT);
15985
15986 mutex_enter(&dtrace_lock);
15987
15988 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15989 mutex_exit(&dtrace_lock);
15990 return (EINVAL);
15991 }
15992
15993 if (ecb->dte_probe == NULL) {
15994 mutex_exit(&dtrace_lock);
15995 return (EINVAL);
15996 }
15997
15998 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15999 epdesc.dtepd_uarg = ecb->dte_uarg;
16000 epdesc.dtepd_size = ecb->dte_size;
16001
16002 nrecs = epdesc.dtepd_nrecs;
16003 epdesc.dtepd_nrecs = 0;
16004 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16005 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16006 continue;
16007
16008 epdesc.dtepd_nrecs++;
16009 }
16010
16011 /*
16012 * Now that we have the size, we need to allocate a temporary
16013 * buffer in which to store the complete description. We need
16014 * the temporary buffer to be able to drop dtrace_lock()
16015 * across the copyout(), below.
16016 */
16017 size = sizeof (dtrace_eprobedesc_t) +
16018 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
16019
16020 buf = kmem_alloc(size, KM_SLEEP);
16021 dest = (uintptr_t)buf;
16022
16023 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
16024 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
16025
16026 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16027 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16028 continue;
16029
16030 if (nrecs-- == 0)
16031 break;
16032
16033 bcopy(&act->dta_rec, (void *)dest,
16034 sizeof (dtrace_recdesc_t));
16035 dest += sizeof (dtrace_recdesc_t);
16036 }
16037
16038 mutex_exit(&dtrace_lock);
16039
16040 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16041 kmem_free(buf, size);
16042 return (EFAULT);
16043 }
16044
16045 kmem_free(buf, size);
16046 return (0);
16047 }
16048
16049 case DTRACEIOC_AGGDESC: {
16050 dtrace_aggdesc_t aggdesc;
16051 dtrace_action_t *act;
16052 dtrace_aggregation_t *agg;
16053 int nrecs;
16054 uint32_t offs;
16055 dtrace_recdesc_t *lrec;
16056 void *buf;
16057 size_t size;
16058 uintptr_t dest;
16059
16060 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16061 return (EFAULT);
16062
16063 mutex_enter(&dtrace_lock);
16064
16065 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16066 mutex_exit(&dtrace_lock);
16067 return (EINVAL);
16068 }
16069
16070 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16071
16072 nrecs = aggdesc.dtagd_nrecs;
16073 aggdesc.dtagd_nrecs = 0;
16074
16075 offs = agg->dtag_base;
16076 lrec = &agg->dtag_action.dta_rec;
16077 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16078
16079 for (act = agg->dtag_first; ; act = act->dta_next) {
16080 ASSERT(act->dta_intuple ||
16081 DTRACEACT_ISAGG(act->dta_kind));
16082
16083 /*
16084 * If this action has a record size of zero, it
16085 * denotes an argument to the aggregating action.
16086 * Because the presence of this record doesn't (or
16087 * shouldn't) affect the way the data is interpreted,
16088 * we don't copy it out to save user-level the
16089 * confusion of dealing with a zero-length record.
16090 */
16091 if (act->dta_rec.dtrd_size == 0) {
16092 ASSERT(agg->dtag_hasarg);
16093 continue;
16094 }
16095
16096 aggdesc.dtagd_nrecs++;
16097
16098 if (act == &agg->dtag_action)
16099 break;
16100 }
16101
16102 /*
16103 * Now that we have the size, we need to allocate a temporary
16104 * buffer in which to store the complete description. We need
16105 * the temporary buffer to be able to drop dtrace_lock()
16106 * across the copyout(), below.
16107 */
16108 size = sizeof (dtrace_aggdesc_t) +
16109 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16110
16111 buf = kmem_alloc(size, KM_SLEEP);
16112 dest = (uintptr_t)buf;
16113
16114 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16115 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16116
16117 for (act = agg->dtag_first; ; act = act->dta_next) {
16118 dtrace_recdesc_t rec = act->dta_rec;
16119
16120 /*
16121 * See the comment in the above loop for why we pass
16122 * over zero-length records.
