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 */
26
27 /*
28 * DTrace - Dynamic Tracing for Solaris
29 *
30 * This is the implementation of the Solaris Dynamic Tracing framework
31 * (DTrace). The user-visible interface to DTrace is described at length in
32 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
33 * library, the in-kernel DTrace framework, and the DTrace providers are
34 * described in the block comments in the <sys/dtrace.h> header file. The
35 * internal architecture of DTrace is described in the block comments in the
36 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
37 * implementation very much assume mastery of all of these sources; if one has
38 * an unanswered question about the implementation, one should consult them
39 * first.
40 *
41 * The functions here are ordered roughly as follows:
42 *
43 * - Probe context functions
44 * - Probe hashing functions
45 * - Non-probe context utility functions
46 * - Matching functions
47 * - Provider-to-Framework API functions
48 * - Probe management functions
49 * - DIF object functions
50 * - Format functions
51 * - Predicate functions
52 * - ECB functions
53 * - Buffer functions
54 * - Enabling functions
55 * - DOF functions
56 * - Anonymous enabling functions
57 * - Consumer state functions
58 * - Helper functions
59 * - Hook functions
60 * - Driver cookbook functions
61 *
62 * Each group of functions begins with a block comment labelled the "DTrace
63 * [Group] Functions", allowing one to find each block by searching forward
64 * on capital-f functions.
65 */
66 #include <sys/errno.h>
67 #include <sys/stat.h>
68 #include <sys/modctl.h>
69 #include <sys/conf.h>
70 #include <sys/systm.h>
71 #include <sys/ddi.h>
72 #include <sys/sunddi.h>
73 #include <sys/cpuvar.h>
74 #include <sys/kmem.h>
75 #include <sys/strsubr.h>
76 #include <sys/sysmacros.h>
77 #include <sys/dtrace_impl.h>
78 #include <sys/atomic.h>
79 #include <sys/cmn_err.h>
80 #include <sys/mutex_impl.h>
81 #include <sys/rwlock_impl.h>
82 #include <sys/ctf_api.h>
83 #include <sys/panic.h>
84 #include <sys/priv_impl.h>
85 #include <sys/policy.h>
86 #include <sys/cred_impl.h>
87 #include <sys/procfs_isa.h>
88 #include <sys/taskq.h>
89 #include <sys/mkdev.h>
90 #include <sys/kdi.h>
91 #include <sys/zone.h>
92 #include <sys/socket.h>
93 #include <netinet/in.h>
94
95 /*
96 * DTrace Tunable Variables
97 *
98 * The following variables may be tuned by adding a line to /etc/system that
99 * includes both the name of the DTrace module ("dtrace") and the name of the
100 * variable. For example:
101 *
102 * set dtrace:dtrace_destructive_disallow = 1
103 *
104 * In general, the only variables that one should be tuning this way are those
105 * that affect system-wide DTrace behavior, and for which the default behavior
106 * is undesirable. Most of these variables are tunable on a per-consumer
107 * basis using DTrace options, and need not be tuned on a system-wide basis.
108 * When tuning these variables, avoid pathological values; while some attempt
109 * is made to verify the integrity of these variables, they are not considered
110 * part of the supported interface to DTrace, and they are therefore not
111 * checked comprehensively. Further, these variables should not be tuned
112 * dynamically via "mdb -kw" or other means; they should only be tuned via
113 * /etc/system.
114 */
115 int dtrace_destructive_disallow = 0;
116 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
117 size_t dtrace_difo_maxsize = (256 * 1024);
118 dtrace_optval_t dtrace_dof_maxsize = (256 * 1024);
119 size_t dtrace_global_maxsize = (16 * 1024);
120 size_t dtrace_actions_max = (16 * 1024);
121 size_t dtrace_retain_max = 1024;
122 dtrace_optval_t dtrace_helper_actions_max = 1024;
123 dtrace_optval_t dtrace_helper_providers_max = 32;
124 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
125 size_t dtrace_strsize_default = 256;
126 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
127 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
128 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
129 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
130 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
131 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
132 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
133 dtrace_optval_t dtrace_nspec_default = 1;
134 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
135 dtrace_optval_t dtrace_stackframes_default = 20;
136 dtrace_optval_t dtrace_ustackframes_default = 20;
137 dtrace_optval_t dtrace_jstackframes_default = 50;
138 dtrace_optval_t dtrace_jstackstrsize_default = 512;
139 int dtrace_msgdsize_max = 128;
140 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
141 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
142 int dtrace_devdepth_max = 32;
143 int dtrace_err_verbose;
144 hrtime_t dtrace_deadman_interval = NANOSEC;
145 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
146 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
147 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
148
149 /*
150 * DTrace External Variables
151 *
152 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
153 * available to DTrace consumers via the backtick (`) syntax. One of these,
154 * dtrace_zero, is made deliberately so: it is provided as a source of
155 * well-known, zero-filled memory. While this variable is not documented,
156 * it is used by some translators as an implementation detail.
157 */
158 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
159
160 /*
161 * DTrace Internal Variables
162 */
163 static dev_info_t *dtrace_devi; /* device info */
164 static vmem_t *dtrace_arena; /* probe ID arena */
165 static vmem_t *dtrace_minor; /* minor number arena */
166 static taskq_t *dtrace_taskq; /* task queue */
167 static dtrace_probe_t **dtrace_probes; /* array of all probes */
168 static int dtrace_nprobes; /* number of probes */
169 static dtrace_provider_t *dtrace_provider; /* provider list */
170 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
171 static int dtrace_opens; /* number of opens */
172 static int dtrace_helpers; /* number of helpers */
173 static int dtrace_getf; /* number of unpriv getf()s */
174 static void *dtrace_softstate; /* softstate pointer */
175 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
176 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
177 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
178 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
179 static int dtrace_toxranges; /* number of toxic ranges */
180 static int dtrace_toxranges_max; /* size of toxic range array */
181 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
182 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
183 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
184 static kthread_t *dtrace_panicked; /* panicking thread */
185 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
186 static dtrace_genid_t dtrace_probegen; /* current probe generation */
187 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
188 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
189 static dtrace_genid_t dtrace_retained_gen; /* current retained enab gen */
190 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
191 static int dtrace_dynvar_failclean; /* dynvars failed to clean */
192
193 /*
194 * DTrace Locking
195 * DTrace is protected by three (relatively coarse-grained) locks:
196 *
197 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
198 * including enabling state, probes, ECBs, consumer state, helper state,
199 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
200 * probe context is lock-free -- synchronization is handled via the
201 * dtrace_sync() cross call mechanism.
202 *
203 * (2) dtrace_provider_lock is required when manipulating provider state, or
204 * when provider state must be held constant.
205 *
206 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
207 * when meta provider state must be held constant.
208 *
209 * The lock ordering between these three locks is dtrace_meta_lock before
210 * dtrace_provider_lock before dtrace_lock. (In particular, there are
211 * several places where dtrace_provider_lock is held by the framework as it
212 * calls into the providers -- which then call back into the framework,
213 * grabbing dtrace_lock.)
214 *
215 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
216 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
217 * role as a coarse-grained lock; it is acquired before both of these locks.
218 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
219 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
220 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
221 * acquired _between_ dtrace_provider_lock and dtrace_lock.
222 */
223 static kmutex_t dtrace_lock; /* probe state lock */
224 static kmutex_t dtrace_provider_lock; /* provider state lock */
225 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
226
227 /*
228 * DTrace Provider Variables
229 *
230 * These are the variables relating to DTrace as a provider (that is, the
231 * provider of the BEGIN, END, and ERROR probes).
232 */
233 static dtrace_pattr_t dtrace_provider_attr = {
234 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
235 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
236 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
237 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
238 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
239 };
240
241 static void
242 dtrace_nullop(void)
243 {}
244
245 static int
246 dtrace_enable_nullop(void)
247 {
248 return (0);
249 }
250
251 static dtrace_pops_t dtrace_provider_ops = {
252 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
253 (void (*)(void *, struct modctl *))dtrace_nullop,
254 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop,
255 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
256 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
257 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
258 NULL,
259 NULL,
260 NULL,
261 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
262 };
263
264 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
265 static dtrace_id_t dtrace_probeid_end; /* special END probe */
266 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
267
268 /*
269 * DTrace Helper Tracing Variables
270 */
271 uint32_t dtrace_helptrace_next = 0;
272 uint32_t dtrace_helptrace_nlocals;
273 char *dtrace_helptrace_buffer;
274 int dtrace_helptrace_bufsize = 512 * 1024;
275
276 #ifdef DEBUG
277 int dtrace_helptrace_enabled = 1;
278 #else
279 int dtrace_helptrace_enabled = 0;
280 #endif
281
282 /*
283 * DTrace Error Hashing
284 *
285 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
286 * table. This is very useful for checking coverage of tests that are
287 * expected to induce DIF or DOF processing errors, and may be useful for
288 * debugging problems in the DIF code generator or in DOF generation . The
289 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
290 */
291 #ifdef DEBUG
292 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
293 static const char *dtrace_errlast;
294 static kthread_t *dtrace_errthread;
295 static kmutex_t dtrace_errlock;
296 #endif
297
298 /*
299 * DTrace Macros and Constants
300 *
301 * These are various macros that are useful in various spots in the
302 * implementation, along with a few random constants that have no meaning
303 * outside of the implementation. There is no real structure to this cpp
304 * mishmash -- but is there ever?
305 */
306 #define DTRACE_HASHSTR(hash, probe) \
307 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
308
309 #define DTRACE_HASHNEXT(hash, probe) \
310 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
311
312 #define DTRACE_HASHPREV(hash, probe) \
313 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
314
315 #define DTRACE_HASHEQ(hash, lhs, rhs) \
316 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
317 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
318
319 #define DTRACE_AGGHASHSIZE_SLEW 17
320
321 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
322
323 /*
324 * The key for a thread-local variable consists of the lower 61 bits of the
325 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
326 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
327 * equal to a variable identifier. This is necessary (but not sufficient) to
328 * assure that global associative arrays never collide with thread-local
329 * variables. To guarantee that they cannot collide, we must also define the
330 * order for keying dynamic variables. That order is:
331 *
332 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
333 *
334 * Because the variable-key and the tls-key are in orthogonal spaces, there is
335 * no way for a global variable key signature to match a thread-local key
336 * signature.
337 */
338 #define DTRACE_TLS_THRKEY(where) { \
339 uint_t intr = 0; \
340 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
341 for (; actv; actv >>= 1) \
342 intr++; \
343 ASSERT(intr < (1 << 3)); \
344 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
345 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
346 }
347
348 #define DT_BSWAP_8(x) ((x) & 0xff)
349 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
350 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
351 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
352
353 #define DT_MASK_LO 0x00000000FFFFFFFFULL
354
355 #define DTRACE_STORE(type, tomax, offset, what) \
356 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
357
358 #ifndef __i386
359 #define DTRACE_ALIGNCHECK(addr, size, flags) \
360 if (addr & (size - 1)) { \
361 *flags |= CPU_DTRACE_BADALIGN; \
362 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
363 return (0); \
364 }
365 #else
366 #define DTRACE_ALIGNCHECK(addr, size, flags)
367 #endif
368
369 /*
370 * Test whether a range of memory starting at testaddr of size testsz falls
371 * within the range of memory described by addr, sz. We take care to avoid
372 * problems with overflow and underflow of the unsigned quantities, and
373 * disallow all negative sizes. Ranges of size 0 are allowed.
374 */
375 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
376 ((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
377 (testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
378 (testaddr) + (testsz) >= (testaddr))
379
380 /*
381 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
382 * alloc_sz on the righthand side of the comparison in order to avoid overflow
383 * or underflow in the comparison with it. This is simpler than the INRANGE
384 * check above, because we know that the dtms_scratch_ptr is valid in the
385 * range. Allocations of size zero are allowed.
386 */
387 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
388 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
389 (mstate)->dtms_scratch_ptr >= (alloc_sz))
390
391 #define DTRACE_LOADFUNC(bits) \
392 /*CSTYLED*/ \
393 uint##bits##_t \
394 dtrace_load##bits(uintptr_t addr) \
395 { \
396 size_t size = bits / NBBY; \
397 /*CSTYLED*/ \
398 uint##bits##_t rval; \
399 int i; \
400 volatile uint16_t *flags = (volatile uint16_t *) \
401 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; \
402 \
403 DTRACE_ALIGNCHECK(addr, size, flags); \
404 \
405 for (i = 0; i < dtrace_toxranges; i++) { \
406 if (addr >= dtrace_toxrange[i].dtt_limit) \
407 continue; \
408 \
409 if (addr + size <= dtrace_toxrange[i].dtt_base) \
410 continue; \
411 \
412 /* \
413 * This address falls within a toxic region; return 0. \
414 */ \
415 *flags |= CPU_DTRACE_BADADDR; \
416 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
417 return (0); \
418 } \
419 \
420 *flags |= CPU_DTRACE_NOFAULT; \
421 /*CSTYLED*/ \
422 rval = *((volatile uint##bits##_t *)addr); \
423 *flags &= ~CPU_DTRACE_NOFAULT; \
424 \
425 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
426 }
427
428 #ifdef _LP64
429 #define dtrace_loadptr dtrace_load64
430 #else
431 #define dtrace_loadptr dtrace_load32
432 #endif
433
434 #define DTRACE_DYNHASH_FREE 0
435 #define DTRACE_DYNHASH_SINK 1
436 #define DTRACE_DYNHASH_VALID 2
437
438 #define DTRACE_MATCH_FAIL -1
439 #define DTRACE_MATCH_NEXT 0
440 #define DTRACE_MATCH_DONE 1
441 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
442 #define DTRACE_STATE_ALIGN 64
443
444 #define DTRACE_FLAGS2FLT(flags) \
445 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
446 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
447 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
448 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
449 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
450 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
451 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
452 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
453 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
454 DTRACEFLT_UNKNOWN)
455
456 #define DTRACEACT_ISSTRING(act) \
457 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
458 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
459
460 static size_t dtrace_strlen(const char *, size_t);
461 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
462 static void dtrace_enabling_provide(dtrace_provider_t *);
463 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
464 static void dtrace_enabling_matchall(void);
465 static void dtrace_enabling_reap(void);
466 static dtrace_state_t *dtrace_anon_grab(void);
467 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
468 dtrace_state_t *, uint64_t, uint64_t);
469 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
470 static void dtrace_buffer_drop(dtrace_buffer_t *);
471 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
472 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
473 dtrace_state_t *, dtrace_mstate_t *);
474 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
475 dtrace_optval_t);
476 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
477 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
478 static int dtrace_priv_proc(dtrace_state_t *, dtrace_mstate_t *);
479 static void dtrace_getf_barrier(void);
480
481 /*
482 * DTrace Probe Context Functions
483 *
484 * These functions are called from probe context. Because probe context is
485 * any context in which C may be called, arbitrarily locks may be held,
486 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
487 * As a result, functions called from probe context may only call other DTrace
488 * support functions -- they may not interact at all with the system at large.
489 * (Note that the ASSERT macro is made probe-context safe by redefining it in
490 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
491 * loads are to be performed from probe context, they _must_ be in terms of
492 * the safe dtrace_load*() variants.
493 *
494 * Some functions in this block are not actually called from probe context;
495 * for these functions, there will be a comment above the function reading
496 * "Note: not called from probe context."
497 */
498 void
499 dtrace_panic(const char *format, ...)
500 {
501 va_list alist;
502
503 va_start(alist, format);
504 dtrace_vpanic(format, alist);
505 va_end(alist);
506 }
507
508 int
509 dtrace_assfail(const char *a, const char *f, int l)
510 {
511 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
512
513 /*
514 * We just need something here that even the most clever compiler
515 * cannot optimize away.
516 */
517 return (a[(uintptr_t)f]);
518 }
519
520 /*
521 * Atomically increment a specified error counter from probe context.
522 */
523 static void
524 dtrace_error(uint32_t *counter)
525 {
526 /*
527 * Most counters stored to in probe context are per-CPU counters.
528 * However, there are some error conditions that are sufficiently
529 * arcane that they don't merit per-CPU storage. If these counters
530 * are incremented concurrently on different CPUs, scalability will be
531 * adversely affected -- but we don't expect them to be white-hot in a
532 * correctly constructed enabling...
533 */
534 uint32_t oval, nval;
535
536 do {
537 oval = *counter;
538
539 if ((nval = oval + 1) == 0) {
540 /*
541 * If the counter would wrap, set it to 1 -- assuring
542 * that the counter is never zero when we have seen
543 * errors. (The counter must be 32-bits because we
544 * aren't guaranteed a 64-bit compare&swap operation.)
545 * To save this code both the infamy of being fingered
546 * by a priggish news story and the indignity of being
547 * the target of a neo-puritan witch trial, we're
548 * carefully avoiding any colorful description of the
549 * likelihood of this condition -- but suffice it to
550 * say that it is only slightly more likely than the
551 * overflow of predicate cache IDs, as discussed in
552 * dtrace_predicate_create().
553 */
554 nval = 1;
555 }
556 } while (dtrace_cas32(counter, oval, nval) != oval);
557 }
558
559 /*
560 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
561 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
562 */
563 DTRACE_LOADFUNC(8)
564 DTRACE_LOADFUNC(16)
565 DTRACE_LOADFUNC(32)
566 DTRACE_LOADFUNC(64)
567
568 static int
569 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
570 {
571 if (dest < mstate->dtms_scratch_base)
572 return (0);
573
574 if (dest + size < dest)
575 return (0);
576
577 if (dest + size > mstate->dtms_scratch_ptr)
578 return (0);
579
580 return (1);
581 }
582
583 static int
584 dtrace_canstore_statvar(uint64_t addr, size_t sz,
585 dtrace_statvar_t **svars, int nsvars)
586 {
587 int i;
588
589 for (i = 0; i < nsvars; i++) {
590 dtrace_statvar_t *svar = svars[i];
591
592 if (svar == NULL || svar->dtsv_size == 0)
593 continue;
594
595 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
596 return (1);
597 }
598
599 return (0);
600 }
601
602 /*
603 * Check to see if the address is within a memory region to which a store may
604 * be issued. This includes the DTrace scratch areas, and any DTrace variable
605 * region. The caller of dtrace_canstore() is responsible for performing any
606 * alignment checks that are needed before stores are actually executed.
607 */
608 static int
609 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
610 dtrace_vstate_t *vstate)
611 {
612 /*
613 * First, check to see if the address is in scratch space...
614 */
615 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
616 mstate->dtms_scratch_size))
617 return (1);
618
619 /*
620 * Now check to see if it's a dynamic variable. This check will pick
621 * up both thread-local variables and any global dynamically-allocated
622 * variables.
623 */
624 if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
625 vstate->dtvs_dynvars.dtds_size)) {
626 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
627 uintptr_t base = (uintptr_t)dstate->dtds_base +
628 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
629 uintptr_t chunkoffs;
630
631 /*
632 * Before we assume that we can store here, we need to make
633 * sure that it isn't in our metadata -- storing to our
634 * dynamic variable metadata would corrupt our state. For
635 * the range to not include any dynamic variable metadata,
636 * it must:
637 *
638 * (1) Start above the hash table that is at the base of
639 * the dynamic variable space
640 *
641 * (2) Have a starting chunk offset that is beyond the
642 * dtrace_dynvar_t that is at the base of every chunk
643 *
644 * (3) Not span a chunk boundary
645 *
646 */
647 if (addr < base)
648 return (0);
649
650 chunkoffs = (addr - base) % dstate->dtds_chunksize;
651
652 if (chunkoffs < sizeof (dtrace_dynvar_t))
653 return (0);
654
655 if (chunkoffs + sz > dstate->dtds_chunksize)
656 return (0);
657
658 return (1);
659 }
660
661 /*
662 * Finally, check the static local and global variables. These checks
663 * take the longest, so we perform them last.
664 */
665 if (dtrace_canstore_statvar(addr, sz,
666 vstate->dtvs_locals, vstate->dtvs_nlocals))
667 return (1);
668
669 if (dtrace_canstore_statvar(addr, sz,
670 vstate->dtvs_globals, vstate->dtvs_nglobals))
671 return (1);
672
673 return (0);
674 }
675
676
677 /*
678 * Convenience routine to check to see if the address is within a memory
679 * region in which a load may be issued given the user's privilege level;
680 * if not, it sets the appropriate error flags and loads 'addr' into the
681 * illegal value slot.
682 *
683 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
684 * appropriate memory access protection.
685 */
686 static int
687 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
688 dtrace_vstate_t *vstate)
689 {
690 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
691 file_t *fp;
692
693 /*
694 * If we hold the privilege to read from kernel memory, then
695 * everything is readable.
696 */
697 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
698 return (1);
699
700 /*
701 * You can obviously read that which you can store.
702 */
703 if (dtrace_canstore(addr, sz, mstate, vstate))
704 return (1);
705
706 /*
707 * We're allowed to read from our own string table.
708 */
709 if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
710 mstate->dtms_difo->dtdo_strlen))
711 return (1);
712
713 if (vstate->dtvs_state != NULL &&
714 dtrace_priv_proc(vstate->dtvs_state, mstate)) {
715 proc_t *p;
716
717 /*
718 * When we have privileges to the current process, there are
719 * several context-related kernel structures that are safe to
720 * read, even absent the privilege to read from kernel memory.
721 * These reads are safe because these structures contain only
722 * state that (1) we're permitted to read, (2) is harmless or
723 * (3) contains pointers to additional kernel state that we're
724 * not permitted to read (and as such, do not present an
725 * opportunity for privilege escalation). Finally (and
726 * critically), because of the nature of their relation with
727 * the current thread context, the memory associated with these
728 * structures cannot change over the duration of probe context,
729 * and it is therefore impossible for this memory to be
730 * deallocated and reallocated as something else while it's
731 * being operated upon.
732 */
733 if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t)))
734 return (1);
735
736 if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
737 sz, curthread->t_procp, sizeof (proc_t))) {
738 return (1);
739 }
740
741 if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
742 curthread->t_cred, sizeof (cred_t))) {
743 return (1);
744 }
745
746 if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
747 &(p->p_pidp->pid_id), sizeof (pid_t))) {
748 return (1);
749 }
750
751 if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
752 curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
753 return (1);
754 }
755 }
756
757 if ((fp = mstate->dtms_getf) != NULL) {
758 uintptr_t psz = sizeof (void *);
759 vnode_t *vp;
760 vnodeops_t *op;
761
762 /*
763 * When getf() returns a file_t, the enabling is implicitly
764 * granted the (transient) right to read the returned file_t
765 * as well as the v_path and v_op->vnop_name of the underlying
766 * vnode. These accesses are allowed after a successful
767 * getf() because the members that they refer to cannot change
768 * once set -- and the barrier logic in the kernel's closef()
769 * path assures that the file_t and its referenced vode_t
770 * cannot themselves be stale (that is, it impossible for
771 * either dtms_getf itself or its f_vnode member to reference
772 * freed memory).
773 */
774 if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t)))
775 return (1);
776
777 if ((vp = fp->f_vnode) != NULL) {
778 if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz))
779 return (1);
780
781 if (vp->v_path != NULL && DTRACE_INRANGE(addr, sz,
782 vp->v_path, strlen(vp->v_path) + 1)) {
783 return (1);
784 }
785
786 if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz))
787 return (1);
788
789 if ((op = vp->v_op) != NULL &&
790 DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
791 return (1);
792 }
793
794 if (op != NULL && op->vnop_name != NULL &&
795 DTRACE_INRANGE(addr, sz, op->vnop_name,
796 strlen(op->vnop_name) + 1)) {
797 return (1);
798 }
799 }
800 }
801
802 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
803 *illval = addr;
804 return (0);
805 }
806
807 /*
808 * Convenience routine to check to see if a given string is within a memory
809 * region in which a load may be issued given the user's privilege level;
810 * this exists so that we don't need to issue unnecessary dtrace_strlen()
811 * calls in the event that the user has all privileges.
812 */
813 static int
814 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
815 dtrace_vstate_t *vstate)
816 {
817 size_t strsz;
818
819 /*
820 * If we hold the privilege to read from kernel memory, then
821 * everything is readable.
822 */
823 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
824 return (1);
825
826 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
827 if (dtrace_canload(addr, strsz, mstate, vstate))
828 return (1);
829
830 return (0);
831 }
832
833 /*
834 * Convenience routine to check to see if a given variable is within a memory
835 * region in which a load may be issued given the user's privilege level.
836 */
837 static int
838 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
839 dtrace_vstate_t *vstate)
840 {
841 size_t sz;
842 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
843
844 /*
845 * If we hold the privilege to read from kernel memory, then
846 * everything is readable.
847 */
848 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
849 return (1);
850
851 if (type->dtdt_kind == DIF_TYPE_STRING)
852 sz = dtrace_strlen(src,
853 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
854 else
855 sz = type->dtdt_size;
856
857 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
858 }
859
860 /*
861 * Compare two strings using safe loads.
862 */
863 static int
864 dtrace_strncmp(char *s1, char *s2, size_t limit)
865 {
866 uint8_t c1, c2;
867 volatile uint16_t *flags;
868
869 if (s1 == s2 || limit == 0)
870 return (0);
871
872 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
873
874 do {
875 if (s1 == NULL) {
876 c1 = '\0';
877 } else {
878 c1 = dtrace_load8((uintptr_t)s1++);
879 }
880
881 if (s2 == NULL) {
882 c2 = '\0';
883 } else {
884 c2 = dtrace_load8((uintptr_t)s2++);
885 }
886
887 if (c1 != c2)
888 return (c1 - c2);
889 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
890
891 return (0);
892 }
893
894 /*
895 * Compute strlen(s) for a string using safe memory accesses. The additional
896 * len parameter is used to specify a maximum length to ensure completion.
897 */
898 static size_t
899 dtrace_strlen(const char *s, size_t lim)
900 {
901 uint_t len;
902
903 for (len = 0; len != lim; len++) {
904 if (dtrace_load8((uintptr_t)s++) == '\0')
905 break;
906 }
907
908 return (len);
909 }
910
911 /*
912 * Check if an address falls within a toxic region.
913 */
914 static int
915 dtrace_istoxic(uintptr_t kaddr, size_t size)
916 {
917 uintptr_t taddr, tsize;
918 int i;
919
920 for (i = 0; i < dtrace_toxranges; i++) {
921 taddr = dtrace_toxrange[i].dtt_base;
922 tsize = dtrace_toxrange[i].dtt_limit - taddr;
923
924 if (kaddr - taddr < tsize) {
925 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
926 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
927 return (1);
928 }
929
930 if (taddr - kaddr < size) {
931 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
932 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
933 return (1);
934 }
935 }
936
937 return (0);
938 }
939
940 /*
941 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
942 * memory specified by the DIF program. The dst is assumed to be safe memory
943 * that we can store to directly because it is managed by DTrace. As with
944 * standard bcopy, overlapping copies are handled properly.
945 */
946 static void
947 dtrace_bcopy(const void *src, void *dst, size_t len)
948 {
949 if (len != 0) {
950 uint8_t *s1 = dst;
951 const uint8_t *s2 = src;
952
953 if (s1 <= s2) {
954 do {
955 *s1++ = dtrace_load8((uintptr_t)s2++);
956 } while (--len != 0);
957 } else {
958 s2 += len;
959 s1 += len;
960
961 do {
962 *--s1 = dtrace_load8((uintptr_t)--s2);
963 } while (--len != 0);
964 }
965 }
966 }
967
968 /*
969 * Copy src to dst using safe memory accesses, up to either the specified
970 * length, or the point that a nul byte is encountered. The src is assumed to
971 * be unsafe memory specified by the DIF program. The dst is assumed to be
972 * safe memory that we can store to directly because it is managed by DTrace.
973 * Unlike dtrace_bcopy(), overlapping regions are not handled.
974 */
975 static void
976 dtrace_strcpy(const void *src, void *dst, size_t len)
977 {
978 if (len != 0) {
979 uint8_t *s1 = dst, c;
980 const uint8_t *s2 = src;
981
982 do {
983 *s1++ = c = dtrace_load8((uintptr_t)s2++);
984 } while (--len != 0 && c != '\0');
985 }
986 }
987
988 /*
989 * Copy src to dst, deriving the size and type from the specified (BYREF)
990 * variable type. The src is assumed to be unsafe memory specified by the DIF
991 * program. The dst is assumed to be DTrace variable memory that is of the
992 * specified type; we assume that we can store to directly.
993 */
994 static void
995 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
996 {
997 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
998
999 if (type->dtdt_kind == DIF_TYPE_STRING) {
1000 dtrace_strcpy(src, dst, type->dtdt_size);
1001 } else {
1002 dtrace_bcopy(src, dst, type->dtdt_size);
1003 }
1004 }
1005
1006 /*
1007 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1008 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1009 * safe memory that we can access directly because it is managed by DTrace.
1010 */
1011 static int
1012 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1013 {
1014 volatile uint16_t *flags;
1015
1016 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1017
1018 if (s1 == s2)
1019 return (0);
1020
1021 if (s1 == NULL || s2 == NULL)
1022 return (1);
1023
1024 if (s1 != s2 && len != 0) {
1025 const uint8_t *ps1 = s1;
1026 const uint8_t *ps2 = s2;
1027
1028 do {
1029 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1030 return (1);
1031 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1032 }
1033 return (0);
1034 }
1035
1036 /*
1037 * Zero the specified region using a simple byte-by-byte loop. Note that this
1038 * is for safe DTrace-managed memory only.
1039 */
1040 static void
1041 dtrace_bzero(void *dst, size_t len)
1042 {
1043 uchar_t *cp;
1044
1045 for (cp = dst; len != 0; len--)
1046 *cp++ = 0;
1047 }
1048
1049 static void
1050 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1051 {
1052 uint64_t result[2];
1053
1054 result[0] = addend1[0] + addend2[0];
1055 result[1] = addend1[1] + addend2[1] +
1056 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1057
1058 sum[0] = result[0];
1059 sum[1] = result[1];
1060 }
1061
1062 /*
1063 * Shift the 128-bit value in a by b. If b is positive, shift left.
1064 * If b is negative, shift right.
1065 */
1066 static void
1067 dtrace_shift_128(uint64_t *a, int b)
1068 {
1069 uint64_t mask;
1070
1071 if (b == 0)
1072 return;
1073
1074 if (b < 0) {
1075 b = -b;
1076 if (b >= 64) {
1077 a[0] = a[1] >> (b - 64);
1078 a[1] = 0;
1079 } else {
1080 a[0] >>= b;
1081 mask = 1LL << (64 - b);
1082 mask -= 1;
1083 a[0] |= ((a[1] & mask) << (64 - b));
1084 a[1] >>= b;
1085 }
1086 } else {
1087 if (b >= 64) {
1088 a[1] = a[0] << (b - 64);
1089 a[0] = 0;
1090 } else {
1091 a[1] <<= b;
1092 mask = a[0] >> (64 - b);
1093 a[1] |= mask;
1094 a[0] <<= b;
1095 }
1096 }
1097 }
1098
1099 /*
1100 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1101 * use native multiplication on those, and then re-combine into the
1102 * resulting 128-bit value.
1103 *
1104 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1105 * hi1 * hi2 << 64 +
1106 * hi1 * lo2 << 32 +
1107 * hi2 * lo1 << 32 +
1108 * lo1 * lo2
1109 */
1110 static void
1111 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1112 {
1113 uint64_t hi1, hi2, lo1, lo2;
1114 uint64_t tmp[2];
1115
1116 hi1 = factor1 >> 32;
1117 hi2 = factor2 >> 32;
1118
1119 lo1 = factor1 & DT_MASK_LO;
1120 lo2 = factor2 & DT_MASK_LO;
1121
1122 product[0] = lo1 * lo2;
1123 product[1] = hi1 * hi2;
1124
1125 tmp[0] = hi1 * lo2;
1126 tmp[1] = 0;
1127 dtrace_shift_128(tmp, 32);
1128 dtrace_add_128(product, tmp, product);
1129
1130 tmp[0] = hi2 * lo1;
1131 tmp[1] = 0;
1132 dtrace_shift_128(tmp, 32);
1133 dtrace_add_128(product, tmp, product);
1134 }
1135
1136 /*
1137 * This privilege check should be used by actions and subroutines to
1138 * verify that the user credentials of the process that enabled the
1139 * invoking ECB match the target credentials
1140 */
1141 static int
1142 dtrace_priv_proc_common_user(dtrace_state_t *state)
1143 {
1144 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1145
1146 /*
1147 * We should always have a non-NULL state cred here, since if cred
1148 * is null (anonymous tracing), we fast-path bypass this routine.
1149 */
1150 ASSERT(s_cr != NULL);
1151
1152 if ((cr = CRED()) != NULL &&
1153 s_cr->cr_uid == cr->cr_uid &&
1154 s_cr->cr_uid == cr->cr_ruid &&
1155 s_cr->cr_uid == cr->cr_suid &&
1156 s_cr->cr_gid == cr->cr_gid &&
1157 s_cr->cr_gid == cr->cr_rgid &&
1158 s_cr->cr_gid == cr->cr_sgid)
1159 return (1);
1160
1161 return (0);
1162 }
1163
1164 /*
1165 * This privilege check should be used by actions and subroutines to
1166 * verify that the zone of the process that enabled the invoking ECB
1167 * matches the target credentials
1168 */
1169 static int
1170 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1171 {
1172 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1173
1174 /*
1175 * We should always have a non-NULL state cred here, since if cred
1176 * is null (anonymous tracing), we fast-path bypass this routine.
1177 */
1178 ASSERT(s_cr != NULL);
1179
1180 if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1181 return (1);
1182
1183 return (0);
1184 }
1185
1186 /*
1187 * This privilege check should be used by actions and subroutines to
1188 * verify that the process has not setuid or changed credentials.
1189 */
1190 static int
1191 dtrace_priv_proc_common_nocd()
1192 {
1193 proc_t *proc;
1194
1195 if ((proc = ttoproc(curthread)) != NULL &&
1196 !(proc->p_flag & SNOCD))
1197 return (1);
1198
1199 return (0);
1200 }
1201
1202 static int
1203 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1204 {
1205 int action = state->dts_cred.dcr_action;
1206
1207 if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1208 goto bad;
1209
1210 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1211 dtrace_priv_proc_common_zone(state) == 0)
1212 goto bad;
1213
1214 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1215 dtrace_priv_proc_common_user(state) == 0)
1216 goto bad;
1217
1218 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1219 dtrace_priv_proc_common_nocd() == 0)
1220 goto bad;
1221
1222 return (1);
1223
1224 bad:
1225 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1226
1227 return (0);
1228 }
1229
1230 static int
1231 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1232 {
1233 if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1234 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1235 return (1);
1236
1237 if (dtrace_priv_proc_common_zone(state) &&
1238 dtrace_priv_proc_common_user(state) &&
1239 dtrace_priv_proc_common_nocd())
1240 return (1);
1241 }
1242
1243 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1244
1245 return (0);
1246 }
1247
1248 static int
1249 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1250 {
1251 if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1252 (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1253 return (1);
1254
1255 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1256
1257 return (0);
1258 }
1259
1260 static int
1261 dtrace_priv_kernel(dtrace_state_t *state)
1262 {
1263 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1264 return (1);
1265
1266 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1267
1268 return (0);
1269 }
1270
1271 static int
1272 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1273 {
1274 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1275 return (1);
1276
1277 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1278
1279 return (0);
1280 }
1281
1282 /*
1283 * Determine if the dte_cond of the specified ECB allows for processing of
1284 * the current probe to continue. Note that this routine may allow continued
1285 * processing, but with access(es) stripped from the mstate's dtms_access
1286 * field.
1287 */
1288 static int
1289 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1290 dtrace_ecb_t *ecb)
1291 {
1292 dtrace_probe_t *probe = ecb->dte_probe;
1293 dtrace_provider_t *prov = probe->dtpr_provider;
1294 dtrace_pops_t *pops = &prov->dtpv_pops;
1295 int mode = DTRACE_MODE_NOPRIV_DROP;
1296
1297 ASSERT(ecb->dte_cond);
1298
1299 if (pops->dtps_mode != NULL) {
1300 mode = pops->dtps_mode(prov->dtpv_arg,
1301 probe->dtpr_id, probe->dtpr_arg);
1302
1303 ASSERT((mode & DTRACE_MODE_USER) ||
1304 (mode & DTRACE_MODE_KERNEL));
1305 ASSERT((mode & DTRACE_MODE_NOPRIV_RESTRICT) ||
1306 (mode & DTRACE_MODE_NOPRIV_DROP));
1307 }
1308
1309 /*
1310 * If the dte_cond bits indicate that this consumer is only allowed to
1311 * see user-mode firings of this probe, call the provider's dtps_mode()
1312 * entry point to check that the probe was fired while in a user
1313 * context. If that's not the case, use the policy specified by the
1314 * provider to determine if we drop the probe or merely restrict
1315 * operation.
1316 */
1317 if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1318 ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1319
1320 if (!(mode & DTRACE_MODE_USER)) {
1321 if (mode & DTRACE_MODE_NOPRIV_DROP)
1322 return (0);
1323
1324 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1325 }
1326 }
1327
1328 /*
1329 * This is more subtle than it looks. We have to be absolutely certain
1330 * that CRED() isn't going to change out from under us so it's only
1331 * legit to examine that structure if we're in constrained situations.
1332 * Currently, the only times we'll this check is if a non-super-user
1333 * has enabled the profile or syscall providers -- providers that
1334 * allow visibility of all processes. For the profile case, the check
1335 * above will ensure that we're examining a user context.
1336 */
1337 if (ecb->dte_cond & DTRACE_COND_OWNER) {
1338 cred_t *cr;
1339 cred_t *s_cr = state->dts_cred.dcr_cred;
1340 proc_t *proc;
1341
1342 ASSERT(s_cr != NULL);
1343
1344 if ((cr = CRED()) == NULL ||
1345 s_cr->cr_uid != cr->cr_uid ||
1346 s_cr->cr_uid != cr->cr_ruid ||
1347 s_cr->cr_uid != cr->cr_suid ||
1348 s_cr->cr_gid != cr->cr_gid ||
1349 s_cr->cr_gid != cr->cr_rgid ||
1350 s_cr->cr_gid != cr->cr_sgid ||
1351 (proc = ttoproc(curthread)) == NULL ||
1352 (proc->p_flag & SNOCD)) {
1353 if (mode & DTRACE_MODE_NOPRIV_DROP)
1354 return (0);
1355
1356 mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1357 }
1358 }
1359
1360 /*
1361 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1362 * in our zone, check to see if our mode policy is to restrict rather
1363 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1364 * and DTRACE_ACCESS_ARGS
1365 */
1366 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1367 cred_t *cr;
1368 cred_t *s_cr = state->dts_cred.dcr_cred;
1369
1370 ASSERT(s_cr != NULL);
1371
1372 if ((cr = CRED()) == NULL ||
1373 s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1374 if (mode & DTRACE_MODE_NOPRIV_DROP)
1375 return (0);
1376
1377 mstate->dtms_access &=
1378 ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1379 }
1380 }
1381
1382 return (1);
1383 }
1384
1385 /*
1386 * Note: not called from probe context. This function is called
1387 * asynchronously (and at a regular interval) from outside of probe context to
1388 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1389 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1390 */
1391 void
1392 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1393 {
1394 dtrace_dynvar_t *dirty;
1395 dtrace_dstate_percpu_t *dcpu;
1396 dtrace_dynvar_t **rinsep;
1397 int i, j, work = 0;
1398
1399 for (i = 0; i < NCPU; i++) {
1400 dcpu = &dstate->dtds_percpu[i];
1401 rinsep = &dcpu->dtdsc_rinsing;
1402
1403 /*
1404 * If the dirty list is NULL, there is no dirty work to do.
1405 */
1406 if (dcpu->dtdsc_dirty == NULL)
1407 continue;
1408
1409 if (dcpu->dtdsc_rinsing != NULL) {
1410 /*
1411 * If the rinsing list is non-NULL, then it is because
1412 * this CPU was selected to accept another CPU's
1413 * dirty list -- and since that time, dirty buffers
1414 * have accumulated. This is a highly unlikely
1415 * condition, but we choose to ignore the dirty
1416 * buffers -- they'll be picked up a future cleanse.
1417 */
1418 continue;
1419 }
1420
1421 if (dcpu->dtdsc_clean != NULL) {
1422 /*
1423 * If the clean list is non-NULL, then we're in a
1424 * situation where a CPU has done deallocations (we
1425 * have a non-NULL dirty list) but no allocations (we
1426 * also have a non-NULL clean list). We can't simply
1427 * move the dirty list into the clean list on this
1428 * CPU, yet we also don't want to allow this condition
1429 * to persist, lest a short clean list prevent a
1430 * massive dirty list from being cleaned (which in
1431 * turn could lead to otherwise avoidable dynamic
1432 * drops). To deal with this, we look for some CPU
1433 * with a NULL clean list, NULL dirty list, and NULL
1434 * rinsing list -- and then we borrow this CPU to
1435 * rinse our dirty list.
1436 */
1437 for (j = 0; j < NCPU; j++) {
1438 dtrace_dstate_percpu_t *rinser;
1439
1440 rinser = &dstate->dtds_percpu[j];
1441
1442 if (rinser->dtdsc_rinsing != NULL)
1443 continue;
1444
1445 if (rinser->dtdsc_dirty != NULL)
1446 continue;
1447
1448 if (rinser->dtdsc_clean != NULL)
1449 continue;
1450
1451 rinsep = &rinser->dtdsc_rinsing;
1452 break;
1453 }
1454
1455 if (j == NCPU) {
1456 /*
1457 * We were unable to find another CPU that
1458 * could accept this dirty list -- we are
1459 * therefore unable to clean it now.
1460 */
1461 dtrace_dynvar_failclean++;
1462 continue;
1463 }
1464 }
1465
1466 work = 1;
1467
1468 /*
1469 * Atomically move the dirty list aside.
1470 */
1471 do {
1472 dirty = dcpu->dtdsc_dirty;
1473
1474 /*
1475 * Before we zap the dirty list, set the rinsing list.
1476 * (This allows for a potential assertion in
1477 * dtrace_dynvar(): if a free dynamic variable appears
1478 * on a hash chain, either the dirty list or the
1479 * rinsing list for some CPU must be non-NULL.)
1480 */
1481 *rinsep = dirty;
1482 dtrace_membar_producer();
1483 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1484 dirty, NULL) != dirty);
1485 }
1486
1487 if (!work) {
1488 /*
1489 * We have no work to do; we can simply return.
1490 */
1491 return;
1492 }
1493
1494 dtrace_sync();
1495
1496 for (i = 0; i < NCPU; i++) {
1497 dcpu = &dstate->dtds_percpu[i];
1498
1499 if (dcpu->dtdsc_rinsing == NULL)
1500 continue;
1501
1502 /*
1503 * We are now guaranteed that no hash chain contains a pointer
1504 * into this dirty list; we can make it clean.
1505 */
1506 ASSERT(dcpu->dtdsc_clean == NULL);
1507 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1508 dcpu->dtdsc_rinsing = NULL;
1509 }
1510
1511 /*
1512 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1513 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1514 * This prevents a race whereby a CPU incorrectly decides that
1515 * the state should be something other than DTRACE_DSTATE_CLEAN
1516 * after dtrace_dynvar_clean() has completed.
1517 */
1518 dtrace_sync();
1519
1520 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1521 }
1522
1523 /*
1524 * Depending on the value of the op parameter, this function looks-up,
1525 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1526 * allocation is requested, this function will return a pointer to a
1527 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1528 * variable can be allocated. If NULL is returned, the appropriate counter
1529 * will be incremented.
1530 */
1531 dtrace_dynvar_t *
1532 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1533 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1534 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1535 {
1536 uint64_t hashval = DTRACE_DYNHASH_VALID;
1537 dtrace_dynhash_t *hash = dstate->dtds_hash;
1538 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1539 processorid_t me = CPU->cpu_id, cpu = me;
1540 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1541 size_t bucket, ksize;
1542 size_t chunksize = dstate->dtds_chunksize;
1543 uintptr_t kdata, lock, nstate;
1544 uint_t i;
1545
1546 ASSERT(nkeys != 0);
1547
1548 /*
1549 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1550 * algorithm. For the by-value portions, we perform the algorithm in
1551 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1552 * bit, and seems to have only a minute effect on distribution. For
1553 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1554 * over each referenced byte. It's painful to do this, but it's much
1555 * better than pathological hash distribution. The efficacy of the
1556 * hashing algorithm (and a comparison with other algorithms) may be
1557 * found by running the ::dtrace_dynstat MDB dcmd.
1558 */
1559 for (i = 0; i < nkeys; i++) {
1560 if (key[i].dttk_size == 0) {
1561 uint64_t val = key[i].dttk_value;
1562
1563 hashval += (val >> 48) & 0xffff;
1564 hashval += (hashval << 10);
1565 hashval ^= (hashval >> 6);
1566
1567 hashval += (val >> 32) & 0xffff;
1568 hashval += (hashval << 10);
1569 hashval ^= (hashval >> 6);
1570
1571 hashval += (val >> 16) & 0xffff;
1572 hashval += (hashval << 10);
1573 hashval ^= (hashval >> 6);
1574
1575 hashval += val & 0xffff;
1576 hashval += (hashval << 10);
1577 hashval ^= (hashval >> 6);
1578 } else {
1579 /*
1580 * This is incredibly painful, but it beats the hell
1581 * out of the alternative.