16123 */
16124 if (rec.dtrd_size == 0) {
16125 ASSERT(agg->dtag_hasarg);
16126 continue;
16127 }
16128
16129 if (nrecs-- == 0)
16130 break;
16131
16132 rec.dtrd_offset -= offs;
16133 bcopy(&rec, (void *)dest, sizeof (rec));
16134 dest += sizeof (dtrace_recdesc_t);
16135
16136 if (act == &agg->dtag_action)
16137 break;
16138 }
16139
16140 mutex_exit(&dtrace_lock);
16141
16142 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16143 kmem_free(buf, size);
16144 return (EFAULT);
16145 }
16146
16147 kmem_free(buf, size);
16148 return (0);
16149 }
16150
16151 case DTRACEIOC_ENABLE: {
16152 dof_hdr_t *dof;
16153 dtrace_enabling_t *enab = NULL;
16154 dtrace_vstate_t *vstate;
16155 int err = 0;
16156
16157 *rv = 0;
16158
16159 /*
16160 * If a NULL argument has been passed, we take this as our
16161 * cue to reevaluate our enablings.
16162 */
16163 if (arg == NULL) {
16164 dtrace_enabling_matchall();
16165
16166 return (0);
16167 }
16168
16169 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16170 return (rval);
16171
16172 mutex_enter(&cpu_lock);
16173 mutex_enter(&dtrace_lock);
16174 vstate = &state->dts_vstate;
16175
16176 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16177 mutex_exit(&dtrace_lock);
16178 mutex_exit(&cpu_lock);
16179 dtrace_dof_destroy(dof);
16180 return (EBUSY);
16181 }
16182
16183 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16184 mutex_exit(&dtrace_lock);
16185 mutex_exit(&cpu_lock);
16186 dtrace_dof_destroy(dof);
16187 return (EINVAL);
16188 }
16189
16190 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16191 dtrace_enabling_destroy(enab);
16192 mutex_exit(&dtrace_lock);
16193 mutex_exit(&cpu_lock);
16194 dtrace_dof_destroy(dof);
16195 return (rval);
16196 }
16197
16198 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16199 err = dtrace_enabling_retain(enab);
16200 } else {
16201 dtrace_enabling_destroy(enab);
16202 }
16203
16204 mutex_exit(&cpu_lock);
16205 mutex_exit(&dtrace_lock);
16206 dtrace_dof_destroy(dof);
16207
16208 return (err);
16209 }
16210
16211 case DTRACEIOC_REPLICATE: {
16212 dtrace_repldesc_t desc;
16213 dtrace_probedesc_t *match = &desc.dtrpd_match;
16214 dtrace_probedesc_t *create = &desc.dtrpd_create;
16215 int err;
16216
16217 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16218 return (EFAULT);
16219
16220 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16221 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16222 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16223 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16224
16225 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16226 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16227 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16228 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16229
16230 mutex_enter(&dtrace_lock);
16231 err = dtrace_enabling_replicate(state, match, create);
16232 mutex_exit(&dtrace_lock);
16233
16234 return (err);
16235 }
16236
16237 case DTRACEIOC_PROBEMATCH:
16238 case DTRACEIOC_PROBES: {
16239 dtrace_probe_t *probe = NULL;
16240 dtrace_probedesc_t desc;
16241 dtrace_probekey_t pkey;
16242 dtrace_id_t i;
16243 int m = 0;
16244 uint32_t priv;
16245 uid_t uid;
16246 zoneid_t zoneid;
16247
16248 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16249 return (EFAULT);
16250
16251 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16252 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16253 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16254 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16255
16256 /*
16257 * Before we attempt to match this probe, we want to give
16258 * all providers the opportunity to provide it.
16259 */
16260 if (desc.dtpd_id == DTRACE_IDNONE) {
16261 mutex_enter(&dtrace_provider_lock);
16262 dtrace_probe_provide(&desc, NULL);
16263 mutex_exit(&dtrace_provider_lock);
16264 desc.dtpd_id++;
16265 }
16266
16267 if (cmd == DTRACEIOC_PROBEMATCH) {
16268 dtrace_probekey(&desc, &pkey);
16269 pkey.dtpk_id = DTRACE_IDNONE;
16270 }
16271
16272 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16273
16274 mutex_enter(&dtrace_lock);
16275
16276 if (cmd == DTRACEIOC_PROBEMATCH) {
16277 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16278 if ((probe = dtrace_probes[i - 1]) != NULL &&
16279 (m = dtrace_match_probe(probe, &pkey,
16280 priv, uid, zoneid)) != 0)
16281 break;
16282 }
16283
16284 if (m < 0) {
16285 mutex_exit(&dtrace_lock);
16286 return (EINVAL);
16287 }
16288
16289 } else {
16290 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16291 if ((probe = dtrace_probes[i - 1]) != NULL &&
16292 dtrace_match_priv(probe, priv, uid, zoneid))
16293 break;
16294 }
16295 }
16296
16297 if (probe == NULL) {
16298 mutex_exit(&dtrace_lock);
16299 return (ESRCH);
16300 }
16301
16302 dtrace_probe_description(probe, &desc);
16303 mutex_exit(&dtrace_lock);
16304
16305 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16306 return (EFAULT);
16307
16308 return (0);
16309 }
16310
16311 case DTRACEIOC_PROBEARG: {
16312 dtrace_argdesc_t desc;
16313 dtrace_probe_t *probe;
16314 dtrace_provider_t *prov;
16315
16316 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16317 return (EFAULT);
16318
16319 if (desc.dtargd_id == DTRACE_IDNONE)
16320 return (EINVAL);
16321
16322 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16323 return (EINVAL);
16324
16325 mutex_enter(&dtrace_provider_lock);
16326 mutex_enter(&mod_lock);
16327 mutex_enter(&dtrace_lock);
16328
16329 if (desc.dtargd_id > dtrace_nprobes) {
16330 mutex_exit(&dtrace_lock);
16331 mutex_exit(&mod_lock);
16332 mutex_exit(&dtrace_provider_lock);
16333 return (EINVAL);
16334 }
16335
16336 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16337 mutex_exit(&dtrace_lock);
16338 mutex_exit(&mod_lock);
16339 mutex_exit(&dtrace_provider_lock);
16340 return (EINVAL);
16341 }
16342
16343 mutex_exit(&dtrace_lock);
16344
16345 prov = probe->dtpr_provider;
16346
16347 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16348 /*
16349 * There isn't any typed information for this probe.
16350 * Set the argument number to DTRACE_ARGNONE.