1582 */
1583 uint64_t j, size = key[i].dttk_size;
1584 uintptr_t base = (uintptr_t)key[i].dttk_value;
1585
1586 if (!dtrace_canload(base, size, mstate, vstate))
1587 break;
1588
1589 for (j = 0; j < size; j++) {
1590 hashval += dtrace_load8(base + j);
1591 hashval += (hashval << 10);
1592 hashval ^= (hashval >> 6);
1593 }
1594 }
1595 }
1596
1597 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1598 return (NULL);
1599
1600 hashval += (hashval << 3);
1601 hashval ^= (hashval >> 11);
1602 hashval += (hashval << 15);
1603
1604 /*
1605 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1606 * comes out to be one of our two sentinel hash values. If this
1607 * actually happens, we set the hashval to be a value known to be a
1608 * non-sentinel value.
1609 */
1610 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1611 hashval = DTRACE_DYNHASH_VALID;
1612
1613 /*
1614 * Yes, it's painful to do a divide here. If the cycle count becomes
1615 * important here, tricks can be pulled to reduce it. (However, it's
1616 * critical that hash collisions be kept to an absolute minimum;
1617 * they're much more painful than a divide.) It's better to have a
1618 * solution that generates few collisions and still keeps things
1619 * relatively simple.
1620 */
1621 bucket = hashval % dstate->dtds_hashsize;
1622
1623 if (op == DTRACE_DYNVAR_DEALLOC) {
1624 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1625
1626 for (;;) {
1627 while ((lock = *lockp) & 1)
1628 continue;
1629
1630 if (dtrace_casptr((void *)lockp,
1631 (void *)lock, (void *)(lock + 1)) == (void *)lock)
1632 break;
1633 }
1634
1635 dtrace_membar_producer();
1636 }
1637
1638 top:
1639 prev = NULL;
1640 lock = hash[bucket].dtdh_lock;
1641
1642 dtrace_membar_consumer();
1643
1644 start = hash[bucket].dtdh_chain;
1645 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1646 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1647 op != DTRACE_DYNVAR_DEALLOC));
1648
1649 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1650 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1651 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1652
1653 if (dvar->dtdv_hashval != hashval) {
1654 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1655 /*
1656 * We've reached the sink, and therefore the
1657 * end of the hash chain; we can kick out of
1658 * the loop knowing that we have seen a valid
1659 * snapshot of state.
1660 */
1661 ASSERT(dvar->dtdv_next == NULL);
1662 ASSERT(dvar == &dtrace_dynhash_sink);
1663 break;
1664 }
1665
1666 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1667 /*
1668 * We've gone off the rails: somewhere along
1669 * the line, one of the members of this hash
1670 * chain was deleted. Note that we could also
1671 * detect this by simply letting this loop run
1672 * to completion, as we would eventually hit
1673 * the end of the dirty list. However, we
1674 * want to avoid running the length of the
1675 * dirty list unnecessarily (it might be quite
1676 * long), so we catch this as early as
1677 * possible by detecting the hash marker. In
1678 * this case, we simply set dvar to NULL and
1679 * break; the conditional after the loop will
1680 * send us back to top.
1681 */
1682 dvar = NULL;
1683 break;
1684 }
1685
1686 goto next;
1687 }
1688
1689 if (dtuple->dtt_nkeys != nkeys)
1690 goto next;
1691
1692 for (i = 0; i < nkeys; i++, dkey++) {
1693 if (dkey->dttk_size != key[i].dttk_size)
1694 goto next; /* size or type mismatch */
1695
1696 if (dkey->dttk_size != 0) {
1697 if (dtrace_bcmp(
1698 (void *)(uintptr_t)key[i].dttk_value,
1699 (void *)(uintptr_t)dkey->dttk_value,
1700 dkey->dttk_size))
1701 goto next;
1702 } else {
1703 if (dkey->dttk_value != key[i].dttk_value)
1704 goto next;
1705 }
1706 }
1707
1708 if (op != DTRACE_DYNVAR_DEALLOC)
1709 return (dvar);
1710
1711 ASSERT(dvar->dtdv_next == NULL ||
1712 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1713
1714 if (prev != NULL) {
1715 ASSERT(hash[bucket].dtdh_chain != dvar);
1716 ASSERT(start != dvar);
1717 ASSERT(prev->dtdv_next == dvar);
1718 prev->dtdv_next = dvar->dtdv_next;
1719 } else {
1720 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1721 start, dvar->dtdv_next) != start) {
1722 /*
1723 * We have failed to atomically swing the
1724 * hash table head pointer, presumably because
1725 * of a conflicting allocation on another CPU.
1726 * We need to reread the hash chain and try
1727 * again.
1728 */
1729 goto top;
1730 }
1731 }
1732
1733 dtrace_membar_producer();
1734
1735 /*
1736 * Now set the hash value to indicate that it's free.
1737 */
1738 ASSERT(hash[bucket].dtdh_chain != dvar);
1739 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1740
1741 dtrace_membar_producer();
1742
1743 /*
1744 * Set the next pointer to point at the dirty list, and
1745 * atomically swing the dirty pointer to the newly freed dvar.
1746 */
1747 do {
1748 next = dcpu->dtdsc_dirty;
1749 dvar->dtdv_next = next;
1750 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1751
1752 /*
1753 * Finally, unlock this hash bucket.
1754 */
1755 ASSERT(hash[bucket].dtdh_lock == lock);
1756 ASSERT(lock & 1);
1757 hash[bucket].dtdh_lock++;
1758
1759 return (NULL);
1760 next:
1761 prev = dvar;
1762 continue;
1763 }
1764
1765 if (dvar == NULL) {
1766 /*
1767 * If dvar is NULL, it is because we went off the rails:
1768 * one of the elements that we traversed in the hash chain
1769 * was deleted while we were traversing it. In this case,
1770 * we assert that we aren't doing a dealloc (deallocs lock
1771 * the hash bucket to prevent themselves from racing with
1772 * one another), and retry the hash chain traversal.
1773 */
1774 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1775 goto top;
1776 }
1777
1778 if (op != DTRACE_DYNVAR_ALLOC) {
1779 /*
1780 * If we are not to allocate a new variable, we want to
1781 * return NULL now. Before we return, check that the value
1782 * of the lock word hasn't changed. If it has, we may have
1783 * seen an inconsistent snapshot.
1784 */
1785 if (op == DTRACE_DYNVAR_NOALLOC) {
1786 if (hash[bucket].dtdh_lock != lock)
1787 goto top;
1788 } else {
1789 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1790 ASSERT(hash[bucket].dtdh_lock == lock);
1791 ASSERT(lock & 1);
1792 hash[bucket].dtdh_lock++;
1793 }
1794
1795 return (NULL);
1796 }
1797
1798 /*
1799 * We need to allocate a new dynamic variable. The size we need is the
1800 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1801 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1802 * the size of any referred-to data (dsize). We then round the final
1803 * size up to the chunksize for allocation.
1804 */
1805 for (ksize = 0, i = 0; i < nkeys; i++)
1806 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1807
1808 /*
1809 * This should be pretty much impossible, but could happen if, say,
1810 * strange DIF specified the tuple. Ideally, this should be an
1811 * assertion and not an error condition -- but that requires that the
1812 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1813 * bullet-proof. (That is, it must not be able to be fooled by
1814 * malicious DIF.) Given the lack of backwards branches in DIF,
1815 * solving this would presumably not amount to solving the Halting
1816 * Problem -- but it still seems awfully hard.
1817 */
1818 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1819 ksize + dsize > chunksize) {
1820 dcpu->dtdsc_drops++;
1821 return (NULL);
1822 }
1823
1824 nstate = DTRACE_DSTATE_EMPTY;
1825
1826 do {
1827 retry:
1828 free = dcpu->dtdsc_free;
1829
1830 if (free == NULL) {
1831 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1832 void *rval;
1833
1834 if (clean == NULL) {
1835 /*
1836 * We're out of dynamic variable space on
1837 * this CPU. Unless we have tried all CPUs,
1838 * we'll try to allocate from a different
1839 * CPU.
1840 */
1841 switch (dstate->dtds_state) {
1842 case DTRACE_DSTATE_CLEAN: {
1843 void *sp = &dstate->dtds_state;
1844
1845 if (++cpu >= NCPU)
1846 cpu = 0;
1847
1848 if (dcpu->dtdsc_dirty != NULL &&
1849 nstate == DTRACE_DSTATE_EMPTY)
1850 nstate = DTRACE_DSTATE_DIRTY;
1851
1852 if (dcpu->dtdsc_rinsing != NULL)
1853 nstate = DTRACE_DSTATE_RINSING;
1854
1855 dcpu = &dstate->dtds_percpu[cpu];
1856
1857 if (cpu != me)
1858 goto retry;
1859
1860 (void) dtrace_cas32(sp,
1861 DTRACE_DSTATE_CLEAN, nstate);
1862
1863 /*
1864 * To increment the correct bean
1865 * counter, take another lap.
1866 */
1867 goto retry;
1868 }
1869
1870 case DTRACE_DSTATE_DIRTY:
1871 dcpu->dtdsc_dirty_drops++;
1872 break;
1873
1874 case DTRACE_DSTATE_RINSING:
1875 dcpu->dtdsc_rinsing_drops++;
1876 break;
1877
1878 case DTRACE_DSTATE_EMPTY:
1879 dcpu->dtdsc_drops++;
1880 break;
1881 }
1882
1883 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1884 return (NULL);
1885 }
1886
1887 /*
1888 * The clean list appears to be non-empty. We want to
1889 * move the clean list to the free list; we start by
1890 * moving the clean pointer aside.
1891 */
1892 if (dtrace_casptr(&dcpu->dtdsc_clean,
1893 clean, NULL) != clean) {
1894 /*
1895 * We are in one of two situations:
1896 *
1897 * (a) The clean list was switched to the
1898 * free list by another CPU.
1899 *
1900 * (b) The clean list was added to by the
1901 * cleansing cyclic.
1902 *
1903 * In either of these situations, we can
1904 * just reattempt the free list allocation.
1905 */
1906 goto retry;
1907 }
1908
1909 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1910
1911 /*
1912 * Now we'll move the clean list to our free list.
1913 * It's impossible for this to fail: the only way
1914 * the free list can be updated is through this
1915 * code path, and only one CPU can own the clean list.
1916 * Thus, it would only be possible for this to fail if
1917 * this code were racing with dtrace_dynvar_clean().
1918 * (That is, if dtrace_dynvar_clean() updated the clean
1919 * list, and we ended up racing to update the free
1920 * list.) This race is prevented by the dtrace_sync()
1921 * in dtrace_dynvar_clean() -- which flushes the
1922 * owners of the clean lists out before resetting
1923 * the clean lists.
1924 */
1925 dcpu = &dstate->dtds_percpu[me];
1926 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1927 ASSERT(rval == NULL);
1928 goto retry;
1929 }
1930
1931 dvar = free;
1932 new_free = dvar->dtdv_next;
1933 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1934
1935 /*
1936 * We have now allocated a new chunk. We copy the tuple keys into the
1937 * tuple array and copy any referenced key data into the data space
1938 * following the tuple array. As we do this, we relocate dttk_value
1939 * in the final tuple to point to the key data address in the chunk.
1940 */
1941 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1942 dvar->dtdv_data = (void *)(kdata + ksize);
1943 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1944
1945 for (i = 0; i < nkeys; i++) {
1946 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1947 size_t kesize = key[i].dttk_size;
1948
1949 if (kesize != 0) {
1950 dtrace_bcopy(
1951 (const void *)(uintptr_t)key[i].dttk_value,
1952 (void *)kdata, kesize);
1953 dkey->dttk_value = kdata;
1954 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1955 } else {
1956 dkey->dttk_value = key[i].dttk_value;
1957 }
1958
1959 dkey->dttk_size = kesize;
1960 }
1961
1962 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1963 dvar->dtdv_hashval = hashval;
1964 dvar->dtdv_next = start;
1965
1966 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1967 return (dvar);
1968
1969 /*
1970 * The cas has failed. Either another CPU is adding an element to
1971 * this hash chain, or another CPU is deleting an element from this
1972 * hash chain. The simplest way to deal with both of these cases
1973 * (though not necessarily the most efficient) is to free our
1974 * allocated block and tail-call ourselves. Note that the free is
1975 * to the dirty list and _not_ to the free list. This is to prevent
1976 * races with allocators, above.
1977 */
1978 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1979
1980 dtrace_membar_producer();
1981
1982 do {
1983 free = dcpu->dtdsc_dirty;
1984 dvar->dtdv_next = free;
1985 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1986
1987 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1988 }
1989
1990 /*ARGSUSED*/
1991 static void
1992 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1993 {
1994 if ((int64_t)nval < (int64_t)*oval)
1995 *oval = nval;
1996 }
1997
1998 /*ARGSUSED*/
1999 static void
2000 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2001 {
2002 if ((int64_t)nval > (int64_t)*oval)
2003 *oval = nval;
2004 }
2005
2006 static void
2007 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2008 {
2009 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2010 int64_t val = (int64_t)nval;
2011
2012 if (val < 0) {
2013 for (i = 0; i < zero; i++) {
2014 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2015 quanta[i] += incr;
2016 return;
2017 }
2018 }
2019 } else {
2020 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2021 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2022 quanta[i - 1] += incr;
2023 return;
2024 }
2025 }
2026
2027 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2028 return;
2029 }
2030
2031 ASSERT(0);
2032 }
2033
2034 static void
2035 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2036 {
2037 uint64_t arg = *lquanta++;
2038 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2039 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2040 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2041 int32_t val = (int32_t)nval, level;
2042
2043 ASSERT(step != 0);
2044 ASSERT(levels != 0);
2045
2046 if (val < base) {
2047 /*
2048 * This is an underflow.
2049 */
2050 lquanta[0] += incr;
2051 return;
2052 }
2053
2054 level = (val - base) / step;
2055
2056 if (level < levels) {
2057 lquanta[level + 1] += incr;
2058 return;
2059 }
2060
2061 /*
2062 * This is an overflow.
2063 */
2064 lquanta[levels + 1] += incr;
2065 }
2066
2067 static int
2068 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2069 uint16_t high, uint16_t nsteps, int64_t value)
2070 {
2071 int64_t this = 1, last, next;
2072 int base = 1, order;
2073
2074 ASSERT(factor <= nsteps);
2075 ASSERT(nsteps % factor == 0);
2076
2077 for (order = 0; order < low; order++)
2078 this *= factor;
2079
2080 /*
2081 * If our value is less than our factor taken to the power of the
2082 * low order of magnitude, it goes into the zeroth bucket.
2083 */
2084 if (value < (last = this))
2085 return (0);
2086
2087 for (this *= factor; order <= high; order++) {
2088 int nbuckets = this > nsteps ? nsteps : this;
2089
2090 if ((next = this * factor) < this) {
2091 /*
2092 * We should not generally get log/linear quantizations
2093 * with a high magnitude that allows 64-bits to
2094 * overflow, but we nonetheless protect against this
2095 * by explicitly checking for overflow, and clamping
2096 * our value accordingly.
2097 */
2098 value = this - 1;
2099 }
2100
2101 if (value < this) {
2102 /*
2103 * If our value lies within this order of magnitude,
2104 * determine its position by taking the offset within
2105 * the order of magnitude, dividing by the bucket
2106 * width, and adding to our (accumulated) base.
2107 */
2108 return (base + (value - last) / (this / nbuckets));
2109 }
2110
2111 base += nbuckets - (nbuckets / factor);
2112 last = this;
2113 this = next;
2114 }
2115
2116 /*
2117 * Our value is greater than or equal to our factor taken to the
2118 * power of one plus the high magnitude -- return the top bucket.
2119 */
2120 return (base);
2121 }
2122
2123 static void
2124 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2125 {
2126 uint64_t arg = *llquanta++;
2127 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2128 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2129 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2130 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2131
2132 llquanta[dtrace_aggregate_llquantize_bucket(factor,
2133 low, high, nsteps, nval)] += incr;
2134 }
2135
2136 /*ARGSUSED*/
2137 static void
2138 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2139 {
2140 data[0]++;
2141 data[1] += nval;
2142 }
2143
2144 /*ARGSUSED*/
2145 static void
2146 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2147 {
2148 int64_t snval = (int64_t)nval;
2149 uint64_t tmp[2];
2150
2151 data[0]++;
2152 data[1] += nval;
2153
2154 /*
2155 * What we want to say here is:
2156 *
2157 * data[2] += nval * nval;
2158 *
2159 * But given that nval is 64-bit, we could easily overflow, so
2160 * we do this as 128-bit arithmetic.
2161 */
2162 if (snval < 0)
2163 snval = -snval;
2164
2165 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2166 dtrace_add_128(data + 2, tmp, data + 2);
2167 }
2168
2169 /*ARGSUSED*/
2170 static void
2171 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2172 {
2173 *oval = *oval + 1;
2174 }
2175
2176 /*ARGSUSED*/
2177 static void
2178 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2179 {
2180 *oval += nval;
2181 }
2182
2183 /*
2184 * Aggregate given the tuple in the principal data buffer, and the aggregating
2185 * action denoted by the specified dtrace_aggregation_t. The aggregation
2186 * buffer is specified as the buf parameter. This routine does not return
2187 * failure; if there is no space in the aggregation buffer, the data will be
2188 * dropped, and a corresponding counter incremented.
2189 */
2190 static void
2191 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2192 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2193 {
2194 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2195 uint32_t i, ndx, size, fsize;
2196 uint32_t align = sizeof (uint64_t) - 1;
2197 dtrace_aggbuffer_t *agb;
2198 dtrace_aggkey_t *key;
2199 uint32_t hashval = 0, limit, isstr;
2200 caddr_t tomax, data, kdata;
2201 dtrace_actkind_t action;
2202 dtrace_action_t *act;
2203 uintptr_t offs;
2204
2205 if (buf == NULL)
2206 return;
2207
2208 if (!agg->dtag_hasarg) {
2209 /*
2210 * Currently, only quantize() and lquantize() take additional
2211 * arguments, and they have the same semantics: an increment
2212 * value that defaults to 1 when not present. If additional
2213 * aggregating actions take arguments, the setting of the
2214 * default argument value will presumably have to become more
2215 * sophisticated...
2216 */
2217 arg = 1;
2218 }
2219
2220 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2221 size = rec->dtrd_offset - agg->dtag_base;
2222 fsize = size + rec->dtrd_size;
2223
2224 ASSERT(dbuf->dtb_tomax != NULL);
2225 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2226
2227 if ((tomax = buf->dtb_tomax) == NULL) {
2228 dtrace_buffer_drop(buf);
2229 return;
2230 }
2231
2232 /*
2233 * The metastructure is always at the bottom of the buffer.
2234 */
2235 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2236 sizeof (dtrace_aggbuffer_t));
2237
2238 if (buf->dtb_offset == 0) {
2239 /*
2240 * We just kludge up approximately 1/8th of the size to be
2241 * buckets. If this guess ends up being routinely
2242 * off-the-mark, we may need to dynamically readjust this
2243 * based on past performance.
2244 */
2245 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2246
2247 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2248 (uintptr_t)tomax || hashsize == 0) {
2249 /*
2250 * We've been given a ludicrously small buffer;
2251 * increment our drop count and leave.
2252 */
2253 dtrace_buffer_drop(buf);
2254 return;
2255 }
2256
2257 /*
2258 * And now, a pathetic attempt to try to get a an odd (or
2259 * perchance, a prime) hash size for better hash distribution.
2260 */
2261 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2262 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2263
2264 agb->dtagb_hashsize = hashsize;
2265 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2266 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2267 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2268
2269 for (i = 0; i < agb->dtagb_hashsize; i++)
2270 agb->dtagb_hash[i] = NULL;
2271 }
2272
2273 ASSERT(agg->dtag_first != NULL);
2274 ASSERT(agg->dtag_first->dta_intuple);
2275
2276 /*
2277 * Calculate the hash value based on the key. Note that we _don't_
2278 * include the aggid in the hashing (but we will store it as part of
2279 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2280 * algorithm: a simple, quick algorithm that has no known funnels, and
2281 * gets good distribution in practice. The efficacy of the hashing
2282 * algorithm (and a comparison with other algorithms) may be found by
2283 * running the ::dtrace_aggstat MDB dcmd.
2284 */
2285 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2286 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2287 limit = i + act->dta_rec.dtrd_size;
2288 ASSERT(limit <= size);
2289 isstr = DTRACEACT_ISSTRING(act);
2290
2291 for (; i < limit; i++) {
2292 hashval += data[i];
2293 hashval += (hashval << 10);
2294 hashval ^= (hashval >> 6);
2295
2296 if (isstr && data[i] == '\0')
2297 break;
2298 }
2299 }
2300
2301 hashval += (hashval << 3);
2302 hashval ^= (hashval >> 11);
2303 hashval += (hashval << 15);
2304
2305 /*
2306 * Yes, the divide here is expensive -- but it's generally the least
2307 * of the performance issues given the amount of data that we iterate
2308 * over to compute hash values, compare data, etc.
2309 */
2310 ndx = hashval % agb->dtagb_hashsize;
2311
2312 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2313 ASSERT((caddr_t)key >= tomax);
2314 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2315
2316 if (hashval != key->dtak_hashval || key->dtak_size != size)
2317 continue;
2318
2319 kdata = key->dtak_data;
2320 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2321
2322 for (act = agg->dtag_first; act->dta_intuple;
2323 act = act->dta_next) {
2324 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2325 limit = i + act->dta_rec.dtrd_size;
2326 ASSERT(limit <= size);
2327 isstr = DTRACEACT_ISSTRING(act);
2328
2329 for (; i < limit; i++) {
2330 if (kdata[i] != data[i])
2331 goto next;
2332
2333 if (isstr && data[i] == '\0')
2334 break;
2335 }
2336 }
2337
2338 if (action != key->dtak_action) {
2339 /*
2340 * We are aggregating on the same value in the same
2341 * aggregation with two different aggregating actions.
2342 * (This should have been picked up in the compiler,
2343 * so we may be dealing with errant or devious DIF.)
2344 * This is an error condition; we indicate as much,
2345 * and return.
2346 */
2347 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2348 return;
2349 }
2350
2351 /*
2352 * This is a hit: we need to apply the aggregator to
2353 * the value at this key.
2354 */
2355 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2356 return;
2357 next:
2358 continue;
2359 }
2360
2361 /*
2362 * We didn't find it. We need to allocate some zero-filled space,
2363 * link it into the hash table appropriately, and apply the aggregator
2364 * to the (zero-filled) value.
2365 */
2366 offs = buf->dtb_offset;
2367 while (offs & (align - 1))
2368 offs += sizeof (uint32_t);
2369
2370 /*
2371 * If we don't have enough room to both allocate a new key _and_
2372 * its associated data, increment the drop count and return.
2373 */
2374 if ((uintptr_t)tomax + offs + fsize >
2375 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2376 dtrace_buffer_drop(buf);
2377 return;
2378 }
2379
2380 /*CONSTCOND*/
2381 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2382 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2383 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2384
2385 key->dtak_data = kdata = tomax + offs;
2386 buf->dtb_offset = offs + fsize;
2387
2388 /*
2389 * Now copy the data across.
2390 */
2391 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2392
2393 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2394 kdata[i] = data[i];
2395
2396 /*
2397 * Because strings are not zeroed out by default, we need to iterate
2398 * looking for actions that store strings, and we need to explicitly
2399 * pad these strings out with zeroes.
2400 */
2401 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2402 int nul;
2403
2404 if (!DTRACEACT_ISSTRING(act))
2405 continue;
2406
2407 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2408 limit = i + act->dta_rec.dtrd_size;
2409 ASSERT(limit <= size);
2410
2411 for (nul = 0; i < limit; i++) {
2412 if (nul) {
2413 kdata[i] = '\0';
2414 continue;
2415 }
2416
2417 if (data[i] != '\0')
2418 continue;
2419
2420 nul = 1;
2421 }
2422 }
2423
2424 for (i = size; i < fsize; i++)
2425 kdata[i] = 0;
2426
2427 key->dtak_hashval = hashval;
2428 key->dtak_size = size;
2429 key->dtak_action = action;
2430 key->dtak_next = agb->dtagb_hash[ndx];
2431 agb->dtagb_hash[ndx] = key;
2432
2433 /*
2434 * Finally, apply the aggregator.
2435 */
2436 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2437 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2438 }
2439
2440 /*
2441 * Given consumer state, this routine finds a speculation in the INACTIVE
2442 * state and transitions it into the ACTIVE state. If there is no speculation
2443 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2444 * incremented -- it is up to the caller to take appropriate action.
2445 */
2446 static int
2447 dtrace_speculation(dtrace_state_t *state)
2448 {
2449 int i = 0;
2450 dtrace_speculation_state_t current;
2451 uint32_t *stat = &state->dts_speculations_unavail, count;
2452
2453 while (i < state->dts_nspeculations) {
2454 dtrace_speculation_t *spec = &state->dts_speculations[i];
2455
2456 current = spec->dtsp_state;
2457
2458 if (current != DTRACESPEC_INACTIVE) {
2459 if (current == DTRACESPEC_COMMITTINGMANY ||
2460 current == DTRACESPEC_COMMITTING ||
2461 current == DTRACESPEC_DISCARDING)
2462 stat = &state->dts_speculations_busy;
2463 i++;
2464 continue;
2465 }
2466
2467 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2468 current, DTRACESPEC_ACTIVE) == current)
2469 return (i + 1);
2470 }
2471
2472 /*
2473 * We couldn't find a speculation. If we found as much as a single
2474 * busy speculation buffer, we'll attribute this failure as "busy"
2475 * instead of "unavail".
2476 */
2477 do {
2478 count = *stat;
2479 } while (dtrace_cas32(stat, count, count + 1) != count);
2480
2481 return (0);
2482 }
2483
2484 /*
2485 * This routine commits an active speculation. If the specified speculation
2486 * is not in a valid state to perform a commit(), this routine will silently do
2487 * nothing. The state of the specified speculation is transitioned according
2488 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2489 */
2490 static void
2491 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2492 dtrace_specid_t which)
2493 {
2494 dtrace_speculation_t *spec;
2495 dtrace_buffer_t *src, *dest;
2496 uintptr_t daddr, saddr, dlimit;
2497 dtrace_speculation_state_t current, new;
2498 intptr_t offs;
2499
2500 if (which == 0)
2501 return;
2502
2503 if (which > state->dts_nspeculations) {
2504 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2505 return;
2506 }
2507
2508 spec = &state->dts_speculations[which - 1];
2509 src = &spec->dtsp_buffer[cpu];
2510 dest = &state->dts_buffer[cpu];
2511
2512 do {
2513 current = spec->dtsp_state;
2514
2515 if (current == DTRACESPEC_COMMITTINGMANY)
2516 break;
2517
2518 switch (current) {
2519 case DTRACESPEC_INACTIVE:
2520 case DTRACESPEC_DISCARDING:
2521 return;
2522
2523 case DTRACESPEC_COMMITTING:
2524 /*
2525 * This is only possible if we are (a) commit()'ing
2526 * without having done a prior speculate() on this CPU
2527 * and (b) racing with another commit() on a different
2528 * CPU. There's nothing to do -- we just assert that
2529 * our offset is 0.
2530 */
2531 ASSERT(src->dtb_offset == 0);
2532 return;
2533
2534 case DTRACESPEC_ACTIVE:
2535 new = DTRACESPEC_COMMITTING;
2536 break;
2537
2538 case DTRACESPEC_ACTIVEONE:
2539 /*
2540 * This speculation is active on one CPU. If our
2541 * buffer offset is non-zero, we know that the one CPU
2542 * must be us. Otherwise, we are committing on a
2543 * different CPU from the speculate(), and we must
2544 * rely on being asynchronously cleaned.
2545 */
2546 if (src->dtb_offset != 0) {
2547 new = DTRACESPEC_COMMITTING;
2548 break;
2549 }
2550 /*FALLTHROUGH*/
2551
2552 case DTRACESPEC_ACTIVEMANY:
2553 new = DTRACESPEC_COMMITTINGMANY;
2554 break;
2555
2556 default:
2557 ASSERT(0);
2558 }
2559 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2560 current, new) != current);
2561
2562 /*
2563 * We have set the state to indicate that we are committing this
2564 * speculation. Now reserve the necessary space in the destination
2565 * buffer.
2566 */
2567 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2568 sizeof (uint64_t), state, NULL)) < 0) {
2569 dtrace_buffer_drop(dest);
2570 goto out;
2571 }
2572
2573 /*
2574 * We have the space; copy the buffer across. (Note that this is a
2575 * highly subobtimal bcopy(); in the unlikely event that this becomes
2576 * a serious performance issue, a high-performance DTrace-specific
2577 * bcopy() should obviously be invented.)
2578 */
2579 daddr = (uintptr_t)dest->dtb_tomax + offs;
2580 dlimit = daddr + src->dtb_offset;
2581 saddr = (uintptr_t)src->dtb_tomax;
2582
2583 /*
2584 * First, the aligned portion.
2585 */
2586 while (dlimit - daddr >= sizeof (uint64_t)) {
2587 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2588
2589 daddr += sizeof (uint64_t);
2590 saddr += sizeof (uint64_t);
2591 }
2592
2593 /*
2594 * Now any left-over bit...
2595 */
2596 while (dlimit - daddr)
2597 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2598
2599 /*
2600 * Finally, commit the reserved space in the destination buffer.
2601 */
2602 dest->dtb_offset = offs + src->dtb_offset;
2603
2604 out:
2605 /*
2606 * If we're lucky enough to be the only active CPU on this speculation
2607 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2608 */
2609 if (current == DTRACESPEC_ACTIVE ||
2610 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2611 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2612 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2613
2614 ASSERT(rval == DTRACESPEC_COMMITTING);
2615 }
2616
2617 src->dtb_offset = 0;
2618 src->dtb_xamot_drops += src->dtb_drops;
2619 src->dtb_drops = 0;
2620 }
2621
2622 /*
2623 * This routine discards an active speculation. If the specified speculation
2624 * is not in a valid state to perform a discard(), this routine will silently
2625 * do nothing. The state of the specified speculation is transitioned
2626 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2627 */
2628 static void
2629 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2630 dtrace_specid_t which)
2631 {
2632 dtrace_speculation_t *spec;
2633 dtrace_speculation_state_t current, new;
2634 dtrace_buffer_t *buf;
2635
2636 if (which == 0)
2637 return;
2638
2639 if (which > state->dts_nspeculations) {
2640 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2641 return;
2642 }
2643
2644 spec = &state->dts_speculations[which - 1];
2645 buf = &spec->dtsp_buffer[cpu];
2646
2647 do {
2648 current = spec->dtsp_state;
2649
2650 switch (current) {
2651 case DTRACESPEC_INACTIVE:
2652 case DTRACESPEC_COMMITTINGMANY:
2653 case DTRACESPEC_COMMITTING:
2654 case DTRACESPEC_DISCARDING:
2655 return;
2656
2657 case DTRACESPEC_ACTIVE:
2658 case DTRACESPEC_ACTIVEMANY:
2659 new = DTRACESPEC_DISCARDING;
2660 break;
2661
2662 case DTRACESPEC_ACTIVEONE:
2663 if (buf->dtb_offset != 0) {
2664 new = DTRACESPEC_INACTIVE;
2665 } else {
2666 new = DTRACESPEC_DISCARDING;
2667 }
2668 break;
2669
2670 default:
2671 ASSERT(0);
2672 }
2673 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2674 current, new) != current);
2675
2676 buf->dtb_offset = 0;
2677 buf->dtb_drops = 0;
2678 }
2679
2680 /*
2681 * Note: not called from probe context. This function is called
2682 * asynchronously from cross call context to clean any speculations that are
2683 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2684 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2685 * speculation.
2686 */
2687 static void
2688 dtrace_speculation_clean_here(dtrace_state_t *state)
2689 {
2690 dtrace_icookie_t cookie;
2691 processorid_t cpu = CPU->cpu_id;
2692 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2693 dtrace_specid_t i;
2694
2695 cookie = dtrace_interrupt_disable();
2696
2697 if (dest->dtb_tomax == NULL) {
2698 dtrace_interrupt_enable(cookie);
2699 return;
2700 }
2701
2702 for (i = 0; i < state->dts_nspeculations; i++) {
2703 dtrace_speculation_t *spec = &state->dts_speculations[i];
2704 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2705
2706 if (src->dtb_tomax == NULL)
2707 continue;
2708
2709 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2710 src->dtb_offset = 0;
2711 continue;
2712 }
2713
2714 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2715 continue;
2716
2717 if (src->dtb_offset == 0)
2718 continue;
2719
2720 dtrace_speculation_commit(state, cpu, i + 1);
2721 }
2722
2723 dtrace_interrupt_enable(cookie);
2724 }
2725
2726 /*
2727 * Note: not called from probe context. This function is called
2728 * asynchronously (and at a regular interval) to clean any speculations that
2729 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2730 * is work to be done, it cross calls all CPUs to perform that work;
2731 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2732 * INACTIVE state until they have been cleaned by all CPUs.
2733 */
2734 static void
2735 dtrace_speculation_clean(dtrace_state_t *state)
2736 {
2737 int work = 0, rv;
2738 dtrace_specid_t i;
2739
2740 for (i = 0; i < state->dts_nspeculations; i++) {
2741 dtrace_speculation_t *spec = &state->dts_speculations[i];
2742
2743 ASSERT(!spec->dtsp_cleaning);
2744
2745 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2746 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2747 continue;
2748
2749 work++;
2750 spec->dtsp_cleaning = 1;
2751 }
2752
2753 if (!work)
2754 return;
2755
2756 dtrace_xcall(DTRACE_CPUALL,
2757 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2758
2759 /*
2760 * We now know that all CPUs have committed or discarded their
2761 * speculation buffers, as appropriate. We can now set the state
2762 * to inactive.
2763 */
2764 for (i = 0; i < state->dts_nspeculations; i++) {
2765 dtrace_speculation_t *spec = &state->dts_speculations[i];
2766 dtrace_speculation_state_t current, new;
2767
2768 if (!spec->dtsp_cleaning)
2769 continue;
2770
2771 current = spec->dtsp_state;
2772 ASSERT(current == DTRACESPEC_DISCARDING ||
2773 current == DTRACESPEC_COMMITTINGMANY);
2774
2775 new = DTRACESPEC_INACTIVE;
2776
2777 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2778 ASSERT(rv == current);
2779 spec->dtsp_cleaning = 0;
2780 }
2781 }
2782
2783 /*
2784 * Called as part of a speculate() to get the speculative buffer associated
2785 * with a given speculation. Returns NULL if the specified speculation is not
2786 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2787 * the active CPU is not the specified CPU -- the speculation will be
2788 * atomically transitioned into the ACTIVEMANY state.
2789 */
2790 static dtrace_buffer_t *
2791 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2792 dtrace_specid_t which)
2793 {
2794 dtrace_speculation_t *spec;
2795 dtrace_speculation_state_t current, new;
2796 dtrace_buffer_t *buf;
2797
2798 if (which == 0)
2799 return (NULL);
2800
2801 if (which > state->dts_nspeculations) {
2802 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2803 return (NULL);
2804 }
2805
2806 spec = &state->dts_speculations[which - 1];
2807 buf = &spec->dtsp_buffer[cpuid];
2808
2809 do {
2810 current = spec->dtsp_state;
2811
2812 switch (current) {
2813 case DTRACESPEC_INACTIVE:
2814 case DTRACESPEC_COMMITTINGMANY:
2815 case DTRACESPEC_DISCARDING:
2816 return (NULL);
2817
2818 case DTRACESPEC_COMMITTING:
2819 ASSERT(buf->dtb_offset == 0);
2820 return (NULL);
2821
2822 case DTRACESPEC_ACTIVEONE:
2823 /*
2824 * This speculation is currently active on one CPU.
2825 * Check the offset in the buffer; if it's non-zero,
2826 * that CPU must be us (and we leave the state alone).
2827 * If it's zero, assume that we're starting on a new
2828 * CPU -- and change the state to indicate that the
2829 * speculation is active on more than one CPU.
2830 */
2831 if (buf->dtb_offset != 0)
2832 return (buf);
2833
2834 new = DTRACESPEC_ACTIVEMANY;
2835 break;
2836
2837 case DTRACESPEC_ACTIVEMANY:
2838 return (buf);
2839
2840 case DTRACESPEC_ACTIVE:
2841 new = DTRACESPEC_ACTIVEONE;
2842 break;
2843
2844 default:
2845 ASSERT(0);
2846 }
2847 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2848 current, new) != current);
2849
2850 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2851 return (buf);
2852 }
2853
2854 /*
2855 * Return a string. In the event that the user lacks the privilege to access
2856 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2857 * don't fail access checking.
2858 *
2859 * dtrace_dif_variable() uses this routine as a helper for various
2860 * builtin values such as 'execname' and 'probefunc.'
2861 */
2862 uintptr_t
2863 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2864 dtrace_mstate_t *mstate)
2865 {
2866 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2867 uintptr_t ret;
2868 size_t strsz;
2869
2870 /*
2871 * The easy case: this probe is allowed to read all of memory, so
2872 * we can just return this as a vanilla pointer.
2873 */
2874 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2875 return (addr);
2876
2877 /*
2878 * This is the tougher case: we copy the string in question from
2879 * kernel memory into scratch memory and return it that way: this
2880 * ensures that we won't trip up when access checking tests the
2881 * BYREF return value.
2882 */
2883 strsz = dtrace_strlen((char *)addr, size) + 1;
2884
2885 if (mstate->dtms_scratch_ptr + strsz >
2886 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2887 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2888 return (NULL);
2889 }
2890
2891 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2892 strsz);
2893 ret = mstate->dtms_scratch_ptr;
2894 mstate->dtms_scratch_ptr += strsz;
2895 return (ret);
2896 }
2897
2898 /*
2899 * This function implements the DIF emulator's variable lookups. The emulator
2900 * passes a reserved variable identifier and optional built-in array index.
2901 */
2902 static uint64_t
2903 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2904 uint64_t ndx)
2905 {
2906 /*
2907 * If we're accessing one of the uncached arguments, we'll turn this
2908 * into a reference in the args array.
2909 */
2910 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2911 ndx = v - DIF_VAR_ARG0;
2912 v = DIF_VAR_ARGS;
2913 }
2914
2915 switch (v) {
2916 case DIF_VAR_ARGS:
2917 if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
2918 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
2919 CPU_DTRACE_KPRIV;
2920 return (0);
2921 }
2922
2923 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2924 if (ndx >= sizeof (mstate->dtms_arg) /
2925 sizeof (mstate->dtms_arg[0])) {
2926 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2927 dtrace_provider_t *pv;
2928 uint64_t val;
2929
2930 pv = mstate->dtms_probe->dtpr_provider;
2931 if (pv->dtpv_pops.dtps_getargval != NULL)
2932 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2933 mstate->dtms_probe->dtpr_id,
2934 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2935 else
2936 val = dtrace_getarg(ndx, aframes);
2937
2938 /*
2939 * This is regrettably required to keep the compiler
2940 * from tail-optimizing the call to dtrace_getarg().
2941 * The condition always evaluates to true, but the
2942 * compiler has no way of figuring that out a priori.
2943 * (None of this would be necessary if the compiler
2944 * could be relied upon to _always_ tail-optimize
2945 * the call to dtrace_getarg() -- but it can't.)
2946 */
2947 if (mstate->dtms_probe != NULL)
2948 return (val);
2949
2950 ASSERT(0);
2951 }
2952
2953 return (mstate->dtms_arg[ndx]);
2954
2955 case DIF_VAR_UREGS: {
2956 klwp_t *lwp;
2957
2958 if (!dtrace_priv_proc(state, mstate))
2959 return (0);
2960
2961 if ((lwp = curthread->t_lwp) == NULL) {
2962 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2963 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
2964 return (0);
2965 }
2966
2967 return (dtrace_getreg(lwp->lwp_regs, ndx));
2968 }
2969
2970 case DIF_VAR_VMREGS: {
2971 uint64_t rval;
2972
2973 if (!dtrace_priv_kernel(state))
2974 return (0);
2975
2976 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2977
2978 rval = dtrace_getvmreg(ndx,
2979 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
2980
2981 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2982
2983 return (rval);
2984 }
2985
2986 case DIF_VAR_CURTHREAD:
2987 if (!dtrace_priv_proc(state, mstate))
2988 return (0);
2989 return ((uint64_t)(uintptr_t)curthread);
2990
2991 case DIF_VAR_TIMESTAMP:
2992 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2993 mstate->dtms_timestamp = dtrace_gethrtime();
2994 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2995 }
2996 return (mstate->dtms_timestamp);
2997
2998 case DIF_VAR_VTIMESTAMP:
2999 ASSERT(dtrace_vtime_references != 0);
3000 return (curthread->t_dtrace_vtime);
3001
3002 case DIF_VAR_WALLTIMESTAMP:
3003 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3004 mstate->dtms_walltimestamp = dtrace_gethrestime();
3005 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3006 }
3007 return (mstate->dtms_walltimestamp);
3008
3009 case DIF_VAR_IPL:
3010 if (!dtrace_priv_kernel(state))
3011 return (0);
3012 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3013 mstate->dtms_ipl = dtrace_getipl();
3014 mstate->dtms_present |= DTRACE_MSTATE_IPL;
3015 }
3016 return (mstate->dtms_ipl);
3017
3018 case DIF_VAR_EPID:
3019 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3020 return (mstate->dtms_epid);
3021
3022 case DIF_VAR_ID:
3023 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3024 return (mstate->dtms_probe->dtpr_id);
3025
3026 case DIF_VAR_STACKDEPTH:
3027 if (!dtrace_priv_kernel(state))
3028 return (0);
3029 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3030 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3031
3032 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3033 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3034 }
3035 return (mstate->dtms_stackdepth);
3036
3037 case DIF_VAR_USTACKDEPTH:
3038 if (!dtrace_priv_proc(state, mstate))
3039 return (0);
3040 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3041 /*
3042 * See comment in DIF_VAR_PID.
3043 */
3044 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3045 CPU_ON_INTR(CPU)) {
3046 mstate->dtms_ustackdepth = 0;
3047 } else {
3048 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3049 mstate->dtms_ustackdepth =
3050 dtrace_getustackdepth();
3051 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3052 }
3053 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3054 }
3055 return (mstate->dtms_ustackdepth);
3056
3057 case DIF_VAR_CALLER:
3058 if (!dtrace_priv_kernel(state))
3059 return (0);
3060 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3061 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3062
3063 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3064 /*
3065 * If this is an unanchored probe, we are
3066 * required to go through the slow path:
3067 * dtrace_caller() only guarantees correct
3068 * results for anchored probes.
3069 */
3070 pc_t caller[2];
3071
3072 dtrace_getpcstack(caller, 2, aframes,
3073 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3074 mstate->dtms_caller = caller[1];
3075 } else if ((mstate->dtms_caller =
3076 dtrace_caller(aframes)) == -1) {
3077 /*
3078 * We have failed to do this the quick way;
3079 * we must resort to the slower approach of
3080 * calling dtrace_getpcstack().
3081 */
3082 pc_t caller;
3083
3084 dtrace_getpcstack(&caller, 1, aframes, NULL);
3085 mstate->dtms_caller = caller;
3086 }
3087
3088 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3089 }
3090 return (mstate->dtms_caller);
3091
3092 case DIF_VAR_UCALLER:
3093 if (!dtrace_priv_proc(state, mstate))
3094 return (0);
3095
3096 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3097 uint64_t ustack[3];
3098
3099 /*
3100 * dtrace_getupcstack() fills in the first uint64_t
3101 * with the current PID. The second uint64_t will
3102 * be the program counter at user-level. The third
3103 * uint64_t will contain the caller, which is what
3104 * we're after.
3105 */
3106 ustack[2] = NULL;
3107 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3108 dtrace_getupcstack(ustack, 3);
3109 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3110 mstate->dtms_ucaller = ustack[2];
3111 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3112 }
3113
3114 return (mstate->dtms_ucaller);
3115
3116 case DIF_VAR_PROBEPROV:
3117 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3118 return (dtrace_dif_varstr(
3119 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3120 state, mstate));
3121
3122 case DIF_VAR_PROBEMOD:
3123 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3124 return (dtrace_dif_varstr(
3125 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3126 state, mstate));
3127
3128 case DIF_VAR_PROBEFUNC:
3129 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3130 return (dtrace_dif_varstr(
3131 (uintptr_t)mstate->dtms_probe->dtpr_func,
3132 state, mstate));
3133
3134 case DIF_VAR_PROBENAME:
3135 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3136 return (dtrace_dif_varstr(
3137 (uintptr_t)mstate->dtms_probe->dtpr_name,
3138 state, mstate));
3139
3140 case DIF_VAR_PID:
3141 if (!dtrace_priv_proc(state, mstate))
3142 return (0);
3143
3144 /*
3145 * Note that we are assuming that an unanchored probe is
3146 * always due to a high-level interrupt. (And we're assuming
3147 * that there is only a single high level interrupt.)