16351 */
16352 desc.dtargd_ndx = DTRACE_ARGNONE;
16353 } else {
16354 desc.dtargd_native[0] = '\0';
16355 desc.dtargd_xlate[0] = '\0';
16356 desc.dtargd_mapping = desc.dtargd_ndx;
16357
16358 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16359 probe->dtpr_id, probe->dtpr_arg, &desc);
16360 }
16361
16362 mutex_exit(&mod_lock);
16363 mutex_exit(&dtrace_provider_lock);
16364
16365 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16366 return (EFAULT);
16367
16368 return (0);
16369 }
16370
16371 case DTRACEIOC_GO: {
16372 processorid_t cpuid;
16373 rval = dtrace_state_go(state, &cpuid);
16374
16375 if (rval != 0)
16376 return (rval);
16377
16378 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16379 return (EFAULT);
16380
16381 return (0);
16382 }
16383
16384 case DTRACEIOC_STOP: {
16385 processorid_t cpuid;
16386
16387 mutex_enter(&dtrace_lock);
16388 rval = dtrace_state_stop(state, &cpuid);
16389 mutex_exit(&dtrace_lock);
16390
16391 if (rval != 0)
16392 return (rval);
16393
16394 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16395 return (EFAULT);
16396
16397 return (0);
16398 }
16399
16400 case DTRACEIOC_DOFGET: {
16401 dof_hdr_t hdr, *dof;
16402 uint64_t len;
16403
16404 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16405 return (EFAULT);
16406
16407 mutex_enter(&dtrace_lock);
16408 dof = dtrace_dof_create(state);
16409 mutex_exit(&dtrace_lock);
16410
16411 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16412 rval = copyout(dof, (void *)arg, len);
16413 dtrace_dof_destroy(dof);
16414
16415 return (rval == 0 ? 0 : EFAULT);
16416 }
16417
16418 case DTRACEIOC_AGGSNAP:
16419 case DTRACEIOC_BUFSNAP: {
16420 dtrace_bufdesc_t desc;
16421 caddr_t cached;
16422 dtrace_buffer_t *buf;
16423
16424 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16425 return (EFAULT);
16426
16427 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16428 return (EINVAL);
16429
16430 mutex_enter(&dtrace_lock);
16431
16432 if (cmd == DTRACEIOC_BUFSNAP) {
16433 buf = &state->dts_buffer[desc.dtbd_cpu];
16434 } else {
16435 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16436 }
16437
16438 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16439 size_t sz = buf->dtb_offset;
16440
16441 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16442 mutex_exit(&dtrace_lock);
16443 return (EBUSY);
16444 }
16445
16446 /*
16447 * If this buffer has already been consumed, we're
16448 * going to indicate that there's nothing left here
16449 * to consume.
16450 */
16451 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16452 mutex_exit(&dtrace_lock);
16453
16454 desc.dtbd_size = 0;
16455 desc.dtbd_drops = 0;
16456 desc.dtbd_errors = 0;
16457 desc.dtbd_oldest = 0;
16458 sz = sizeof (desc);
16459
16460 if (copyout(&desc, (void *)arg, sz) != 0)
16461 return (EFAULT);
16462
16463 return (0);
16464 }
16465
16466 /*
16467 * If this is a ring buffer that has wrapped, we want
16468 * to copy the whole thing out.
16469 */
16470 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16471 dtrace_buffer_polish(buf);
16472 sz = buf->dtb_size;
16473 }
16474
16475 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16476 mutex_exit(&dtrace_lock);
16477 return (EFAULT);
16478 }
16479
16480 desc.dtbd_size = sz;
16481 desc.dtbd_drops = buf->dtb_drops;
16482 desc.dtbd_errors = buf->dtb_errors;
16483 desc.dtbd_oldest = buf->dtb_xamot_offset;
16484 desc.dtbd_timestamp = dtrace_gethrtime();
16485
16486 mutex_exit(&dtrace_lock);
16487
16488 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16489 return (EFAULT);
16490
16491 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16492
16493 return (0);
16494 }
16495
16496 if (buf->dtb_tomax == NULL) {
16497 ASSERT(buf->dtb_xamot == NULL);
16498 mutex_exit(&dtrace_lock);
16499 return (ENOENT);
16500 }
16501
16502 cached = buf->dtb_tomax;
16503 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16504
16505 dtrace_xcall(desc.dtbd_cpu,
16506 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16507
16508 state->dts_errors += buf->dtb_xamot_errors;
16509
16510 /*
16511 * If the buffers did not actually switch, then the cross call
16512 * did not take place -- presumably because the given CPU is
16513 * not in the ready set. If this is the case, we'll return
16514 * ENOENT.
16515 */
16516 if (buf->dtb_tomax == cached) {
16517 ASSERT(buf->dtb_xamot != cached);
16518 mutex_exit(&dtrace_lock);
16519 return (ENOENT);
16520 }
16521
16522 ASSERT(cached == buf->dtb_xamot);
16523
16524 /*
16525 * We have our snapshot; now copy it out.
16526 */
16527 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16528 buf->dtb_xamot_offset) != 0) {
16529 mutex_exit(&dtrace_lock);
16530 return (EFAULT);
16531 }
16532
16533 desc.dtbd_size = buf->dtb_xamot_offset;
16534 desc.dtbd_drops = buf->dtb_xamot_drops;
16535 desc.dtbd_errors = buf->dtb_xamot_errors;
16536 desc.dtbd_oldest = 0;
16537 desc.dtbd_timestamp = buf->dtb_switched;
16538
16539 mutex_exit(&dtrace_lock);
16540
16541 /*
16542 * Finally, copy out the buffer description.
16543 */
16544 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16545 return (EFAULT);
16546
16547 return (0);
16548 }
16549
16550 case DTRACEIOC_CONF: {
16551 dtrace_conf_t conf;
16552
16553 bzero(&conf, sizeof (conf));
16554 conf.dtc_difversion = DIF_VERSION;
16555 conf.dtc_difintregs = DIF_DIR_NREGS;
16556 conf.dtc_diftupregs = DIF_DTR_NREGS;
16557 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16558
16559 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16560 return (EFAULT);
16561
16562 return (0);
16563 }
16564
16565 case DTRACEIOC_STATUS: {
16566 dtrace_status_t stat;
16567 dtrace_dstate_t *dstate;
16568 int i, j;
16569 uint64_t nerrs;
16570
16571 /*
16572 * See the comment in dtrace_state_deadman() for the reason
16573 * for setting dts_laststatus to INT64_MAX before setting
16574 * it to the correct value.
16575 */
16576 state->dts_laststatus = INT64_MAX;
16577 dtrace_membar_producer();
16578 state->dts_laststatus = dtrace_gethrtime();
16579
16580 bzero(&stat, sizeof (stat));
16581
16582 mutex_enter(&dtrace_lock);
16583
16584 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16585 mutex_exit(&dtrace_lock);
16586 return (ENOENT);
16587 }
16588
16589 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16590 stat.dtst_exiting = 1;
16591
16592 nerrs = state->dts_errors;
16593 dstate = &state->dts_vstate.dtvs_dynvars;
16594
16595 for (i = 0; i < NCPU; i++) {
16596 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16597
16598 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16599 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16600 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16601
16602 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16603 stat.dtst_filled++;
16604
16605 nerrs += state->dts_buffer[i].dtb_errors;
16606
16607 for (j = 0; j < state->dts_nspeculations; j++) {
16608 dtrace_speculation_t *spec;
16609 dtrace_buffer_t *buf;
16610
16611 spec = &state->dts_speculations[j];
16612 buf = &spec->dtsp_buffer[i];
16613 stat.dtst_specdrops += buf->dtb_xamot_drops;
16614 }
16615 }
16616
16617 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16618 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16619 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16620 stat.dtst_dblerrors = state->dts_dblerrors;
16621 stat.dtst_killed =
16622 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16623 stat.dtst_errors = nerrs;
16624
16625 mutex_exit(&dtrace_lock);
16626
16627 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16628 return (EFAULT);
16629
16630 return (0);
16631 }
16632
16633 case DTRACEIOC_FORMAT: {
16634 dtrace_fmtdesc_t fmt;
16635 char *str;
16636 int len;
16637
16638 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16639 return (EFAULT);
16640
16641 mutex_enter(&dtrace_lock);
16642
16643 if (fmt.dtfd_format == 0 ||
16644 fmt.dtfd_format > state->dts_nformats) {
16645 mutex_exit(&dtrace_lock);
16646 return (EINVAL);
16647 }
16648
16649 /*
16650 * Format strings are allocated contiguously and they are
16651 * never freed; if a format index is less than the number
16652 * of formats, we can assert that the format map is non-NULL
16653 * and that the format for the specified index is non-NULL.