3148 */
3149 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3150 return (pid0.pid_id);
3151
3152 /*
3153 * It is always safe to dereference one's own t_procp pointer:
3154 * it always points to a valid, allocated proc structure.
3155 * Further, it is always safe to dereference the p_pidp member
3156 * of one's own proc structure. (These are truisms becuase
3157 * threads and processes don't clean up their own state --
3158 * they leave that task to whomever reaps them.)
3159 */
3160 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3161
3162 case DIF_VAR_PPID:
3163 if (!dtrace_priv_proc(state, mstate))
3164 return (0);
3165
3166 /*
3167 * See comment in DIF_VAR_PID.
3168 */
3169 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3170 return (pid0.pid_id);
3171
3172 /*
3173 * It is always safe to dereference one's own t_procp pointer:
3174 * it always points to a valid, allocated proc structure.
3175 * (This is true because threads don't clean up their own
3176 * state -- they leave that task to whomever reaps them.)
3177 */
3178 return ((uint64_t)curthread->t_procp->p_ppid);
3179
3180 case DIF_VAR_TID:
3181 /*
3182 * See comment in DIF_VAR_PID.
3183 */
3184 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3185 return (0);
3186
3187 return ((uint64_t)curthread->t_tid);
3188
3189 case DIF_VAR_EXECNAME:
3190 if (!dtrace_priv_proc(state, mstate))
3191 return (0);
3192
3193 /*
3194 * See comment in DIF_VAR_PID.
3195 */
3196 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3197 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3198
3199 /*
3200 * It is always safe to dereference one's own t_procp pointer:
3201 * it always points to a valid, allocated proc structure.
3202 * (This is true because threads don't clean up their own
3203 * state -- they leave that task to whomever reaps them.)
3204 */
3205 return (dtrace_dif_varstr(
3206 (uintptr_t)curthread->t_procp->p_user.u_comm,
3207 state, mstate));
3208
3209 case DIF_VAR_ZONENAME:
3210 if (!dtrace_priv_proc(state, mstate))
3211 return (0);
3212
3213 /*
3214 * See comment in DIF_VAR_PID.
3215 */
3216 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3217 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3218
3219 /*
3220 * It is always safe to dereference one's own t_procp pointer:
3221 * it always points to a valid, allocated proc structure.
3222 * (This is true because threads don't clean up their own
3223 * state -- they leave that task to whomever reaps them.)
3224 */
3225 return (dtrace_dif_varstr(
3226 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3227 state, mstate));
3228
3229 case DIF_VAR_UID:
3230 if (!dtrace_priv_proc(state, mstate))
3231 return (0);
3232
3233 /*
3234 * See comment in DIF_VAR_PID.
3235 */
3236 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3237 return ((uint64_t)p0.p_cred->cr_uid);
3238
3239 /*
3240 * It is always safe to dereference one's own t_procp pointer:
3241 * it always points to a valid, allocated proc structure.
3242 * (This is true because threads don't clean up their own
3243 * state -- they leave that task to whomever reaps them.)
3244 *
3245 * Additionally, it is safe to dereference one's own process
3246 * credential, since this is never NULL after process birth.
3247 */
3248 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3249
3250 case DIF_VAR_GID:
3251 if (!dtrace_priv_proc(state, mstate))
3252 return (0);
3253
3254 /*
3255 * See comment in DIF_VAR_PID.
3256 */
3257 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3258 return ((uint64_t)p0.p_cred->cr_gid);
3259
3260 /*
3261 * It is always safe to dereference one's own t_procp pointer:
3262 * it always points to a valid, allocated proc structure.
3263 * (This is true because threads don't clean up their own
3264 * state -- they leave that task to whomever reaps them.)
3265 *
3266 * Additionally, it is safe to dereference one's own process
3267 * credential, since this is never NULL after process birth.
3268 */
3269 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3270
3271 case DIF_VAR_ERRNO: {
3272 klwp_t *lwp;
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 (0);
3281
3282 /*
3283 * It is always safe to dereference one's own t_lwp pointer in
3284 * the event that this pointer is non-NULL. (This is true
3285 * because threads and lwps don't clean up their own state --
3286 * they leave that task to whomever reaps them.)
3287 */
3288 if ((lwp = curthread->t_lwp) == NULL)
3289 return (0);
3290
3291 return ((uint64_t)lwp->lwp_errno);
3292 }
3293 default:
3294 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3295 return (0);
3296 }
3297 }
3298
3299 /*
3300 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3301 * Notice that we don't bother validating the proper number of arguments or
3302 * their types in the tuple stack. This isn't needed because all argument
3303 * interpretation is safe because of our load safety -- the worst that can
3304 * happen is that a bogus program can obtain bogus results.
3305 */
3306 static void
3307 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3308 dtrace_key_t *tupregs, int nargs,
3309 dtrace_mstate_t *mstate, dtrace_state_t *state)
3310 {
3311 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
3312 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
3313 dtrace_vstate_t *vstate = &state->dts_vstate;
3314
3315 union {
3316 mutex_impl_t mi;
3317 uint64_t mx;
3318 } m;
3319
3320 union {
3321 krwlock_t ri;
3322 uintptr_t rw;
3323 } r;
3324
3325 switch (subr) {
3326 case DIF_SUBR_RAND:
3327 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3328 break;
3329
3330 case DIF_SUBR_MUTEX_OWNED:
3331 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3332 mstate, vstate)) {
3333 regs[rd] = NULL;
3334 break;
3335 }
3336
3337 m.mx = dtrace_load64(tupregs[0].dttk_value);
3338 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3339 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3340 else
3341 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3342 break;
3343
3344 case DIF_SUBR_MUTEX_OWNER:
3345 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3346 mstate, vstate)) {
3347 regs[rd] = NULL;
3348 break;
3349 }
3350
3351 m.mx = dtrace_load64(tupregs[0].dttk_value);
3352 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3353 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3354 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3355 else
3356 regs[rd] = 0;
3357 break;
3358
3359 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3360 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3361 mstate, vstate)) {
3362 regs[rd] = NULL;
3363 break;
3364 }
3365
3366 m.mx = dtrace_load64(tupregs[0].dttk_value);
3367 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3368 break;
3369
3370 case DIF_SUBR_MUTEX_TYPE_SPIN:
3371 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3372 mstate, vstate)) {
3373 regs[rd] = NULL;
3374 break;
3375 }
3376
3377 m.mx = dtrace_load64(tupregs[0].dttk_value);
3378 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3379 break;
3380
3381 case DIF_SUBR_RW_READ_HELD: {
3382 uintptr_t tmp;
3383
3384 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3385 mstate, vstate)) {
3386 regs[rd] = NULL;
3387 break;
3388 }
3389
3390 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3391 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3392 break;
3393 }
3394
3395 case DIF_SUBR_RW_WRITE_HELD:
3396 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3397 mstate, vstate)) {
3398 regs[rd] = NULL;
3399 break;
3400 }
3401
3402 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3403 regs[rd] = _RW_WRITE_HELD(&r.ri);
3404 break;
3405
3406 case DIF_SUBR_RW_ISWRITER:
3407 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3408 mstate, vstate)) {
3409 regs[rd] = NULL;
3410 break;
3411 }
3412
3413 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3414 regs[rd] = _RW_ISWRITER(&r.ri);
3415 break;
3416
3417 case DIF_SUBR_BCOPY: {
3418 /*
3419 * We need to be sure that the destination is in the scratch
3420 * region -- no other region is allowed.
3421 */
3422 uintptr_t src = tupregs[0].dttk_value;
3423 uintptr_t dest = tupregs[1].dttk_value;
3424 size_t size = tupregs[2].dttk_value;
3425
3426 if (!dtrace_inscratch(dest, size, mstate)) {
3427 *flags |= CPU_DTRACE_BADADDR;
3428 *illval = regs[rd];
3429 break;
3430 }
3431
3432 if (!dtrace_canload(src, size, mstate, vstate)) {
3433 regs[rd] = NULL;
3434 break;
3435 }
3436
3437 dtrace_bcopy((void *)src, (void *)dest, size);
3438 break;
3439 }
3440
3441 case DIF_SUBR_ALLOCA:
3442 case DIF_SUBR_COPYIN: {
3443 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3444 uint64_t size =
3445 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3446 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3447
3448 /*
3449 * This action doesn't require any credential checks since
3450 * probes will not activate in user contexts to which the
3451 * enabling user does not have permissions.
3452 */
3453
3454 /*
3455 * Rounding up the user allocation size could have overflowed
3456 * a large, bogus allocation (like -1ULL) to 0.
3457 */
3458 if (scratch_size < size ||
3459 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3460 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3461 regs[rd] = NULL;
3462 break;
3463 }
3464
3465 if (subr == DIF_SUBR_COPYIN) {
3466 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3467 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3468 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3469 }
3470
3471 mstate->dtms_scratch_ptr += scratch_size;
3472 regs[rd] = dest;
3473 break;
3474 }
3475
3476 case DIF_SUBR_COPYINTO: {
3477 uint64_t size = tupregs[1].dttk_value;
3478 uintptr_t dest = tupregs[2].dttk_value;
3479
3480 /*
3481 * This action doesn't require any credential checks since
3482 * probes will not activate in user contexts to which the
3483 * enabling user does not have permissions.
3484 */
3485 if (!dtrace_inscratch(dest, size, mstate)) {
3486 *flags |= CPU_DTRACE_BADADDR;
3487 *illval = regs[rd];
3488 break;
3489 }
3490
3491 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3492 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3493 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3494 break;
3495 }
3496
3497 case DIF_SUBR_COPYINSTR: {
3498 uintptr_t dest = mstate->dtms_scratch_ptr;
3499 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3500
3501 if (nargs > 1 && tupregs[1].dttk_value < size)
3502 size = tupregs[1].dttk_value + 1;
3503
3504 /*
3505 * This action doesn't require any credential checks since
3506 * probes will not activate in user contexts to which the
3507 * enabling user does not have permissions.
3508 */
3509 if (!DTRACE_INSCRATCH(mstate, size)) {
3510 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3511 regs[rd] = NULL;
3512 break;
3513 }
3514
3515 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3516 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3517 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3518
3519 ((char *)dest)[size - 1] = '\0';
3520 mstate->dtms_scratch_ptr += size;
3521 regs[rd] = dest;
3522 break;
3523 }
3524
3525 case DIF_SUBR_MSGSIZE:
3526 case DIF_SUBR_MSGDSIZE: {
3527 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3528 uintptr_t wptr, rptr;
3529 size_t count = 0;
3530 int cont = 0;
3531
3532 while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3533
3534 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3535 vstate)) {
3536 regs[rd] = NULL;
3537 break;
3538 }
3539
3540 wptr = dtrace_loadptr(baddr +
3541 offsetof(mblk_t, b_wptr));
3542
3543 rptr = dtrace_loadptr(baddr +
3544 offsetof(mblk_t, b_rptr));
3545
3546 if (wptr < rptr) {
3547 *flags |= CPU_DTRACE_BADADDR;
3548 *illval = tupregs[0].dttk_value;
3549 break;
3550 }
3551
3552 daddr = dtrace_loadptr(baddr +
3553 offsetof(mblk_t, b_datap));
3554
3555 baddr = dtrace_loadptr(baddr +
3556 offsetof(mblk_t, b_cont));
3557
3558 /*
3559 * We want to prevent against denial-of-service here,
3560 * so we're only going to search the list for
3561 * dtrace_msgdsize_max mblks.
3562 */
3563 if (cont++ > dtrace_msgdsize_max) {
3564 *flags |= CPU_DTRACE_ILLOP;
3565 break;
3566 }
3567
3568 if (subr == DIF_SUBR_MSGDSIZE) {
3569 if (dtrace_load8(daddr +
3570 offsetof(dblk_t, db_type)) != M_DATA)
3571 continue;
3572 }
3573
3574 count += wptr - rptr;
3575 }
3576
3577 if (!(*flags & CPU_DTRACE_FAULT))
3578 regs[rd] = count;
3579
3580 break;
3581 }
3582
3583 case DIF_SUBR_PROGENYOF: {
3584 pid_t pid = tupregs[0].dttk_value;
3585 proc_t *p;
3586 int rval = 0;
3587
3588 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3589
3590 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3591 if (p->p_pidp->pid_id == pid) {
3592 rval = 1;
3593 break;
3594 }
3595 }
3596
3597 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3598
3599 regs[rd] = rval;
3600 break;
3601 }
3602
3603 case DIF_SUBR_SPECULATION:
3604 regs[rd] = dtrace_speculation(state);
3605 break;
3606
3607 case DIF_SUBR_COPYOUT: {
3608 uintptr_t kaddr = tupregs[0].dttk_value;
3609 uintptr_t uaddr = tupregs[1].dttk_value;
3610 uint64_t size = tupregs[2].dttk_value;
3611
3612 if (!dtrace_destructive_disallow &&
3613 dtrace_priv_proc_control(state, mstate) &&
3614 !dtrace_istoxic(kaddr, size)) {
3615 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3616 dtrace_copyout(kaddr, uaddr, size, flags);
3617 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3618 }
3619 break;
3620 }
3621
3622 case DIF_SUBR_COPYOUTSTR: {
3623 uintptr_t kaddr = tupregs[0].dttk_value;
3624 uintptr_t uaddr = tupregs[1].dttk_value;
3625 uint64_t size = tupregs[2].dttk_value;
3626
3627 if (!dtrace_destructive_disallow &&
3628 dtrace_priv_proc_control(state, mstate) &&
3629 !dtrace_istoxic(kaddr, size)) {
3630 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3631 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3632 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3633 }
3634 break;
3635 }
3636
3637 case DIF_SUBR_STRLEN: {
3638 size_t sz;
3639 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3640 sz = dtrace_strlen((char *)addr,
3641 state->dts_options[DTRACEOPT_STRSIZE]);
3642
3643 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3644 regs[rd] = NULL;
3645 break;
3646 }
3647
3648 regs[rd] = sz;
3649
3650 break;
3651 }
3652
3653 case DIF_SUBR_STRCHR:
3654 case DIF_SUBR_STRRCHR: {
3655 /*
3656 * We're going to iterate over the string looking for the
3657 * specified character. We will iterate until we have reached
3658 * the string length or we have found the character. If this
3659 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3660 * of the specified character instead of the first.
3661 */
3662 uintptr_t saddr = tupregs[0].dttk_value;
3663 uintptr_t addr = tupregs[0].dttk_value;
3664 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3665 char c, target = (char)tupregs[1].dttk_value;
3666
3667 for (regs[rd] = NULL; addr < limit; addr++) {
3668 if ((c = dtrace_load8(addr)) == target) {
3669 regs[rd] = addr;
3670
3671 if (subr == DIF_SUBR_STRCHR)
3672 break;
3673 }
3674
3675 if (c == '\0')
3676 break;
3677 }
3678
3679 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3680 regs[rd] = NULL;
3681 break;
3682 }
3683
3684 break;
3685 }
3686
3687 case DIF_SUBR_STRSTR:
3688 case DIF_SUBR_INDEX:
3689 case DIF_SUBR_RINDEX: {
3690 /*
3691 * We're going to iterate over the string looking for the
3692 * specified string. We will iterate until we have reached
3693 * the string length or we have found the string. (Yes, this
3694 * is done in the most naive way possible -- but considering
3695 * that the string we're searching for is likely to be
3696 * relatively short, the complexity of Rabin-Karp or similar
3697 * hardly seems merited.)
3698 */
3699 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3700 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3701 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3702 size_t len = dtrace_strlen(addr, size);
3703 size_t sublen = dtrace_strlen(substr, size);
3704 char *limit = addr + len, *orig = addr;
3705 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3706 int inc = 1;
3707
3708 regs[rd] = notfound;
3709
3710 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3711 regs[rd] = NULL;
3712 break;
3713 }
3714
3715 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3716 vstate)) {
3717 regs[rd] = NULL;
3718 break;
3719 }
3720
3721 /*
3722 * strstr() and index()/rindex() have similar semantics if
3723 * both strings are the empty string: strstr() returns a
3724 * pointer to the (empty) string, and index() and rindex()
3725 * both return index 0 (regardless of any position argument).
3726 */
3727 if (sublen == 0 && len == 0) {
3728 if (subr == DIF_SUBR_STRSTR)
3729 regs[rd] = (uintptr_t)addr;
3730 else
3731 regs[rd] = 0;
3732 break;
3733 }
3734
3735 if (subr != DIF_SUBR_STRSTR) {
3736 if (subr == DIF_SUBR_RINDEX) {
3737 limit = orig - 1;
3738 addr += len;
3739 inc = -1;
3740 }
3741
3742 /*
3743 * Both index() and rindex() take an optional position
3744 * argument that denotes the starting position.
3745 */
3746 if (nargs == 3) {
3747 int64_t pos = (int64_t)tupregs[2].dttk_value;
3748
3749 /*
3750 * If the position argument to index() is
3751 * negative, Perl implicitly clamps it at
3752 * zero. This semantic is a little surprising
3753 * given the special meaning of negative
3754 * positions to similar Perl functions like
3755 * substr(), but it appears to reflect a
3756 * notion that index() can start from a
3757 * negative index and increment its way up to
3758 * the string. Given this notion, Perl's
3759 * rindex() is at least self-consistent in
3760 * that it implicitly clamps positions greater
3761 * than the string length to be the string
3762 * length. Where Perl completely loses
3763 * coherence, however, is when the specified
3764 * substring is the empty string (""). In
3765 * this case, even if the position is
3766 * negative, rindex() returns 0 -- and even if
3767 * the position is greater than the length,
3768 * index() returns the string length. These
3769 * semantics violate the notion that index()
3770 * should never return a value less than the
3771 * specified position and that rindex() should
3772 * never return a value greater than the
3773 * specified position. (One assumes that
3774 * these semantics are artifacts of Perl's
3775 * implementation and not the results of
3776 * deliberate design -- it beggars belief that
3777 * even Larry Wall could desire such oddness.)
3778 * While in the abstract one would wish for
3779 * consistent position semantics across
3780 * substr(), index() and rindex() -- or at the
3781 * very least self-consistent position
3782 * semantics for index() and rindex() -- we
3783 * instead opt to keep with the extant Perl
3784 * semantics, in all their broken glory. (Do
3785 * we have more desire to maintain Perl's
3786 * semantics than Perl does? Probably.)
3787 */
3788 if (subr == DIF_SUBR_RINDEX) {
3789 if (pos < 0) {
3790 if (sublen == 0)
3791 regs[rd] = 0;
3792 break;
3793 }
3794
3795 if (pos > len)
3796 pos = len;
3797 } else {
3798 if (pos < 0)
3799 pos = 0;
3800
3801 if (pos >= len) {
3802 if (sublen == 0)
3803 regs[rd] = len;
3804 break;
3805 }
3806 }
3807
3808 addr = orig + pos;
3809 }
3810 }
3811
3812 for (regs[rd] = notfound; addr != limit; addr += inc) {
3813 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3814 if (subr != DIF_SUBR_STRSTR) {
3815 /*
3816 * As D index() and rindex() are
3817 * modeled on Perl (and not on awk),
3818 * we return a zero-based (and not a
3819 * one-based) index. (For you Perl
3820 * weenies: no, we're not going to add
3821 * $[ -- and shouldn't you be at a con
3822 * or something?)
3823 */
3824 regs[rd] = (uintptr_t)(addr - orig);
3825 break;
3826 }
3827
3828 ASSERT(subr == DIF_SUBR_STRSTR);
3829 regs[rd] = (uintptr_t)addr;
3830 break;
3831 }
3832 }
3833
3834 break;
3835 }
3836
3837 case DIF_SUBR_STRTOK: {
3838 uintptr_t addr = tupregs[0].dttk_value;
3839 uintptr_t tokaddr = tupregs[1].dttk_value;
3840 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3841 uintptr_t limit, toklimit = tokaddr + size;
3842 uint8_t c, tokmap[32]; /* 256 / 8 */
3843 char *dest = (char *)mstate->dtms_scratch_ptr;
3844 int i;
3845
3846 /*
3847 * Check both the token buffer and (later) the input buffer,
3848 * since both could be non-scratch addresses.
3849 */
3850 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3851 regs[rd] = NULL;
3852 break;
3853 }
3854
3855 if (!DTRACE_INSCRATCH(mstate, size)) {
3856 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3857 regs[rd] = NULL;
3858 break;
3859 }
3860
3861 if (addr == NULL) {
3862 /*
3863 * If the address specified is NULL, we use our saved
3864 * strtok pointer from the mstate. Note that this
3865 * means that the saved strtok pointer is _only_
3866 * valid within multiple enablings of the same probe --
3867 * it behaves like an implicit clause-local variable.
3868 */
3869 addr = mstate->dtms_strtok;
3870 } else {
3871 /*
3872 * If the user-specified address is non-NULL we must
3873 * access check it. This is the only time we have
3874 * a chance to do so, since this address may reside
3875 * in the string table of this clause-- future calls
3876 * (when we fetch addr from mstate->dtms_strtok)
3877 * would fail this access check.
3878 */
3879 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3880 regs[rd] = NULL;
3881 break;
3882 }
3883 }
3884
3885 /*
3886 * First, zero the token map, and then process the token
3887 * string -- setting a bit in the map for every character
3888 * found in the token string.
3889 */
3890 for (i = 0; i < sizeof (tokmap); i++)
3891 tokmap[i] = 0;
3892
3893 for (; tokaddr < toklimit; tokaddr++) {
3894 if ((c = dtrace_load8(tokaddr)) == '\0')
3895 break;
3896
3897 ASSERT((c >> 3) < sizeof (tokmap));
3898 tokmap[c >> 3] |= (1 << (c & 0x7));
3899 }
3900
3901 for (limit = addr + size; addr < limit; addr++) {
3902 /*
3903 * We're looking for a character that is _not_ contained
3904 * in the token string.
3905 */
3906 if ((c = dtrace_load8(addr)) == '\0')
3907 break;
3908
3909 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3910 break;
3911 }
3912
3913 if (c == '\0') {
3914 /*
3915 * We reached the end of the string without finding
3916 * any character that was not in the token string.
3917 * We return NULL in this case, and we set the saved
3918 * address to NULL as well.
3919 */
3920 regs[rd] = NULL;
3921 mstate->dtms_strtok = NULL;
3922 break;
3923 }
3924
3925 /*
3926 * From here on, we're copying into the destination string.
3927 */
3928 for (i = 0; addr < limit && i < size - 1; addr++) {
3929 if ((c = dtrace_load8(addr)) == '\0')
3930 break;
3931
3932 if (tokmap[c >> 3] & (1 << (c & 0x7)))
3933 break;
3934
3935 ASSERT(i < size);
3936 dest[i++] = c;
3937 }
3938
3939 ASSERT(i < size);
3940 dest[i] = '\0';
3941 regs[rd] = (uintptr_t)dest;
3942 mstate->dtms_scratch_ptr += size;
3943 mstate->dtms_strtok = addr;
3944 break;
3945 }
3946
3947 case DIF_SUBR_SUBSTR: {
3948 uintptr_t s = tupregs[0].dttk_value;
3949 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3950 char *d = (char *)mstate->dtms_scratch_ptr;
3951 int64_t index = (int64_t)tupregs[1].dttk_value;
3952 int64_t remaining = (int64_t)tupregs[2].dttk_value;
3953 size_t len = dtrace_strlen((char *)s, size);
3954 int64_t i;
3955
3956 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3957 regs[rd] = NULL;
3958 break;
3959 }
3960
3961 if (!DTRACE_INSCRATCH(mstate, size)) {
3962 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3963 regs[rd] = NULL;
3964 break;
3965 }
3966
3967 if (nargs <= 2)
3968 remaining = (int64_t)size;
3969
3970 if (index < 0) {
3971 index += len;
3972
3973 if (index < 0 && index + remaining > 0) {
3974 remaining += index;
3975 index = 0;
3976 }
3977 }
3978
3979 if (index >= len || index < 0) {
3980 remaining = 0;
3981 } else if (remaining < 0) {
3982 remaining += len - index;
3983 } else if (index + remaining > size) {
3984 remaining = size - index;
3985 }
3986
3987 for (i = 0; i < remaining; i++) {
3988 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
3989 break;
3990 }
3991
3992 d[i] = '\0';
3993
3994 mstate->dtms_scratch_ptr += size;
3995 regs[rd] = (uintptr_t)d;
3996 break;
3997 }
3998
3999 case DIF_SUBR_TOUPPER:
4000 case DIF_SUBR_TOLOWER: {
4001 uintptr_t s = tupregs[0].dttk_value;
4002 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4003 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4004 size_t len = dtrace_strlen((char *)s, size);
4005 char lower, upper, convert;
4006 int64_t i;
4007
4008 if (subr == DIF_SUBR_TOUPPER) {
4009 lower = 'a';
4010 upper = 'z';
4011 convert = 'A';
4012 } else {
4013 lower = 'A';
4014 upper = 'Z';
4015 convert = 'a';
4016 }
4017
4018 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4019 regs[rd] = NULL;
4020 break;
4021 }
4022
4023 if (!DTRACE_INSCRATCH(mstate, size)) {
4024 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4025 regs[rd] = NULL;
4026 break;
4027 }
4028
4029 for (i = 0; i < size - 1; i++) {
4030 if ((c = dtrace_load8(s + i)) == '\0')
4031 break;
4032
4033 if (c >= lower && c <= upper)
4034 c = convert + (c - lower);
4035
4036 dest[i] = c;
4037 }
4038
4039 ASSERT(i < size);
4040 dest[i] = '\0';
4041 regs[rd] = (uintptr_t)dest;
4042 mstate->dtms_scratch_ptr += size;
4043 break;
4044 }
4045
4046 case DIF_SUBR_GETMAJOR:
4047 #ifdef _LP64
4048 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4049 #else
4050 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4051 #endif
4052 break;
4053
4054 case DIF_SUBR_GETMINOR:
4055 #ifdef _LP64
4056 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4057 #else
4058 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4059 #endif
4060 break;
4061
4062 case DIF_SUBR_DDI_PATHNAME: {
4063 /*
4064 * This one is a galactic mess. We are going to roughly
4065 * emulate ddi_pathname(), but it's made more complicated
4066 * by the fact that we (a) want to include the minor name and
4067 * (b) must proceed iteratively instead of recursively.
4068 */
4069 uintptr_t dest = mstate->dtms_scratch_ptr;
4070 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4071 char *start = (char *)dest, *end = start + size - 1;
4072 uintptr_t daddr = tupregs[0].dttk_value;
4073 int64_t minor = (int64_t)tupregs[1].dttk_value;
4074 char *s;
4075 int i, len, depth = 0;
4076
4077 /*
4078 * Due to all the pointer jumping we do and context we must
4079 * rely upon, we just mandate that the user must have kernel
4080 * read privileges to use this routine.
4081 */
4082 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4083 *flags |= CPU_DTRACE_KPRIV;
4084 *illval = daddr;
4085 regs[rd] = NULL;
4086 }
4087
4088 if (!DTRACE_INSCRATCH(mstate, size)) {
4089 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4090 regs[rd] = NULL;
4091 break;
4092 }
4093
4094 *end = '\0';
4095
4096 /*
4097 * We want to have a name for the minor. In order to do this,
4098 * we need to walk the minor list from the devinfo. We want
4099 * to be sure that we don't infinitely walk a circular list,
4100 * so we check for circularity by sending a scout pointer
4101 * ahead two elements for every element that we iterate over;
4102 * if the list is circular, these will ultimately point to the
4103 * same element. You may recognize this little trick as the
4104 * answer to a stupid interview question -- one that always
4105 * seems to be asked by those who had to have it laboriously
4106 * explained to them, and who can't even concisely describe
4107 * the conditions under which one would be forced to resort to
4108 * this technique. Needless to say, those conditions are
4109 * found here -- and probably only here. Is this the only use
4110 * of this infamous trick in shipping, production code? If it
4111 * isn't, it probably should be...
4112 */
4113 if (minor != -1) {
4114 uintptr_t maddr = dtrace_loadptr(daddr +
4115 offsetof(struct dev_info, devi_minor));
4116
4117 uintptr_t next = offsetof(struct ddi_minor_data, next);
4118 uintptr_t name = offsetof(struct ddi_minor_data,
4119 d_minor) + offsetof(struct ddi_minor, name);
4120 uintptr_t dev = offsetof(struct ddi_minor_data,
4121 d_minor) + offsetof(struct ddi_minor, dev);
4122 uintptr_t scout;
4123
4124 if (maddr != NULL)
4125 scout = dtrace_loadptr(maddr + next);
4126
4127 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4128 uint64_t m;
4129 #ifdef _LP64
4130 m = dtrace_load64(maddr + dev) & MAXMIN64;
4131 #else
4132 m = dtrace_load32(maddr + dev) & MAXMIN;
4133 #endif
4134 if (m != minor) {
4135 maddr = dtrace_loadptr(maddr + next);
4136
4137 if (scout == NULL)
4138 continue;
4139
4140 scout = dtrace_loadptr(scout + next);
4141
4142 if (scout == NULL)
4143 continue;
4144
4145 scout = dtrace_loadptr(scout + next);
4146
4147 if (scout == NULL)
4148 continue;
4149
4150 if (scout == maddr) {
4151 *flags |= CPU_DTRACE_ILLOP;
4152 break;
4153 }
4154
4155 continue;
4156 }
4157
4158 /*
4159 * We have the minor data. Now we need to
4160 * copy the minor's name into the end of the
4161 * pathname.
4162 */
4163 s = (char *)dtrace_loadptr(maddr + name);
4164 len = dtrace_strlen(s, size);
4165
4166 if (*flags & CPU_DTRACE_FAULT)
4167 break;
4168
4169 if (len != 0) {
4170 if ((end -= (len + 1)) < start)
4171 break;
4172
4173 *end = ':';
4174 }
4175
4176 for (i = 1; i <= len; i++)
4177 end[i] = dtrace_load8((uintptr_t)s++);
4178 break;
4179 }
4180 }
4181
4182 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4183 ddi_node_state_t devi_state;
4184
4185 devi_state = dtrace_load32(daddr +
4186 offsetof(struct dev_info, devi_node_state));
4187
4188 if (*flags & CPU_DTRACE_FAULT)
4189 break;
4190
4191 if (devi_state >= DS_INITIALIZED) {
4192 s = (char *)dtrace_loadptr(daddr +
4193 offsetof(struct dev_info, devi_addr));
4194 len = dtrace_strlen(s, size);
4195
4196 if (*flags & CPU_DTRACE_FAULT)
4197 break;
4198
4199 if (len != 0) {
4200 if ((end -= (len + 1)) < start)
4201 break;
4202
4203 *end = '@';
4204 }
4205
4206 for (i = 1; i <= len; i++)
4207 end[i] = dtrace_load8((uintptr_t)s++);
4208 }
4209
4210 /*
4211 * Now for the node name...
4212 */
4213 s = (char *)dtrace_loadptr(daddr +
4214 offsetof(struct dev_info, devi_node_name));
4215
4216 daddr = dtrace_loadptr(daddr +
4217 offsetof(struct dev_info, devi_parent));
4218
4219 /*
4220 * If our parent is NULL (that is, if we're the root
4221 * node), we're going to use the special path
4222 * "devices".
4223 */
4224 if (daddr == NULL)
4225 s = "devices";
4226
4227 len = dtrace_strlen(s, size);
4228 if (*flags & CPU_DTRACE_FAULT)
4229 break;
4230
4231 if ((end -= (len + 1)) < start)
4232 break;
4233
4234 for (i = 1; i <= len; i++)
4235 end[i] = dtrace_load8((uintptr_t)s++);
4236 *end = '/';
4237
4238 if (depth++ > dtrace_devdepth_max) {
4239 *flags |= CPU_DTRACE_ILLOP;
4240 break;
4241 }
4242 }
4243
4244 if (end < start)
4245 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4246
4247 if (daddr == NULL) {
4248 regs[rd] = (uintptr_t)end;
4249 mstate->dtms_scratch_ptr += size;
4250 }
4251
4252 break;
4253 }
4254
4255 case DIF_SUBR_STRJOIN: {
4256 char *d = (char *)mstate->dtms_scratch_ptr;
4257 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4258 uintptr_t s1 = tupregs[0].dttk_value;
4259 uintptr_t s2 = tupregs[1].dttk_value;
4260 int i = 0;
4261
4262 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4263 !dtrace_strcanload(s2, size, mstate, vstate)) {
4264 regs[rd] = NULL;
4265 break;
4266 }
4267
4268 if (!DTRACE_INSCRATCH(mstate, size)) {
4269 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4270 regs[rd] = NULL;
4271 break;
4272 }
4273
4274 for (;;) {
4275 if (i >= size) {
4276 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4277 regs[rd] = NULL;
4278 break;
4279 }
4280
4281 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4282 i--;
4283 break;
4284 }
4285 }
4286
4287 for (;;) {
4288 if (i >= size) {
4289 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4290 regs[rd] = NULL;
4291 break;
4292 }
4293
4294 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4295 break;
4296 }
4297
4298 if (i < size) {
4299 mstate->dtms_scratch_ptr += i;
4300 regs[rd] = (uintptr_t)d;
4301 }
4302
4303 break;
4304 }
4305
4306 case DIF_SUBR_LLTOSTR: {
4307 int64_t i = (int64_t)tupregs[0].dttk_value;
4308 uint64_t val, digit;
4309 uint64_t size = 65; /* enough room for 2^64 in binary */
4310 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4311 int base = 10;
4312
4313 if (nargs > 1) {
4314 if ((base = tupregs[1].dttk_value) <= 1 ||
4315 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4316 *flags |= CPU_DTRACE_ILLOP;
4317 break;
4318 }
4319 }
4320
4321 val = (base == 10 && i < 0) ? i * -1 : i;
4322
4323 if (!DTRACE_INSCRATCH(mstate, size)) {
4324 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4325 regs[rd] = NULL;
4326 break;
4327 }
4328
4329 for (*end-- = '\0'; val; val /= base) {
4330 if ((digit = val % base) <= '9' - '0') {
4331 *end-- = '0' + digit;
4332 } else {
4333 *end-- = 'a' + (digit - ('9' - '0') - 1);
4334 }
4335 }
4336
4337 if (i == 0 && base == 16)
4338 *end-- = '0';
4339
4340 if (base == 16)
4341 *end-- = 'x';
4342
4343 if (i == 0 || base == 8 || base == 16)
4344 *end-- = '0';
4345
4346 if (i < 0 && base == 10)
4347 *end-- = '-';
4348
4349 regs[rd] = (uintptr_t)end + 1;
4350 mstate->dtms_scratch_ptr += size;
4351 break;
4352 }
4353
4354 case DIF_SUBR_HTONS:
4355 case DIF_SUBR_NTOHS:
4356 #ifdef _BIG_ENDIAN
4357 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4358 #else
4359 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4360 #endif
4361 break;
4362
4363
4364 case DIF_SUBR_HTONL:
4365 case DIF_SUBR_NTOHL:
4366 #ifdef _BIG_ENDIAN
4367 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4368 #else
4369 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4370 #endif
4371 break;
4372
4373
4374 case DIF_SUBR_HTONLL:
4375 case DIF_SUBR_NTOHLL:
4376 #ifdef _BIG_ENDIAN
4377 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4378 #else
4379 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4380 #endif
4381 break;
4382
4383
4384 case DIF_SUBR_DIRNAME:
4385 case DIF_SUBR_BASENAME: {
4386 char *dest = (char *)mstate->dtms_scratch_ptr;
4387 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4388 uintptr_t src = tupregs[0].dttk_value;
4389 int i, j, len = dtrace_strlen((char *)src, size);
4390 int lastbase = -1, firstbase = -1, lastdir = -1;
4391 int start, end;
4392
4393 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4394 regs[rd] = NULL;
4395 break;
4396 }
4397
4398 if (!DTRACE_INSCRATCH(mstate, size)) {
4399 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4400 regs[rd] = NULL;
4401 break;
4402 }
4403
4404 /*
4405 * The basename and dirname for a zero-length string is
4406 * defined to be "."
4407 */
4408 if (len == 0) {
4409 len = 1;
4410 src = (uintptr_t)".";
4411 }
4412
4413 /*
4414 * Start from the back of the string, moving back toward the
4415 * front until we see a character that isn't a slash. That
4416 * character is the last character in the basename.
4417 */
4418 for (i = len - 1; i >= 0; i--) {
4419 if (dtrace_load8(src + i) != '/')
4420 break;
4421 }
4422
4423 if (i >= 0)
4424 lastbase = i;
4425
4426 /*
4427 * Starting from the last character in the basename, move
4428 * towards the front until we find a slash. The character
4429 * that we processed immediately before that is the first
4430 * character in the basename.
4431 */
4432 for (; i >= 0; i--) {
4433 if (dtrace_load8(src + i) == '/')
4434 break;
4435 }
4436
4437 if (i >= 0)
4438 firstbase = i + 1;
4439
4440 /*
4441 * Now keep going until we find a non-slash character. That
4442 * character is the last character in the dirname.
4443 */
4444 for (; i >= 0; i--) {
4445 if (dtrace_load8(src + i) != '/')
4446 break;
4447 }
4448
4449 if (i >= 0)
4450 lastdir = i;
4451
4452 ASSERT(!(lastbase == -1 && firstbase != -1));
4453 ASSERT(!(firstbase == -1 && lastdir != -1));
4454
4455 if (lastbase == -1) {
4456 /*
4457 * We didn't find a non-slash character. We know that
4458 * the length is non-zero, so the whole string must be
4459 * slashes. In either the dirname or the basename
4460 * case, we return '/'.
4461 */
4462 ASSERT(firstbase == -1);
4463 firstbase = lastbase = lastdir = 0;
4464 }
4465
4466 if (firstbase == -1) {
4467 /*
4468 * The entire string consists only of a basename
4469 * component. If we're looking for dirname, we need
4470 * to change our string to be just "."; if we're
4471 * looking for a basename, we'll just set the first
4472 * character of the basename to be 0.
4473 */
4474 if (subr == DIF_SUBR_DIRNAME) {
4475 ASSERT(lastdir == -1);
4476 src = (uintptr_t)".";
4477 lastdir = 0;
4478 } else {
4479 firstbase = 0;
4480 }
4481 }
4482
4483 if (subr == DIF_SUBR_DIRNAME) {
4484 if (lastdir == -1) {
4485 /*
4486 * We know that we have a slash in the name --
4487 * or lastdir would be set to 0, above. And
4488 * because lastdir is -1, we know that this
4489 * slash must be the first character. (That
4490 * is, the full string must be of the form
4491 * "/basename".) In this case, the last
4492 * character of the directory name is 0.
4493 */
4494 lastdir = 0;
4495 }
4496
4497 start = 0;
4498 end = lastdir;
4499 } else {
4500 ASSERT(subr == DIF_SUBR_BASENAME);
4501 ASSERT(firstbase != -1 && lastbase != -1);
4502 start = firstbase;
4503 end = lastbase;
4504 }
4505
4506 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4507 dest[j] = dtrace_load8(src + i);
4508
4509 dest[j] = '\0';
4510 regs[rd] = (uintptr_t)dest;
4511 mstate->dtms_scratch_ptr += size;
4512 break;
4513 }
4514
4515 case DIF_SUBR_GETF: {
4516 uintptr_t fd = tupregs[0].dttk_value;
4517 uf_info_t *finfo = &curthread->t_procp->p_user.u_finfo;
4518 file_t *fp;
4519
4520 if (!dtrace_priv_proc(state, mstate)) {
4521 regs[rd] = NULL;
4522 break;
4523 }
4524
4525 /*
4526 * This is safe because fi_nfiles only increases, and the
4527 * fi_list array is not freed when the array size doubles.
4528 * (See the comment in flist_grow() for details on the
4529 * management of the u_finfo structure.)
4530 */
4531 fp = fd < finfo->fi_nfiles ? finfo->fi_list[fd].uf_file : NULL;
4532
4533 mstate->dtms_getf = fp;
4534 regs[rd] = (uintptr_t)fp;
4535 break;
4536 }
4537
4538 case DIF_SUBR_CLEANPATH: {
4539 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4540 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4541 uintptr_t src = tupregs[0].dttk_value;
4542 int i = 0, j = 0;
4543 zone_t *z;
4544
4545 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4546 regs[rd] = NULL;
4547 break;
4548 }
4549
4550 if (!DTRACE_INSCRATCH(mstate, size)) {
4551 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4552 regs[rd] = NULL;
4553 break;
4554 }
4555
4556 /*
4557 * Move forward, loading each character.
4558 */
4559 do {
4560 c = dtrace_load8(src + i++);
4561 next:
4562 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4563 break;
4564
4565 if (c != '/') {
4566 dest[j++] = c;
4567 continue;
4568 }
4569
4570 c = dtrace_load8(src + i++);
4571
4572 if (c == '/') {
4573 /*
4574 * We have two slashes -- we can just advance
4575 * to the next character.
4576 */
4577 goto next;
4578 }
4579
4580 if (c != '.') {
4581 /*
4582 * This is not "." and it's not ".." -- we can
4583 * just store the "/" and this character and
4584 * drive on.
4585 */
4586 dest[j++] = '/';
4587 dest[j++] = c;
4588 continue;
4589 }
4590
4591 c = dtrace_load8(src + i++);
4592
4593 if (c == '/') {
4594 /*
4595 * This is a "/./" component. We're not going
4596 * to store anything in the destination buffer;
4597 * we're just going to go to the next component.
4598 */
4599 goto next;
4600 }
4601
4602 if (c != '.') {
4603 /*
4604 * This is not ".." -- we can just store the
4605 * "/." and this character and continue
4606 * processing.
4607 */
4608 dest[j++] = '/';
4609 dest[j++] = '.';
4610 dest[j++] = c;
4611 continue;
4612 }
4613
4614 c = dtrace_load8(src + i++);
4615
4616 if (c != '/' && c != '\0') {
4617 /*
4618 * This is not ".." -- it's "..[mumble]".
4619 * We'll store the "/.." and this character
4620 * and continue processing.
4621 */
4622 dest[j++] = '/';
4623 dest[j++] = '.';
4624 dest[j++] = '.';
4625 dest[j++] = c;
4626 continue;
4627 }
4628
4629 /*
4630 * This is "/../" or "/..\0". We need to back up
4631 * our destination pointer until we find a "/".
4632 */
4633 i--;
4634 while (j != 0 && dest[--j] != '/')
4635 continue;
4636
4637 if (c == '\0')
4638 dest[++j] = '/';
4639 } while (c != '\0');
4640
4641 dest[j] = '\0';
4642
4643 if (mstate->dtms_getf != NULL &&
4644 !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
4645 (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
4646 /*
4647 * If we've done a getf() as a part of this ECB and we
4648 * don't have kernel access (and we're not in the global
4649 * zone), check if the path we cleaned up begins with
4650 * the zone's root path, and trim it off if so. Note
4651 * that this is an output cleanliness issue, not a
4652 * security issue: knowing one's zone root path does
4653 * not enable privilege escalation.
4654 */
4655 if (strstr(dest, z->zone_rootpath) == dest)
4656 dest += strlen(z->zone_rootpath) - 1;
4657 }
4658
4659 regs[rd] = (uintptr_t)dest;
4660 mstate->dtms_scratch_ptr += size;
4661 break;
4662 }
4663
4664 case DIF_SUBR_INET_NTOA:
4665 case DIF_SUBR_INET_NTOA6:
4666 case DIF_SUBR_INET_NTOP: {
4667 size_t size;
4668 int af, argi, i;
4669 char *base, *end;
4670
4671 if (subr == DIF_SUBR_INET_NTOP) {
4672 af = (int)tupregs[0].dttk_value;
4673 argi = 1;
4674 } else {
4675 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4676 argi = 0;
4677 }
4678
4679 if (af == AF_INET) {
4680 ipaddr_t ip4;
4681 uint8_t *ptr8, val;
4682
4683 /*
4684 * Safely load the IPv4 address.
4685 */
4686 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4687
4688 /*
4689 * Check an IPv4 string will fit in scratch.
4690 */
4691 size = INET_ADDRSTRLEN;
4692 if (!DTRACE_INSCRATCH(mstate, size)) {
4693 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4694 regs[rd] = NULL;
4695 break;
4696 }
4697 base = (char *)mstate->dtms_scratch_ptr;
4698 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4699
4700 /*
4701 * Stringify as a dotted decimal quad.
4702 */
4703 *end-- = '\0';
4704 ptr8 = (uint8_t *)&ip4;
4705 for (i = 3; i >= 0; i--) {
4706 val = ptr8[i];
4707
4708 if (val == 0) {
4709 *end-- = '0';
4710 } else {
4711 for (; val; val /= 10) {
4712 *end-- = '0' + (val % 10);
4713 }
4714 }
4715
4716 if (i > 0)
4717 *end-- = '.';
4718 }
4719 ASSERT(end + 1 >= base);
4720
4721 } else if (af == AF_INET6) {
4722 struct in6_addr ip6;
4723 int firstzero, tryzero, numzero, v6end;
4724 uint16_t val;
4725 const char digits[] = "0123456789abcdef";
4726
4727 /*
4728 * Stringify using RFC 1884 convention 2 - 16 bit
4729 * hexadecimal values with a zero-run compression.