16654 */
16655 ASSERT(state->dts_formats != NULL);
16656 str = state->dts_formats[fmt.dtfd_format - 1];
16657 ASSERT(str != NULL);
16658
16659 len = strlen(str) + 1;
16660
16661 if (len > fmt.dtfd_length) {
16662 fmt.dtfd_length = len;
16663
16664 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16665 mutex_exit(&dtrace_lock);
16666 return (EINVAL);
16667 }
16668 } else {
16669 if (copyout(str, fmt.dtfd_string, len) != 0) {
16670 mutex_exit(&dtrace_lock);
16671 return (EINVAL);
16672 }
16673 }
16674
16675 mutex_exit(&dtrace_lock);
16676 return (0);
16677 }
16678
16679 default:
16680 break;
16681 }
16682
16683 return (ENOTTY);
16684 }
16685
16686 /*ARGSUSED*/
16687 static int
16688 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16689 {
16690 dtrace_state_t *state;
16691
16692 switch (cmd) {
16693 case DDI_DETACH:
16694 break;
16695
16696 case DDI_SUSPEND:
16697 return (DDI_SUCCESS);
16698
16699 default:
16700 return (DDI_FAILURE);
16701 }
16702
16703 mutex_enter(&cpu_lock);
16704 mutex_enter(&dtrace_provider_lock);
16705 mutex_enter(&dtrace_lock);
16706
16707 ASSERT(dtrace_opens == 0);
16708
16709 if (dtrace_helpers > 0) {
16710 mutex_exit(&dtrace_provider_lock);
16711 mutex_exit(&dtrace_lock);
16712 mutex_exit(&cpu_lock);
16713 return (DDI_FAILURE);
16714 }
16715
16716 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16717 mutex_exit(&dtrace_provider_lock);
16718 mutex_exit(&dtrace_lock);
16719 mutex_exit(&cpu_lock);
16720 return (DDI_FAILURE);
16721 }
16722
16723 dtrace_provider = NULL;
16724
16725 if ((state = dtrace_anon_grab()) != NULL) {
16726 /*
16727 * If there were ECBs on this state, the provider should
16728 * have not been allowed to detach; assert that there is
16729 * none.
16730 */
16731 ASSERT(state->dts_necbs == 0);
16732 dtrace_state_destroy(state);
16733
16734 /*
16735 * If we're being detached with anonymous state, we need to
16736 * indicate to the kernel debugger that DTrace is now inactive.
16737 */
16738 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16739 }
16740
16741 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16742 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16743 dtrace_cpu_init = NULL;
16744 dtrace_helpers_cleanup = NULL;
16745 dtrace_helpers_fork = NULL;
16746 dtrace_cpustart_init = NULL;
16747 dtrace_cpustart_fini = NULL;
16748 dtrace_debugger_init = NULL;
16749 dtrace_debugger_fini = NULL;
16750 dtrace_modload = NULL;
16751 dtrace_modunload = NULL;
16752
16753 ASSERT(dtrace_getf == 0);
16754 ASSERT(dtrace_closef == NULL);
16755
16756 mutex_exit(&cpu_lock);
16757
16758 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16759 dtrace_probes = NULL;
16760 dtrace_nprobes = 0;
16761
16762 dtrace_hash_destroy(dtrace_bymod);
16763 dtrace_hash_destroy(dtrace_byfunc);
16764 dtrace_hash_destroy(dtrace_byname);
16765 dtrace_bymod = NULL;
16766 dtrace_byfunc = NULL;
16767 dtrace_byname = NULL;
16768
16769 kmem_cache_destroy(dtrace_state_cache);
16770 vmem_destroy(dtrace_minor);
16771 vmem_destroy(dtrace_arena);
16772
16773 if (dtrace_toxrange != NULL) {
16774 kmem_free(dtrace_toxrange,
16775 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16776 dtrace_toxrange = NULL;
16777 dtrace_toxranges = 0;
16778 dtrace_toxranges_max = 0;
16779 }
16780
16781 ddi_remove_minor_node(dtrace_devi, NULL);
16782 dtrace_devi = NULL;
16783
16784 ddi_soft_state_fini(&dtrace_softstate);
16785
16786 ASSERT(dtrace_vtime_references == 0);
16787 ASSERT(dtrace_opens == 0);
16788 ASSERT(dtrace_retained == NULL);
16789
16790 mutex_exit(&dtrace_lock);
16791 mutex_exit(&dtrace_provider_lock);
16792
16793 /*
16794 * We don't destroy the task queue until after we have dropped our
16795 * locks (taskq_destroy() may block on running tasks). To prevent
16796 * attempting to do work after we have effectively detached but before
16797 * the task queue has been destroyed, all tasks dispatched via the
16798 * task queue must check that DTrace is still attached before
16799 * performing any operation.
16800 */
16801 taskq_destroy(dtrace_taskq);
16802 dtrace_taskq = NULL;
16803
16804 return (DDI_SUCCESS);
16805 }
16806
16807 /*ARGSUSED*/
16808 static int
16809 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16810 {
16811 int error;
16812
16813 switch (infocmd) {
16814 case DDI_INFO_DEVT2DEVINFO:
16815 *result = (void *)dtrace_devi;
16816 error = DDI_SUCCESS;
16817 break;
16818 case DDI_INFO_DEVT2INSTANCE:
16819 *result = (void *)0;
16820 error = DDI_SUCCESS;
16821 break;
16822 default:
16823 error = DDI_FAILURE;
16824 }
16825 return (error);
16826 }
16827
16828 static struct cb_ops dtrace_cb_ops = {
16829 dtrace_open, /* open */
16830 dtrace_close, /* close */
16831 nulldev, /* strategy */
16832 nulldev, /* print */
16833 nodev, /* dump */
16834 nodev, /* read */
16835 nodev, /* write */
16836 dtrace_ioctl, /* ioctl */
16837 nodev, /* devmap */
16838 nodev, /* mmap */
16839 nodev, /* segmap */
16840 nochpoll, /* poll */
16841 ddi_prop_op, /* cb_prop_op */
16842 0, /* streamtab */
16843 D_NEW | D_MP /* Driver compatibility flag */
16844 };
16845
16846 static struct dev_ops dtrace_ops = {
16847 DEVO_REV, /* devo_rev */
16848 0, /* refcnt */
16849 dtrace_info, /* get_dev_info */
16850 nulldev, /* identify */
16851 nulldev, /* probe */
16852 dtrace_attach, /* attach */
16853 dtrace_detach, /* detach */
16854 nodev, /* reset */
16855 &dtrace_cb_ops, /* driver operations */
16856 NULL, /* bus operations */
16857 nodev, /* dev power */
16858 ddi_quiesce_not_needed, /* quiesce */
16859 };
16860
16861 static struct modldrv modldrv = {
16862 &mod_driverops, /* module type (this is a pseudo driver) */
16863 "Dynamic Tracing", /* name of module */
16864 &dtrace_ops, /* driver ops */
16865 };
16866
16867 static struct modlinkage modlinkage = {
16868 MODREV_1,
16869 (void *)&modldrv,
16870 NULL
16871 };
16872
16873 int
16874 _init(void)
16875 {
16876 return (mod_install(&modlinkage));
16877 }
16878
16879 int
16880 _info(struct modinfo *modinfop)
16881 {
16882 return (mod_info(&modlinkage, modinfop));
16883 }
16884
16885 int
16886 _fini(void)
16887 {
16888 return (mod_remove(&modlinkage));
16889 }