4730 * Lower case hexadecimal digits are used.
4731 * eg, fe80::214:4fff:fe0b:76c8.
4732 * The IPv4 embedded form is returned for inet_ntop,
4733 * just the IPv4 string is returned for inet_ntoa6.
4734 */
4735
4736 /*
4737 * Safely load the IPv6 address.
4738 */
4739 dtrace_bcopy(
4740 (void *)(uintptr_t)tupregs[argi].dttk_value,
4741 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4742
4743 /*
4744 * Check an IPv6 string will fit in scratch.
4745 */
4746 size = INET6_ADDRSTRLEN;
4747 if (!DTRACE_INSCRATCH(mstate, size)) {
4748 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4749 regs[rd] = NULL;
4750 break;
4751 }
4752 base = (char *)mstate->dtms_scratch_ptr;
4753 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4754 *end-- = '\0';
4755
4756 /*
4757 * Find the longest run of 16 bit zero values
4758 * for the single allowed zero compression - "::".
4759 */
4760 firstzero = -1;
4761 tryzero = -1;
4762 numzero = 1;
4763 for (i = 0; i < sizeof (struct in6_addr); i++) {
4764 if (ip6._S6_un._S6_u8[i] == 0 &&
4765 tryzero == -1 && i % 2 == 0) {
4766 tryzero = i;
4767 continue;
4768 }
4769
4770 if (tryzero != -1 &&
4771 (ip6._S6_un._S6_u8[i] != 0 ||
4772 i == sizeof (struct in6_addr) - 1)) {
4773
4774 if (i - tryzero <= numzero) {
4775 tryzero = -1;
4776 continue;
4777 }
4778
4779 firstzero = tryzero;
4780 numzero = i - i % 2 - tryzero;
4781 tryzero = -1;
4782
4783 if (ip6._S6_un._S6_u8[i] == 0 &&
4784 i == sizeof (struct in6_addr) - 1)
4785 numzero += 2;
4786 }
4787 }
4788 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4789
4790 /*
4791 * Check for an IPv4 embedded address.
4792 */
4793 v6end = sizeof (struct in6_addr) - 2;
4794 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4795 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4796 for (i = sizeof (struct in6_addr) - 1;
4797 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4798 ASSERT(end >= base);
4799
4800 val = ip6._S6_un._S6_u8[i];
4801
4802 if (val == 0) {
4803 *end-- = '0';
4804 } else {
4805 for (; val; val /= 10) {
4806 *end-- = '0' + val % 10;
4807 }
4808 }
4809
4810 if (i > DTRACE_V4MAPPED_OFFSET)
4811 *end-- = '.';
4812 }
4813
4814 if (subr == DIF_SUBR_INET_NTOA6)
4815 goto inetout;
4816
4817 /*
4818 * Set v6end to skip the IPv4 address that
4819 * we have already stringified.
4820 */
4821 v6end = 10;
4822 }
4823
4824 /*
4825 * Build the IPv6 string by working through the
4826 * address in reverse.
4827 */
4828 for (i = v6end; i >= 0; i -= 2) {
4829 ASSERT(end >= base);
4830
4831 if (i == firstzero + numzero - 2) {
4832 *end-- = ':';
4833 *end-- = ':';
4834 i -= numzero - 2;
4835 continue;
4836 }
4837
4838 if (i < 14 && i != firstzero - 2)
4839 *end-- = ':';
4840
4841 val = (ip6._S6_un._S6_u8[i] << 8) +
4842 ip6._S6_un._S6_u8[i + 1];
4843
4844 if (val == 0) {
4845 *end-- = '0';
4846 } else {
4847 for (; val; val /= 16) {
4848 *end-- = digits[val % 16];
4849 }
4850 }
4851 }
4852 ASSERT(end + 1 >= base);
4853
4854 } else {
4855 /*
4856 * The user didn't use AH_INET or AH_INET6.
4857 */
4858 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4859 regs[rd] = NULL;
4860 break;
4861 }
4862
4863 inetout: regs[rd] = (uintptr_t)end + 1;
4864 mstate->dtms_scratch_ptr += size;
4865 break;
4866 }
4867
4868 }
4869 }
4870
4871 /*
4872 * Emulate the execution of DTrace IR instructions specified by the given
4873 * DIF object. This function is deliberately void of assertions as all of
4874 * the necessary checks are handled by a call to dtrace_difo_validate().
4875 */
4876 static uint64_t
4877 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4878 dtrace_vstate_t *vstate, dtrace_state_t *state)
4879 {
4880 const dif_instr_t *text = difo->dtdo_buf;
4881 const uint_t textlen = difo->dtdo_len;
4882 const char *strtab = difo->dtdo_strtab;
4883 const uint64_t *inttab = difo->dtdo_inttab;
4884
4885 uint64_t rval = 0;
4886 dtrace_statvar_t *svar;
4887 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4888 dtrace_difv_t *v;
4889 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4890 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
4891
4892 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4893 uint64_t regs[DIF_DIR_NREGS];
4894 uint64_t *tmp;
4895
4896 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4897 int64_t cc_r;
4898 uint_t pc = 0, id, opc;
4899 uint8_t ttop = 0;
4900 dif_instr_t instr;
4901 uint_t r1, r2, rd;
4902
4903 /*
4904 * We stash the current DIF object into the machine state: we need it
4905 * for subsequent access checking.
4906 */
4907 mstate->dtms_difo = difo;
4908
4909 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
4910
4911 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4912 opc = pc;
4913
4914 instr = text[pc++];
4915 r1 = DIF_INSTR_R1(instr);
4916 r2 = DIF_INSTR_R2(instr);
4917 rd = DIF_INSTR_RD(instr);
4918
4919 switch (DIF_INSTR_OP(instr)) {
4920 case DIF_OP_OR:
4921 regs[rd] = regs[r1] | regs[r2];
4922 break;
4923 case DIF_OP_XOR:
4924 regs[rd] = regs[r1] ^ regs[r2];
4925 break;
4926 case DIF_OP_AND:
4927 regs[rd] = regs[r1] & regs[r2];
4928 break;
4929 case DIF_OP_SLL:
4930 regs[rd] = regs[r1] << regs[r2];
4931 break;
4932 case DIF_OP_SRL:
4933 regs[rd] = regs[r1] >> regs[r2];
4934 break;
4935 case DIF_OP_SUB:
4936 regs[rd] = regs[r1] - regs[r2];
4937 break;
4938 case DIF_OP_ADD:
4939 regs[rd] = regs[r1] + regs[r2];
4940 break;
4941 case DIF_OP_MUL:
4942 regs[rd] = regs[r1] * regs[r2];
4943 break;
4944 case DIF_OP_SDIV:
4945 if (regs[r2] == 0) {
4946 regs[rd] = 0;
4947 *flags |= CPU_DTRACE_DIVZERO;
4948 } else {
4949 regs[rd] = (int64_t)regs[r1] /
4950 (int64_t)regs[r2];
4951 }
4952 break;
4953
4954 case DIF_OP_UDIV:
4955 if (regs[r2] == 0) {
4956 regs[rd] = 0;
4957 *flags |= CPU_DTRACE_DIVZERO;
4958 } else {
4959 regs[rd] = regs[r1] / regs[r2];
4960 }
4961 break;
4962
4963 case DIF_OP_SREM:
4964 if (regs[r2] == 0) {
4965 regs[rd] = 0;
4966 *flags |= CPU_DTRACE_DIVZERO;
4967 } else {
4968 regs[rd] = (int64_t)regs[r1] %
4969 (int64_t)regs[r2];
4970 }
4971 break;
4972
4973 case DIF_OP_UREM:
4974 if (regs[r2] == 0) {
4975 regs[rd] = 0;
4976 *flags |= CPU_DTRACE_DIVZERO;
4977 } else {
4978 regs[rd] = regs[r1] % regs[r2];
4979 }
4980 break;
4981
4982 case DIF_OP_NOT:
4983 regs[rd] = ~regs[r1];
4984 break;
4985 case DIF_OP_MOV:
4986 regs[rd] = regs[r1];
4987 break;
4988 case DIF_OP_CMP:
4989 cc_r = regs[r1] - regs[r2];
4990 cc_n = cc_r < 0;
4991 cc_z = cc_r == 0;
4992 cc_v = 0;
4993 cc_c = regs[r1] < regs[r2];
4994 break;
4995 case DIF_OP_TST:
4996 cc_n = cc_v = cc_c = 0;
4997 cc_z = regs[r1] == 0;
4998 break;
4999 case DIF_OP_BA:
5000 pc = DIF_INSTR_LABEL(instr);
5001 break;
5002 case DIF_OP_BE:
5003 if (cc_z)
5004 pc = DIF_INSTR_LABEL(instr);
5005 break;
5006 case DIF_OP_BNE:
5007 if (cc_z == 0)
5008 pc = DIF_INSTR_LABEL(instr);
5009 break;
5010 case DIF_OP_BG:
5011 if ((cc_z | (cc_n ^ cc_v)) == 0)
5012 pc = DIF_INSTR_LABEL(instr);
5013 break;
5014 case DIF_OP_BGU:
5015 if ((cc_c | cc_z) == 0)
5016 pc = DIF_INSTR_LABEL(instr);
5017 break;
5018 case DIF_OP_BGE:
5019 if ((cc_n ^ cc_v) == 0)
5020 pc = DIF_INSTR_LABEL(instr);
5021 break;
5022 case DIF_OP_BGEU:
5023 if (cc_c == 0)
5024 pc = DIF_INSTR_LABEL(instr);
5025 break;
5026 case DIF_OP_BL:
5027 if (cc_n ^ cc_v)
5028 pc = DIF_INSTR_LABEL(instr);
5029 break;
5030 case DIF_OP_BLU:
5031 if (cc_c)
5032 pc = DIF_INSTR_LABEL(instr);
5033 break;
5034 case DIF_OP_BLE:
5035 if (cc_z | (cc_n ^ cc_v))
5036 pc = DIF_INSTR_LABEL(instr);
5037 break;
5038 case DIF_OP_BLEU:
5039 if (cc_c | cc_z)
5040 pc = DIF_INSTR_LABEL(instr);
5041 break;
5042 case DIF_OP_RLDSB:
5043 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5044 break;
5045 /*FALLTHROUGH*/
5046 case DIF_OP_LDSB:
5047 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5048 break;
5049 case DIF_OP_RLDSH:
5050 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5051 break;
5052 /*FALLTHROUGH*/
5053 case DIF_OP_LDSH:
5054 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5055 break;
5056 case DIF_OP_RLDSW:
5057 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5058 break;
5059 /*FALLTHROUGH*/
5060 case DIF_OP_LDSW:
5061 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5062 break;
5063 case DIF_OP_RLDUB:
5064 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5065 break;
5066 /*FALLTHROUGH*/
5067 case DIF_OP_LDUB:
5068 regs[rd] = dtrace_load8(regs[r1]);
5069 break;
5070 case DIF_OP_RLDUH:
5071 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5072 break;
5073 /*FALLTHROUGH*/
5074 case DIF_OP_LDUH:
5075 regs[rd] = dtrace_load16(regs[r1]);
5076 break;
5077 case DIF_OP_RLDUW:
5078 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5079 break;
5080 /*FALLTHROUGH*/
5081 case DIF_OP_LDUW:
5082 regs[rd] = dtrace_load32(regs[r1]);
5083 break;
5084 case DIF_OP_RLDX:
5085 if (!dtrace_canload(regs[r1], 8, mstate, vstate))
5086 break;
5087 /*FALLTHROUGH*/
5088 case DIF_OP_LDX:
5089 regs[rd] = dtrace_load64(regs[r1]);
5090 break;
5091 case DIF_OP_ULDSB:
5092 regs[rd] = (int8_t)
5093 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5094 break;
5095 case DIF_OP_ULDSH:
5096 regs[rd] = (int16_t)
5097 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5098 break;
5099 case DIF_OP_ULDSW:
5100 regs[rd] = (int32_t)
5101 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5102 break;
5103 case DIF_OP_ULDUB:
5104 regs[rd] =
5105 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5106 break;
5107 case DIF_OP_ULDUH:
5108 regs[rd] =
5109 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5110 break;
5111 case DIF_OP_ULDUW:
5112 regs[rd] =
5113 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5114 break;
5115 case DIF_OP_ULDX:
5116 regs[rd] =
5117 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5118 break;
5119 case DIF_OP_RET:
5120 rval = regs[rd];
5121 pc = textlen;
5122 break;
5123 case DIF_OP_NOP:
5124 break;
5125 case DIF_OP_SETX:
5126 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5127 break;
5128 case DIF_OP_SETS:
5129 regs[rd] = (uint64_t)(uintptr_t)
5130 (strtab + DIF_INSTR_STRING(instr));
5131 break;
5132 case DIF_OP_SCMP: {
5133 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5134 uintptr_t s1 = regs[r1];
5135 uintptr_t s2 = regs[r2];
5136
5137 if (s1 != NULL &&
5138 !dtrace_strcanload(s1, sz, mstate, vstate))
5139 break;
5140 if (s2 != NULL &&
5141 !dtrace_strcanload(s2, sz, mstate, vstate))
5142 break;
5143
5144 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5145
5146 cc_n = cc_r < 0;
5147 cc_z = cc_r == 0;
5148 cc_v = cc_c = 0;
5149 break;
5150 }
5151 case DIF_OP_LDGA:
5152 regs[rd] = dtrace_dif_variable(mstate, state,
5153 r1, regs[r2]);
5154 break;
5155 case DIF_OP_LDGS:
5156 id = DIF_INSTR_VAR(instr);
5157
5158 if (id >= DIF_VAR_OTHER_UBASE) {
5159 uintptr_t a;
5160
5161 id -= DIF_VAR_OTHER_UBASE;
5162 svar = vstate->dtvs_globals[id];
5163 ASSERT(svar != NULL);
5164 v = &svar->dtsv_var;
5165
5166 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5167 regs[rd] = svar->dtsv_data;
5168 break;
5169 }
5170
5171 a = (uintptr_t)svar->dtsv_data;
5172
5173 if (*(uint8_t *)a == UINT8_MAX) {
5174 /*
5175 * If the 0th byte is set to UINT8_MAX
5176 * then this is to be treated as a
5177 * reference to a NULL variable.
5178 */
5179 regs[rd] = NULL;
5180 } else {
5181 regs[rd] = a + sizeof (uint64_t);
5182 }
5183
5184 break;
5185 }
5186
5187 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5188 break;
5189
5190 case DIF_OP_STGS:
5191 id = DIF_INSTR_VAR(instr);
5192
5193 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5194 id -= DIF_VAR_OTHER_UBASE;
5195
5196 svar = vstate->dtvs_globals[id];
5197 ASSERT(svar != NULL);
5198 v = &svar->dtsv_var;
5199
5200 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5201 uintptr_t a = (uintptr_t)svar->dtsv_data;
5202
5203 ASSERT(a != NULL);
5204 ASSERT(svar->dtsv_size != 0);
5205
5206 if (regs[rd] == NULL) {
5207 *(uint8_t *)a = UINT8_MAX;
5208 break;
5209 } else {
5210 *(uint8_t *)a = 0;
5211 a += sizeof (uint64_t);
5212 }
5213 if (!dtrace_vcanload(
5214 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5215 mstate, vstate))
5216 break;
5217
5218 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5219 (void *)a, &v->dtdv_type);
5220 break;
5221 }
5222
5223 svar->dtsv_data = regs[rd];
5224 break;
5225
5226 case DIF_OP_LDTA:
5227 /*
5228 * There are no DTrace built-in thread-local arrays at
5229 * present. This opcode is saved for future work.
5230 */
5231 *flags |= CPU_DTRACE_ILLOP;
5232 regs[rd] = 0;
5233 break;
5234
5235 case DIF_OP_LDLS:
5236 id = DIF_INSTR_VAR(instr);
5237
5238 if (id < DIF_VAR_OTHER_UBASE) {
5239 /*
5240 * For now, this has no meaning.
5241 */
5242 regs[rd] = 0;
5243 break;
5244 }
5245
5246 id -= DIF_VAR_OTHER_UBASE;
5247
5248 ASSERT(id < vstate->dtvs_nlocals);
5249 ASSERT(vstate->dtvs_locals != NULL);
5250
5251 svar = vstate->dtvs_locals[id];
5252 ASSERT(svar != NULL);
5253 v = &svar->dtsv_var;
5254
5255 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5256 uintptr_t a = (uintptr_t)svar->dtsv_data;
5257 size_t sz = v->dtdv_type.dtdt_size;
5258
5259 sz += sizeof (uint64_t);
5260 ASSERT(svar->dtsv_size == NCPU * sz);
5261 a += CPU->cpu_id * sz;
5262
5263 if (*(uint8_t *)a == UINT8_MAX) {
5264 /*
5265 * If the 0th byte is set to UINT8_MAX
5266 * then this is to be treated as a
5267 * reference to a NULL variable.
5268 */
5269 regs[rd] = NULL;
5270 } else {
5271 regs[rd] = a + sizeof (uint64_t);
5272 }
5273
5274 break;
5275 }
5276
5277 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5278 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5279 regs[rd] = tmp[CPU->cpu_id];
5280 break;
5281
5282 case DIF_OP_STLS:
5283 id = DIF_INSTR_VAR(instr);
5284
5285 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5286 id -= DIF_VAR_OTHER_UBASE;
5287 ASSERT(id < vstate->dtvs_nlocals);
5288
5289 ASSERT(vstate->dtvs_locals != NULL);
5290 svar = vstate->dtvs_locals[id];
5291 ASSERT(svar != NULL);
5292 v = &svar->dtsv_var;
5293
5294 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5295 uintptr_t a = (uintptr_t)svar->dtsv_data;
5296 size_t sz = v->dtdv_type.dtdt_size;
5297
5298 sz += sizeof (uint64_t);
5299 ASSERT(svar->dtsv_size == NCPU * sz);
5300 a += CPU->cpu_id * sz;
5301
5302 if (regs[rd] == NULL) {
5303 *(uint8_t *)a = UINT8_MAX;
5304 break;
5305 } else {
5306 *(uint8_t *)a = 0;
5307 a += sizeof (uint64_t);
5308 }
5309
5310 if (!dtrace_vcanload(
5311 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5312 mstate, vstate))
5313 break;
5314
5315 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5316 (void *)a, &v->dtdv_type);
5317 break;
5318 }
5319
5320 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5321 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5322 tmp[CPU->cpu_id] = regs[rd];
5323 break;
5324
5325 case DIF_OP_LDTS: {
5326 dtrace_dynvar_t *dvar;
5327 dtrace_key_t *key;
5328
5329 id = DIF_INSTR_VAR(instr);
5330 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5331 id -= DIF_VAR_OTHER_UBASE;
5332 v = &vstate->dtvs_tlocals[id];
5333
5334 key = &tupregs[DIF_DTR_NREGS];
5335 key[0].dttk_value = (uint64_t)id;
5336 key[0].dttk_size = 0;
5337 DTRACE_TLS_THRKEY(key[1].dttk_value);
5338 key[1].dttk_size = 0;
5339
5340 dvar = dtrace_dynvar(dstate, 2, key,
5341 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5342 mstate, vstate);
5343
5344 if (dvar == NULL) {
5345 regs[rd] = 0;
5346 break;
5347 }
5348
5349 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5350 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5351 } else {
5352 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5353 }
5354
5355 break;
5356 }
5357
5358 case DIF_OP_STTS: {
5359 dtrace_dynvar_t *dvar;
5360 dtrace_key_t *key;
5361
5362 id = DIF_INSTR_VAR(instr);
5363 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5364 id -= DIF_VAR_OTHER_UBASE;
5365
5366 key = &tupregs[DIF_DTR_NREGS];
5367 key[0].dttk_value = (uint64_t)id;
5368 key[0].dttk_size = 0;
5369 DTRACE_TLS_THRKEY(key[1].dttk_value);
5370 key[1].dttk_size = 0;
5371 v = &vstate->dtvs_tlocals[id];
5372
5373 dvar = dtrace_dynvar(dstate, 2, key,
5374 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5375 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5376 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5377 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5378
5379 /*
5380 * Given that we're storing to thread-local data,
5381 * we need to flush our predicate cache.
5382 */
5383 curthread->t_predcache = NULL;
5384
5385 if (dvar == NULL)
5386 break;
5387
5388 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5389 if (!dtrace_vcanload(
5390 (void *)(uintptr_t)regs[rd],
5391 &v->dtdv_type, mstate, vstate))
5392 break;
5393
5394 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5395 dvar->dtdv_data, &v->dtdv_type);
5396 } else {
5397 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5398 }
5399
5400 break;
5401 }
5402
5403 case DIF_OP_SRA:
5404 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5405 break;
5406
5407 case DIF_OP_CALL:
5408 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5409 regs, tupregs, ttop, mstate, state);
5410 break;
5411
5412 case DIF_OP_PUSHTR:
5413 if (ttop == DIF_DTR_NREGS) {
5414 *flags |= CPU_DTRACE_TUPOFLOW;
5415 break;
5416 }
5417
5418 if (r1 == DIF_TYPE_STRING) {
5419 /*
5420 * If this is a string type and the size is 0,
5421 * we'll use the system-wide default string
5422 * size. Note that we are _not_ looking at
5423 * the value of the DTRACEOPT_STRSIZE option;
5424 * had this been set, we would expect to have
5425 * a non-zero size value in the "pushtr".
5426 */
5427 tupregs[ttop].dttk_size =
5428 dtrace_strlen((char *)(uintptr_t)regs[rd],
5429 regs[r2] ? regs[r2] :
5430 dtrace_strsize_default) + 1;
5431 } else {
5432 tupregs[ttop].dttk_size = regs[r2];
5433 }
5434
5435 tupregs[ttop++].dttk_value = regs[rd];
5436 break;
5437
5438 case DIF_OP_PUSHTV:
5439 if (ttop == DIF_DTR_NREGS) {
5440 *flags |= CPU_DTRACE_TUPOFLOW;
5441 break;
5442 }
5443
5444 tupregs[ttop].dttk_value = regs[rd];
5445 tupregs[ttop++].dttk_size = 0;
5446 break;
5447
5448 case DIF_OP_POPTS:
5449 if (ttop != 0)
5450 ttop--;
5451 break;
5452
5453 case DIF_OP_FLUSHTS:
5454 ttop = 0;
5455 break;
5456
5457 case DIF_OP_LDGAA:
5458 case DIF_OP_LDTAA: {
5459 dtrace_dynvar_t *dvar;
5460 dtrace_key_t *key = tupregs;
5461 uint_t nkeys = ttop;
5462
5463 id = DIF_INSTR_VAR(instr);
5464 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5465 id -= DIF_VAR_OTHER_UBASE;
5466
5467 key[nkeys].dttk_value = (uint64_t)id;
5468 key[nkeys++].dttk_size = 0;
5469
5470 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5471 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5472 key[nkeys++].dttk_size = 0;
5473 v = &vstate->dtvs_tlocals[id];
5474 } else {
5475 v = &vstate->dtvs_globals[id]->dtsv_var;
5476 }
5477
5478 dvar = dtrace_dynvar(dstate, nkeys, key,
5479 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5480 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5481 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5482
5483 if (dvar == NULL) {
5484 regs[rd] = 0;
5485 break;
5486 }
5487
5488 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5489 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5490 } else {
5491 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5492 }
5493
5494 break;
5495 }
5496
5497 case DIF_OP_STGAA:
5498 case DIF_OP_STTAA: {
5499 dtrace_dynvar_t *dvar;
5500 dtrace_key_t *key = tupregs;
5501 uint_t nkeys = ttop;
5502
5503 id = DIF_INSTR_VAR(instr);
5504 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5505 id -= DIF_VAR_OTHER_UBASE;
5506
5507 key[nkeys].dttk_value = (uint64_t)id;
5508 key[nkeys++].dttk_size = 0;
5509
5510 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5511 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5512 key[nkeys++].dttk_size = 0;
5513 v = &vstate->dtvs_tlocals[id];
5514 } else {
5515 v = &vstate->dtvs_globals[id]->dtsv_var;
5516 }
5517
5518 dvar = dtrace_dynvar(dstate, nkeys, key,
5519 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5520 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5521 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5522 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5523
5524 if (dvar == NULL)
5525 break;
5526
5527 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5528 if (!dtrace_vcanload(
5529 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5530 mstate, vstate))
5531 break;
5532
5533 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5534 dvar->dtdv_data, &v->dtdv_type);
5535 } else {
5536 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5537 }
5538
5539 break;
5540 }
5541
5542 case DIF_OP_ALLOCS: {
5543 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5544 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5545
5546 /*
5547 * Rounding up the user allocation size could have
5548 * overflowed large, bogus allocations (like -1ULL) to
5549 * 0.
5550 */
5551 if (size < regs[r1] ||
5552 !DTRACE_INSCRATCH(mstate, size)) {
5553 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5554 regs[rd] = NULL;
5555 break;
5556 }
5557
5558 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5559 mstate->dtms_scratch_ptr += size;
5560 regs[rd] = ptr;
5561 break;
5562 }
5563
5564 case DIF_OP_COPYS:
5565 if (!dtrace_canstore(regs[rd], regs[r2],
5566 mstate, vstate)) {
5567 *flags |= CPU_DTRACE_BADADDR;
5568 *illval = regs[rd];
5569 break;
5570 }
5571
5572 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5573 break;
5574
5575 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5576 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5577 break;
5578
5579 case DIF_OP_STB:
5580 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5581 *flags |= CPU_DTRACE_BADADDR;
5582 *illval = regs[rd];
5583 break;
5584 }
5585 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5586 break;
5587
5588 case DIF_OP_STH:
5589 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5590 *flags |= CPU_DTRACE_BADADDR;
5591 *illval = regs[rd];
5592 break;
5593 }
5594 if (regs[rd] & 1) {
5595 *flags |= CPU_DTRACE_BADALIGN;
5596 *illval = regs[rd];
5597 break;
5598 }
5599 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5600 break;
5601
5602 case DIF_OP_STW:
5603 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5604 *flags |= CPU_DTRACE_BADADDR;
5605 *illval = regs[rd];
5606 break;
5607 }
5608 if (regs[rd] & 3) {
5609 *flags |= CPU_DTRACE_BADALIGN;
5610 *illval = regs[rd];
5611 break;
5612 }
5613 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5614 break;
5615
5616 case DIF_OP_STX:
5617 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5618 *flags |= CPU_DTRACE_BADADDR;
5619 *illval = regs[rd];
5620 break;
5621 }
5622 if (regs[rd] & 7) {
5623 *flags |= CPU_DTRACE_BADALIGN;
5624 *illval = regs[rd];
5625 break;
5626 }
5627 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5628 break;
5629 }
5630 }
5631
5632 if (!(*flags & CPU_DTRACE_FAULT))
5633 return (rval);
5634
5635 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5636 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5637
5638 return (0);
5639 }
5640
5641 static void
5642 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5643 {
5644 dtrace_probe_t *probe = ecb->dte_probe;
5645 dtrace_provider_t *prov = probe->dtpr_provider;
5646 char c[DTRACE_FULLNAMELEN + 80], *str;
5647 char *msg = "dtrace: breakpoint action at probe ";
5648 char *ecbmsg = " (ecb ";
5649 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5650 uintptr_t val = (uintptr_t)ecb;
5651 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5652
5653 if (dtrace_destructive_disallow)
5654 return;
5655
5656 /*
5657 * It's impossible to be taking action on the NULL probe.
5658 */
5659 ASSERT(probe != NULL);
5660
5661 /*
5662 * This is a poor man's (destitute man's?) sprintf(): we want to
5663 * print the provider name, module name, function name and name of
5664 * the probe, along with the hex address of the ECB with the breakpoint
5665 * action -- all of which we must place in the character buffer by
5666 * hand.
5667 */
5668 while (*msg != '\0')
5669 c[i++] = *msg++;
5670
5671 for (str = prov->dtpv_name; *str != '\0'; str++)
5672 c[i++] = *str;
5673 c[i++] = ':';
5674
5675 for (str = probe->dtpr_mod; *str != '\0'; str++)
5676 c[i++] = *str;
5677 c[i++] = ':';
5678
5679 for (str = probe->dtpr_func; *str != '\0'; str++)
5680 c[i++] = *str;
5681 c[i++] = ':';
5682
5683 for (str = probe->dtpr_name; *str != '\0'; str++)
5684 c[i++] = *str;
5685
5686 while (*ecbmsg != '\0')
5687 c[i++] = *ecbmsg++;
5688
5689 while (shift >= 0) {
5690 mask = (uintptr_t)0xf << shift;
5691
5692 if (val >= ((uintptr_t)1 << shift))
5693 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5694 shift -= 4;
5695 }
5696
5697 c[i++] = ')';
5698 c[i] = '\0';
5699
5700 debug_enter(c);
5701 }
5702
5703 static void
5704 dtrace_action_panic(dtrace_ecb_t *ecb)
5705 {
5706 dtrace_probe_t *probe = ecb->dte_probe;
5707
5708 /*
5709 * It's impossible to be taking action on the NULL probe.
5710 */
5711 ASSERT(probe != NULL);
5712
5713 if (dtrace_destructive_disallow)
5714 return;
5715
5716 if (dtrace_panicked != NULL)
5717 return;
5718
5719 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5720 return;
5721
5722 /*
5723 * We won the right to panic. (We want to be sure that only one
5724 * thread calls panic() from dtrace_probe(), and that panic() is
5725 * called exactly once.)
5726 */
5727 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5728 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5729 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5730 }
5731
5732 static void
5733 dtrace_action_raise(uint64_t sig)
5734 {
5735 if (dtrace_destructive_disallow)
5736 return;
5737
5738 if (sig >= NSIG) {
5739 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5740 return;
5741 }
5742
5743 /*
5744 * raise() has a queue depth of 1 -- we ignore all subsequent
5745 * invocations of the raise() action.
5746 */
5747 if (curthread->t_dtrace_sig == 0)
5748 curthread->t_dtrace_sig = (uint8_t)sig;
5749
5750 curthread->t_sig_check = 1;
5751 aston(curthread);
5752 }
5753
5754 static void
5755 dtrace_action_stop(void)
5756 {
5757 if (dtrace_destructive_disallow)
5758 return;
5759
5760 if (!curthread->t_dtrace_stop) {
5761 curthread->t_dtrace_stop = 1;
5762 curthread->t_sig_check = 1;
5763 aston(curthread);
5764 }
5765 }
5766
5767 static void
5768 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5769 {
5770 hrtime_t now;
5771 volatile uint16_t *flags;
5772 cpu_t *cpu = CPU;
5773
5774 if (dtrace_destructive_disallow)
5775 return;
5776
5777 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5778
5779 now = dtrace_gethrtime();
5780
5781 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5782 /*
5783 * We need to advance the mark to the current time.
5784 */
5785 cpu->cpu_dtrace_chillmark = now;
5786 cpu->cpu_dtrace_chilled = 0;
5787 }
5788
5789 /*
5790 * Now check to see if the requested chill time would take us over
5791 * the maximum amount of time allowed in the chill interval. (Or
5792 * worse, if the calculation itself induces overflow.)
5793 */
5794 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5795 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5796 *flags |= CPU_DTRACE_ILLOP;
5797 return;
5798 }
5799
5800 while (dtrace_gethrtime() - now < val)
5801 continue;
5802
5803 /*
5804 * Normally, we assure that the value of the variable "timestamp" does
5805 * not change within an ECB. The presence of chill() represents an
5806 * exception to this rule, however.
5807 */
5808 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5809 cpu->cpu_dtrace_chilled += val;
5810 }
5811
5812 static void
5813 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5814 uint64_t *buf, uint64_t arg)
5815 {
5816 int nframes = DTRACE_USTACK_NFRAMES(arg);
5817 int strsize = DTRACE_USTACK_STRSIZE(arg);
5818 uint64_t *pcs = &buf[1], *fps;
5819 char *str = (char *)&pcs[nframes];
5820 int size, offs = 0, i, j;
5821 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5822 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5823 char *sym;
5824
5825 /*
5826 * Should be taking a faster path if string space has not been
5827 * allocated.
5828 */
5829 ASSERT(strsize != 0);
5830
5831 /*
5832 * We will first allocate some temporary space for the frame pointers.
5833 */
5834 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5835 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5836 (nframes * sizeof (uint64_t));
5837
5838 if (!DTRACE_INSCRATCH(mstate, size)) {
5839 /*
5840 * Not enough room for our frame pointers -- need to indicate
5841 * that we ran out of scratch space.
5842 */
5843 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5844 return;
5845 }
5846
5847 mstate->dtms_scratch_ptr += size;
5848 saved = mstate->dtms_scratch_ptr;
5849
5850 /*
5851 * Now get a stack with both program counters and frame pointers.
5852 */
5853 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5854 dtrace_getufpstack(buf, fps, nframes + 1);
5855 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5856
5857 /*
5858 * If that faulted, we're cooked.
5859 */
5860 if (*flags & CPU_DTRACE_FAULT)
5861 goto out;
5862
5863 /*
5864 * Now we want to walk up the stack, calling the USTACK helper. For
5865 * each iteration, we restore the scratch pointer.
5866 */
5867 for (i = 0; i < nframes; i++) {
5868 mstate->dtms_scratch_ptr = saved;
5869
5870 if (offs >= strsize)
5871 break;
5872
5873 sym = (char *)(uintptr_t)dtrace_helper(
5874 DTRACE_HELPER_ACTION_USTACK,
5875 mstate, state, pcs[i], fps[i]);
5876
5877 /*
5878 * If we faulted while running the helper, we're going to
5879 * clear the fault and null out the corresponding string.
5880 */
5881 if (*flags & CPU_DTRACE_FAULT) {
5882 *flags &= ~CPU_DTRACE_FAULT;
5883 str[offs++] = '\0';
5884 continue;
5885 }
5886
5887 if (sym == NULL) {
5888 str[offs++] = '\0';
5889 continue;
5890 }
5891
5892 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5893
5894 /*
5895 * Now copy in the string that the helper returned to us.
5896 */
5897 for (j = 0; offs + j < strsize; j++) {
5898 if ((str[offs + j] = sym[j]) == '\0')
5899 break;
5900 }
5901
5902 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5903
5904 offs += j + 1;
5905 }
5906
5907 if (offs >= strsize) {
5908 /*
5909 * If we didn't have room for all of the strings, we don't
5910 * abort processing -- this needn't be a fatal error -- but we
5911 * still want to increment a counter (dts_stkstroverflows) to
5912 * allow this condition to be warned about. (If this is from
5913 * a jstack() action, it is easily tuned via jstackstrsize.)
5914 */
5915 dtrace_error(&state->dts_stkstroverflows);
5916 }
5917
5918 while (offs < strsize)
5919 str[offs++] = '\0';
5920
5921 out:
5922 mstate->dtms_scratch_ptr = old;
5923 }
5924
5925 /*
5926 * If you're looking for the epicenter of DTrace, you just found it. This
5927 * is the function called by the provider to fire a probe -- from which all
5928 * subsequent probe-context DTrace activity emanates.
5929 */
5930 void
5931 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5932 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5933 {
5934 processorid_t cpuid;
5935 dtrace_icookie_t cookie;
5936 dtrace_probe_t *probe;
5937 dtrace_mstate_t mstate;
5938 dtrace_ecb_t *ecb;
5939 dtrace_action_t *act;
5940 intptr_t offs;
5941 size_t size;
5942 int vtime, onintr;
5943 volatile uint16_t *flags;
5944 hrtime_t now;
5945
5946 /*
5947 * Kick out immediately if this CPU is still being born (in which case
5948 * curthread will be set to -1) or the current thread can't allow
5949 * probes in its current context.
5950 */
5951 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
5952 return;
5953
5954 cookie = dtrace_interrupt_disable();
5955 probe = dtrace_probes[id - 1];
5956 cpuid = CPU->cpu_id;
5957 onintr = CPU_ON_INTR(CPU);
5958
5959 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
5960 probe->dtpr_predcache == curthread->t_predcache) {
5961 /*
5962 * We have hit in the predicate cache; we know that
5963 * this predicate would evaluate to be false.
5964 */
5965 dtrace_interrupt_enable(cookie);
5966 return;
5967 }
5968
5969 if (panic_quiesce) {
5970 /*
5971 * We don't trace anything if we're panicking.
5972 */
5973 dtrace_interrupt_enable(cookie);
5974 return;
5975 }
5976
5977 now = dtrace_gethrtime();
5978 vtime = dtrace_vtime_references != 0;
5979
5980 if (vtime && curthread->t_dtrace_start)
5981 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
5982
5983 mstate.dtms_difo = NULL;
5984 mstate.dtms_probe = probe;
5985 mstate.dtms_strtok = NULL;
5986 mstate.dtms_arg[0] = arg0;
5987 mstate.dtms_arg[1] = arg1;
5988 mstate.dtms_arg[2] = arg2;
5989 mstate.dtms_arg[3] = arg3;
5990 mstate.dtms_arg[4] = arg4;
5991
5992 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
5993
5994 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
5995 dtrace_predicate_t *pred = ecb->dte_predicate;
5996 dtrace_state_t *state = ecb->dte_state;
5997 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
5998 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
5999 dtrace_vstate_t *vstate = &state->dts_vstate;
6000 dtrace_provider_t *prov = probe->dtpr_provider;
6001 uint64_t tracememsize = 0;
6002 int committed = 0;
6003 caddr_t tomax;
6004
6005 /*
6006 * A little subtlety with the following (seemingly innocuous)
6007 * declaration of the automatic 'val': by looking at the
6008 * code, you might think that it could be declared in the
6009 * action processing loop, below. (That is, it's only used in
6010 * the action processing loop.) However, it must be declared
6011 * out of that scope because in the case of DIF expression
6012 * arguments to aggregating actions, one iteration of the
6013 * action loop will use the last iteration's value.
6014 */
6015 #ifdef lint
6016 uint64_t val = 0;
6017 #else
6018 uint64_t val;
6019 #endif
6020
6021 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6022 mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
6023 mstate.dtms_getf = NULL;
6024
6025 *flags &= ~CPU_DTRACE_ERROR;
6026
6027 if (prov == dtrace_provider) {
6028 /*
6029 * If dtrace itself is the provider of this probe,
6030 * we're only going to continue processing the ECB if
6031 * arg0 (the dtrace_state_t) is equal to the ECB's
6032 * creating state. (This prevents disjoint consumers
6033 * from seeing one another's metaprobes.)
6034 */
6035 if (arg0 != (uint64_t)(uintptr_t)state)
6036 continue;
6037 }
6038
6039 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6040 /*
6041 * We're not currently active. If our provider isn't
6042 * the dtrace pseudo provider, we're not interested.
6043 */
6044 if (prov != dtrace_provider)
6045 continue;
6046
6047 /*
6048 * Now we must further check if we are in the BEGIN
6049 * probe. If we are, we will only continue processing
6050 * if we're still in WARMUP -- if one BEGIN enabling
6051 * has invoked the exit() action, we don't want to
6052 * evaluate subsequent BEGIN enablings.
6053 */
6054 if (probe->dtpr_id == dtrace_probeid_begin &&
6055 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6056 ASSERT(state->dts_activity ==
6057 DTRACE_ACTIVITY_DRAINING);
6058 continue;
6059 }
6060 }
6061
6062 if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
6063 continue;
6064
6065 if (now - state->dts_alive > dtrace_deadman_timeout) {
6066 /*
6067 * We seem to be dead. Unless we (a) have kernel
6068 * destructive permissions (b) have expicitly enabled
6069 * destructive actions and (c) destructive actions have
6070 * not been disabled, we're going to transition into
6071 * the KILLED state, from which no further processing
6072 * on this state will be performed.
6073 */
6074 if (!dtrace_priv_kernel_destructive(state) ||
6075 !state->dts_cred.dcr_destructive ||
6076 dtrace_destructive_disallow) {
6077 void *activity = &state->dts_activity;
6078 dtrace_activity_t current;
6079
6080 do {
6081 current = state->dts_activity;
6082 } while (dtrace_cas32(activity, current,
6083 DTRACE_ACTIVITY_KILLED) != current);
6084
6085 continue;
6086 }
6087 }
6088
6089 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6090 ecb->dte_alignment, state, &mstate)) < 0)
6091 continue;
6092
6093 tomax = buf->dtb_tomax;
6094 ASSERT(tomax != NULL);
6095
6096 if (ecb->dte_size != 0)
6097 DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
6098
6099 mstate.dtms_epid = ecb->dte_epid;
6100 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6101
6102 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6103 mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
6104
6105 if (pred != NULL) {
6106 dtrace_difo_t *dp = pred->dtp_difo;
6107 int rval;
6108
6109 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6110
6111 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6112 dtrace_cacheid_t cid = probe->dtpr_predcache;
6113
6114 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6115 /*
6116 * Update the predicate cache...
6117 */
6118 ASSERT(cid == pred->dtp_cacheid);
6119 curthread->t_predcache = cid;
6120 }
6121
6122 continue;
6123 }
6124 }
6125
6126 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6127 act != NULL; act = act->dta_next) {
6128 size_t valoffs;
6129 dtrace_difo_t *dp;
6130 dtrace_recdesc_t *rec = &act->dta_rec;
6131
6132 size = rec->dtrd_size;
6133 valoffs = offs + rec->dtrd_offset;
6134
6135 if (DTRACEACT_ISAGG(act->dta_kind)) {
6136 uint64_t v = 0xbad;
6137 dtrace_aggregation_t *agg;
6138
6139 agg = (dtrace_aggregation_t *)act;
6140
6141 if ((dp = act->dta_difo) != NULL)
6142 v = dtrace_dif_emulate(dp,
6143 &mstate, vstate, state);
6144
6145 if (*flags & CPU_DTRACE_ERROR)
6146 continue;
6147
6148 /*
6149 * Note that we always pass the expression
6150 * value from the previous iteration of the
6151 * action loop. This value will only be used
6152 * if there is an expression argument to the
6153 * aggregating action, denoted by the
6154 * dtag_hasarg field.
6155 */
6156 dtrace_aggregate(agg, buf,
6157 offs, aggbuf, v, val);
6158 continue;
6159 }
6160
6161 switch (act->dta_kind) {
6162 case DTRACEACT_STOP:
6163 if (dtrace_priv_proc_destructive(state,
6164 &mstate))
6165 dtrace_action_stop();
6166 continue;
6167
6168 case DTRACEACT_BREAKPOINT:
6169 if (dtrace_priv_kernel_destructive(state))
6170 dtrace_action_breakpoint(ecb);
6171 continue;
6172
6173 case DTRACEACT_PANIC:
6174 if (dtrace_priv_kernel_destructive(state))
6175 dtrace_action_panic(ecb);
6176 continue;
6177
6178 case DTRACEACT_STACK:
6179 if (!dtrace_priv_kernel(state))
6180 continue;
6181
6182 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6183 size / sizeof (pc_t), probe->dtpr_aframes,
6184 DTRACE_ANCHORED(probe) ? NULL :
6185 (uint32_t *)arg0);
6186
6187 continue;
6188
6189 case DTRACEACT_JSTACK:
6190 case DTRACEACT_USTACK:
6191 if (!dtrace_priv_proc(state, &mstate))
6192 continue;
6193
6194 /*
6195 * See comment in DIF_VAR_PID.
6196 */
6197 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6198 CPU_ON_INTR(CPU)) {
6199 int depth = DTRACE_USTACK_NFRAMES(
6200 rec->dtrd_arg) + 1;
6201
6202 dtrace_bzero((void *)(tomax + valoffs),
6203 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6204 + depth * sizeof (uint64_t));
6205
6206 continue;
6207 }
6208
6209 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6210 curproc->p_dtrace_helpers != NULL) {
6211 /*
6212 * This is the slow path -- we have
6213 * allocated string space, and we're
6214 * getting the stack of a process that
6215 * has helpers. Call into a separate
6216 * routine to perform this processing.
6217 */
6218 dtrace_action_ustack(&mstate, state,
6219 (uint64_t *)(tomax + valoffs),
6220 rec->dtrd_arg);
6221 continue;
6222 }
6223
6224 /*
6225 * Clear the string space, since there's no
6226 * helper to do it for us.
6227 */
6228 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
6229 int depth = DTRACE_USTACK_NFRAMES(
6230 rec->dtrd_arg);
6231 size_t strsize = DTRACE_USTACK_STRSIZE(
6232 rec->dtrd_arg);
6233 uint64_t *buf = (uint64_t *)(tomax +
6234 valoffs);
6235 void *strspace = &buf[depth + 1];
6236
6237 dtrace_bzero(strspace,
6238 MIN(depth, strsize));
6239 }
6240
6241 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6242 dtrace_getupcstack((uint64_t *)
6243 (tomax + valoffs),
6244 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6245 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6246 continue;
6247
6248 default:
6249 break;
6250 }
6251
6252 dp = act->dta_difo;
6253 ASSERT(dp != NULL);
6254
6255 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6256
6257 if (*flags & CPU_DTRACE_ERROR)
6258 continue;
6259
6260 switch (act->dta_kind) {
6261 case DTRACEACT_SPECULATE:
6262 ASSERT(buf == &state->dts_buffer[cpuid]);
6263 buf = dtrace_speculation_buffer(state,
6264 cpuid, val);
6265
6266 if (buf == NULL) {
6267 *flags |= CPU_DTRACE_DROP;
6268 continue;
6269 }
6270
6271 offs = dtrace_buffer_reserve(buf,
6272 ecb->dte_needed, ecb->dte_alignment,
6273 state, NULL);
6274
6275 if (offs < 0) {
6276 *flags |= CPU_DTRACE_DROP;
6277 continue;
6278 }
6279
6280 tomax = buf->dtb_tomax;
6281 ASSERT(tomax != NULL);
6282
6283 if (ecb->dte_size != 0)
6284 DTRACE_STORE(uint32_t, tomax, offs,
6285 ecb->dte_epid);
6286 continue;
6287
6288 case DTRACEACT_CHILL:
6289 if (dtrace_priv_kernel_destructive(state))
6290 dtrace_action_chill(&mstate, val);
6291 continue;
6292
6293 case DTRACEACT_RAISE:
6294 if (dtrace_priv_proc_destructive(state,
6295 &mstate))
6296 dtrace_action_raise(val);
6297 continue;
6298
6299 case DTRACEACT_COMMIT:
6300 ASSERT(!committed);
6301
6302 /*
6303 * We need to commit our buffer state.
6304 */
6305 if (ecb->dte_size)
6306 buf->dtb_offset = offs + ecb->dte_size;
6307 buf = &state->dts_buffer[cpuid];
6308 dtrace_speculation_commit(state, cpuid, val);
6309 committed = 1;
6310 continue;
6311
6312 case DTRACEACT_DISCARD:
6313 dtrace_speculation_discard(state, cpuid, val);
6314 continue;
6315
6316 case DTRACEACT_DIFEXPR:
6317 case DTRACEACT_LIBACT:
6318 case DTRACEACT_PRINTF:
6319 case DTRACEACT_PRINTA:
6320 case DTRACEACT_SYSTEM:
6321 case DTRACEACT_FREOPEN:
6322 case DTRACEACT_TRACEMEM:
6323 break;
6324
6325 case DTRACEACT_TRACEMEM_DYNSIZE:
6326 tracememsize = val;
6327 break;
6328
6329 case DTRACEACT_SYM:
6330 case DTRACEACT_MOD:
6331 if (!dtrace_priv_kernel(state))
6332 continue;
6333 break;
6334
6335 case DTRACEACT_USYM:
6336 case DTRACEACT_UMOD:
6337 case DTRACEACT_UADDR: {
6338 struct pid *pid = curthread->t_procp->p_pidp;
6339
6340 if (!dtrace_priv_proc(state, &mstate))
6341 continue;
6342
6343 DTRACE_STORE(uint64_t, tomax,
6344 valoffs, (uint64_t)pid->pid_id);
6345 DTRACE_STORE(uint64_t, tomax,
6346 valoffs + sizeof (uint64_t), val);
6347
6348 continue;
6349 }
6350
6351 case DTRACEACT_EXIT: {
6352 /*
6353 * For the exit action, we are going to attempt
6354 * to atomically set our activity to be
6355 * draining. If this fails (either because
6356 * another CPU has beat us to the exit action,
6357 * or because our current activity is something
6358 * other than ACTIVE or WARMUP), we will
6359 * continue. This assures that the exit action
6360 * can be successfully recorded at most once
6361 * when we're in the ACTIVE state. If we're
6362 * encountering the exit() action while in
6363 * COOLDOWN, however, we want to honor the new
6364 * status code. (We know that we're the only
6365 * thread in COOLDOWN, so there is no race.)
6366 */
6367 void *activity = &state->dts_activity;
6368 dtrace_activity_t current = state->dts_activity;
6369
6370 if (current == DTRACE_ACTIVITY_COOLDOWN)
6371 break;
6372
6373 if (current != DTRACE_ACTIVITY_WARMUP)
6374 current = DTRACE_ACTIVITY_ACTIVE;
6375
6376 if (dtrace_cas32(activity, current,
6377 DTRACE_ACTIVITY_DRAINING) != current) {
6378 *flags |= CPU_DTRACE_DROP;
6379 continue;
6380 }
6381
6382 break;
6383 }
6384
6385 default:
6386 ASSERT(0);
6387 }
6388
6389 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6390 uintptr_t end = valoffs + size;
6391
6392 if (tracememsize != 0 &&
6393 valoffs + tracememsize < end) {
6394 end = valoffs + tracememsize;
6395 tracememsize = 0;
6396 }
6397
6398 if (!dtrace_vcanload((void *)(uintptr_t)val,
6399 &dp->dtdo_rtype, &mstate, vstate))
6400 continue;
6401
6402 /*
6403 * If this is a string, we're going to only
6404 * load until we find the zero byte -- after
6405 * which we'll store zero bytes.
6406 */
6407 if (dp->dtdo_rtype.dtdt_kind ==
6408 DIF_TYPE_STRING) {
6409 char c = '\0' + 1;
6410 int intuple = act->dta_intuple;
6411 size_t s;
6412
6413 for (s = 0; s < size; s++) {
6414 if (c != '\0')
6415 c = dtrace_load8(val++);
6416
6417 DTRACE_STORE(uint8_t, tomax,
6418 valoffs++, c);
6419
6420 if (c == '\0' && intuple)
6421 break;
6422 }
6423
6424 continue;
6425 }
6426
6427 while (valoffs < end) {
6428 DTRACE_STORE(uint8_t, tomax, valoffs++,
6429 dtrace_load8(val++));
6430 }
6431
6432 continue;
6433 }
6434
6435 switch (size) {
6436 case 0:
6437 break;
6438
6439 case sizeof (uint8_t):
6440 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6441 break;
6442 case sizeof (uint16_t):
6443 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6444 break;
6445 case sizeof (uint32_t):
6446 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6447 break;
6448 case sizeof (uint64_t):
6449 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6450 break;
6451 default:
6452 /*
6453 * Any other size should have been returned by
6454 * reference, not by value.
6455 */
6456 ASSERT(0);
6457 break;
6458 }
6459 }
6460
6461 if (*flags & CPU_DTRACE_DROP)
6462 continue;
6463
6464 if (*flags & CPU_DTRACE_FAULT) {
6465 int ndx;
6466 dtrace_action_t *err;
6467
6468 buf->dtb_errors++;
6469
6470 if (probe->dtpr_id == dtrace_probeid_error) {
6471 /*
6472 * There's nothing we can do -- we had an
6473 * error on the error probe. We bump an
6474 * error counter to at least indicate that
6475 * this condition happened.
6476 */
6477 dtrace_error(&state->dts_dblerrors);
6478 continue;
6479 }
6480
6481 if (vtime) {
6482 /*
6483 * Before recursing on dtrace_probe(), we
6484 * need to explicitly clear out our start
6485 * time to prevent it from being accumulated
6486 * into t_dtrace_vtime.
6487 */
6488 curthread->t_dtrace_start = 0;
6489 }
6490
6491 /*
6492 * Iterate over the actions to figure out which action
6493 * we were processing when we experienced the error.
6494 * Note that act points _past_ the faulting action; if
6495 * act is ecb->dte_action, the fault was in the
6496 * predicate, if it's ecb->dte_action->dta_next it's
6497 * in action #1, and so on.
6498 */
6499 for (err = ecb->dte_action, ndx = 0;
6500 err != act; err = err->dta_next, ndx++)
6501 continue;
6502
6503 dtrace_probe_error(state, ecb->dte_epid, ndx,
6504 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6505 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6506 cpu_core[cpuid].cpuc_dtrace_illval);
6507
6508 continue;
6509 }
6510
6511 if (!committed)
6512 buf->dtb_offset = offs + ecb->dte_size;
6513 }
6514
6515 if (vtime)
6516 curthread->t_dtrace_start = dtrace_gethrtime();
6517
6518 dtrace_interrupt_enable(cookie);
6519 }
6520
6521 /*
6522 * DTrace Probe Hashing Functions
6523 *
6524 * The functions in this section (and indeed, the functions in remaining
6525 * sections) are not _called_ from probe context. (Any exceptions to this are
6526 * marked with a "Note:".) Rather, they are called from elsewhere in the
6527 * DTrace framework to look-up probes in, add probes to and remove probes from
6528 * the DTrace probe hashes. (Each probe is hashed by each element of the
6529 * probe tuple -- allowing for fast lookups, regardless of what was
6530 * specified.)
6531 */
6532 static uint_t
6533 dtrace_hash_str(char *p)
6534 {
6535 unsigned int g;
6536 uint_t hval = 0;
6537
6538 while (*p) {
6539 hval = (hval << 4) + *p++;
6540 if ((g = (hval & 0xf0000000)) != 0)
6541 hval ^= g >> 24;
6542 hval &= ~g;
6543 }
6544 return (hval);
6545 }
6546
6547 static dtrace_hash_t *
6548 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6549 {
6550 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6551
6552 hash->dth_stroffs = stroffs;
6553 hash->dth_nextoffs = nextoffs;
6554 hash->dth_prevoffs = prevoffs;
6555
6556 hash->dth_size = 1;
6557 hash->dth_mask = hash->dth_size - 1;
6558
6559 hash->dth_tab = kmem_zalloc(hash->dth_size *
6560 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6561
6562 return (hash);
6563 }
6564
6565 static void
6566 dtrace_hash_destroy(dtrace_hash_t *hash)
6567 {
6568 #ifdef DEBUG
6569 int i;
6570
6571 for (i = 0; i < hash->dth_size; i++)
6572 ASSERT(hash->dth_tab[i] == NULL);
6573 #endif
6574
6575 kmem_free(hash->dth_tab,
6576 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6577 kmem_free(hash, sizeof (dtrace_hash_t));
6578 }
6579
6580 static void
6581 dtrace_hash_resize(dtrace_hash_t *hash)
6582 {
6583 int size = hash->dth_size, i, ndx;
6584 int new_size = hash->dth_size << 1;
6585 int new_mask = new_size - 1;
6586 dtrace_hashbucket_t **new_tab, *bucket, *next;
6587
6588 ASSERT((new_size & new_mask) == 0);
6589
6590 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6591
6592 for (i = 0; i < size; i++) {
6593 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6594 dtrace_probe_t *probe = bucket->dthb_chain;
6595
6596 ASSERT(probe != NULL);
6597 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6598
6599 next = bucket->dthb_next;
6600 bucket->dthb_next = new_tab[ndx];
6601 new_tab[ndx] = bucket;
6602 }
6603 }
6604
6605 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6606 hash->dth_tab = new_tab;
6607 hash->dth_size = new_size;
6608 hash->dth_mask = new_mask;
6609 }
6610
6611 static void
6612 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6613 {
6614 int hashval = DTRACE_HASHSTR(hash, new);
6615 int ndx = hashval & hash->dth_mask;
6616 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6617 dtrace_probe_t **nextp, **prevp;
6618
6619 for (; bucket != NULL; bucket = bucket->dthb_next) {
6620 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6621 goto add;
6622 }
6623
6624 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6625 dtrace_hash_resize(hash);
6626 dtrace_hash_add(hash, new);
6627 return;
6628 }
6629
6630 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6631 bucket->dthb_next = hash->dth_tab[ndx];
6632 hash->dth_tab[ndx] = bucket;
6633 hash->dth_nbuckets++;
6634
6635 add:
6636 nextp = DTRACE_HASHNEXT(hash, new);
6637 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6638 *nextp = bucket->dthb_chain;
6639
6640 if (bucket->dthb_chain != NULL) {
6641 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6642 ASSERT(*prevp == NULL);
6643 *prevp = new;
6644 }
6645
6646 bucket->dthb_chain = new;
6647 bucket->dthb_len++;
6648 }
6649
6650 static dtrace_probe_t *
6651 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6652 {
6653 int hashval = DTRACE_HASHSTR(hash, template);
6654 int ndx = hashval & hash->dth_mask;
6655 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6656
6657 for (; bucket != NULL; bucket = bucket->dthb_next) {
6658 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6659 return (bucket->dthb_chain);
6660 }
6661
6662 return (NULL);
6663 }
6664
6665 static int
6666 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6667 {
6668 int hashval = DTRACE_HASHSTR(hash, template);
6669 int ndx = hashval & hash->dth_mask;
6670 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6671
6672 for (; bucket != NULL; bucket = bucket->dthb_next) {
6673 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6674 return (bucket->dthb_len);
6675 }
6676
6677 return (NULL);
6678 }
6679
6680 static void
6681 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6682 {
6683 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6684 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6685
6686 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6687 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6688
6689 /*
6690 * Find the bucket that we're removing this probe from.
6691 */
6692 for (; bucket != NULL; bucket = bucket->dthb_next) {
6693 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6694 break;
6695 }
6696
6697 ASSERT(bucket != NULL);
6698
6699 if (*prevp == NULL) {
6700 if (*nextp == NULL) {
6701 /*
6702 * The removed probe was the only probe on this
6703 * bucket; we need to remove the bucket.
6704 */
6705 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6706
6707 ASSERT(bucket->dthb_chain == probe);
6708 ASSERT(b != NULL);
6709
6710 if (b == bucket) {
6711 hash->dth_tab[ndx] = bucket->dthb_next;
6712 } else {
6713 while (b->dthb_next != bucket)
6714 b = b->dthb_next;
6715 b->dthb_next = bucket->dthb_next;
6716 }
6717
6718 ASSERT(hash->dth_nbuckets > 0);
6719 hash->dth_nbuckets--;
6720 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6721 return;
6722 }
6723
6724 bucket->dthb_chain = *nextp;
6725 } else {
6726 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6727 }
6728
6729 if (*nextp != NULL)
6730 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6731 }
6732
6733 /*
6734 * DTrace Utility Functions
6735 *
6736 * These are random utility functions that are _not_ called from probe context.
6737 */
6738 static int
6739 dtrace_badattr(const dtrace_attribute_t *a)
6740 {
6741 return (a->dtat_name > DTRACE_STABILITY_MAX ||
6742 a->dtat_data > DTRACE_STABILITY_MAX ||
6743 a->dtat_class > DTRACE_CLASS_MAX);
6744 }
6745
6746 /*
6747 * Return a duplicate copy of a string. If the specified string is NULL,
6748 * this function returns a zero-length string.
6749 */
6750 static char *
6751 dtrace_strdup(const char *str)
6752 {
6753 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6754
6755 if (str != NULL)
6756 (void) strcpy(new, str);
6757
6758 return (new);
6759 }
6760
6761 #define DTRACE_ISALPHA(c) \
6762 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6763
6764 static int
6765 dtrace_badname(const char *s)
6766 {
6767 char c;
6768
6769 if (s == NULL || (c = *s++) == '\0')
6770 return (0);
6771
6772 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6773 return (1);
6774
6775 while ((c = *s++) != '\0') {
6776 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6777 c != '-' && c != '_' && c != '.' && c != '`')
6778 return (1);
6779 }
6780
6781 return (0);
6782 }
6783
6784 static void
6785 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6786 {
6787 uint32_t priv;
6788
6789 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6790 /*
6791 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6792 */
6793 priv = DTRACE_PRIV_ALL;
6794 } else {
6795 *uidp = crgetuid(cr);
6796 *zoneidp = crgetzoneid(cr);
6797
6798 priv = 0;
6799 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6800 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6801 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6802 priv |= DTRACE_PRIV_USER;
6803 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6804 priv |= DTRACE_PRIV_PROC;
6805 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6806 priv |= DTRACE_PRIV_OWNER;
6807 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6808 priv |= DTRACE_PRIV_ZONEOWNER;
6809 }
6810
6811 *privp = priv;
6812 }
6813
6814 #ifdef DTRACE_ERRDEBUG
6815 static void
6816 dtrace_errdebug(const char *str)
6817 {
6818 int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
6819 int occupied = 0;
6820
6821 mutex_enter(&dtrace_errlock);
6822 dtrace_errlast = str;
6823 dtrace_errthread = curthread;
6824
6825 while (occupied++ < DTRACE_ERRHASHSZ) {
6826 if (dtrace_errhash[hval].dter_msg == str) {
6827 dtrace_errhash[hval].dter_count++;
6828 goto out;
6829 }
6830
6831 if (dtrace_errhash[hval].dter_msg != NULL) {
6832 hval = (hval + 1) % DTRACE_ERRHASHSZ;
6833 continue;
6834 }
6835
6836 dtrace_errhash[hval].dter_msg = str;
6837 dtrace_errhash[hval].dter_count = 1;
6838 goto out;
6839 }
6840
6841 panic("dtrace: undersized error hash");
6842 out:
6843 mutex_exit(&dtrace_errlock);
6844 }
6845 #endif
6846
6847 /*
6848 * DTrace Matching Functions
6849 *
6850 * These functions are used to match groups of probes, given some elements of
6851 * a probe tuple, or some globbed expressions for elements of a probe tuple.
6852 */
6853 static int
6854 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
6855 zoneid_t zoneid)
6856 {
6857 if (priv != DTRACE_PRIV_ALL) {
6858 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
6859 uint32_t match = priv & ppriv;
6860
6861 /*
6862 * No PRIV_DTRACE_* privileges...
6863 */
6864 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
6865 DTRACE_PRIV_KERNEL)) == 0)
6866 return (0);
6867
6868 /*
6869 * No matching bits, but there were bits to match...
6870 */
6871 if (match == 0 && ppriv != 0)
6872 return (0);
6873
6874 /*
6875 * Need to have permissions to the process, but don't...
6876 */
6877 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
6878 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
6879 return (0);
6880 }
6881
6882 /*
6883 * Need to be in the same zone unless we possess the
6884 * privilege to examine all zones.
6885 */
6886 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
6887 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
6888 return (0);
6889 }
6890 }
6891
6892 return (1);
6893 }
6894
6895 /*
6896 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
6897 * consists of input pattern strings and an ops-vector to evaluate them.
6898 * This function returns >0 for match, 0 for no match, and <0 for error.
6899 */
6900 static int
6901 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
6902 uint32_t priv, uid_t uid, zoneid_t zoneid)
6903 {
6904 dtrace_provider_t *pvp = prp->dtpr_provider;
6905 int rv;
6906
6907 if (pvp->dtpv_defunct)
6908 return (0);
6909
6910 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
6911 return (rv);
6912
6913 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
6914 return (rv);
6915
6916 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
6917 return (rv);
6918
6919 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
6920 return (rv);
6921
6922 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
6923 return (0);
6924
6925 return (rv);
6926 }
6927
6928 /*
6929 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
6930 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
6931 * libc's version, the kernel version only applies to 8-bit ASCII strings.
6932 * In addition, all of the recursion cases except for '*' matching have been
6933 * unwound. For '*', we still implement recursive evaluation, but a depth
6934 * counter is maintained and matching is aborted if we recurse too deep.
6935 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
6936 */
6937 static int
6938 dtrace_match_glob(const char *s, const char *p, int depth)
6939 {
6940 const char *olds;
6941 char s1, c;
6942 int gs;
6943
6944 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
6945 return (-1);
6946
6947 if (s == NULL)
6948 s = ""; /* treat NULL as empty string */
6949
6950 top:
6951 olds = s;
6952 s1 = *s++;
6953
6954 if (p == NULL)
6955 return (0);
6956
6957 if ((c = *p++) == '\0')
6958 return (s1 == '\0');
6959
6960 switch (c) {
6961 case '[': {
6962 int ok = 0, notflag = 0;
6963 char lc = '\0';
6964
6965 if (s1 == '\0')
6966 return (0);
6967
6968 if (*p == '!') {
6969 notflag = 1;
6970 p++;
6971 }
6972
6973 if ((c = *p++) == '\0')
6974 return (0);
6975
6976 do {
6977 if (c == '-' && lc != '\0' && *p != ']') {
6978 if ((c = *p++) == '\0')
6979 return (0);
6980 if (c == '\\' && (c = *p++) == '\0')
6981 return (0);
6982
6983 if (notflag) {
6984 if (s1 < lc || s1 > c)
6985 ok++;
6986 else
6987 return (0);
6988 } else if (lc <= s1 && s1 <= c)
6989 ok++;
6990
6991 } else if (c == '\\' && (c = *p++) == '\0')
6992 return (0);
6993
6994 lc = c; /* save left-hand 'c' for next iteration */
6995
6996 if (notflag) {
6997 if (s1 != c)
6998 ok++;
6999 else
7000 return (0);
7001 } else if (s1 == c)
7002 ok++;
7003
7004 if ((c = *p++) == '\0')
7005 return (0);
7006
7007 } while (c != ']');
7008
7009 if (ok)
7010 goto top;
7011
7012 return (0);
7013 }
7014
7015 case '\\':
7016 if ((c = *p++) == '\0')
7017 return (0);
7018 /*FALLTHRU*/
7019
7020 default:
7021 if (c != s1)
7022 return (0);
7023 /*FALLTHRU*/
7024
7025 case '?':
7026 if (s1 != '\0')
7027 goto top;
7028 return (0);
7029
7030 case '*':
7031 while (*p == '*')
7032 p++; /* consecutive *'s are identical to a single one */
7033
7034 if (*p == '\0')
7035 return (1);
7036
7037 for (s = olds; *s != '\0'; s++) {
7038 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7039 return (gs);
7040 }
7041
7042 return (0);
7043 }
7044 }
7045
7046 /*ARGSUSED*/
7047 static int
7048 dtrace_match_string(const char *s, const char *p, int depth)
7049 {
7050 return (s != NULL && strcmp(s, p) == 0);
7051 }
7052
7053 /*ARGSUSED*/
7054 static int
7055 dtrace_match_nul(const char *s, const char *p, int depth)
7056 {
7057 return (1); /* always match the empty pattern */
7058 }
7059
7060 /*ARGSUSED*/
7061 static int
7062 dtrace_match_nonzero(const char *s, const char *p, int depth)
7063 {
7064 return (s != NULL && s[0] != '\0');
7065 }
7066
7067 static int
7068 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7069 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7070 {
7071 dtrace_probe_t template, *probe;
7072 dtrace_hash_t *hash = NULL;
7073 int len, rc, best = INT_MAX, nmatched = 0;
7074 dtrace_id_t i;
7075
7076 ASSERT(MUTEX_HELD(&dtrace_lock));
7077
7078 /*
7079 * If the probe ID is specified in the key, just lookup by ID and
7080 * invoke the match callback once if a matching probe is found.
7081 */
7082 if (pkp->dtpk_id != DTRACE_IDNONE) {
7083 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7084 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7085 if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
7086 return (DTRACE_MATCH_FAIL);
7087 nmatched++;
7088 }
7089 return (nmatched);
7090 }
7091
7092 template.dtpr_mod = (char *)pkp->dtpk_mod;
7093 template.dtpr_func = (char *)pkp->dtpk_func;
7094 template.dtpr_name = (char *)pkp->dtpk_name;
7095
7096 /*
7097 * We want to find the most distinct of the module name, function
7098 * name, and name. So for each one that is not a glob pattern or
7099 * empty string, we perform a lookup in the corresponding hash and
7100 * use the hash table with the fewest collisions to do our search.
7101 */
7102 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7103 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7104 best = len;
7105 hash = dtrace_bymod;
7106 }
7107
7108 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7109 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7110 best = len;
7111 hash = dtrace_byfunc;
7112 }
7113
7114 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7115 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7116 best = len;
7117 hash = dtrace_byname;
7118 }
7119
7120 /*
7121 * If we did not select a hash table, iterate over every probe and
7122 * invoke our callback for each one that matches our input probe key.
7123 */
7124 if (hash == NULL) {
7125 for (i = 0; i < dtrace_nprobes; i++) {
7126 if ((probe = dtrace_probes[i]) == NULL ||
7127 dtrace_match_probe(probe, pkp, priv, uid,
7128 zoneid) <= 0)
7129 continue;
7130
7131 nmatched++;
7132
7133 if ((rc = (*matched)(probe, arg)) !=
7134 DTRACE_MATCH_NEXT) {
7135 if (rc == DTRACE_MATCH_FAIL)
7136 return (DTRACE_MATCH_FAIL);
7137 break;
7138 }
7139 }
7140
7141 return (nmatched);
7142 }
7143
7144 /*
7145 * If we selected a hash table, iterate over each probe of the same key
7146 * name and invoke the callback for every probe that matches the other
7147 * attributes of our input probe key.
7148 */
7149 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7150 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7151
7152 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7153 continue;
7154
7155 nmatched++;
7156
7157 if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
7158 if (rc == DTRACE_MATCH_FAIL)
7159 return (DTRACE_MATCH_FAIL);
7160 break;
7161 }
7162 }
7163
7164 return (nmatched);
7165 }
7166
7167 /*
7168 * Return the function pointer dtrace_probecmp() should use to compare the
7169 * specified pattern with a string. For NULL or empty patterns, we select
7170 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7171 * For non-empty non-glob strings, we use dtrace_match_string().
7172 */
7173 static dtrace_probekey_f *
7174 dtrace_probekey_func(const char *p)
7175 {
7176 char c;
7177
7178 if (p == NULL || *p == '\0')
7179 return (&dtrace_match_nul);
7180
7181 while ((c = *p++) != '\0') {
7182 if (c == '[' || c == '?' || c == '*' || c == '\\')
7183 return (&dtrace_match_glob);
7184 }
7185
7186 return (&dtrace_match_string);
7187 }
7188
7189 /*
7190 * Build a probe comparison key for use with dtrace_match_probe() from the
7191 * given probe description. By convention, a null key only matches anchored
7192 * probes: if each field is the empty string, reset dtpk_fmatch to
7193 * dtrace_match_nonzero().
7194 */
7195 static void
7196 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7197 {
7198 pkp->dtpk_prov = pdp->dtpd_provider;
7199 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7200
7201 pkp->dtpk_mod = pdp->dtpd_mod;
7202 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7203
7204 pkp->dtpk_func = pdp->dtpd_func;
7205 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7206
7207 pkp->dtpk_name = pdp->dtpd_name;
7208 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7209
7210 pkp->dtpk_id = pdp->dtpd_id;
7211
7212 if (pkp->dtpk_id == DTRACE_IDNONE &&
7213 pkp->dtpk_pmatch == &dtrace_match_nul &&
7214 pkp->dtpk_mmatch == &dtrace_match_nul &&
7215 pkp->dtpk_fmatch == &dtrace_match_nul &&
7216 pkp->dtpk_nmatch == &dtrace_match_nul)
7217 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7218 }
7219
7220 /*
7221 * DTrace Provider-to-Framework API Functions
7222 *
7223 * These functions implement much of the Provider-to-Framework API, as
7224 * described in <sys/dtrace.h>. The parts of the API not in this section are
7225 * the functions in the API for probe management (found below), and
7226 * dtrace_probe() itself (found above).
7227 */
7228
7229 /*
7230 * Register the calling provider with the DTrace framework. This should
7231 * generally be called by DTrace providers in their attach(9E) entry point.
7232 */
7233 int
7234 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7235 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7236 {
7237 dtrace_provider_t *provider;
7238
7239 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7240 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7241 "arguments", name ? name : "<NULL>");
7242 return (EINVAL);
7243 }
7244
7245 if (name[0] == '\0' || dtrace_badname(name)) {
7246 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7247 "provider name", name);
7248 return (EINVAL);
7249 }
7250
7251 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7252 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7253 pops->dtps_destroy == NULL ||
7254 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7255 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7256 "provider ops", name);
7257 return (EINVAL);
7258 }
7259
7260 if (dtrace_badattr(&pap->dtpa_provider) ||
7261 dtrace_badattr(&pap->dtpa_mod) ||
7262 dtrace_badattr(&pap->dtpa_func) ||
7263 dtrace_badattr(&pap->dtpa_name) ||
7264 dtrace_badattr(&pap->dtpa_args)) {
7265 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7266 "provider attributes", name);
7267 return (EINVAL);
7268 }
7269
7270 if (priv & ~DTRACE_PRIV_ALL) {
7271 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7272 "privilege attributes", name);
7273 return (EINVAL);
7274 }
7275
7276 if ((priv & DTRACE_PRIV_KERNEL) &&
7277 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7278 pops->dtps_mode == NULL) {
7279 cmn_err(CE_WARN, "failed to register provider '%s': need "
7280 "dtps_mode() op for given privilege attributes", name);
7281 return (EINVAL);
7282 }
7283
7284 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7285 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7286 (void) strcpy(provider->dtpv_name, name);
7287
7288 provider->dtpv_attr = *pap;
7289 provider->dtpv_priv.dtpp_flags = priv;
7290 if (cr != NULL) {
7291 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7292 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7293 }
7294 provider->dtpv_pops = *pops;
7295
7296 if (pops->dtps_provide == NULL) {
7297 ASSERT(pops->dtps_provide_module != NULL);
7298 provider->dtpv_pops.dtps_provide =
7299 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
7300 }
7301
7302 if (pops->dtps_provide_module == NULL) {
7303 ASSERT(pops->dtps_provide != NULL);
7304 provider->dtpv_pops.dtps_provide_module =
7305 (void (*)(void *, struct modctl *))dtrace_nullop;
7306 }
7307
7308 if (pops->dtps_suspend == NULL) {
7309 ASSERT(pops->dtps_resume == NULL);
7310 provider->dtpv_pops.dtps_suspend =
7311 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7312 provider->dtpv_pops.dtps_resume =
7313 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7314 }
7315
7316 provider->dtpv_arg = arg;
7317 *idp = (dtrace_provider_id_t)provider;
7318
7319 if (pops == &dtrace_provider_ops) {
7320 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7321 ASSERT(MUTEX_HELD(&dtrace_lock));
7322 ASSERT(dtrace_anon.dta_enabling == NULL);
7323
7324 /*
7325 * We make sure that the DTrace provider is at the head of
7326 * the provider chain.
7327 */
7328 provider->dtpv_next = dtrace_provider;
7329 dtrace_provider = provider;
7330 return (0);
7331 }
7332
7333 mutex_enter(&dtrace_provider_lock);
7334 mutex_enter(&dtrace_lock);
7335
7336 /*
7337 * If there is at least one provider registered, we'll add this
7338 * provider after the first provider.
7339 */
7340 if (dtrace_provider != NULL) {
7341 provider->dtpv_next = dtrace_provider->dtpv_next;
7342 dtrace_provider->dtpv_next = provider;
7343 } else {
7344 dtrace_provider = provider;
7345 }
7346
7347 if (dtrace_retained != NULL) {
7348 dtrace_enabling_provide(provider);
7349
7350 /*
7351 * Now we need to call dtrace_enabling_matchall() -- which
7352 * will acquire cpu_lock and dtrace_lock. We therefore need
7353 * to drop all of our locks before calling into it...
7354 */
7355 mutex_exit(&dtrace_lock);
7356 mutex_exit(&dtrace_provider_lock);
7357 dtrace_enabling_matchall();
7358
7359 return (0);
7360 }
7361
7362 mutex_exit(&dtrace_lock);
7363 mutex_exit(&dtrace_provider_lock);
7364
7365 return (0);
7366 }
7367
7368 /*
7369 * Unregister the specified provider from the DTrace framework. This should
7370 * generally be called by DTrace providers in their detach(9E) entry point.
7371 */
7372 int
7373 dtrace_unregister(dtrace_provider_id_t id)
7374 {
7375 dtrace_provider_t *old = (dtrace_provider_t *)id;
7376 dtrace_provider_t *prev = NULL;
7377 int i, self = 0, noreap = 0;
7378 dtrace_probe_t *probe, *first = NULL;
7379
7380 if (old->dtpv_pops.dtps_enable ==
7381 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
7382 /*
7383 * If DTrace itself is the provider, we're called with locks
7384 * already held.
7385 */
7386 ASSERT(old == dtrace_provider);
7387 ASSERT(dtrace_devi != NULL);
7388 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7389 ASSERT(MUTEX_HELD(&dtrace_lock));
7390 self = 1;
7391
7392 if (dtrace_provider->dtpv_next != NULL) {
7393 /*
7394 * There's another provider here; return failure.
7395 */
7396 return (EBUSY);
7397 }
7398 } else {
7399 mutex_enter(&dtrace_provider_lock);
7400 mutex_enter(&mod_lock);
7401 mutex_enter(&dtrace_lock);
7402 }
7403
7404 /*
7405 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7406 * probes, we refuse to let providers slither away, unless this
7407 * provider has already been explicitly invalidated.
7408 */
7409 if (!old->dtpv_defunct &&
7410 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7411 dtrace_anon.dta_state->dts_necbs > 0))) {
7412 if (!self) {
7413 mutex_exit(&dtrace_lock);
7414 mutex_exit(&mod_lock);
7415 mutex_exit(&dtrace_provider_lock);
7416 }
7417 return (EBUSY);
7418 }
7419
7420 /*
7421 * Attempt to destroy the probes associated with this provider.
7422 */
7423 for (i = 0; i < dtrace_nprobes; i++) {
7424 if ((probe = dtrace_probes[i]) == NULL)
7425 continue;
7426
7427 if (probe->dtpr_provider != old)
7428 continue;
7429
7430 if (probe->dtpr_ecb == NULL)
7431 continue;
7432
7433 /*
7434 * If we are trying to unregister a defunct provider, and the
7435 * provider was made defunct within the interval dictated by
7436 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7437 * attempt to reap our enablings. To denote that the provider
7438 * should reattempt to unregister itself at some point in the
7439 * future, we will return a differentiable error code (EAGAIN
7440 * instead of EBUSY) in this case.
7441 */
7442 if (dtrace_gethrtime() - old->dtpv_defunct >
7443 dtrace_unregister_defunct_reap)
7444 noreap = 1;
7445
7446 if (!self) {
7447 mutex_exit(&dtrace_lock);
7448 mutex_exit(&mod_lock);
7449 mutex_exit(&dtrace_provider_lock);
7450 }
7451
7452 if (noreap)
7453 return (EBUSY);
7454
7455 (void) taskq_dispatch(dtrace_taskq,
7456 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7457
7458 return (EAGAIN);
7459 }
7460
7461 /*
7462 * All of the probes for this provider are disabled; we can safely
7463 * remove all of them from their hash chains and from the probe array.
7464 */
7465 for (i = 0; i < dtrace_nprobes; i++) {
7466 if ((probe = dtrace_probes[i]) == NULL)
7467 continue;
7468
7469 if (probe->dtpr_provider != old)
7470 continue;
7471
7472 dtrace_probes[i] = NULL;
7473
7474 dtrace_hash_remove(dtrace_bymod, probe);
7475 dtrace_hash_remove(dtrace_byfunc, probe);
7476 dtrace_hash_remove(dtrace_byname, probe);
7477
7478 if (first == NULL) {
7479 first = probe;
7480 probe->dtpr_nextmod = NULL;
7481 } else {
7482 probe->dtpr_nextmod = first;
7483 first = probe;
7484 }
7485 }
7486
7487 /*
7488 * The provider's probes have been removed from the hash chains and
7489 * from the probe array. Now issue a dtrace_sync() to be sure that
7490 * everyone has cleared out from any probe array processing.
7491 */
7492 dtrace_sync();
7493
7494 for (probe = first; probe != NULL; probe = first) {
7495 first = probe->dtpr_nextmod;
7496
7497 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7498 probe->dtpr_arg);
7499 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7500 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7501 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7502 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7503 kmem_free(probe, sizeof (dtrace_probe_t));
7504 }
7505
7506 if ((prev = dtrace_provider) == old) {
7507 ASSERT(self || dtrace_devi == NULL);
7508 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7509 dtrace_provider = old->dtpv_next;
7510 } else {
7511 while (prev != NULL && prev->dtpv_next != old)
7512 prev = prev->dtpv_next;
7513
7514 if (prev == NULL) {
7515 panic("attempt to unregister non-existent "
7516 "dtrace provider %p\n", (void *)id);
7517 }
7518
7519 prev->dtpv_next = old->dtpv_next;
7520 }
7521
7522 if (!self) {
7523 mutex_exit(&dtrace_lock);
7524 mutex_exit(&mod_lock);
7525 mutex_exit(&dtrace_provider_lock);
7526 }
7527
7528 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7529 kmem_free(old, sizeof (dtrace_provider_t));
7530
7531 return (0);
7532 }
7533
7534 /*
7535 * Invalidate the specified provider. All subsequent probe lookups for the
7536 * specified provider will fail, but its probes will not be removed.
7537 */
7538 void
7539 dtrace_invalidate(dtrace_provider_id_t id)
7540 {
7541 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7542
7543 ASSERT(pvp->dtpv_pops.dtps_enable !=
7544 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7545
7546 mutex_enter(&dtrace_provider_lock);
7547 mutex_enter(&dtrace_lock);
7548
7549 pvp->dtpv_defunct = dtrace_gethrtime();
7550
7551 mutex_exit(&dtrace_lock);
7552 mutex_exit(&dtrace_provider_lock);
7553 }
7554
7555 /*
7556 * Indicate whether or not DTrace has attached.
7557 */
7558 int
7559 dtrace_attached(void)
7560 {
7561 /*
7562 * dtrace_provider will be non-NULL iff the DTrace driver has
7563 * attached. (It's non-NULL because DTrace is always itself a
7564 * provider.)
7565 */
7566 return (dtrace_provider != NULL);
7567 }
7568
7569 /*
7570 * Remove all the unenabled probes for the given provider. This function is
7571 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7572 * -- just as many of its associated probes as it can.
7573 */
7574 int
7575 dtrace_condense(dtrace_provider_id_t id)
7576 {
7577 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7578 int i;
7579 dtrace_probe_t *probe;
7580
7581 /*
7582 * Make sure this isn't the dtrace provider itself.
7583 */
7584 ASSERT(prov->dtpv_pops.dtps_enable !=
7585 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7586
7587 mutex_enter(&dtrace_provider_lock);
7588 mutex_enter(&dtrace_lock);
7589
7590 /*
7591 * Attempt to destroy the probes associated with this provider.
7592 */
7593 for (i = 0; i < dtrace_nprobes; i++) {
7594 if ((probe = dtrace_probes[i]) == NULL)
7595 continue;
7596
7597 if (probe->dtpr_provider != prov)
7598 continue;
7599
7600 if (probe->dtpr_ecb != NULL)
7601 continue;
7602
7603 dtrace_probes[i] = NULL;
7604
7605 dtrace_hash_remove(dtrace_bymod, probe);
7606 dtrace_hash_remove(dtrace_byfunc, probe);
7607 dtrace_hash_remove(dtrace_byname, probe);
7608
7609 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7610 probe->dtpr_arg);
7611 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7612 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7613 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7614 kmem_free(probe, sizeof (dtrace_probe_t));
7615 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7616 }
7617
7618 mutex_exit(&dtrace_lock);
7619 mutex_exit(&dtrace_provider_lock);
7620
7621 return (0);
7622 }
7623
7624 /*
7625 * DTrace Probe Management Functions
7626 *
7627 * The functions in this section perform the DTrace probe management,
7628 * including functions to create probes, look-up probes, and call into the
7629 * providers to request that probes be provided. Some of these functions are
7630 * in the Provider-to-Framework API; these functions can be identified by the
7631 * fact that they are not declared "static".
7632 */
7633
7634 /*
7635 * Create a probe with the specified module name, function name, and name.
7636 */
7637 dtrace_id_t
7638 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7639 const char *func, const char *name, int aframes, void *arg)
7640 {
7641 dtrace_probe_t *probe, **probes;
7642 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7643 dtrace_id_t id;
7644
7645 if (provider == dtrace_provider) {
7646 ASSERT(MUTEX_HELD(&dtrace_lock));
7647 } else {
7648 mutex_enter(&dtrace_lock);
7649 }
7650
7651 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7652 VM_BESTFIT | VM_SLEEP);
7653 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7654
7655 probe->dtpr_id = id;
7656 probe->dtpr_gen = dtrace_probegen++;
7657 probe->dtpr_mod = dtrace_strdup(mod);
7658 probe->dtpr_func = dtrace_strdup(func);
7659 probe->dtpr_name = dtrace_strdup(name);
7660 probe->dtpr_arg = arg;
7661 probe->dtpr_aframes = aframes;
7662 probe->dtpr_provider = provider;
7663
7664 dtrace_hash_add(dtrace_bymod, probe);
7665 dtrace_hash_add(dtrace_byfunc, probe);
7666 dtrace_hash_add(dtrace_byname, probe);
7667
7668 if (id - 1 >= dtrace_nprobes) {
7669 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7670 size_t nsize = osize << 1;
7671
7672 if (nsize == 0) {
7673 ASSERT(osize == 0);
7674 ASSERT(dtrace_probes == NULL);
7675 nsize = sizeof (dtrace_probe_t *);
7676 }
7677
7678 probes = kmem_zalloc(nsize, KM_SLEEP);
7679
7680 if (dtrace_probes == NULL) {
7681 ASSERT(osize == 0);
7682 dtrace_probes = probes;
7683 dtrace_nprobes = 1;
7684 } else {
7685 dtrace_probe_t **oprobes = dtrace_probes;
7686
7687 bcopy(oprobes, probes, osize);
7688 dtrace_membar_producer();
7689 dtrace_probes = probes;
7690
7691 dtrace_sync();
7692
7693 /*
7694 * All CPUs are now seeing the new probes array; we can
7695 * safely free the old array.
7696 */
7697 kmem_free(oprobes, osize);
7698 dtrace_nprobes <<= 1;
7699 }
7700
7701 ASSERT(id - 1 < dtrace_nprobes);
7702 }
7703
7704 ASSERT(dtrace_probes[id - 1] == NULL);
7705 dtrace_probes[id - 1] = probe;
7706
7707 if (provider != dtrace_provider)
7708 mutex_exit(&dtrace_lock);
7709
7710 return (id);
7711 }
7712
7713 static dtrace_probe_t *
7714 dtrace_probe_lookup_id(dtrace_id_t id)
7715 {
7716 ASSERT(MUTEX_HELD(&dtrace_lock));
7717
7718 if (id == 0 || id > dtrace_nprobes)
7719 return (NULL);
7720
7721 return (dtrace_probes[id - 1]);
7722 }
7723
7724 static int
7725 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7726 {
7727 *((dtrace_id_t *)arg) = probe->dtpr_id;
7728
7729 return (DTRACE_MATCH_DONE);
7730 }
7731
7732 /*
7733 * Look up a probe based on provider and one or more of module name, function
7734 * name and probe name.
7735 */
7736 dtrace_id_t
7737 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
7738 const char *func, const char *name)
7739 {
7740 dtrace_probekey_t pkey;
7741 dtrace_id_t id;
7742 int match;
7743
7744 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7745 pkey.dtpk_pmatch = &dtrace_match_string;
7746 pkey.dtpk_mod = mod;
7747 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7748 pkey.dtpk_func = func;
7749 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7750 pkey.dtpk_name = name;
7751 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7752 pkey.dtpk_id = DTRACE_IDNONE;
7753
7754 mutex_enter(&dtrace_lock);
7755 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7756 dtrace_probe_lookup_match, &id);
7757 mutex_exit(&dtrace_lock);
7758
7759 ASSERT(match == 1 || match == 0);
7760 return (match ? id : 0);
7761 }
7762
7763 /*
7764 * Returns the probe argument associated with the specified probe.
7765 */
7766 void *
7767 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7768 {
7769 dtrace_probe_t *probe;
7770 void *rval = NULL;
7771
7772 mutex_enter(&dtrace_lock);
7773
7774 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7775 probe->dtpr_provider == (dtrace_provider_t *)id)
7776 rval = probe->dtpr_arg;
7777
7778 mutex_exit(&dtrace_lock);
7779
7780 return (rval);
7781 }
7782
7783 /*
7784 * Copy a probe into a probe description.
7785 */
7786 static void
7787 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7788 {
7789 bzero(pdp, sizeof (dtrace_probedesc_t));
7790 pdp->dtpd_id = prp->dtpr_id;
7791
7792 (void) strncpy(pdp->dtpd_provider,
7793 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7794
7795 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7796 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7797 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7798 }
7799
7800 /*
7801 * Called to indicate that a probe -- or probes -- should be provided by a
7802 * specfied provider. If the specified description is NULL, the provider will
7803 * be told to provide all of its probes. (This is done whenever a new
7804 * consumer comes along, or whenever a retained enabling is to be matched.) If
7805 * the specified description is non-NULL, the provider is given the
7806 * opportunity to dynamically provide the specified probe, allowing providers
7807 * to support the creation of probes on-the-fly. (So-called _autocreated_
7808 * probes.) If the provider is NULL, the operations will be applied to all
7809 * providers; if the provider is non-NULL the operations will only be applied
7810 * to the specified provider. The dtrace_provider_lock must be held, and the
7811 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7812 * will need to grab the dtrace_lock when it reenters the framework through
7813 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7814 */
7815 static void
7816 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7817 {
7818 struct modctl *ctl;
7819 int all = 0;
7820
7821 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7822
7823 if (prv == NULL) {
7824 all = 1;
7825 prv = dtrace_provider;
7826 }
7827
7828 do {
7829 /*
7830 * First, call the blanket provide operation.
7831 */
7832 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7833
7834 /*
7835 * Now call the per-module provide operation. We will grab
7836 * mod_lock to prevent the list from being modified. Note
7837 * that this also prevents the mod_busy bits from changing.
7838 * (mod_busy can only be changed with mod_lock held.)
7839 */
7840 mutex_enter(&mod_lock);
7841
7842 ctl = &modules;
7843 do {
7844 if (ctl->mod_busy || ctl->mod_mp == NULL)
7845 continue;
7846
7847 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
7848
7849 } while ((ctl = ctl->mod_next) != &modules);
7850
7851 mutex_exit(&mod_lock);
7852 } while (all && (prv = prv->dtpv_next) != NULL);
7853 }
7854
7855 /*
7856 * Iterate over each probe, and call the Framework-to-Provider API function
7857 * denoted by offs.
7858 */
7859 static void
7860 dtrace_probe_foreach(uintptr_t offs)
7861 {
7862 dtrace_provider_t *prov;
7863 void (*func)(void *, dtrace_id_t, void *);
7864 dtrace_probe_t *probe;
7865 dtrace_icookie_t cookie;
7866 int i;
7867
7868 /*
7869 * We disable interrupts to walk through the probe array. This is
7870 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
7871 * won't see stale data.
7872 */
7873 cookie = dtrace_interrupt_disable();
7874
7875 for (i = 0; i < dtrace_nprobes; i++) {
7876 if ((probe = dtrace_probes[i]) == NULL)
7877 continue;
7878
7879 if (probe->dtpr_ecb == NULL) {
7880 /*
7881 * This probe isn't enabled -- don't call the function.
7882 */
7883 continue;
7884 }
7885
7886 prov = probe->dtpr_provider;
7887 func = *((void(**)(void *, dtrace_id_t, void *))
7888 ((uintptr_t)&prov->dtpv_pops + offs));
7889
7890 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
7891 }
7892
7893 dtrace_interrupt_enable(cookie);
7894 }
7895
7896 static int
7897 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
7898 {
7899 dtrace_probekey_t pkey;
7900 uint32_t priv;
7901 uid_t uid;
7902 zoneid_t zoneid;
7903
7904 ASSERT(MUTEX_HELD(&dtrace_lock));
7905 dtrace_ecb_create_cache = NULL;
7906
7907 if (desc == NULL) {
7908 /*
7909 * If we're passed a NULL description, we're being asked to
7910 * create an ECB with a NULL probe.
7911 */
7912 (void) dtrace_ecb_create_enable(NULL, enab);
7913 return (0);
7914 }
7915
7916 dtrace_probekey(desc, &pkey);
7917 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
7918 &priv, &uid, &zoneid);
7919
7920 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
7921 enab));
7922 }
7923
7924 /*
7925 * DTrace Helper Provider Functions
7926 */
7927 static void
7928 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
7929 {
7930 attr->dtat_name = DOF_ATTR_NAME(dofattr);
7931 attr->dtat_data = DOF_ATTR_DATA(dofattr);
7932 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
7933 }
7934
7935 static void
7936 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
7937 const dof_provider_t *dofprov, char *strtab)
7938 {
7939 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
7940 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
7941 dofprov->dofpv_provattr);
7942 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
7943 dofprov->dofpv_modattr);
7944 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
7945 dofprov->dofpv_funcattr);
7946 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
7947 dofprov->dofpv_nameattr);
7948 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
7949 dofprov->dofpv_argsattr);
7950 }
7951
7952 static void
7953 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
7954 {
7955 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7956 dof_hdr_t *dof = (dof_hdr_t *)daddr;
7957 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
7958 dof_provider_t *provider;
7959 dof_probe_t *probe;
7960 uint32_t *off, *enoff;
7961 uint8_t *arg;
7962 char *strtab;
7963 uint_t i, nprobes;
7964 dtrace_helper_provdesc_t dhpv;
7965 dtrace_helper_probedesc_t dhpb;
7966 dtrace_meta_t *meta = dtrace_meta_pid;
7967 dtrace_mops_t *mops = &meta->dtm_mops;
7968 void *parg;
7969
7970 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
7971 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7972 provider->dofpv_strtab * dof->dofh_secsize);
7973 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7974 provider->dofpv_probes * dof->dofh_secsize);
7975 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7976 provider->dofpv_prargs * dof->dofh_secsize);
7977 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7978 provider->dofpv_proffs * dof->dofh_secsize);
7979
7980 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
7981 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
7982 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
7983 enoff = NULL;
7984
7985 /*
7986 * See dtrace_helper_provider_validate().
7987 */
7988 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
7989 provider->dofpv_prenoffs != DOF_SECT_NONE) {
7990 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7991 provider->dofpv_prenoffs * dof->dofh_secsize);
7992 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
7993 }
7994
7995 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
7996
7997 /*
7998 * Create the provider.
7999 */
8000 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8001
8002 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8003 return;
8004
8005 meta->dtm_count++;
8006
8007 /*
8008 * Create the probes.
8009 */
8010 for (i = 0; i < nprobes; i++) {
8011 probe = (dof_probe_t *)(uintptr_t)(daddr +
8012 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8013
8014 dhpb.dthpb_mod = dhp->dofhp_mod;
8015 dhpb.dthpb_func = strtab + probe->dofpr_func;
8016 dhpb.dthpb_name = strtab + probe->dofpr_name;
8017 dhpb.dthpb_base = probe->dofpr_addr;
8018 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8019 dhpb.dthpb_noffs = probe->dofpr_noffs;
8020 if (enoff != NULL) {
8021 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8022 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8023 } else {
8024 dhpb.dthpb_enoffs = NULL;
8025 dhpb.dthpb_nenoffs = 0;
8026 }
8027 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8028 dhpb.dthpb_nargc = probe->dofpr_nargc;
8029 dhpb.dthpb_xargc = probe->dofpr_xargc;
8030 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8031 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8032
8033 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8034 }
8035 }
8036
8037 static void
8038 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8039 {
8040 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8041 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8042 int i;
8043
8044 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8045
8046 for (i = 0; i < dof->dofh_secnum; i++) {
8047 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8048 dof->dofh_secoff + i * dof->dofh_secsize);
8049
8050 if (sec->dofs_type != DOF_SECT_PROVIDER)
8051 continue;
8052
8053 dtrace_helper_provide_one(dhp, sec, pid);
8054 }
8055
8056 /*
8057 * We may have just created probes, so we must now rematch against
8058 * any retained enablings. Note that this call will acquire both
8059 * cpu_lock and dtrace_lock; the fact that we are holding
8060 * dtrace_meta_lock now is what defines the ordering with respect to
8061 * these three locks.
8062 */
8063 dtrace_enabling_matchall();
8064 }
8065
8066 static void
8067 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8068 {
8069 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8070 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8071 dof_sec_t *str_sec;
8072 dof_provider_t *provider;
8073 char *strtab;
8074 dtrace_helper_provdesc_t dhpv;
8075 dtrace_meta_t *meta = dtrace_meta_pid;
8076 dtrace_mops_t *mops = &meta->dtm_mops;
8077
8078 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8079 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8080 provider->dofpv_strtab * dof->dofh_secsize);
8081
8082 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8083
8084 /*
8085 * Create the provider.
8086 */
8087 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8088
8089 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8090
8091 meta->dtm_count--;
8092 }
8093
8094 static void
8095 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8096 {
8097 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8098 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8099 int i;
8100
8101 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8102
8103 for (i = 0; i < dof->dofh_secnum; i++) {
8104 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8105 dof->dofh_secoff + i * dof->dofh_secsize);
8106
8107 if (sec->dofs_type != DOF_SECT_PROVIDER)
8108 continue;
8109
8110 dtrace_helper_provider_remove_one(dhp, sec, pid);
8111 }
8112 }
8113
8114 /*
8115 * DTrace Meta Provider-to-Framework API Functions
8116 *
8117 * These functions implement the Meta Provider-to-Framework API, as described
8118 * in <sys/dtrace.h>.
8119 */
8120 int
8121 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8122 dtrace_meta_provider_id_t *idp)
8123 {
8124 dtrace_meta_t *meta;
8125 dtrace_helpers_t *help, *next;
8126 int i;
8127
8128 *idp = DTRACE_METAPROVNONE;
8129
8130 /*
8131 * We strictly don't need the name, but we hold onto it for
8132 * debuggability. All hail error queues!
8133 */
8134 if (name == NULL) {
8135 cmn_err(CE_WARN, "failed to register meta-provider: "
8136 "invalid name");
8137 return (EINVAL);
8138 }
8139
8140 if (mops == NULL ||
8141 mops->dtms_create_probe == NULL ||
8142 mops->dtms_provide_pid == NULL ||
8143 mops->dtms_remove_pid == NULL) {
8144 cmn_err(CE_WARN, "failed to register meta-register %s: "
8145 "invalid ops", name);
8146 return (EINVAL);
8147 }
8148
8149 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8150 meta->dtm_mops = *mops;
8151 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8152 (void) strcpy(meta->dtm_name, name);
8153 meta->dtm_arg = arg;
8154
8155 mutex_enter(&dtrace_meta_lock);
8156 mutex_enter(&dtrace_lock);
8157
8158 if (dtrace_meta_pid != NULL) {
8159 mutex_exit(&dtrace_lock);
8160 mutex_exit(&dtrace_meta_lock);
8161 cmn_err(CE_WARN, "failed to register meta-register %s: "
8162 "user-land meta-provider exists", name);
8163 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8164 kmem_free(meta, sizeof (dtrace_meta_t));
8165 return (EINVAL);
8166 }
8167
8168 dtrace_meta_pid = meta;
8169 *idp = (dtrace_meta_provider_id_t)meta;
8170
8171 /*
8172 * If there are providers and probes ready to go, pass them
8173 * off to the new meta provider now.
8174 */
8175
8176 help = dtrace_deferred_pid;
8177 dtrace_deferred_pid = NULL;
8178
8179 mutex_exit(&dtrace_lock);
8180
8181 while (help != NULL) {
8182 for (i = 0; i < help->dthps_nprovs; i++) {
8183 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8184 help->dthps_pid);
8185 }
8186
8187 next = help->dthps_next;
8188 help->dthps_next = NULL;
8189 help->dthps_prev = NULL;
8190 help->dthps_deferred = 0;
8191 help = next;
8192 }
8193
8194 mutex_exit(&dtrace_meta_lock);
8195
8196 return (0);
8197 }
8198
8199 int
8200 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8201 {
8202 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8203
8204 mutex_enter(&dtrace_meta_lock);
8205 mutex_enter(&dtrace_lock);
8206
8207 if (old == dtrace_meta_pid) {
8208 pp = &dtrace_meta_pid;
8209 } else {
8210 panic("attempt to unregister non-existent "
8211 "dtrace meta-provider %p\n", (void *)old);
8212 }
8213
8214 if (old->dtm_count != 0) {
8215 mutex_exit(&dtrace_lock);
8216 mutex_exit(&dtrace_meta_lock);
8217 return (EBUSY);
8218 }
8219
8220 *pp = NULL;
8221
8222 mutex_exit(&dtrace_lock);
8223 mutex_exit(&dtrace_meta_lock);
8224
8225 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8226 kmem_free(old, sizeof (dtrace_meta_t));
8227
8228 return (0);
8229 }
8230
8231
8232 /*
8233 * DTrace DIF Object Functions
8234 */
8235 static int
8236 dtrace_difo_err(uint_t pc, const char *format, ...)
8237 {
8238 if (dtrace_err_verbose) {
8239 va_list alist;
8240
8241 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8242 va_start(alist, format);
8243 (void) vuprintf(format, alist);
8244 va_end(alist);
8245 }
8246
8247 #ifdef DTRACE_ERRDEBUG
8248 dtrace_errdebug(format);
8249 #endif
8250 return (1);
8251 }
8252
8253 /*
8254 * Validate a DTrace DIF object by checking the IR instructions. The following
8255 * rules are currently enforced by dtrace_difo_validate():
8256 *
8257 * 1. Each instruction must have a valid opcode
8258 * 2. Each register, string, variable, or subroutine reference must be valid
8259 * 3. No instruction can modify register %r0 (must be zero)
8260 * 4. All instruction reserved bits must be set to zero
8261 * 5. The last instruction must be a "ret" instruction
8262 * 6. All branch targets must reference a valid instruction _after_ the branch
8263 */
8264 static int
8265 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8266 cred_t *cr)
8267 {
8268 int err = 0, i;
8269 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8270 int kcheckload;
8271 uint_t pc;
8272
8273 kcheckload = cr == NULL ||
8274 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8275
8276 dp->dtdo_destructive = 0;
8277
8278 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8279 dif_instr_t instr = dp->dtdo_buf[pc];
8280
8281 uint_t r1 = DIF_INSTR_R1(instr);
8282 uint_t r2 = DIF_INSTR_R2(instr);
8283 uint_t rd = DIF_INSTR_RD(instr);
8284 uint_t rs = DIF_INSTR_RS(instr);
8285 uint_t label = DIF_INSTR_LABEL(instr);
8286 uint_t v = DIF_INSTR_VAR(instr);
8287 uint_t subr = DIF_INSTR_SUBR(instr);
8288 uint_t type = DIF_INSTR_TYPE(instr);
8289 uint_t op = DIF_INSTR_OP(instr);
8290
8291 switch (op) {
8292 case DIF_OP_OR:
8293 case DIF_OP_XOR:
8294 case DIF_OP_AND:
8295 case DIF_OP_SLL:
8296 case DIF_OP_SRL:
8297 case DIF_OP_SRA:
8298 case DIF_OP_SUB:
8299 case DIF_OP_ADD:
8300 case DIF_OP_MUL:
8301 case DIF_OP_SDIV:
8302 case DIF_OP_UDIV:
8303 case DIF_OP_SREM:
8304 case DIF_OP_UREM:
8305 case DIF_OP_COPYS:
8306 if (r1 >= nregs)
8307 err += efunc(pc, "invalid register %u\n", r1);
8308 if (r2 >= nregs)
8309 err += efunc(pc, "invalid register %u\n", r2);
8310 if (rd >= nregs)
8311 err += efunc(pc, "invalid register %u\n", rd);
8312 if (rd == 0)
8313 err += efunc(pc, "cannot write to %r0\n");
8314 break;
8315 case DIF_OP_NOT:
8316 case DIF_OP_MOV:
8317 case DIF_OP_ALLOCS:
8318 if (r1 >= nregs)
8319 err += efunc(pc, "invalid register %u\n", r1);
8320 if (r2 != 0)
8321 err += efunc(pc, "non-zero reserved bits\n");
8322 if (rd >= nregs)
8323 err += efunc(pc, "invalid register %u\n", rd);
8324 if (rd == 0)
8325 err += efunc(pc, "cannot write to %r0\n");
8326 break;
8327 case DIF_OP_LDSB:
8328 case DIF_OP_LDSH:
8329 case DIF_OP_LDSW:
8330 case DIF_OP_LDUB:
8331 case DIF_OP_LDUH:
8332 case DIF_OP_LDUW:
8333 case DIF_OP_LDX:
8334 if (r1 >= nregs)
8335 err += efunc(pc, "invalid register %u\n", r1);
8336 if (r2 != 0)
8337 err += efunc(pc, "non-zero reserved bits\n");
8338 if (rd >= nregs)
8339 err += efunc(pc, "invalid register %u\n", rd);
8340 if (rd == 0)
8341 err += efunc(pc, "cannot write to %r0\n");
8342 if (kcheckload)
8343 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8344 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8345 break;
8346 case DIF_OP_RLDSB:
8347 case DIF_OP_RLDSH:
8348 case DIF_OP_RLDSW:
8349 case DIF_OP_RLDUB:
8350 case DIF_OP_RLDUH:
8351 case DIF_OP_RLDUW:
8352 case DIF_OP_RLDX:
8353 if (r1 >= nregs)
8354 err += efunc(pc, "invalid register %u\n", r1);
8355 if (r2 != 0)
8356 err += efunc(pc, "non-zero reserved bits\n");
8357 if (rd >= nregs)
8358 err += efunc(pc, "invalid register %u\n", rd);
8359 if (rd == 0)
8360 err += efunc(pc, "cannot write to %r0\n");
8361 break;
8362 case DIF_OP_ULDSB:
8363 case DIF_OP_ULDSH:
8364 case DIF_OP_ULDSW:
8365 case DIF_OP_ULDUB:
8366 case DIF_OP_ULDUH:
8367 case DIF_OP_ULDUW:
8368 case DIF_OP_ULDX:
8369 if (r1 >= nregs)
8370 err += efunc(pc, "invalid register %u\n", r1);
8371 if (r2 != 0)
8372 err += efunc(pc, "non-zero reserved bits\n");
8373 if (rd >= nregs)
8374 err += efunc(pc, "invalid register %u\n", rd);
8375 if (rd == 0)
8376 err += efunc(pc, "cannot write to %r0\n");
8377 break;
8378 case DIF_OP_STB:
8379 case DIF_OP_STH:
8380 case DIF_OP_STW:
8381 case DIF_OP_STX:
8382 if (r1 >= nregs)
8383 err += efunc(pc, "invalid register %u\n", r1);
8384 if (r2 != 0)
8385 err += efunc(pc, "non-zero reserved bits\n");
8386 if (rd >= nregs)
8387 err += efunc(pc, "invalid register %u\n", rd);
8388 if (rd == 0)
8389 err += efunc(pc, "cannot write to 0 address\n");
8390 break;
8391 case DIF_OP_CMP:
8392 case DIF_OP_SCMP:
8393 if (r1 >= nregs)
8394 err += efunc(pc, "invalid register %u\n", r1);
8395 if (r2 >= nregs)
8396 err += efunc(pc, "invalid register %u\n", r2);
8397 if (rd != 0)
8398 err += efunc(pc, "non-zero reserved bits\n");
8399 break;
8400 case DIF_OP_TST:
8401 if (r1 >= nregs)
8402 err += efunc(pc, "invalid register %u\n", r1);
8403 if (r2 != 0 || rd != 0)
8404 err += efunc(pc, "non-zero reserved bits\n");
8405 break;
8406 case DIF_OP_BA:
8407 case DIF_OP_BE:
8408 case DIF_OP_BNE:
8409 case DIF_OP_BG:
8410 case DIF_OP_BGU:
8411 case DIF_OP_BGE:
8412 case DIF_OP_BGEU:
8413 case DIF_OP_BL:
8414 case DIF_OP_BLU:
8415 case DIF_OP_BLE:
8416 case DIF_OP_BLEU:
8417 if (label >= dp->dtdo_len) {
8418 err += efunc(pc, "invalid branch target %u\n",
8419 label);
8420 }
8421 if (label <= pc) {
8422 err += efunc(pc, "backward branch to %u\n",
8423 label);
8424 }
8425 break;
8426 case DIF_OP_RET:
8427 if (r1 != 0 || r2 != 0)
8428 err += efunc(pc, "non-zero reserved bits\n");
8429 if (rd >= nregs)
8430 err += efunc(pc, "invalid register %u\n", rd);
8431 break;
8432 case DIF_OP_NOP:
8433 case DIF_OP_POPTS:
8434 case DIF_OP_FLUSHTS:
8435 if (r1 != 0 || r2 != 0 || rd != 0)
8436 err += efunc(pc, "non-zero reserved bits\n");
8437 break;
8438 case DIF_OP_SETX:
8439 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8440 err += efunc(pc, "invalid integer ref %u\n",
8441 DIF_INSTR_INTEGER(instr));
8442 }
8443 if (rd >= nregs)
8444 err += efunc(pc, "invalid register %u\n", rd);
8445 if (rd == 0)
8446 err += efunc(pc, "cannot write to %r0\n");
8447 break;
8448 case DIF_OP_SETS:
8449 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8450 err += efunc(pc, "invalid string ref %u\n",
8451 DIF_INSTR_STRING(instr));
8452 }
8453 if (rd >= nregs)
8454 err += efunc(pc, "invalid register %u\n", rd);
8455 if (rd == 0)
8456 err += efunc(pc, "cannot write to %r0\n");
8457 break;
8458 case DIF_OP_LDGA:
8459 case DIF_OP_LDTA:
8460 if (r1 > DIF_VAR_ARRAY_MAX)
8461 err += efunc(pc, "invalid array %u\n", r1);
8462 if (r2 >= nregs)
8463 err += efunc(pc, "invalid register %u\n", r2);
8464 if (rd >= nregs)
8465 err += efunc(pc, "invalid register %u\n", rd);
8466 if (rd == 0)
8467 err += efunc(pc, "cannot write to %r0\n");
8468 break;
8469 case DIF_OP_LDGS:
8470 case DIF_OP_LDTS:
8471 case DIF_OP_LDLS:
8472 case DIF_OP_LDGAA:
8473 case DIF_OP_LDTAA:
8474 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8475 err += efunc(pc, "invalid variable %u\n", v);
8476 if (rd >= nregs)
8477 err += efunc(pc, "invalid register %u\n", rd);
8478 if (rd == 0)
8479 err += efunc(pc, "cannot write to %r0\n");
8480 break;
8481 case DIF_OP_STGS:
8482 case DIF_OP_STTS:
8483 case DIF_OP_STLS:
8484 case DIF_OP_STGAA:
8485 case DIF_OP_STTAA:
8486 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8487 err += efunc(pc, "invalid variable %u\n", v);
8488 if (rs >= nregs)
8489 err += efunc(pc, "invalid register %u\n", rd);
8490 break;
8491 case DIF_OP_CALL:
8492 if (subr > DIF_SUBR_MAX)
8493 err += efunc(pc, "invalid subr %u\n", subr);
8494 if (rd >= nregs)
8495 err += efunc(pc, "invalid register %u\n", rd);
8496 if (rd == 0)
8497 err += efunc(pc, "cannot write to %r0\n");
8498
8499 if (subr == DIF_SUBR_COPYOUT ||
8500 subr == DIF_SUBR_COPYOUTSTR) {
8501 dp->dtdo_destructive = 1;
8502 }
8503
8504 if (subr == DIF_SUBR_GETF) {
8505 /*
8506 * If we have a getf() we need to record that
8507 * in our state. Note that our state can be
8508 * NULL if this is a helper -- but in that
8509 * case, the call to getf() is itself illegal,
8510 * and will be caught (slightly later) when
8511 * the helper is validated.
8512 */
8513 if (vstate->dtvs_state != NULL)
8514 vstate->dtvs_state->dts_getf++;
8515 }
8516
8517 break;
8518 case DIF_OP_PUSHTR:
8519 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8520 err += efunc(pc, "invalid ref type %u\n", type);
8521 if (r2 >= nregs)
8522 err += efunc(pc, "invalid register %u\n", r2);
8523 if (rs >= nregs)
8524 err += efunc(pc, "invalid register %u\n", rs);
8525 break;
8526 case DIF_OP_PUSHTV:
8527 if (type != DIF_TYPE_CTF)
8528 err += efunc(pc, "invalid val type %u\n", type);
8529 if (r2 >= nregs)
8530 err += efunc(pc, "invalid register %u\n", r2);
8531 if (rs >= nregs)
8532 err += efunc(pc, "invalid register %u\n", rs);
8533 break;
8534 default:
8535 err += efunc(pc, "invalid opcode %u\n",
8536 DIF_INSTR_OP(instr));
8537 }
8538 }
8539
8540 if (dp->dtdo_len != 0 &&
8541 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8542 err += efunc(dp->dtdo_len - 1,
8543 "expected 'ret' as last DIF instruction\n");
8544 }
8545
8546 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8547 /*
8548 * If we're not returning by reference, the size must be either
8549 * 0 or the size of one of the base types.
8550 */
8551 switch (dp->dtdo_rtype.dtdt_size) {
8552 case 0:
8553 case sizeof (uint8_t):
8554 case sizeof (uint16_t):
8555 case sizeof (uint32_t):
8556 case sizeof (uint64_t):
8557 break;
8558
8559 default:
8560 err += efunc(dp->dtdo_len - 1, "bad return size\n");
8561 }
8562 }
8563
8564 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8565 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8566 dtrace_diftype_t *vt, *et;
8567 uint_t id, ndx;
8568
8569 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8570 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8571 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8572 err += efunc(i, "unrecognized variable scope %d\n",
8573 v->dtdv_scope);
8574 break;
8575 }
8576
8577 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8578 v->dtdv_kind != DIFV_KIND_SCALAR) {
8579 err += efunc(i, "unrecognized variable type %d\n",
8580 v->dtdv_kind);
8581 break;
8582 }
8583
8584 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8585 err += efunc(i, "%d exceeds variable id limit\n", id);
8586 break;
8587 }
8588
8589 if (id < DIF_VAR_OTHER_UBASE)
8590 continue;
8591
8592 /*
8593 * For user-defined variables, we need to check that this
8594 * definition is identical to any previous definition that we
8595 * encountered.
8596 */
8597 ndx = id - DIF_VAR_OTHER_UBASE;
8598
8599 switch (v->dtdv_scope) {
8600 case DIFV_SCOPE_GLOBAL:
8601 if (ndx < vstate->dtvs_nglobals) {
8602 dtrace_statvar_t *svar;
8603
8604 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8605 existing = &svar->dtsv_var;
8606 }
8607
8608 break;
8609
8610 case DIFV_SCOPE_THREAD:
8611 if (ndx < vstate->dtvs_ntlocals)
8612 existing = &vstate->dtvs_tlocals[ndx];
8613 break;
8614
8615 case DIFV_SCOPE_LOCAL:
8616 if (ndx < vstate->dtvs_nlocals) {
8617 dtrace_statvar_t *svar;
8618
8619 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8620 existing = &svar->dtsv_var;
8621 }
8622
8623 break;
8624 }
8625
8626 vt = &v->dtdv_type;
8627
8628 if (vt->dtdt_flags & DIF_TF_BYREF) {
8629 if (vt->dtdt_size == 0) {
8630 err += efunc(i, "zero-sized variable\n");
8631 break;
8632 }
8633
8634 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8635 vt->dtdt_size > dtrace_global_maxsize) {
8636 err += efunc(i, "oversized by-ref global\n");
8637 break;
8638 }
8639 }
8640
8641 if (existing == NULL || existing->dtdv_id == 0)
8642 continue;
8643
8644 ASSERT(existing->dtdv_id == v->dtdv_id);
8645 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8646
8647 if (existing->dtdv_kind != v->dtdv_kind)
8648 err += efunc(i, "%d changed variable kind\n", id);
8649
8650 et = &existing->dtdv_type;
8651
8652 if (vt->dtdt_flags != et->dtdt_flags) {
8653 err += efunc(i, "%d changed variable type flags\n", id);
8654 break;
8655 }
8656
8657 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8658 err += efunc(i, "%d changed variable type size\n", id);
8659 break;
8660 }
8661 }
8662
8663 return (err);
8664 }
8665
8666 /*
8667 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
8668 * are much more constrained than normal DIFOs. Specifically, they may
8669 * not:
8670 *
8671 * 1. Make calls to subroutines other than copyin(), copyinstr() or
8672 * miscellaneous string routines
8673 * 2. Access DTrace variables other than the args[] array, and the
8674 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8675 * 3. Have thread-local variables.
8676 * 4. Have dynamic variables.
8677 */
8678 static int
8679 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8680 {
8681 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8682 int err = 0;
8683 uint_t pc;
8684
8685 for (pc = 0; pc < dp->dtdo_len; pc++) {
8686 dif_instr_t instr = dp->dtdo_buf[pc];
8687
8688 uint_t v = DIF_INSTR_VAR(instr);
8689 uint_t subr = DIF_INSTR_SUBR(instr);
8690 uint_t op = DIF_INSTR_OP(instr);
8691
8692 switch (op) {
8693 case DIF_OP_OR:
8694 case DIF_OP_XOR:
8695 case DIF_OP_AND:
8696 case DIF_OP_SLL:
8697 case DIF_OP_SRL:
8698 case DIF_OP_SRA:
8699 case DIF_OP_SUB:
8700 case DIF_OP_ADD:
8701 case DIF_OP_MUL:
8702 case DIF_OP_SDIV:
8703 case DIF_OP_UDIV:
8704 case DIF_OP_SREM:
8705 case DIF_OP_UREM:
8706 case DIF_OP_COPYS:
8707 case DIF_OP_NOT:
8708 case DIF_OP_MOV:
8709 case DIF_OP_RLDSB:
8710 case DIF_OP_RLDSH:
8711 case DIF_OP_RLDSW:
8712 case DIF_OP_RLDUB:
8713 case DIF_OP_RLDUH:
8714 case DIF_OP_RLDUW:
8715 case DIF_OP_RLDX:
8716 case DIF_OP_ULDSB:
8717 case DIF_OP_ULDSH:
8718 case DIF_OP_ULDSW:
8719 case DIF_OP_ULDUB:
8720 case DIF_OP_ULDUH:
8721 case DIF_OP_ULDUW:
8722 case DIF_OP_ULDX:
8723 case DIF_OP_STB:
8724 case DIF_OP_STH:
8725 case DIF_OP_STW:
8726 case DIF_OP_STX:
8727 case DIF_OP_ALLOCS:
8728 case DIF_OP_CMP:
8729 case DIF_OP_SCMP:
8730 case DIF_OP_TST:
8731 case DIF_OP_BA:
8732 case DIF_OP_BE:
8733 case DIF_OP_BNE:
8734 case DIF_OP_BG:
8735 case DIF_OP_BGU:
8736 case DIF_OP_BGE:
8737 case DIF_OP_BGEU:
8738 case DIF_OP_BL:
8739 case DIF_OP_BLU:
8740 case DIF_OP_BLE:
8741 case DIF_OP_BLEU:
8742 case DIF_OP_RET:
8743 case DIF_OP_NOP:
8744 case DIF_OP_POPTS:
8745 case DIF_OP_FLUSHTS:
8746 case DIF_OP_SETX:
8747 case DIF_OP_SETS:
8748 case DIF_OP_LDGA:
8749 case DIF_OP_LDLS:
8750 case DIF_OP_STGS:
8751 case DIF_OP_STLS:
8752 case DIF_OP_PUSHTR:
8753 case DIF_OP_PUSHTV:
8754 break;
8755
8756 case DIF_OP_LDGS:
8757 if (v >= DIF_VAR_OTHER_UBASE)
8758 break;
8759
8760 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8761 break;
8762
8763 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8764 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8765 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8766 v == DIF_VAR_UID || v == DIF_VAR_GID)
8767 break;
8768
8769 err += efunc(pc, "illegal variable %u\n", v);
8770 break;
8771
8772 case DIF_OP_LDTA:
8773 case DIF_OP_LDTS:
8774 case DIF_OP_LDGAA:
8775 case DIF_OP_LDTAA:
8776 err += efunc(pc, "illegal dynamic variable load\n");
8777 break;
8778
8779 case DIF_OP_STTS:
8780 case DIF_OP_STGAA:
8781 case DIF_OP_STTAA:
8782 err += efunc(pc, "illegal dynamic variable store\n");
8783 break;
8784
8785 case DIF_OP_CALL:
8786 if (subr == DIF_SUBR_ALLOCA ||
8787 subr == DIF_SUBR_BCOPY ||
8788 subr == DIF_SUBR_COPYIN ||
8789 subr == DIF_SUBR_COPYINTO ||
8790 subr == DIF_SUBR_COPYINSTR ||
8791 subr == DIF_SUBR_INDEX ||
8792 subr == DIF_SUBR_INET_NTOA ||
8793 subr == DIF_SUBR_INET_NTOA6 ||
8794 subr == DIF_SUBR_INET_NTOP ||
8795 subr == DIF_SUBR_LLTOSTR ||
8796 subr == DIF_SUBR_RINDEX ||
8797 subr == DIF_SUBR_STRCHR ||
8798 subr == DIF_SUBR_STRJOIN ||
8799 subr == DIF_SUBR_STRRCHR ||
8800 subr == DIF_SUBR_STRSTR ||
8801 subr == DIF_SUBR_HTONS ||
8802 subr == DIF_SUBR_HTONL ||
8803 subr == DIF_SUBR_HTONLL ||
8804 subr == DIF_SUBR_NTOHS ||
8805 subr == DIF_SUBR_NTOHL ||
8806 subr == DIF_SUBR_NTOHLL)
8807 break;
8808
8809 err += efunc(pc, "invalid subr %u\n", subr);
8810 break;
8811
8812 default:
8813 err += efunc(pc, "invalid opcode %u\n",
8814 DIF_INSTR_OP(instr));
8815 }
8816 }
8817
8818 return (err);
8819 }
8820
8821 /*
8822 * Returns 1 if the expression in the DIF object can be cached on a per-thread
8823 * basis; 0 if not.
8824 */
8825 static int
8826 dtrace_difo_cacheable(dtrace_difo_t *dp)
8827 {
8828 int i;
8829
8830 if (dp == NULL)
8831 return (0);
8832
8833 for (i = 0; i < dp->dtdo_varlen; i++) {
8834 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8835
8836 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8837 continue;
8838
8839 switch (v->dtdv_id) {
8840 case DIF_VAR_CURTHREAD:
8841 case DIF_VAR_PID:
8842 case DIF_VAR_TID:
8843 case DIF_VAR_EXECNAME:
8844 case DIF_VAR_ZONENAME:
8845 break;
8846
8847 default:
8848 return (0);
8849 }
8850 }
8851
8852 /*
8853 * This DIF object may be cacheable. Now we need to look for any
8854 * array loading instructions, any memory loading instructions, or
8855 * any stores to thread-local variables.
8856 */
8857 for (i = 0; i < dp->dtdo_len; i++) {
8858 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
8859
8860 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
8861 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
8862 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
8863 op == DIF_OP_LDGA || op == DIF_OP_STTS)
8864 return (0);
8865 }
8866
8867 return (1);
8868 }
8869
8870 static void
8871 dtrace_difo_hold(dtrace_difo_t *dp)
8872 {
8873 int i;
8874
8875 ASSERT(MUTEX_HELD(&dtrace_lock));
8876
8877 dp->dtdo_refcnt++;
8878 ASSERT(dp->dtdo_refcnt != 0);
8879
8880 /*
8881 * We need to check this DIF object for references to the variable
8882 * DIF_VAR_VTIMESTAMP.
8883 */
8884 for (i = 0; i < dp->dtdo_varlen; i++) {
8885 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8886
8887 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8888 continue;
8889
8890 if (dtrace_vtime_references++ == 0)
8891 dtrace_vtime_enable();
8892 }
8893 }
8894
8895 /*
8896 * This routine calculates the dynamic variable chunksize for a given DIF
8897 * object. The calculation is not fool-proof, and can probably be tricked by
8898 * malicious DIF -- but it works for all compiler-generated DIF. Because this
8899 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
8900 * if a dynamic variable size exceeds the chunksize.
8901 */
8902 static void
8903 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8904 {
8905 uint64_t sval;
8906 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
8907 const dif_instr_t *text = dp->dtdo_buf;
8908 uint_t pc, srd = 0;
8909 uint_t ttop = 0;
8910 size_t size, ksize;
8911 uint_t id, i;
8912
8913 for (pc = 0; pc < dp->dtdo_len; pc++) {
8914 dif_instr_t instr = text[pc];
8915 uint_t op = DIF_INSTR_OP(instr);
8916 uint_t rd = DIF_INSTR_RD(instr);
8917 uint_t r1 = DIF_INSTR_R1(instr);
8918 uint_t nkeys = 0;
8919 uchar_t scope;
8920
8921 dtrace_key_t *key = tupregs;
8922
8923 switch (op) {
8924 case DIF_OP_SETX:
8925 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
8926 srd = rd;
8927 continue;
8928
8929 case DIF_OP_STTS:
8930 key = &tupregs[DIF_DTR_NREGS];
8931 key[0].dttk_size = 0;
8932 key[1].dttk_size = 0;
8933 nkeys = 2;
8934 scope = DIFV_SCOPE_THREAD;
8935 break;
8936
8937 case DIF_OP_STGAA:
8938 case DIF_OP_STTAA:
8939 nkeys = ttop;
8940
8941 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
8942 key[nkeys++].dttk_size = 0;
8943
8944 key[nkeys++].dttk_size = 0;
8945
8946 if (op == DIF_OP_STTAA) {
8947 scope = DIFV_SCOPE_THREAD;
8948 } else {
8949 scope = DIFV_SCOPE_GLOBAL;
8950 }
8951
8952 break;
8953
8954 case DIF_OP_PUSHTR:
8955 if (ttop == DIF_DTR_NREGS)
8956 return;
8957
8958 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
8959 /*
8960 * If the register for the size of the "pushtr"
8961 * is %r0 (or the value is 0) and the type is
8962 * a string, we'll use the system-wide default
8963 * string size.
8964 */
8965 tupregs[ttop++].dttk_size =
8966 dtrace_strsize_default;
8967 } else {
8968 if (srd == 0)
8969 return;
8970
8971 tupregs[ttop++].dttk_size = sval;
8972 }
8973
8974 break;
8975
8976 case DIF_OP_PUSHTV:
8977 if (ttop == DIF_DTR_NREGS)
8978 return;
8979
8980 tupregs[ttop++].dttk_size = 0;
8981 break;
8982
8983 case DIF_OP_FLUSHTS:
8984 ttop = 0;
8985 break;
8986
8987 case DIF_OP_POPTS:
8988 if (ttop != 0)
8989 ttop--;
8990 break;
8991 }
8992
8993 sval = 0;
8994 srd = 0;
8995
8996 if (nkeys == 0)
8997 continue;
8998
8999 /*
9000 * We have a dynamic variable allocation; calculate its size.
9001 */
9002 for (ksize = 0, i = 0; i < nkeys; i++)
9003 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9004
9005 size = sizeof (dtrace_dynvar_t);
9006 size += sizeof (dtrace_key_t) * (nkeys - 1);
9007 size += ksize;
9008
9009 /*
9010 * Now we need to determine the size of the stored data.
9011 */
9012 id = DIF_INSTR_VAR(instr);
9013
9014 for (i = 0; i < dp->dtdo_varlen; i++) {
9015 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9016
9017 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9018 size += v->dtdv_type.dtdt_size;
9019 break;
9020 }
9021 }
9022
9023 if (i == dp->dtdo_varlen)
9024 return;
9025
9026 /*
9027 * We have the size. If this is larger than the chunk size
9028 * for our dynamic variable state, reset the chunk size.
9029 */
9030 size = P2ROUNDUP(size, sizeof (uint64_t));
9031
9032 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9033 vstate->dtvs_dynvars.dtds_chunksize = size;
9034 }
9035 }
9036
9037 static void
9038 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9039 {
9040 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9041 uint_t id;
9042
9043 ASSERT(MUTEX_HELD(&dtrace_lock));
9044 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9045
9046 for (i = 0; i < dp->dtdo_varlen; i++) {
9047 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9048 dtrace_statvar_t *svar, ***svarp;
9049 size_t dsize = 0;
9050 uint8_t scope = v->dtdv_scope;
9051 int *np;
9052
9053 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9054 continue;
9055
9056 id -= DIF_VAR_OTHER_UBASE;
9057
9058 switch (scope) {
9059 case DIFV_SCOPE_THREAD:
9060 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9061 dtrace_difv_t *tlocals;
9062
9063 if ((ntlocals = (otlocals << 1)) == 0)
9064 ntlocals = 1;
9065
9066 osz = otlocals * sizeof (dtrace_difv_t);
9067 nsz = ntlocals * sizeof (dtrace_difv_t);
9068
9069 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9070
9071 if (osz != 0) {
9072 bcopy(vstate->dtvs_tlocals,
9073 tlocals, osz);
9074 kmem_free(vstate->dtvs_tlocals, osz);
9075 }
9076
9077 vstate->dtvs_tlocals = tlocals;
9078 vstate->dtvs_ntlocals = ntlocals;
9079 }
9080
9081 vstate->dtvs_tlocals[id] = *v;
9082 continue;
9083
9084 case DIFV_SCOPE_LOCAL:
9085 np = &vstate->dtvs_nlocals;
9086 svarp = &vstate->dtvs_locals;
9087
9088 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9089 dsize = NCPU * (v->dtdv_type.dtdt_size +
9090 sizeof (uint64_t));
9091 else
9092 dsize = NCPU * sizeof (uint64_t);
9093
9094 break;
9095
9096 case DIFV_SCOPE_GLOBAL:
9097 np = &vstate->dtvs_nglobals;
9098 svarp = &vstate->dtvs_globals;
9099
9100 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9101 dsize = v->dtdv_type.dtdt_size +
9102 sizeof (uint64_t);
9103
9104 break;
9105
9106 default:
9107 ASSERT(0);
9108 }
9109
9110 while (id >= (oldsvars = *np)) {
9111 dtrace_statvar_t **statics;
9112 int newsvars, oldsize, newsize;
9113
9114 if ((newsvars = (oldsvars << 1)) == 0)
9115 newsvars = 1;
9116
9117 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9118 newsize = newsvars * sizeof (dtrace_statvar_t *);
9119
9120 statics = kmem_zalloc(newsize, KM_SLEEP);
9121
9122 if (oldsize != 0) {
9123 bcopy(*svarp, statics, oldsize);
9124 kmem_free(*svarp, oldsize);
9125 }
9126
9127 *svarp = statics;
9128 *np = newsvars;
9129 }
9130
9131 if ((svar = (*svarp)[id]) == NULL) {
9132 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9133 svar->dtsv_var = *v;
9134
9135 if ((svar->dtsv_size = dsize) != 0) {
9136 svar->dtsv_data = (uint64_t)(uintptr_t)
9137 kmem_zalloc(dsize, KM_SLEEP);
9138 }
9139
9140 (*svarp)[id] = svar;
9141 }
9142
9143 svar->dtsv_refcnt++;
9144 }
9145
9146 dtrace_difo_chunksize(dp, vstate);
9147 dtrace_difo_hold(dp);
9148 }
9149
9150 static dtrace_difo_t *
9151 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9152 {
9153 dtrace_difo_t *new;
9154 size_t sz;
9155
9156 ASSERT(dp->dtdo_buf != NULL);
9157 ASSERT(dp->dtdo_refcnt != 0);
9158
9159 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9160
9161 ASSERT(dp->dtdo_buf != NULL);
9162 sz = dp->dtdo_len * sizeof (dif_instr_t);
9163 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9164 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9165 new->dtdo_len = dp->dtdo_len;
9166
9167 if (dp->dtdo_strtab != NULL) {
9168 ASSERT(dp->dtdo_strlen != 0);
9169 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9170 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9171 new->dtdo_strlen = dp->dtdo_strlen;
9172 }
9173
9174 if (dp->dtdo_inttab != NULL) {
9175 ASSERT(dp->dtdo_intlen != 0);
9176 sz = dp->dtdo_intlen * sizeof (uint64_t);
9177 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9178 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9179 new->dtdo_intlen = dp->dtdo_intlen;
9180 }
9181
9182 if (dp->dtdo_vartab != NULL) {
9183 ASSERT(dp->dtdo_varlen != 0);
9184 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9185 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9186 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9187 new->dtdo_varlen = dp->dtdo_varlen;
9188 }
9189
9190 dtrace_difo_init(new, vstate);
9191 return (new);
9192 }
9193
9194 static void
9195 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9196 {
9197 int i;
9198
9199 ASSERT(dp->dtdo_refcnt == 0);
9200
9201 for (i = 0; i < dp->dtdo_varlen; i++) {
9202 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9203 dtrace_statvar_t *svar, **svarp;
9204 uint_t id;
9205 uint8_t scope = v->dtdv_scope;
9206 int *np;
9207
9208 switch (scope) {
9209 case DIFV_SCOPE_THREAD:
9210 continue;
9211
9212 case DIFV_SCOPE_LOCAL:
9213 np = &vstate->dtvs_nlocals;
9214 svarp = vstate->dtvs_locals;
9215 break;
9216
9217 case DIFV_SCOPE_GLOBAL:
9218 np = &vstate->dtvs_nglobals;
9219 svarp = vstate->dtvs_globals;
9220 break;
9221
9222 default:
9223 ASSERT(0);
9224 }
9225
9226 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9227 continue;
9228
9229 id -= DIF_VAR_OTHER_UBASE;
9230 ASSERT(id < *np);
9231
9232 svar = svarp[id];
9233 ASSERT(svar != NULL);
9234 ASSERT(svar->dtsv_refcnt > 0);
9235
9236 if (--svar->dtsv_refcnt > 0)
9237 continue;
9238
9239 if (svar->dtsv_size != 0) {
9240 ASSERT(svar->dtsv_data != NULL);
9241 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9242 svar->dtsv_size);
9243 }
9244
9245 kmem_free(svar, sizeof (dtrace_statvar_t));
9246 svarp[id] = NULL;
9247 }
9248
9249 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9250 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9251 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9252 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9253
9254 kmem_free(dp, sizeof (dtrace_difo_t));
9255 }
9256
9257 static void
9258 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9259 {
9260 int i;
9261
9262 ASSERT(MUTEX_HELD(&dtrace_lock));
9263 ASSERT(dp->dtdo_refcnt != 0);
9264
9265 for (i = 0; i < dp->dtdo_varlen; i++) {
9266 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9267
9268 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9269 continue;
9270
9271 ASSERT(dtrace_vtime_references > 0);
9272 if (--dtrace_vtime_references == 0)
9273 dtrace_vtime_disable();
9274 }
9275
9276 if (--dp->dtdo_refcnt == 0)
9277 dtrace_difo_destroy(dp, vstate);
9278 }
9279
9280 /*
9281 * DTrace Format Functions
9282 */
9283 static uint16_t
9284 dtrace_format_add(dtrace_state_t *state, char *str)
9285 {
9286 char *fmt, **new;
9287 uint16_t ndx, len = strlen(str) + 1;
9288
9289 fmt = kmem_zalloc(len, KM_SLEEP);
9290 bcopy(str, fmt, len);
9291
9292 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9293 if (state->dts_formats[ndx] == NULL) {
9294 state->dts_formats[ndx] = fmt;
9295 return (ndx + 1);
9296 }
9297 }
9298
9299 if (state->dts_nformats == USHRT_MAX) {
9300 /*
9301 * This is only likely if a denial-of-service attack is being
9302 * attempted. As such, it's okay to fail silently here.
9303 */
9304 kmem_free(fmt, len);
9305 return (0);
9306 }
9307
9308 /*
9309 * For simplicity, we always resize the formats array to be exactly the
9310 * number of formats.
9311 */
9312 ndx = state->dts_nformats++;
9313 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9314
9315 if (state->dts_formats != NULL) {
9316 ASSERT(ndx != 0);
9317 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9318 kmem_free(state->dts_formats, ndx * sizeof (char *));
9319 }
9320
9321 state->dts_formats = new;
9322 state->dts_formats[ndx] = fmt;
9323
9324 return (ndx + 1);
9325 }
9326
9327 static void
9328 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9329 {
9330 char *fmt;
9331
9332 ASSERT(state->dts_formats != NULL);
9333 ASSERT(format <= state->dts_nformats);
9334 ASSERT(state->dts_formats[format - 1] != NULL);
9335
9336 fmt = state->dts_formats[format - 1];
9337 kmem_free(fmt, strlen(fmt) + 1);
9338 state->dts_formats[format - 1] = NULL;
9339 }
9340
9341 static void
9342 dtrace_format_destroy(dtrace_state_t *state)
9343 {
9344 int i;
9345
9346 if (state->dts_nformats == 0) {
9347 ASSERT(state->dts_formats == NULL);
9348 return;
9349 }
9350
9351 ASSERT(state->dts_formats != NULL);
9352
9353 for (i = 0; i < state->dts_nformats; i++) {
9354 char *fmt = state->dts_formats[i];
9355
9356 if (fmt == NULL)
9357 continue;
9358
9359 kmem_free(fmt, strlen(fmt) + 1);
9360 }
9361
9362 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9363 state->dts_nformats = 0;
9364 state->dts_formats = NULL;
9365 }
9366
9367 /*
9368 * DTrace Predicate Functions
9369 */
9370 static dtrace_predicate_t *
9371 dtrace_predicate_create(dtrace_difo_t *dp)
9372 {
9373 dtrace_predicate_t *pred;
9374
9375 ASSERT(MUTEX_HELD(&dtrace_lock));
9376 ASSERT(dp->dtdo_refcnt != 0);
9377
9378 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9379 pred->dtp_difo = dp;
9380 pred->dtp_refcnt = 1;
9381
9382 if (!dtrace_difo_cacheable(dp))
9383 return (pred);
9384
9385 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9386 /*
9387 * This is only theoretically possible -- we have had 2^32
9388 * cacheable predicates on this machine. We cannot allow any
9389 * more predicates to become cacheable: as unlikely as it is,
9390 * there may be a thread caching a (now stale) predicate cache
9391 * ID. (N.B.: the temptation is being successfully resisted to
9392 * have this cmn_err() "Holy shit -- we executed this code!")
9393 */
9394 return (pred);
9395 }
9396
9397 pred->dtp_cacheid = dtrace_predcache_id++;
9398
9399 return (pred);
9400 }
9401
9402 static void
9403 dtrace_predicate_hold(dtrace_predicate_t *pred)
9404 {
9405 ASSERT(MUTEX_HELD(&dtrace_lock));
9406 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9407 ASSERT(pred->dtp_refcnt > 0);
9408
9409 pred->dtp_refcnt++;
9410 }
9411
9412 static void
9413 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9414 {
9415 dtrace_difo_t *dp = pred->dtp_difo;
9416
9417 ASSERT(MUTEX_HELD(&dtrace_lock));
9418 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9419 ASSERT(pred->dtp_refcnt > 0);
9420
9421 if (--pred->dtp_refcnt == 0) {
9422 dtrace_difo_release(pred->dtp_difo, vstate);
9423 kmem_free(pred, sizeof (dtrace_predicate_t));
9424 }
9425 }
9426
9427 /*
9428 * DTrace Action Description Functions
9429 */
9430 static dtrace_actdesc_t *
9431 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9432 uint64_t uarg, uint64_t arg)
9433 {
9434 dtrace_actdesc_t *act;
9435
9436 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9437 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9438
9439 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9440 act->dtad_kind = kind;
9441 act->dtad_ntuple = ntuple;
9442 act->dtad_uarg = uarg;
9443 act->dtad_arg = arg;
9444 act->dtad_refcnt = 1;
9445
9446 return (act);
9447 }
9448
9449 static void
9450 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9451 {
9452 ASSERT(act->dtad_refcnt >= 1);
9453 act->dtad_refcnt++;
9454 }
9455
9456 static void
9457 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9458 {
9459 dtrace_actkind_t kind = act->dtad_kind;
9460 dtrace_difo_t *dp;
9461
9462 ASSERT(act->dtad_refcnt >= 1);
9463
9464 if (--act->dtad_refcnt != 0)
9465 return;
9466
9467 if ((dp = act->dtad_difo) != NULL)
9468 dtrace_difo_release(dp, vstate);
9469
9470 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9471 char *str = (char *)(uintptr_t)act->dtad_arg;
9472
9473 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9474 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9475
9476 if (str != NULL)
9477 kmem_free(str, strlen(str) + 1);
9478 }
9479
9480 kmem_free(act, sizeof (dtrace_actdesc_t));
9481 }
9482
9483 /*
9484 * DTrace ECB Functions
9485 */
9486 static dtrace_ecb_t *
9487 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9488 {
9489 dtrace_ecb_t *ecb;
9490 dtrace_epid_t epid;
9491
9492 ASSERT(MUTEX_HELD(&dtrace_lock));
9493
9494 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9495 ecb->dte_predicate = NULL;
9496 ecb->dte_probe = probe;
9497
9498 /*
9499 * The default size is the size of the default action: recording
9500 * the epid.
9501 */
9502 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9503 ecb->dte_alignment = sizeof (dtrace_epid_t);
9504
9505 epid = state->dts_epid++;
9506
9507 if (epid - 1 >= state->dts_necbs) {
9508 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9509 int necbs = state->dts_necbs << 1;
9510
9511 ASSERT(epid == state->dts_necbs + 1);
9512
9513 if (necbs == 0) {
9514 ASSERT(oecbs == NULL);
9515 necbs = 1;
9516 }
9517
9518 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9519
9520 if (oecbs != NULL)
9521 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9522
9523 dtrace_membar_producer();
9524 state->dts_ecbs = ecbs;
9525
9526 if (oecbs != NULL) {
9527 /*
9528 * If this state is active, we must dtrace_sync()
9529 * before we can free the old dts_ecbs array: we're
9530 * coming in hot, and there may be active ring
9531 * buffer processing (which indexes into the dts_ecbs
9532 * array) on another CPU.
9533 */
9534 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9535 dtrace_sync();
9536
9537 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9538 }
9539
9540 dtrace_membar_producer();
9541 state->dts_necbs = necbs;
9542 }
9543
9544 ecb->dte_state = state;
9545
9546 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9547 dtrace_membar_producer();
9548 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9549
9550 return (ecb);
9551 }
9552
9553 static int
9554 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9555 {
9556 dtrace_probe_t *probe = ecb->dte_probe;
9557
9558 ASSERT(MUTEX_HELD(&cpu_lock));
9559 ASSERT(MUTEX_HELD(&dtrace_lock));
9560 ASSERT(ecb->dte_next == NULL);
9561
9562 if (probe == NULL) {
9563 /*
9564 * This is the NULL probe -- there's nothing to do.
9565 */
9566 return (0);
9567 }
9568
9569 if (probe->dtpr_ecb == NULL) {
9570 dtrace_provider_t *prov = probe->dtpr_provider;
9571
9572 /*
9573 * We're the first ECB on this probe.
9574 */
9575 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9576
9577 if (ecb->dte_predicate != NULL)
9578 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9579
9580 return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9581 probe->dtpr_id, probe->dtpr_arg));
9582 } else {
9583 /*
9584 * This probe is already active. Swing the last pointer to
9585 * point to the new ECB, and issue a dtrace_sync() to assure
9586 * that all CPUs have seen the change.
9587 */
9588 ASSERT(probe->dtpr_ecb_last != NULL);
9589 probe->dtpr_ecb_last->dte_next = ecb;
9590 probe->dtpr_ecb_last = ecb;
9591 probe->dtpr_predcache = 0;
9592
9593 dtrace_sync();
9594 return (0);
9595 }
9596 }
9597
9598 static void
9599 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9600 {
9601 uint32_t maxalign = sizeof (dtrace_epid_t);
9602 uint32_t align = sizeof (uint8_t), offs, diff;
9603 dtrace_action_t *act;
9604 int wastuple = 0;
9605 uint32_t aggbase = UINT32_MAX;
9606 dtrace_state_t *state = ecb->dte_state;
9607
9608 /*
9609 * If we record anything, we always record the epid. (And we always
9610 * record it first.)
9611 */
9612 offs = sizeof (dtrace_epid_t);
9613 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9614
9615 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9616 dtrace_recdesc_t *rec = &act->dta_rec;
9617
9618 if ((align = rec->dtrd_alignment) > maxalign)
9619 maxalign = align;
9620
9621 if (!wastuple && act->dta_intuple) {
9622 /*
9623 * This is the first record in a tuple. Align the
9624 * offset to be at offset 4 in an 8-byte aligned
9625 * block.
9626 */
9627 diff = offs + sizeof (dtrace_aggid_t);
9628
9629 if (diff = (diff & (sizeof (uint64_t) - 1)))
9630 offs += sizeof (uint64_t) - diff;
9631
9632 aggbase = offs - sizeof (dtrace_aggid_t);
9633 ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9634 }
9635
9636 /*LINTED*/
9637 if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9638 /*
9639 * The current offset is not properly aligned; align it.
9640 */
9641 offs += align - diff;
9642 }
9643
9644 rec->dtrd_offset = offs;
9645
9646 if (offs + rec->dtrd_size > ecb->dte_needed) {
9647 ecb->dte_needed = offs + rec->dtrd_size;
9648
9649 if (ecb->dte_needed > state->dts_needed)
9650 state->dts_needed = ecb->dte_needed;
9651 }
9652
9653 if (DTRACEACT_ISAGG(act->dta_kind)) {
9654 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9655 dtrace_action_t *first = agg->dtag_first, *prev;
9656
9657 ASSERT(rec->dtrd_size != 0 && first != NULL);
9658 ASSERT(wastuple);
9659 ASSERT(aggbase != UINT32_MAX);
9660
9661 agg->dtag_base = aggbase;
9662
9663 while ((prev = first->dta_prev) != NULL &&
9664 DTRACEACT_ISAGG(prev->dta_kind)) {
9665 agg = (dtrace_aggregation_t *)prev;
9666 first = agg->dtag_first;
9667 }
9668
9669 if (prev != NULL) {
9670 offs = prev->dta_rec.dtrd_offset +
9671 prev->dta_rec.dtrd_size;
9672 } else {
9673 offs = sizeof (dtrace_epid_t);
9674 }
9675 wastuple = 0;
9676 } else {
9677 if (!act->dta_intuple)
9678 ecb->dte_size = offs + rec->dtrd_size;
9679
9680 offs += rec->dtrd_size;
9681 }
9682
9683 wastuple = act->dta_intuple;
9684 }
9685
9686 if ((act = ecb->dte_action) != NULL &&
9687 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9688 ecb->dte_size == sizeof (dtrace_epid_t)) {
9689 /*
9690 * If the size is still sizeof (dtrace_epid_t), then all
9691 * actions store no data; set the size to 0.
9692 */
9693 ecb->dte_alignment = maxalign;
9694 ecb->dte_size = 0;
9695
9696 /*
9697 * If the needed space is still sizeof (dtrace_epid_t), then
9698 * all actions need no additional space; set the needed
9699 * size to 0.
9700 */
9701 if (ecb->dte_needed == sizeof (dtrace_epid_t))
9702 ecb->dte_needed = 0;
9703
9704 return;
9705 }
9706
9707 /*
9708 * Set our alignment, and make sure that the dte_size and dte_needed
9709 * are aligned to the size of an EPID.
9710 */
9711 ecb->dte_alignment = maxalign;
9712 ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9713 ~(sizeof (dtrace_epid_t) - 1);
9714 ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9715 ~(sizeof (dtrace_epid_t) - 1);
9716 ASSERT(ecb->dte_size <= ecb->dte_needed);
9717 }
9718
9719 static dtrace_action_t *
9720 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9721 {
9722 dtrace_aggregation_t *agg;
9723 size_t size = sizeof (uint64_t);
9724 int ntuple = desc->dtad_ntuple;
9725 dtrace_action_t *act;
9726 dtrace_recdesc_t *frec;
9727 dtrace_aggid_t aggid;
9728 dtrace_state_t *state = ecb->dte_state;
9729
9730 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9731 agg->dtag_ecb = ecb;
9732
9733 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9734
9735 switch (desc->dtad_kind) {
9736 case DTRACEAGG_MIN:
9737 agg->dtag_initial = INT64_MAX;
9738 agg->dtag_aggregate = dtrace_aggregate_min;
9739 break;
9740
9741 case DTRACEAGG_MAX:
9742 agg->dtag_initial = INT64_MIN;
9743 agg->dtag_aggregate = dtrace_aggregate_max;
9744 break;
9745
9746 case DTRACEAGG_COUNT:
9747 agg->dtag_aggregate = dtrace_aggregate_count;
9748 break;
9749
9750 case DTRACEAGG_QUANTIZE:
9751 agg->dtag_aggregate = dtrace_aggregate_quantize;
9752 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9753 sizeof (uint64_t);
9754 break;
9755
9756 case DTRACEAGG_LQUANTIZE: {
9757 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9758 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9759
9760 agg->dtag_initial = desc->dtad_arg;
9761 agg->dtag_aggregate = dtrace_aggregate_lquantize;
9762
9763 if (step == 0 || levels == 0)
9764 goto err;
9765
9766 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9767 break;
9768 }
9769
9770 case DTRACEAGG_LLQUANTIZE: {
9771 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
9772 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
9773 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
9774 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
9775 int64_t v;
9776
9777 agg->dtag_initial = desc->dtad_arg;
9778 agg->dtag_aggregate = dtrace_aggregate_llquantize;
9779
9780 if (factor < 2 || low >= high || nsteps < factor)
9781 goto err;
9782
9783 /*
9784 * Now check that the number of steps evenly divides a power
9785 * of the factor. (This assures both integer bucket size and
9786 * linearity within each magnitude.)
9787 */
9788 for (v = factor; v < nsteps; v *= factor)
9789 continue;
9790
9791 if ((v % nsteps) || (nsteps % factor))
9792 goto err;
9793
9794 size = (dtrace_aggregate_llquantize_bucket(factor,
9795 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
9796 break;
9797 }
9798
9799 case DTRACEAGG_AVG:
9800 agg->dtag_aggregate = dtrace_aggregate_avg;
9801 size = sizeof (uint64_t) * 2;
9802 break;
9803
9804 case DTRACEAGG_STDDEV:
9805 agg->dtag_aggregate = dtrace_aggregate_stddev;
9806 size = sizeof (uint64_t) * 4;
9807 break;
9808
9809 case DTRACEAGG_SUM:
9810 agg->dtag_aggregate = dtrace_aggregate_sum;
9811 break;
9812
9813 default:
9814 goto err;
9815 }
9816
9817 agg->dtag_action.dta_rec.dtrd_size = size;
9818
9819 if (ntuple == 0)
9820 goto err;
9821
9822 /*
9823 * We must make sure that we have enough actions for the n-tuple.
9824 */
9825 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9826 if (DTRACEACT_ISAGG(act->dta_kind))
9827 break;
9828
9829 if (--ntuple == 0) {
9830 /*
9831 * This is the action with which our n-tuple begins.
9832 */
9833 agg->dtag_first = act;
9834 goto success;
9835 }
9836 }
9837
9838 /*
9839 * This n-tuple is short by ntuple elements. Return failure.
9840 */
9841 ASSERT(ntuple != 0);
9842 err:
9843 kmem_free(agg, sizeof (dtrace_aggregation_t));
9844 return (NULL);
9845
9846 success:
9847 /*
9848 * If the last action in the tuple has a size of zero, it's actually
9849 * an expression argument for the aggregating action.
9850 */
9851 ASSERT(ecb->dte_action_last != NULL);
9852 act = ecb->dte_action_last;
9853
9854 if (act->dta_kind == DTRACEACT_DIFEXPR) {
9855 ASSERT(act->dta_difo != NULL);
9856
9857 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
9858 agg->dtag_hasarg = 1;
9859 }
9860
9861 /*
9862 * We need to allocate an id for this aggregation.
9863 */
9864 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
9865 VM_BESTFIT | VM_SLEEP);
9866
9867 if (aggid - 1 >= state->dts_naggregations) {
9868 dtrace_aggregation_t **oaggs = state->dts_aggregations;
9869 dtrace_aggregation_t **aggs;
9870 int naggs = state->dts_naggregations << 1;
9871 int onaggs = state->dts_naggregations;
9872
9873 ASSERT(aggid == state->dts_naggregations + 1);
9874
9875 if (naggs == 0) {
9876 ASSERT(oaggs == NULL);
9877 naggs = 1;
9878 }
9879
9880 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
9881
9882 if (oaggs != NULL) {
9883 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
9884 kmem_free(oaggs, onaggs * sizeof (*aggs));
9885 }
9886
9887 state->dts_aggregations = aggs;
9888 state->dts_naggregations = naggs;
9889 }
9890
9891 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
9892 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
9893
9894 frec = &agg->dtag_first->dta_rec;
9895 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
9896 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
9897
9898 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
9899 ASSERT(!act->dta_intuple);
9900 act->dta_intuple = 1;
9901 }
9902
9903 return (&agg->dtag_action);
9904 }
9905
9906 static void
9907 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
9908 {
9909 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9910 dtrace_state_t *state = ecb->dte_state;
9911 dtrace_aggid_t aggid = agg->dtag_id;
9912
9913 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
9914 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
9915
9916 ASSERT(state->dts_aggregations[aggid - 1] == agg);
9917 state->dts_aggregations[aggid - 1] = NULL;
9918
9919 kmem_free(agg, sizeof (dtrace_aggregation_t));
9920 }
9921
9922 static int
9923 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9924 {
9925 dtrace_action_t *action, *last;
9926 dtrace_difo_t *dp = desc->dtad_difo;
9927 uint32_t size = 0, align = sizeof (uint8_t), mask;
9928 uint16_t format = 0;
9929 dtrace_recdesc_t *rec;
9930 dtrace_state_t *state = ecb->dte_state;
9931 dtrace_optval_t *opt = state->dts_options, nframes, strsize;
9932 uint64_t arg = desc->dtad_arg;
9933
9934 ASSERT(MUTEX_HELD(&dtrace_lock));
9935 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
9936
9937 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
9938 /*
9939 * If this is an aggregating action, there must be neither
9940 * a speculate nor a commit on the action chain.
9941 */
9942 dtrace_action_t *act;
9943
9944 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9945 if (act->dta_kind == DTRACEACT_COMMIT)
9946 return (EINVAL);
9947
9948 if (act->dta_kind == DTRACEACT_SPECULATE)
9949 return (EINVAL);
9950 }
9951
9952 action = dtrace_ecb_aggregation_create(ecb, desc);
9953
9954 if (action == NULL)
9955 return (EINVAL);
9956 } else {
9957 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
9958 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
9959 dp != NULL && dp->dtdo_destructive)) {
9960 state->dts_destructive = 1;
9961 }
9962
9963 switch (desc->dtad_kind) {
9964 case DTRACEACT_PRINTF:
9965 case DTRACEACT_PRINTA:
9966 case DTRACEACT_SYSTEM:
9967 case DTRACEACT_FREOPEN:
9968 case DTRACEACT_DIFEXPR:
9969 /*
9970 * We know that our arg is a string -- turn it into a
9971 * format.
9972 */
9973 if (arg == NULL) {
9974 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
9975 desc->dtad_kind == DTRACEACT_DIFEXPR);
9976 format = 0;
9977 } else {
9978 ASSERT(arg != NULL);
9979 ASSERT(arg > KERNELBASE);
9980 format = dtrace_format_add(state,
9981 (char *)(uintptr_t)arg);
9982 }
9983
9984 /*FALLTHROUGH*/
9985 case DTRACEACT_LIBACT:
9986 case DTRACEACT_TRACEMEM:
9987 case DTRACEACT_TRACEMEM_DYNSIZE:
9988 if (dp == NULL)
9989 return (EINVAL);
9990
9991 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
9992 break;
9993
9994 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
9995 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9996 return (EINVAL);
9997
9998 size = opt[DTRACEOPT_STRSIZE];
9999 }
10000
10001 break;
10002
10003 case DTRACEACT_STACK:
10004 if ((nframes = arg) == 0) {
10005 nframes = opt[DTRACEOPT_STACKFRAMES];
10006 ASSERT(nframes > 0);
10007 arg = nframes;
10008 }
10009
10010 size = nframes * sizeof (pc_t);
10011 break;
10012
10013 case DTRACEACT_JSTACK:
10014 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10015 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10016
10017 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10018 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10019
10020 arg = DTRACE_USTACK_ARG(nframes, strsize);
10021
10022 /*FALLTHROUGH*/
10023 case DTRACEACT_USTACK:
10024 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10025 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10026 strsize = DTRACE_USTACK_STRSIZE(arg);
10027 nframes = opt[DTRACEOPT_USTACKFRAMES];
10028 ASSERT(nframes > 0);
10029 arg = DTRACE_USTACK_ARG(nframes, strsize);
10030 }
10031
10032 /*
10033 * Save a slot for the pid.
10034 */
10035 size = (nframes + 1) * sizeof (uint64_t);
10036 size += DTRACE_USTACK_STRSIZE(arg);
10037 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10038
10039 break;
10040
10041 case DTRACEACT_SYM:
10042 case DTRACEACT_MOD:
10043 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10044 sizeof (uint64_t)) ||
10045 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10046 return (EINVAL);
10047 break;
10048
10049 case DTRACEACT_USYM:
10050 case DTRACEACT_UMOD:
10051 case DTRACEACT_UADDR:
10052 if (dp == NULL ||
10053 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10054 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10055 return (EINVAL);
10056
10057 /*
10058 * We have a slot for the pid, plus a slot for the
10059 * argument. To keep things simple (aligned with
10060 * bitness-neutral sizing), we store each as a 64-bit
10061 * quantity.
10062 */
10063 size = 2 * sizeof (uint64_t);
10064 break;
10065
10066 case DTRACEACT_STOP:
10067 case DTRACEACT_BREAKPOINT:
10068 case DTRACEACT_PANIC:
10069 break;
10070
10071 case DTRACEACT_CHILL:
10072 case DTRACEACT_DISCARD:
10073 case DTRACEACT_RAISE:
10074 if (dp == NULL)
10075 return (EINVAL);
10076 break;
10077
10078 case DTRACEACT_EXIT:
10079 if (dp == NULL ||
10080 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10081 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10082 return (EINVAL);
10083 break;
10084
10085 case DTRACEACT_SPECULATE:
10086 if (ecb->dte_size > sizeof (dtrace_epid_t))
10087 return (EINVAL);
10088
10089 if (dp == NULL)
10090 return (EINVAL);
10091
10092 state->dts_speculates = 1;
10093 break;
10094
10095 case DTRACEACT_COMMIT: {
10096 dtrace_action_t *act = ecb->dte_action;
10097
10098 for (; act != NULL; act = act->dta_next) {
10099 if (act->dta_kind == DTRACEACT_COMMIT)
10100 return (EINVAL);
10101 }
10102
10103 if (dp == NULL)
10104 return (EINVAL);
10105 break;
10106 }
10107
10108 default:
10109 return (EINVAL);
10110 }
10111
10112 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10113 /*
10114 * If this is a data-storing action or a speculate,
10115 * we must be sure that there isn't a commit on the
10116 * action chain.
10117 */
10118 dtrace_action_t *act = ecb->dte_action;
10119
10120 for (; act != NULL; act = act->dta_next) {
10121 if (act->dta_kind == DTRACEACT_COMMIT)
10122 return (EINVAL);
10123 }
10124 }
10125
10126 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10127 action->dta_rec.dtrd_size = size;
10128 }
10129
10130 action->dta_refcnt = 1;
10131 rec = &action->dta_rec;
10132 size = rec->dtrd_size;
10133
10134 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10135 if (!(size & mask)) {
10136 align = mask + 1;
10137 break;
10138 }
10139 }
10140
10141 action->dta_kind = desc->dtad_kind;
10142
10143 if ((action->dta_difo = dp) != NULL)
10144 dtrace_difo_hold(dp);
10145
10146 rec->dtrd_action = action->dta_kind;
10147 rec->dtrd_arg = arg;
10148 rec->dtrd_uarg = desc->dtad_uarg;
10149 rec->dtrd_alignment = (uint16_t)align;
10150 rec->dtrd_format = format;
10151
10152 if ((last = ecb->dte_action_last) != NULL) {
10153 ASSERT(ecb->dte_action != NULL);
10154 action->dta_prev = last;
10155 last->dta_next = action;
10156 } else {
10157 ASSERT(ecb->dte_action == NULL);
10158 ecb->dte_action = action;
10159 }
10160
10161 ecb->dte_action_last = action;
10162
10163 return (0);
10164 }
10165
10166 static void
10167 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10168 {
10169 dtrace_action_t *act = ecb->dte_action, *next;
10170 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10171 dtrace_difo_t *dp;
10172 uint16_t format;
10173
10174 if (act != NULL && act->dta_refcnt > 1) {
10175 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10176 act->dta_refcnt--;
10177 } else {
10178 for (; act != NULL; act = next) {
10179 next = act->dta_next;
10180 ASSERT(next != NULL || act == ecb->dte_action_last);
10181 ASSERT(act->dta_refcnt == 1);
10182
10183 if ((format = act->dta_rec.dtrd_format) != 0)
10184 dtrace_format_remove(ecb->dte_state, format);
10185
10186 if ((dp = act->dta_difo) != NULL)
10187 dtrace_difo_release(dp, vstate);
10188
10189 if (DTRACEACT_ISAGG(act->dta_kind)) {
10190 dtrace_ecb_aggregation_destroy(ecb, act);
10191 } else {
10192 kmem_free(act, sizeof (dtrace_action_t));
10193 }
10194 }
10195 }
10196
10197 ecb->dte_action = NULL;
10198 ecb->dte_action_last = NULL;
10199 ecb->dte_size = sizeof (dtrace_epid_t);
10200 }
10201
10202 static void
10203 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10204 {
10205 /*
10206 * We disable the ECB by removing it from its probe.
10207 */
10208 dtrace_ecb_t *pecb, *prev = NULL;
10209 dtrace_probe_t *probe = ecb->dte_probe;
10210
10211 ASSERT(MUTEX_HELD(&dtrace_lock));
10212
10213 if (probe == NULL) {
10214 /*
10215 * This is the NULL probe; there is nothing to disable.
10216 */
10217 return;
10218 }
10219
10220 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10221 if (pecb == ecb)
10222 break;
10223 prev = pecb;
10224 }
10225
10226 ASSERT(pecb != NULL);
10227
10228 if (prev == NULL) {
10229 probe->dtpr_ecb = ecb->dte_next;
10230 } else {
10231 prev->dte_next = ecb->dte_next;
10232 }
10233
10234 if (ecb == probe->dtpr_ecb_last) {
10235 ASSERT(ecb->dte_next == NULL);
10236 probe->dtpr_ecb_last = prev;
10237 }
10238
10239 /*
10240 * The ECB has been disconnected from the probe; now sync to assure
10241 * that all CPUs have seen the change before returning.
10242 */
10243 dtrace_sync();
10244
10245 if (probe->dtpr_ecb == NULL) {
10246 /*
10247 * That was the last ECB on the probe; clear the predicate
10248 * cache ID for the probe, disable it and sync one more time
10249 * to assure that we'll never hit it again.
10250 */
10251 dtrace_provider_t *prov = probe->dtpr_provider;
10252
10253 ASSERT(ecb->dte_next == NULL);
10254 ASSERT(probe->dtpr_ecb_last == NULL);
10255 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10256 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10257 probe->dtpr_id, probe->dtpr_arg);
10258 dtrace_sync();
10259 } else {
10260 /*
10261 * There is at least one ECB remaining on the probe. If there
10262 * is _exactly_ one, set the probe's predicate cache ID to be
10263 * the predicate cache ID of the remaining ECB.
10264 */
10265 ASSERT(probe->dtpr_ecb_last != NULL);
10266 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10267
10268 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10269 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10270
10271 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10272
10273 if (p != NULL)
10274 probe->dtpr_predcache = p->dtp_cacheid;
10275 }
10276
10277 ecb->dte_next = NULL;
10278 }
10279 }
10280
10281 static void
10282 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10283 {
10284 dtrace_state_t *state = ecb->dte_state;
10285 dtrace_vstate_t *vstate = &state->dts_vstate;
10286 dtrace_predicate_t *pred;
10287 dtrace_epid_t epid = ecb->dte_epid;
10288
10289 ASSERT(MUTEX_HELD(&dtrace_lock));
10290 ASSERT(ecb->dte_next == NULL);
10291 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10292
10293 if ((pred = ecb->dte_predicate) != NULL)
10294 dtrace_predicate_release(pred, vstate);
10295
10296 dtrace_ecb_action_remove(ecb);
10297
10298 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10299 state->dts_ecbs[epid - 1] = NULL;
10300
10301 kmem_free(ecb, sizeof (dtrace_ecb_t));
10302 }
10303
10304 static dtrace_ecb_t *
10305 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10306 dtrace_enabling_t *enab)
10307 {
10308 dtrace_ecb_t *ecb;
10309 dtrace_predicate_t *pred;
10310 dtrace_actdesc_t *act;
10311 dtrace_provider_t *prov;
10312 dtrace_ecbdesc_t *desc = enab->dten_current;
10313
10314 ASSERT(MUTEX_HELD(&dtrace_lock));
10315 ASSERT(state != NULL);
10316
10317 ecb = dtrace_ecb_add(state, probe);
10318 ecb->dte_uarg = desc->dted_uarg;
10319
10320 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10321 dtrace_predicate_hold(pred);
10322 ecb->dte_predicate = pred;
10323 }
10324
10325 if (probe != NULL) {
10326 /*
10327 * If the provider shows more leg than the consumer is old
10328 * enough to see, we need to enable the appropriate implicit
10329 * predicate bits to prevent the ecb from activating at
10330 * revealing times.
10331 *
10332 * Providers specifying DTRACE_PRIV_USER at register time
10333 * are stating that they need the /proc-style privilege
10334 * model to be enforced, and this is what DTRACE_COND_OWNER
10335 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10336 */
10337 prov = probe->dtpr_provider;
10338 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10339 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10340 ecb->dte_cond |= DTRACE_COND_OWNER;
10341
10342 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10343 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10344 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10345
10346 /*
10347 * If the provider shows us kernel innards and the user
10348 * is lacking sufficient privilege, enable the
10349 * DTRACE_COND_USERMODE implicit predicate.
10350 */
10351 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10352 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10353 ecb->dte_cond |= DTRACE_COND_USERMODE;
10354 }
10355
10356 if (dtrace_ecb_create_cache != NULL) {
10357 /*
10358 * If we have a cached ecb, we'll use its action list instead
10359 * of creating our own (saving both time and space).
10360 */
10361 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10362 dtrace_action_t *act = cached->dte_action;
10363
10364 if (act != NULL) {
10365 ASSERT(act->dta_refcnt > 0);
10366 act->dta_refcnt++;
10367 ecb->dte_action = act;
10368 ecb->dte_action_last = cached->dte_action_last;
10369 ecb->dte_needed = cached->dte_needed;
10370 ecb->dte_size = cached->dte_size;
10371 ecb->dte_alignment = cached->dte_alignment;
10372 }
10373
10374 return (ecb);
10375 }
10376
10377 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10378 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10379 dtrace_ecb_destroy(ecb);
10380 return (NULL);
10381 }
10382 }
10383
10384 dtrace_ecb_resize(ecb);
10385
10386 return (dtrace_ecb_create_cache = ecb);
10387 }
10388
10389 static int
10390 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10391 {
10392 dtrace_ecb_t *ecb;
10393 dtrace_enabling_t *enab = arg;
10394 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10395
10396 ASSERT(state != NULL);
10397
10398 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10399 /*
10400 * This probe was created in a generation for which this
10401 * enabling has previously created ECBs; we don't want to
10402 * enable it again, so just kick out.
10403 */
10404 return (DTRACE_MATCH_NEXT);
10405 }
10406
10407 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10408 return (DTRACE_MATCH_DONE);
10409
10410 if (dtrace_ecb_enable(ecb) < 0)
10411 return (DTRACE_MATCH_FAIL);
10412
10413 return (DTRACE_MATCH_NEXT);
10414 }
10415
10416 static dtrace_ecb_t *
10417 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10418 {
10419 dtrace_ecb_t *ecb;
10420
10421 ASSERT(MUTEX_HELD(&dtrace_lock));
10422
10423 if (id == 0 || id > state->dts_necbs)
10424 return (NULL);
10425
10426 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10427 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10428
10429 return (state->dts_ecbs[id - 1]);
10430 }
10431
10432 static dtrace_aggregation_t *
10433 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10434 {
10435 dtrace_aggregation_t *agg;
10436
10437 ASSERT(MUTEX_HELD(&dtrace_lock));
10438
10439 if (id == 0 || id > state->dts_naggregations)
10440 return (NULL);
10441
10442 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10443 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10444 agg->dtag_id == id);
10445
10446 return (state->dts_aggregations[id - 1]);
10447 }
10448
10449 /*
10450 * DTrace Buffer Functions
10451 *
10452 * The following functions manipulate DTrace buffers. Most of these functions
10453 * are called in the context of establishing or processing consumer state;
10454 * exceptions are explicitly noted.
10455 */
10456
10457 /*
10458 * Note: called from cross call context. This function switches the two
10459 * buffers on a given CPU. The atomicity of this operation is assured by
10460 * disabling interrupts while the actual switch takes place; the disabling of
10461 * interrupts serializes the execution with any execution of dtrace_probe() on
10462 * the same CPU.
10463 */
10464 static void
10465 dtrace_buffer_switch(dtrace_buffer_t *buf)
10466 {
10467 caddr_t tomax = buf->dtb_tomax;
10468 caddr_t xamot = buf->dtb_xamot;
10469 dtrace_icookie_t cookie;
10470 hrtime_t now = dtrace_gethrtime();
10471
10472 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10473 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10474
10475 cookie = dtrace_interrupt_disable();
10476 buf->dtb_tomax = xamot;
10477 buf->dtb_xamot = tomax;
10478 buf->dtb_xamot_drops = buf->dtb_drops;
10479 buf->dtb_xamot_offset = buf->dtb_offset;
10480 buf->dtb_xamot_errors = buf->dtb_errors;
10481 buf->dtb_xamot_flags = buf->dtb_flags;
10482 buf->dtb_offset = 0;
10483 buf->dtb_drops = 0;
10484 buf->dtb_errors = 0;
10485 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10486 buf->dtb_interval = now - buf->dtb_switched;
10487 buf->dtb_switched = now;
10488 dtrace_interrupt_enable(cookie);
10489 }
10490
10491 /*
10492 * Note: called from cross call context. This function activates a buffer
10493 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10494 * is guaranteed by the disabling of interrupts.
10495 */
10496 static void
10497 dtrace_buffer_activate(dtrace_state_t *state)
10498 {
10499 dtrace_buffer_t *buf;
10500 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10501
10502 buf = &state->dts_buffer[CPU->cpu_id];
10503
10504 if (buf->dtb_tomax != NULL) {
10505 /*
10506 * We might like to assert that the buffer is marked inactive,
10507 * but this isn't necessarily true: the buffer for the CPU
10508 * that processes the BEGIN probe has its buffer activated
10509 * manually. In this case, we take the (harmless) action
10510 * re-clearing the bit INACTIVE bit.
10511 */
10512 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10513 }
10514
10515 dtrace_interrupt_enable(cookie);
10516 }
10517
10518 static int
10519 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10520 processorid_t cpu, int *factor)
10521 {
10522 cpu_t *cp;
10523 dtrace_buffer_t *buf;
10524 int allocated = 0, desired = 0;
10525
10526 ASSERT(MUTEX_HELD(&cpu_lock));
10527 ASSERT(MUTEX_HELD(&dtrace_lock));
10528
10529 *factor = 1;
10530
10531 if (size > dtrace_nonroot_maxsize &&
10532 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10533 return (EFBIG);
10534
10535 cp = cpu_list;
10536
10537 do {
10538 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10539 continue;
10540
10541 buf = &bufs[cp->cpu_id];
10542
10543 /*
10544 * If there is already a buffer allocated for this CPU, it
10545 * is only possible that this is a DR event. In this case,
10546 * the buffer size must match our specified size.
10547 */
10548 if (buf->dtb_tomax != NULL) {
10549 ASSERT(buf->dtb_size == size);
10550 continue;
10551 }
10552
10553 ASSERT(buf->dtb_xamot == NULL);
10554
10555 if ((buf->dtb_tomax = kmem_zalloc(size,
10556 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10557 goto err;
10558
10559 buf->dtb_size = size;
10560 buf->dtb_flags = flags;
10561 buf->dtb_offset = 0;
10562 buf->dtb_drops = 0;
10563
10564 if (flags & DTRACEBUF_NOSWITCH)
10565 continue;
10566
10567 if ((buf->dtb_xamot = kmem_zalloc(size,
10568 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10569 goto err;
10570 } while ((cp = cp->cpu_next) != cpu_list);
10571
10572 return (0);
10573
10574 err:
10575 cp = cpu_list;
10576
10577 do {
10578 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10579 continue;
10580
10581 buf = &bufs[cp->cpu_id];
10582 desired += 2;
10583
10584 if (buf->dtb_xamot != NULL) {
10585 ASSERT(buf->dtb_tomax != NULL);
10586 ASSERT(buf->dtb_size == size);
10587 kmem_free(buf->dtb_xamot, size);
10588 allocated++;
10589 }
10590
10591 if (buf->dtb_tomax != NULL) {
10592 ASSERT(buf->dtb_size == size);
10593 kmem_free(buf->dtb_tomax, size);
10594 allocated++;
10595 }
10596
10597 buf->dtb_tomax = NULL;
10598 buf->dtb_xamot = NULL;
10599 buf->dtb_size = 0;
10600 } while ((cp = cp->cpu_next) != cpu_list);
10601
10602 *factor = desired / (allocated > 0 ? allocated : 1);
10603
10604 return (ENOMEM);
10605 }
10606
10607 /*
10608 * Note: called from probe context. This function just increments the drop
10609 * count on a buffer. It has been made a function to allow for the
10610 * possibility of understanding the source of mysterious drop counts. (A
10611 * problem for which one may be particularly disappointed that DTrace cannot
10612 * be used to understand DTrace.)
10613 */
10614 static void
10615 dtrace_buffer_drop(dtrace_buffer_t *buf)
10616 {
10617 buf->dtb_drops++;
10618 }
10619
10620 /*
10621 * Note: called from probe context. This function is called to reserve space
10622 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
10623 * mstate. Returns the new offset in the buffer, or a negative value if an
10624 * error has occurred.
10625 */
10626 static intptr_t
10627 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10628 dtrace_state_t *state, dtrace_mstate_t *mstate)
10629 {
10630 intptr_t offs = buf->dtb_offset, soffs;
10631 intptr_t woffs;
10632 caddr_t tomax;
10633 size_t total;
10634
10635 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10636 return (-1);
10637
10638 if ((tomax = buf->dtb_tomax) == NULL) {
10639 dtrace_buffer_drop(buf);
10640 return (-1);
10641 }
10642
10643 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10644 while (offs & (align - 1)) {
10645 /*
10646 * Assert that our alignment is off by a number which
10647 * is itself sizeof (uint32_t) aligned.
10648 */
10649 ASSERT(!((align - (offs & (align - 1))) &
10650 (sizeof (uint32_t) - 1)));
10651 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10652 offs += sizeof (uint32_t);
10653 }
10654
10655 if ((soffs = offs + needed) > buf->dtb_size) {
10656 dtrace_buffer_drop(buf);
10657 return (-1);
10658 }
10659
10660 if (mstate == NULL)
10661 return (offs);
10662
10663 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10664 mstate->dtms_scratch_size = buf->dtb_size - soffs;
10665 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10666
10667 return (offs);
10668 }
10669
10670 if (buf->dtb_flags & DTRACEBUF_FILL) {
10671 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10672 (buf->dtb_flags & DTRACEBUF_FULL))
10673 return (-1);
10674 goto out;
10675 }
10676
10677 total = needed + (offs & (align - 1));
10678
10679 /*
10680 * For a ring buffer, life is quite a bit more complicated. Before
10681 * we can store any padding, we need to adjust our wrapping offset.
10682 * (If we've never before wrapped or we're not about to, no adjustment
10683 * is required.)
10684 */
10685 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10686 offs + total > buf->dtb_size) {
10687 woffs = buf->dtb_xamot_offset;
10688
10689 if (offs + total > buf->dtb_size) {
10690 /*
10691 * We can't fit in the end of the buffer. First, a
10692 * sanity check that we can fit in the buffer at all.
10693 */
10694 if (total > buf->dtb_size) {
10695 dtrace_buffer_drop(buf);
10696 return (-1);
10697 }
10698
10699 /*
10700 * We're going to be storing at the top of the buffer,
10701 * so now we need to deal with the wrapped offset. We
10702 * only reset our wrapped offset to 0 if it is
10703 * currently greater than the current offset. If it
10704 * is less than the current offset, it is because a
10705 * previous allocation induced a wrap -- but the
10706 * allocation didn't subsequently take the space due
10707 * to an error or false predicate evaluation. In this
10708 * case, we'll just leave the wrapped offset alone: if
10709 * the wrapped offset hasn't been advanced far enough
10710 * for this allocation, it will be adjusted in the
10711 * lower loop.
10712 */
10713 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10714 if (woffs >= offs)
10715 woffs = 0;
10716 } else {
10717 woffs = 0;
10718 }
10719
10720 /*
10721 * Now we know that we're going to be storing to the
10722 * top of the buffer and that there is room for us
10723 * there. We need to clear the buffer from the current
10724 * offset to the end (there may be old gunk there).
10725 */
10726 while (offs < buf->dtb_size)
10727 tomax[offs++] = 0;
10728
10729 /*
10730 * We need to set our offset to zero. And because we
10731 * are wrapping, we need to set the bit indicating as
10732 * much. We can also adjust our needed space back
10733 * down to the space required by the ECB -- we know
10734 * that the top of the buffer is aligned.
10735 */
10736 offs = 0;
10737 total = needed;
10738 buf->dtb_flags |= DTRACEBUF_WRAPPED;
10739 } else {
10740 /*
10741 * There is room for us in the buffer, so we simply
10742 * need to check the wrapped offset.
10743 */
10744 if (woffs < offs) {
10745 /*
10746 * The wrapped offset is less than the offset.
10747 * This can happen if we allocated buffer space
10748 * that induced a wrap, but then we didn't
10749 * subsequently take the space due to an error
10750 * or false predicate evaluation. This is
10751 * okay; we know that _this_ allocation isn't
10752 * going to induce a wrap. We still can't
10753 * reset the wrapped offset to be zero,
10754 * however: the space may have been trashed in
10755 * the previous failed probe attempt. But at
10756 * least the wrapped offset doesn't need to
10757 * be adjusted at all...
10758 */
10759 goto out;
10760 }
10761 }
10762
10763 while (offs + total > woffs) {
10764 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
10765 size_t size;
10766
10767 if (epid == DTRACE_EPIDNONE) {
10768 size = sizeof (uint32_t);
10769 } else {
10770 ASSERT(epid <= state->dts_necbs);
10771 ASSERT(state->dts_ecbs[epid - 1] != NULL);
10772
10773 size = state->dts_ecbs[epid - 1]->dte_size;
10774 }
10775
10776 ASSERT(woffs + size <= buf->dtb_size);
10777 ASSERT(size != 0);
10778
10779 if (woffs + size == buf->dtb_size) {
10780 /*
10781 * We've reached the end of the buffer; we want
10782 * to set the wrapped offset to 0 and break
10783 * out. However, if the offs is 0, then we're
10784 * in a strange edge-condition: the amount of
10785 * space that we want to reserve plus the size
10786 * of the record that we're overwriting is
10787 * greater than the size of the buffer. This
10788 * is problematic because if we reserve the
10789 * space but subsequently don't consume it (due
10790 * to a failed predicate or error) the wrapped
10791 * offset will be 0 -- yet the EPID at offset 0
10792 * will not be committed. This situation is
10793 * relatively easy to deal with: if we're in
10794 * this case, the buffer is indistinguishable
10795 * from one that hasn't wrapped; we need only
10796 * finish the job by clearing the wrapped bit,
10797 * explicitly setting the offset to be 0, and
10798 * zero'ing out the old data in the buffer.
10799 */
10800 if (offs == 0) {
10801 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
10802 buf->dtb_offset = 0;
10803 woffs = total;
10804
10805 while (woffs < buf->dtb_size)
10806 tomax[woffs++] = 0;
10807 }
10808
10809 woffs = 0;
10810 break;
10811 }
10812
10813 woffs += size;
10814 }
10815
10816 /*
10817 * We have a wrapped offset. It may be that the wrapped offset
10818 * has become zero -- that's okay.
10819 */
10820 buf->dtb_xamot_offset = woffs;
10821 }
10822
10823 out:
10824 /*
10825 * Now we can plow the buffer with any necessary padding.
10826 */
10827 while (offs & (align - 1)) {
10828 /*
10829 * Assert that our alignment is off by a number which
10830 * is itself sizeof (uint32_t) aligned.
10831 */
10832 ASSERT(!((align - (offs & (align - 1))) &
10833 (sizeof (uint32_t) - 1)));
10834 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10835 offs += sizeof (uint32_t);
10836 }
10837
10838 if (buf->dtb_flags & DTRACEBUF_FILL) {
10839 if (offs + needed > buf->dtb_size - state->dts_reserve) {
10840 buf->dtb_flags |= DTRACEBUF_FULL;
10841 return (-1);
10842 }
10843 }
10844
10845 if (mstate == NULL)
10846 return (offs);
10847
10848 /*
10849 * For ring buffers and fill buffers, the scratch space is always
10850 * the inactive buffer.
10851 */
10852 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
10853 mstate->dtms_scratch_size = buf->dtb_size;
10854 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10855
10856 return (offs);
10857 }
10858
10859 static void
10860 dtrace_buffer_polish(dtrace_buffer_t *buf)
10861 {
10862 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
10863 ASSERT(MUTEX_HELD(&dtrace_lock));
10864
10865 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
10866 return;
10867
10868 /*
10869 * We need to polish the ring buffer. There are three cases:
10870 *
10871 * - The first (and presumably most common) is that there is no gap
10872 * between the buffer offset and the wrapped offset. In this case,
10873 * there is nothing in the buffer that isn't valid data; we can
10874 * mark the buffer as polished and return.
10875 *
10876 * - The second (less common than the first but still more common
10877 * than the third) is that there is a gap between the buffer offset
10878 * and the wrapped offset, and the wrapped offset is larger than the
10879 * buffer offset. This can happen because of an alignment issue, or
10880 * can happen because of a call to dtrace_buffer_reserve() that
10881 * didn't subsequently consume the buffer space. In this case,
10882 * we need to zero the data from the buffer offset to the wrapped
10883 * offset.
10884 *
10885 * - The third (and least common) is that there is a gap between the
10886 * buffer offset and the wrapped offset, but the wrapped offset is
10887 * _less_ than the buffer offset. This can only happen because a
10888 * call to dtrace_buffer_reserve() induced a wrap, but the space
10889 * was not subsequently consumed. In this case, we need to zero the
10890 * space from the offset to the end of the buffer _and_ from the
10891 * top of the buffer to the wrapped offset.
10892 */
10893 if (buf->dtb_offset < buf->dtb_xamot_offset) {
10894 bzero(buf->dtb_tomax + buf->dtb_offset,
10895 buf->dtb_xamot_offset - buf->dtb_offset);
10896 }
10897
10898 if (buf->dtb_offset > buf->dtb_xamot_offset) {
10899 bzero(buf->dtb_tomax + buf->dtb_offset,
10900 buf->dtb_size - buf->dtb_offset);
10901 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
10902 }
10903 }
10904
10905 /*
10906 * This routine determines if data generated at the specified time has likely
10907 * been entirely consumed at user-level. This routine is called to determine
10908 * if an ECB on a defunct probe (but for an active enabling) can be safely
10909 * disabled and destroyed.
10910 */
10911 static int
10912 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
10913 {
10914 int i;
10915
10916 for (i = 0; i < NCPU; i++) {
10917 dtrace_buffer_t *buf = &bufs[i];
10918
10919 if (buf->dtb_size == 0)
10920 continue;
10921
10922 if (buf->dtb_flags & DTRACEBUF_RING)
10923 return (0);
10924
10925 if (!buf->dtb_switched && buf->dtb_offset != 0)
10926 return (0);
10927
10928 if (buf->dtb_switched - buf->dtb_interval < when)
10929 return (0);
10930 }
10931
10932 return (1);
10933 }
10934
10935 static void
10936 dtrace_buffer_free(dtrace_buffer_t *bufs)
10937 {
10938 int i;
10939
10940 for (i = 0; i < NCPU; i++) {
10941 dtrace_buffer_t *buf = &bufs[i];
10942
10943 if (buf->dtb_tomax == NULL) {
10944 ASSERT(buf->dtb_xamot == NULL);
10945 ASSERT(buf->dtb_size == 0);
10946 continue;
10947 }
10948
10949 if (buf->dtb_xamot != NULL) {
10950 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10951 kmem_free(buf->dtb_xamot, buf->dtb_size);
10952 }
10953
10954 kmem_free(buf->dtb_tomax, buf->dtb_size);
10955 buf->dtb_size = 0;
10956 buf->dtb_tomax = NULL;
10957 buf->dtb_xamot = NULL;
10958 }
10959 }
10960
10961 /*
10962 * DTrace Enabling Functions
10963 */
10964 static dtrace_enabling_t *
10965 dtrace_enabling_create(dtrace_vstate_t *vstate)
10966 {
10967 dtrace_enabling_t *enab;
10968
10969 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
10970 enab->dten_vstate = vstate;
10971
10972 return (enab);
10973 }
10974
10975 static void
10976 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
10977 {
10978 dtrace_ecbdesc_t **ndesc;
10979 size_t osize, nsize;
10980
10981 /*
10982 * We can't add to enablings after we've enabled them, or after we've
10983 * retained them.
10984 */
10985 ASSERT(enab->dten_probegen == 0);
10986 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
10987
10988 if (enab->dten_ndesc < enab->dten_maxdesc) {
10989 enab->dten_desc[enab->dten_ndesc++] = ecb;
10990 return;
10991 }
10992
10993 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
10994
10995 if (enab->dten_maxdesc == 0) {
10996 enab->dten_maxdesc = 1;
10997 } else {
10998 enab->dten_maxdesc <<= 1;
10999 }
11000
11001 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11002
11003 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11004 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11005 bcopy(enab->dten_desc, ndesc, osize);
11006 kmem_free(enab->dten_desc, osize);
11007
11008 enab->dten_desc = ndesc;
11009 enab->dten_desc[enab->dten_ndesc++] = ecb;
11010 }
11011
11012 static void
11013 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11014 dtrace_probedesc_t *pd)
11015 {
11016 dtrace_ecbdesc_t *new;
11017 dtrace_predicate_t *pred;
11018 dtrace_actdesc_t *act;
11019
11020 /*
11021 * We're going to create a new ECB description that matches the
11022 * specified ECB in every way, but has the specified probe description.
11023 */
11024 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11025
11026 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11027 dtrace_predicate_hold(pred);
11028
11029 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11030 dtrace_actdesc_hold(act);
11031
11032 new->dted_action = ecb->dted_action;
11033 new->dted_pred = ecb->dted_pred;
11034 new->dted_probe = *pd;
11035 new->dted_uarg = ecb->dted_uarg;
11036
11037 dtrace_enabling_add(enab, new);
11038 }
11039
11040 static void
11041 dtrace_enabling_dump(dtrace_enabling_t *enab)
11042 {
11043 int i;
11044
11045 for (i = 0; i < enab->dten_ndesc; i++) {
11046 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11047
11048 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11049 desc->dtpd_provider, desc->dtpd_mod,
11050 desc->dtpd_func, desc->dtpd_name);
11051 }
11052 }
11053
11054 static void
11055 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11056 {
11057 int i;
11058 dtrace_ecbdesc_t *ep;
11059 dtrace_vstate_t *vstate = enab->dten_vstate;
11060
11061 ASSERT(MUTEX_HELD(&dtrace_lock));
11062
11063 for (i = 0; i < enab->dten_ndesc; i++) {
11064 dtrace_actdesc_t *act, *next;
11065 dtrace_predicate_t *pred;
11066
11067 ep = enab->dten_desc[i];
11068
11069 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11070 dtrace_predicate_release(pred, vstate);
11071
11072 for (act = ep->dted_action; act != NULL; act = next) {
11073 next = act->dtad_next;
11074 dtrace_actdesc_release(act, vstate);
11075 }
11076
11077 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11078 }
11079
11080 kmem_free(enab->dten_desc,
11081 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11082
11083 /*
11084 * If this was a retained enabling, decrement the dts_nretained count
11085 * and take it off of the dtrace_retained list.
11086 */
11087 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11088 dtrace_retained == enab) {
11089 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11090 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11091 enab->dten_vstate->dtvs_state->dts_nretained--;
11092 dtrace_retained_gen++;
11093 }
11094
11095 if (enab->dten_prev == NULL) {
11096 if (dtrace_retained == enab) {
11097 dtrace_retained = enab->dten_next;
11098
11099 if (dtrace_retained != NULL)
11100 dtrace_retained->dten_prev = NULL;
11101 }
11102 } else {
11103 ASSERT(enab != dtrace_retained);
11104 ASSERT(dtrace_retained != NULL);
11105 enab->dten_prev->dten_next = enab->dten_next;
11106 }
11107
11108 if (enab->dten_next != NULL) {
11109 ASSERT(dtrace_retained != NULL);
11110 enab->dten_next->dten_prev = enab->dten_prev;
11111 }
11112
11113 kmem_free(enab, sizeof (dtrace_enabling_t));
11114 }
11115
11116 static int
11117 dtrace_enabling_retain(dtrace_enabling_t *enab)
11118 {
11119 dtrace_state_t *state;
11120
11121 ASSERT(MUTEX_HELD(&dtrace_lock));
11122 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11123 ASSERT(enab->dten_vstate != NULL);
11124
11125 state = enab->dten_vstate->dtvs_state;
11126 ASSERT(state != NULL);
11127
11128 /*
11129 * We only allow each state to retain dtrace_retain_max enablings.
11130 */
11131 if (state->dts_nretained >= dtrace_retain_max)
11132 return (ENOSPC);
11133
11134 state->dts_nretained++;
11135 dtrace_retained_gen++;
11136
11137 if (dtrace_retained == NULL) {
11138 dtrace_retained = enab;
11139 return (0);
11140 }
11141
11142 enab->dten_next = dtrace_retained;
11143 dtrace_retained->dten_prev = enab;
11144 dtrace_retained = enab;
11145
11146 return (0);
11147 }
11148
11149 static int
11150 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11151 dtrace_probedesc_t *create)
11152 {
11153 dtrace_enabling_t *new, *enab;
11154 int found = 0, err = ENOENT;
11155
11156 ASSERT(MUTEX_HELD(&dtrace_lock));
11157 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11158 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11159 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11160 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11161
11162 new = dtrace_enabling_create(&state->dts_vstate);
11163
11164 /*
11165 * Iterate over all retained enablings, looking for enablings that
11166 * match the specified state.
11167 */
11168 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11169 int i;
11170
11171 /*
11172 * dtvs_state can only be NULL for helper enablings -- and
11173 * helper enablings can't be retained.
11174 */
11175 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11176
11177 if (enab->dten_vstate->dtvs_state != state)
11178 continue;
11179
11180 /*
11181 * Now iterate over each probe description; we're looking for
11182 * an exact match to the specified probe description.
11183 */
11184 for (i = 0; i < enab->dten_ndesc; i++) {
11185 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11186 dtrace_probedesc_t *pd = &ep->dted_probe;
11187
11188 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11189 continue;
11190
11191 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11192 continue;
11193
11194 if (strcmp(pd->dtpd_func, match->dtpd_func))
11195 continue;
11196
11197 if (strcmp(pd->dtpd_name, match->dtpd_name))
11198 continue;
11199
11200 /*
11201 * We have a winning probe! Add it to our growing
11202 * enabling.
11203 */
11204 found = 1;
11205 dtrace_enabling_addlike(new, ep, create);
11206 }
11207 }
11208
11209 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11210 dtrace_enabling_destroy(new);
11211 return (err);
11212 }
11213
11214 return (0);
11215 }
11216
11217 static void
11218 dtrace_enabling_retract(dtrace_state_t *state)
11219 {
11220 dtrace_enabling_t *enab, *next;
11221
11222 ASSERT(MUTEX_HELD(&dtrace_lock));
11223
11224 /*
11225 * Iterate over all retained enablings, destroy the enablings retained
11226 * for the specified state.
11227 */
11228 for (enab = dtrace_retained; enab != NULL; enab = next) {
11229 next = enab->dten_next;
11230
11231 /*
11232 * dtvs_state can only be NULL for helper enablings -- and
11233 * helper enablings can't be retained.
11234 */
11235 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11236
11237 if (enab->dten_vstate->dtvs_state == state) {
11238 ASSERT(state->dts_nretained > 0);
11239 dtrace_enabling_destroy(enab);
11240 }
11241 }
11242
11243 ASSERT(state->dts_nretained == 0);
11244 }
11245
11246 static int
11247 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11248 {
11249 int i = 0;
11250 int total_matched = 0, matched = 0;
11251
11252 ASSERT(MUTEX_HELD(&cpu_lock));
11253 ASSERT(MUTEX_HELD(&dtrace_lock));
11254
11255 for (i = 0; i < enab->dten_ndesc; i++) {
11256 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11257
11258 enab->dten_current = ep;
11259 enab->dten_error = 0;
11260
11261 /*
11262 * If a provider failed to enable a probe then get out and
11263 * let the consumer know we failed.
11264 */
11265 if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
11266 return (EBUSY);
11267
11268 total_matched += matched;
11269
11270 if (enab->dten_error != 0) {
11271 /*
11272 * If we get an error half-way through enabling the
11273 * probes, we kick out -- perhaps with some number of
11274 * them enabled. Leaving enabled probes enabled may
11275 * be slightly confusing for user-level, but we expect
11276 * that no one will attempt to actually drive on in
11277 * the face of such errors. If this is an anonymous
11278 * enabling (indicated with a NULL nmatched pointer),
11279 * we cmn_err() a message. We aren't expecting to
11280 * get such an error -- such as it can exist at all,
11281 * it would be a result of corrupted DOF in the driver
11282 * properties.
11283 */
11284 if (nmatched == NULL) {
11285 cmn_err(CE_WARN, "dtrace_enabling_match() "
11286 "error on %p: %d", (void *)ep,
11287 enab->dten_error);
11288 }
11289
11290 return (enab->dten_error);
11291 }
11292 }
11293
11294 enab->dten_probegen = dtrace_probegen;
11295 if (nmatched != NULL)
11296 *nmatched = total_matched;
11297
11298 return (0);
11299 }
11300
11301 static void
11302 dtrace_enabling_matchall(void)
11303 {
11304 dtrace_enabling_t *enab;
11305
11306 mutex_enter(&cpu_lock);
11307 mutex_enter(&dtrace_lock);
11308
11309 /*
11310 * Iterate over all retained enablings to see if any probes match
11311 * against them. We only perform this operation on enablings for which
11312 * we have sufficient permissions by virtue of being in the global zone
11313 * or in the same zone as the DTrace client. Because we can be called
11314 * after dtrace_detach() has been called, we cannot assert that there
11315 * are retained enablings. We can safely load from dtrace_retained,
11316 * however: the taskq_destroy() at the end of dtrace_detach() will
11317 * block pending our completion.
11318 */
11319 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11320 dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
11321 cred_t *cr = dcr->dcr_cred;
11322 zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
11323
11324 if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
11325 (zone == GLOBAL_ZONEID || getzoneid() == zone)))
11326 (void) dtrace_enabling_match(enab, NULL);
11327 }
11328
11329 mutex_exit(&dtrace_lock);
11330 mutex_exit(&cpu_lock);
11331 }
11332
11333 /*
11334 * If an enabling is to be enabled without having matched probes (that is, if
11335 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11336 * enabling must be _primed_ by creating an ECB for every ECB description.
11337 * This must be done to assure that we know the number of speculations, the
11338 * number of aggregations, the minimum buffer size needed, etc. before we
11339 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11340 * enabling any probes, we create ECBs for every ECB decription, but with a
11341 * NULL probe -- which is exactly what this function does.
11342 */
11343 static void
11344 dtrace_enabling_prime(dtrace_state_t *state)
11345 {
11346 dtrace_enabling_t *enab;
11347 int i;
11348
11349 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11350 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11351
11352 if (enab->dten_vstate->dtvs_state != state)
11353 continue;
11354
11355 /*
11356 * We don't want to prime an enabling more than once, lest
11357 * we allow a malicious user to induce resource exhaustion.
11358 * (The ECBs that result from priming an enabling aren't
11359 * leaked -- but they also aren't deallocated until the
11360 * consumer state is destroyed.)
11361 */
11362 if (enab->dten_primed)
11363 continue;
11364
11365 for (i = 0; i < enab->dten_ndesc; i++) {
11366 enab->dten_current = enab->dten_desc[i];
11367 (void) dtrace_probe_enable(NULL, enab);
11368 }
11369
11370 enab->dten_primed = 1;
11371 }
11372 }
11373
11374 /*
11375 * Called to indicate that probes should be provided due to retained
11376 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11377 * must take an initial lap through the enabling calling the dtps_provide()
11378 * entry point explicitly to allow for autocreated probes.
11379 */
11380 static void
11381 dtrace_enabling_provide(dtrace_provider_t *prv)
11382 {
11383 int i, all = 0;
11384 dtrace_probedesc_t desc;
11385 dtrace_genid_t gen;
11386
11387 ASSERT(MUTEX_HELD(&dtrace_lock));
11388 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11389
11390 if (prv == NULL) {
11391 all = 1;
11392 prv = dtrace_provider;
11393 }
11394
11395 do {
11396 dtrace_enabling_t *enab;
11397 void *parg = prv->dtpv_arg;
11398
11399 retry:
11400 gen = dtrace_retained_gen;
11401 for (enab = dtrace_retained; enab != NULL;
11402 enab = enab->dten_next) {
11403 for (i = 0; i < enab->dten_ndesc; i++) {
11404 desc = enab->dten_desc[i]->dted_probe;
11405 mutex_exit(&dtrace_lock);
11406 prv->dtpv_pops.dtps_provide(parg, &desc);
11407 mutex_enter(&dtrace_lock);
11408 /*
11409 * Process the retained enablings again if
11410 * they have changed while we weren't holding
11411 * dtrace_lock.
11412 */
11413 if (gen != dtrace_retained_gen)
11414 goto retry;
11415 }
11416 }
11417 } while (all && (prv = prv->dtpv_next) != NULL);
11418
11419 mutex_exit(&dtrace_lock);
11420 dtrace_probe_provide(NULL, all ? NULL : prv);
11421 mutex_enter(&dtrace_lock);
11422 }
11423
11424 /*
11425 * Called to reap ECBs that are attached to probes from defunct providers.
11426 */
11427 static void
11428 dtrace_enabling_reap(void)
11429 {
11430 dtrace_provider_t *prov;
11431 dtrace_probe_t *probe;
11432 dtrace_ecb_t *ecb;
11433 hrtime_t when;
11434 int i;
11435
11436 mutex_enter(&cpu_lock);
11437 mutex_enter(&dtrace_lock);
11438
11439 for (i = 0; i < dtrace_nprobes; i++) {
11440 if ((probe = dtrace_probes[i]) == NULL)
11441 continue;
11442
11443 if (probe->dtpr_ecb == NULL)
11444 continue;
11445
11446 prov = probe->dtpr_provider;
11447
11448 if ((when = prov->dtpv_defunct) == 0)
11449 continue;
11450
11451 /*
11452 * We have ECBs on a defunct provider: we want to reap these
11453 * ECBs to allow the provider to unregister. The destruction
11454 * of these ECBs must be done carefully: if we destroy the ECB
11455 * and the consumer later wishes to consume an EPID that
11456 * corresponds to the destroyed ECB (and if the EPID metadata
11457 * has not been previously consumed), the consumer will abort
11458 * processing on the unknown EPID. To reduce (but not, sadly,
11459 * eliminate) the possibility of this, we will only destroy an
11460 * ECB for a defunct provider if, for the state that
11461 * corresponds to the ECB:
11462 *
11463 * (a) There is no speculative tracing (which can effectively
11464 * cache an EPID for an arbitrary amount of time).
11465 *
11466 * (b) The principal buffers have been switched twice since the
11467 * provider became defunct.
11468 *
11469 * (c) The aggregation buffers are of zero size or have been
11470 * switched twice since the provider became defunct.
11471 *
11472 * We use dts_speculates to determine (a) and call a function
11473 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
11474 * that as soon as we've been unable to destroy one of the ECBs
11475 * associated with the probe, we quit trying -- reaping is only
11476 * fruitful in as much as we can destroy all ECBs associated
11477 * with the defunct provider's probes.
11478 */
11479 while ((ecb = probe->dtpr_ecb) != NULL) {
11480 dtrace_state_t *state = ecb->dte_state;
11481 dtrace_buffer_t *buf = state->dts_buffer;
11482 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11483
11484 if (state->dts_speculates)
11485 break;
11486
11487 if (!dtrace_buffer_consumed(buf, when))
11488 break;
11489
11490 if (!dtrace_buffer_consumed(aggbuf, when))
11491 break;
11492
11493 dtrace_ecb_disable(ecb);
11494 ASSERT(probe->dtpr_ecb != ecb);
11495 dtrace_ecb_destroy(ecb);
11496 }
11497 }
11498
11499 mutex_exit(&dtrace_lock);
11500 mutex_exit(&cpu_lock);
11501 }
11502
11503 /*
11504 * DTrace DOF Functions
11505 */
11506 /*ARGSUSED*/
11507 static void
11508 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11509 {
11510 if (dtrace_err_verbose)
11511 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11512
11513 #ifdef DTRACE_ERRDEBUG
11514 dtrace_errdebug(str);
11515 #endif
11516 }
11517
11518 /*
11519 * Create DOF out of a currently enabled state. Right now, we only create
11520 * DOF containing the run-time options -- but this could be expanded to create
11521 * complete DOF representing the enabled state.
11522 */
11523 static dof_hdr_t *
11524 dtrace_dof_create(dtrace_state_t *state)
11525 {
11526 dof_hdr_t *dof;
11527 dof_sec_t *sec;
11528 dof_optdesc_t *opt;
11529 int i, len = sizeof (dof_hdr_t) +
11530 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11531 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11532
11533 ASSERT(MUTEX_HELD(&dtrace_lock));
11534
11535 dof = kmem_zalloc(len, KM_SLEEP);
11536 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11537 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11538 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11539 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11540
11541 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11542 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11543 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11544 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11545 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11546 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11547
11548 dof->dofh_flags = 0;
11549 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11550 dof->dofh_secsize = sizeof (dof_sec_t);
11551 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11552 dof->dofh_secoff = sizeof (dof_hdr_t);
11553 dof->dofh_loadsz = len;
11554 dof->dofh_filesz = len;
11555 dof->dofh_pad = 0;
11556
11557 /*
11558 * Fill in the option section header...
11559 */
11560 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11561 sec->dofs_type = DOF_SECT_OPTDESC;
11562 sec->dofs_align = sizeof (uint64_t);
11563 sec->dofs_flags = DOF_SECF_LOAD;
11564 sec->dofs_entsize = sizeof (dof_optdesc_t);
11565
11566 opt = (dof_optdesc_t *)((uintptr_t)sec +
11567 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11568
11569 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11570 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11571
11572 for (i = 0; i < DTRACEOPT_MAX; i++) {
11573 opt[i].dofo_option = i;
11574 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11575 opt[i].dofo_value = state->dts_options[i];
11576 }
11577
11578 return (dof);
11579 }
11580
11581 static dof_hdr_t *
11582 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11583 {
11584 dof_hdr_t hdr, *dof;
11585
11586 ASSERT(!MUTEX_HELD(&dtrace_lock));
11587
11588 /*
11589 * First, we're going to copyin() the sizeof (dof_hdr_t).
11590 */
11591 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11592 dtrace_dof_error(NULL, "failed to copyin DOF header");
11593 *errp = EFAULT;
11594 return (NULL);
11595 }
11596
11597 /*
11598 * Now we'll allocate the entire DOF and copy it in -- provided
11599 * that the length isn't outrageous.
11600 */
11601 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11602 dtrace_dof_error(&hdr, "load size exceeds maximum");
11603 *errp = E2BIG;
11604 return (NULL);
11605 }
11606
11607 if (hdr.dofh_loadsz < sizeof (hdr)) {
11608 dtrace_dof_error(&hdr, "invalid load size");
11609 *errp = EINVAL;
11610 return (NULL);
11611 }
11612
11613 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11614
11615 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
11616 dof->dofh_loadsz != hdr.dofh_loadsz) {
11617 kmem_free(dof, hdr.dofh_loadsz);
11618 *errp = EFAULT;
11619 return (NULL);
11620 }
11621
11622 return (dof);
11623 }
11624
11625 static dof_hdr_t *
11626 dtrace_dof_property(const char *name)
11627 {
11628 uchar_t *buf;
11629 uint64_t loadsz;
11630 unsigned int len, i;
11631 dof_hdr_t *dof;
11632
11633 /*
11634 * Unfortunately, array of values in .conf files are always (and
11635 * only) interpreted to be integer arrays. We must read our DOF
11636 * as an integer array, and then squeeze it into a byte array.
11637 */
11638 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11639 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11640 return (NULL);
11641
11642 for (i = 0; i < len; i++)
11643 buf[i] = (uchar_t)(((int *)buf)[i]);
11644
11645 if (len < sizeof (dof_hdr_t)) {
11646 ddi_prop_free(buf);
11647 dtrace_dof_error(NULL, "truncated header");
11648 return (NULL);
11649 }
11650
11651 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11652 ddi_prop_free(buf);
11653 dtrace_dof_error(NULL, "truncated DOF");
11654 return (NULL);
11655 }
11656
11657 if (loadsz >= dtrace_dof_maxsize) {
11658 ddi_prop_free(buf);
11659 dtrace_dof_error(NULL, "oversized DOF");
11660 return (NULL);
11661 }
11662
11663 dof = kmem_alloc(loadsz, KM_SLEEP);
11664 bcopy(buf, dof, loadsz);
11665 ddi_prop_free(buf);
11666
11667 return (dof);
11668 }
11669
11670 static void
11671 dtrace_dof_destroy(dof_hdr_t *dof)
11672 {
11673 kmem_free(dof, dof->dofh_loadsz);
11674 }
11675
11676 /*
11677 * Return the dof_sec_t pointer corresponding to a given section index. If the
11678 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
11679 * a type other than DOF_SECT_NONE is specified, the header is checked against
11680 * this type and NULL is returned if the types do not match.
11681 */
11682 static dof_sec_t *
11683 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11684 {
11685 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11686 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11687
11688 if (i >= dof->dofh_secnum) {
11689 dtrace_dof_error(dof, "referenced section index is invalid");
11690 return (NULL);
11691 }
11692
11693 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11694 dtrace_dof_error(dof, "referenced section is not loadable");
11695 return (NULL);
11696 }
11697
11698 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11699 dtrace_dof_error(dof, "referenced section is the wrong type");
11700 return (NULL);
11701 }
11702
11703 return (sec);
11704 }
11705
11706 static dtrace_probedesc_t *
11707 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11708 {
11709 dof_probedesc_t *probe;
11710 dof_sec_t *strtab;
11711 uintptr_t daddr = (uintptr_t)dof;
11712 uintptr_t str;
11713 size_t size;
11714
11715 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11716 dtrace_dof_error(dof, "invalid probe section");
11717 return (NULL);
11718 }
11719
11720 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11721 dtrace_dof_error(dof, "bad alignment in probe description");
11722 return (NULL);
11723 }
11724
11725 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11726 dtrace_dof_error(dof, "truncated probe description");
11727 return (NULL);
11728 }
11729
11730 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11731 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11732
11733 if (strtab == NULL)
11734 return (NULL);
11735
11736 str = daddr + strtab->dofs_offset;
11737 size = strtab->dofs_size;
11738
11739 if (probe->dofp_provider >= strtab->dofs_size) {
11740 dtrace_dof_error(dof, "corrupt probe provider");
11741 return (NULL);
11742 }
11743
11744 (void) strncpy(desc->dtpd_provider,
11745 (char *)(str + probe->dofp_provider),
11746 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11747
11748 if (probe->dofp_mod >= strtab->dofs_size) {
11749 dtrace_dof_error(dof, "corrupt probe module");
11750 return (NULL);
11751 }
11752
11753 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11754 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11755
11756 if (probe->dofp_func >= strtab->dofs_size) {
11757 dtrace_dof_error(dof, "corrupt probe function");
11758 return (NULL);
11759 }
11760
11761 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11762 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11763
11764 if (probe->dofp_name >= strtab->dofs_size) {
11765 dtrace_dof_error(dof, "corrupt probe name");
11766 return (NULL);
11767 }
11768
11769 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11770 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11771
11772 return (desc);
11773 }
11774
11775 static dtrace_difo_t *
11776 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11777 cred_t *cr)
11778 {
11779 dtrace_difo_t *dp;
11780 size_t ttl = 0;
11781 dof_difohdr_t *dofd;
11782 uintptr_t daddr = (uintptr_t)dof;
11783 size_t max = dtrace_difo_maxsize;
11784 int i, l, n;
11785
11786 static const struct {
11787 int section;
11788 int bufoffs;
11789 int lenoffs;
11790 int entsize;
11791 int align;
11792 const char *msg;
11793 } difo[] = {
11794 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11795 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11796 sizeof (dif_instr_t), "multiple DIF sections" },
11797
11798 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11799 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11800 sizeof (uint64_t), "multiple integer tables" },
11801
11802 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11803 offsetof(dtrace_difo_t, dtdo_strlen), 0,
11804 sizeof (char), "multiple string tables" },
11805
11806 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11807 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11808 sizeof (uint_t), "multiple variable tables" },
11809
11810 { DOF_SECT_NONE, 0, 0, 0, NULL }
11811 };
11812
11813 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11814 dtrace_dof_error(dof, "invalid DIFO header section");
11815 return (NULL);
11816 }
11817
11818 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11819 dtrace_dof_error(dof, "bad alignment in DIFO header");
11820 return (NULL);
11821 }
11822
11823 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
11824 sec->dofs_size % sizeof (dof_secidx_t)) {
11825 dtrace_dof_error(dof, "bad size in DIFO header");
11826 return (NULL);
11827 }
11828
11829 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11830 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
11831
11832 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
11833 dp->dtdo_rtype = dofd->dofd_rtype;
11834
11835 for (l = 0; l < n; l++) {
11836 dof_sec_t *subsec;
11837 void **bufp;
11838 uint32_t *lenp;
11839
11840 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
11841 dofd->dofd_links[l])) == NULL)
11842 goto err; /* invalid section link */
11843
11844 if (ttl + subsec->dofs_size > max) {
11845 dtrace_dof_error(dof, "exceeds maximum size");
11846 goto err;
11847 }
11848
11849 ttl += subsec->dofs_size;
11850
11851 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
11852 if (subsec->dofs_type != difo[i].section)
11853 continue;
11854
11855 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
11856 dtrace_dof_error(dof, "section not loaded");
11857 goto err;
11858 }
11859
11860 if (subsec->dofs_align != difo[i].align) {
11861 dtrace_dof_error(dof, "bad alignment");
11862 goto err;
11863 }
11864
11865 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
11866 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
11867
11868 if (*bufp != NULL) {
11869 dtrace_dof_error(dof, difo[i].msg);
11870 goto err;
11871 }
11872
11873 if (difo[i].entsize != subsec->dofs_entsize) {
11874 dtrace_dof_error(dof, "entry size mismatch");
11875 goto err;
11876 }
11877
11878 if (subsec->dofs_entsize != 0 &&
11879 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
11880 dtrace_dof_error(dof, "corrupt entry size");
11881 goto err;
11882 }
11883
11884 *lenp = subsec->dofs_size;
11885 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
11886 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
11887 *bufp, subsec->dofs_size);
11888
11889 if (subsec->dofs_entsize != 0)
11890 *lenp /= subsec->dofs_entsize;
11891
11892 break;
11893 }
11894
11895 /*
11896 * If we encounter a loadable DIFO sub-section that is not
11897 * known to us, assume this is a broken program and fail.
11898 */
11899 if (difo[i].section == DOF_SECT_NONE &&
11900 (subsec->dofs_flags & DOF_SECF_LOAD)) {
11901 dtrace_dof_error(dof, "unrecognized DIFO subsection");
11902 goto err;
11903 }
11904 }
11905
11906 if (dp->dtdo_buf == NULL) {
11907 /*
11908 * We can't have a DIF object without DIF text.
11909 */
11910 dtrace_dof_error(dof, "missing DIF text");
11911 goto err;
11912 }
11913
11914 /*
11915 * Before we validate the DIF object, run through the variable table
11916 * looking for the strings -- if any of their size are under, we'll set
11917 * their size to be the system-wide default string size. Note that
11918 * this should _not_ happen if the "strsize" option has been set --
11919 * in this case, the compiler should have set the size to reflect the
11920 * setting of the option.
11921 */
11922 for (i = 0; i < dp->dtdo_varlen; i++) {
11923 dtrace_difv_t *v = &dp->dtdo_vartab[i];
11924 dtrace_diftype_t *t = &v->dtdv_type;
11925
11926 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
11927 continue;
11928
11929 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
11930 t->dtdt_size = dtrace_strsize_default;
11931 }
11932
11933 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
11934 goto err;
11935
11936 dtrace_difo_init(dp, vstate);
11937 return (dp);
11938
11939 err:
11940 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
11941 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
11942 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
11943 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
11944
11945 kmem_free(dp, sizeof (dtrace_difo_t));
11946 return (NULL);
11947 }
11948
11949 static dtrace_predicate_t *
11950 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11951 cred_t *cr)
11952 {
11953 dtrace_difo_t *dp;
11954
11955 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
11956 return (NULL);
11957
11958 return (dtrace_predicate_create(dp));
11959 }
11960
11961 static dtrace_actdesc_t *
11962 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11963 cred_t *cr)
11964 {
11965 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
11966 dof_actdesc_t *desc;
11967 dof_sec_t *difosec;
11968 size_t offs;
11969 uintptr_t daddr = (uintptr_t)dof;
11970 uint64_t arg;
11971 dtrace_actkind_t kind;
11972
11973 if (sec->dofs_type != DOF_SECT_ACTDESC) {
11974 dtrace_dof_error(dof, "invalid action section");
11975 return (NULL);
11976 }
11977
11978 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
11979 dtrace_dof_error(dof, "truncated action description");
11980 return (NULL);
11981 }
11982
11983 if (sec->dofs_align != sizeof (uint64_t)) {
11984 dtrace_dof_error(dof, "bad alignment in action description");
11985 return (NULL);
11986 }
11987
11988 if (sec->dofs_size < sec->dofs_entsize) {
11989 dtrace_dof_error(dof, "section entry size exceeds total size");
11990 return (NULL);
11991 }
11992
11993 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
11994 dtrace_dof_error(dof, "bad entry size in action description");
11995 return (NULL);
11996 }
11997
11998 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
11999 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12000 return (NULL);
12001 }
12002
12003 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12004 desc = (dof_actdesc_t *)(daddr +
12005 (uintptr_t)sec->dofs_offset + offs);
12006 kind = (dtrace_actkind_t)desc->dofa_kind;
12007
12008 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12009 (kind != DTRACEACT_PRINTA ||
12010 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12011 (kind == DTRACEACT_DIFEXPR &&
12012 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12013 dof_sec_t *strtab;
12014 char *str, *fmt;
12015 uint64_t i;
12016
12017 /*
12018 * The argument to these actions is an index into the
12019 * DOF string table. For printf()-like actions, this
12020 * is the format string. For print(), this is the
12021 * CTF type of the expression result.
12022 */
12023 if ((strtab = dtrace_dof_sect(dof,
12024 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12025 goto err;
12026
12027 str = (char *)((uintptr_t)dof +
12028 (uintptr_t)strtab->dofs_offset);
12029
12030 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12031 if (str[i] == '\0')
12032 break;
12033 }
12034
12035 if (i >= strtab->dofs_size) {
12036 dtrace_dof_error(dof, "bogus format string");
12037 goto err;
12038 }
12039
12040 if (i == desc->dofa_arg) {
12041 dtrace_dof_error(dof, "empty format string");
12042 goto err;
12043 }
12044
12045 i -= desc->dofa_arg;
12046 fmt = kmem_alloc(i + 1, KM_SLEEP);
12047 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12048 arg = (uint64_t)(uintptr_t)fmt;
12049 } else {
12050 if (kind == DTRACEACT_PRINTA) {
12051 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12052 arg = 0;
12053 } else {
12054 arg = desc->dofa_arg;
12055 }
12056 }
12057
12058 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12059 desc->dofa_uarg, arg);
12060
12061 if (last != NULL) {
12062 last->dtad_next = act;
12063 } else {
12064 first = act;
12065 }
12066
12067 last = act;
12068
12069 if (desc->dofa_difo == DOF_SECIDX_NONE)
12070 continue;
12071
12072 if ((difosec = dtrace_dof_sect(dof,
12073 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12074 goto err;
12075
12076 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12077
12078 if (act->dtad_difo == NULL)
12079 goto err;
12080 }
12081
12082 ASSERT(first != NULL);
12083 return (first);
12084
12085 err:
12086 for (act = first; act != NULL; act = next) {
12087 next = act->dtad_next;
12088 dtrace_actdesc_release(act, vstate);
12089 }
12090
12091 return (NULL);
12092 }
12093
12094 static dtrace_ecbdesc_t *
12095 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12096 cred_t *cr)
12097 {
12098 dtrace_ecbdesc_t *ep;
12099 dof_ecbdesc_t *ecb;
12100 dtrace_probedesc_t *desc;
12101 dtrace_predicate_t *pred = NULL;
12102
12103 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12104 dtrace_dof_error(dof, "truncated ECB description");
12105 return (NULL);
12106 }
12107
12108 if (sec->dofs_align != sizeof (uint64_t)) {
12109 dtrace_dof_error(dof, "bad alignment in ECB description");
12110 return (NULL);
12111 }
12112
12113 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12114 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12115
12116 if (sec == NULL)
12117 return (NULL);
12118
12119 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12120 ep->dted_uarg = ecb->dofe_uarg;
12121 desc = &ep->dted_probe;
12122
12123 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12124 goto err;
12125
12126 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12127 if ((sec = dtrace_dof_sect(dof,
12128 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12129 goto err;
12130
12131 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12132 goto err;
12133
12134 ep->dted_pred.dtpdd_predicate = pred;
12135 }
12136
12137 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12138 if ((sec = dtrace_dof_sect(dof,
12139 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12140 goto err;
12141
12142 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12143
12144 if (ep->dted_action == NULL)
12145 goto err;
12146 }
12147
12148 return (ep);
12149
12150 err:
12151 if (pred != NULL)
12152 dtrace_predicate_release(pred, vstate);
12153 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12154 return (NULL);
12155 }
12156
12157 /*
12158 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12159 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12160 * site of any user SETX relocations to account for load object base address.
12161 * In the future, if we need other relocations, this function can be extended.
12162 */
12163 static int
12164 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12165 {
12166 uintptr_t daddr = (uintptr_t)dof;
12167 dof_relohdr_t *dofr =
12168 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12169 dof_sec_t *ss, *rs, *ts;
12170 dof_relodesc_t *r;
12171 uint_t i, n;
12172
12173 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12174 sec->dofs_align != sizeof (dof_secidx_t)) {
12175 dtrace_dof_error(dof, "invalid relocation header");
12176 return (-1);
12177 }
12178
12179 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12180 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12181 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12182
12183 if (ss == NULL || rs == NULL || ts == NULL)
12184 return (-1); /* dtrace_dof_error() has been called already */
12185
12186 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12187 rs->dofs_align != sizeof (uint64_t)) {
12188 dtrace_dof_error(dof, "invalid relocation section");
12189 return (-1);
12190 }
12191
12192 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12193 n = rs->dofs_size / rs->dofs_entsize;
12194
12195 for (i = 0; i < n; i++) {
12196 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12197
12198 switch (r->dofr_type) {
12199 case DOF_RELO_NONE:
12200 break;
12201 case DOF_RELO_SETX:
12202 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12203 sizeof (uint64_t) > ts->dofs_size) {
12204 dtrace_dof_error(dof, "bad relocation offset");
12205 return (-1);
12206 }
12207
12208 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12209 dtrace_dof_error(dof, "misaligned setx relo");
12210 return (-1);
12211 }
12212
12213 *(uint64_t *)taddr += ubase;
12214 break;
12215 default:
12216 dtrace_dof_error(dof, "invalid relocation type");
12217 return (-1);
12218 }
12219
12220 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12221 }
12222
12223 return (0);
12224 }
12225
12226 /*
12227 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12228 * header: it should be at the front of a memory region that is at least
12229 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12230 * size. It need not be validated in any other way.
12231 */
12232 static int
12233 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12234 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12235 {
12236 uint64_t len = dof->dofh_loadsz, seclen;
12237 uintptr_t daddr = (uintptr_t)dof;
12238 dtrace_ecbdesc_t *ep;
12239 dtrace_enabling_t *enab;
12240 uint_t i;
12241
12242 ASSERT(MUTEX_HELD(&dtrace_lock));
12243 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12244
12245 /*
12246 * Check the DOF header identification bytes. In addition to checking
12247 * valid settings, we also verify that unused bits/bytes are zeroed so
12248 * we can use them later without fear of regressing existing binaries.
12249 */
12250 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12251 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12252 dtrace_dof_error(dof, "DOF magic string mismatch");
12253 return (-1);
12254 }
12255
12256 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12257 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12258 dtrace_dof_error(dof, "DOF has invalid data model");
12259 return (-1);
12260 }
12261
12262 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12263 dtrace_dof_error(dof, "DOF encoding mismatch");
12264 return (-1);
12265 }
12266
12267 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12268 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12269 dtrace_dof_error(dof, "DOF version mismatch");
12270 return (-1);
12271 }
12272
12273 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12274 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12275 return (-1);
12276 }
12277
12278 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12279 dtrace_dof_error(dof, "DOF uses too many integer registers");
12280 return (-1);
12281 }
12282
12283 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12284 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12285 return (-1);
12286 }
12287
12288 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12289 if (dof->dofh_ident[i] != 0) {
12290 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12291 return (-1);
12292 }
12293 }
12294
12295 if (dof->dofh_flags & ~DOF_FL_VALID) {
12296 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12297 return (-1);
12298 }
12299
12300 if (dof->dofh_secsize == 0) {
12301 dtrace_dof_error(dof, "zero section header size");
12302 return (-1);
12303 }
12304
12305 /*
12306 * Check that the section headers don't exceed the amount of DOF
12307 * data. Note that we cast the section size and number of sections
12308 * to uint64_t's to prevent possible overflow in the multiplication.
12309 */
12310 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12311
12312 if (dof->dofh_secoff > len || seclen > len ||
12313 dof->dofh_secoff + seclen > len) {
12314 dtrace_dof_error(dof, "truncated section headers");
12315 return (-1);
12316 }
12317
12318 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12319 dtrace_dof_error(dof, "misaligned section headers");
12320 return (-1);
12321 }
12322
12323 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12324 dtrace_dof_error(dof, "misaligned section size");
12325 return (-1);
12326 }
12327
12328 /*
12329 * Take an initial pass through the section headers to be sure that
12330 * the headers don't have stray offsets. If the 'noprobes' flag is
12331 * set, do not permit sections relating to providers, probes, or args.
12332 */
12333 for (i = 0; i < dof->dofh_secnum; i++) {
12334 dof_sec_t *sec = (dof_sec_t *)(daddr +
12335 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12336
12337 if (noprobes) {
12338 switch (sec->dofs_type) {
12339 case DOF_SECT_PROVIDER:
12340 case DOF_SECT_PROBES:
12341 case DOF_SECT_PRARGS:
12342 case DOF_SECT_PROFFS:
12343 dtrace_dof_error(dof, "illegal sections "
12344 "for enabling");
12345 return (-1);
12346 }
12347 }
12348
12349 if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
12350 !(sec->dofs_flags & DOF_SECF_LOAD)) {
12351 dtrace_dof_error(dof, "loadable section with load "
12352 "flag unset");
12353 return (-1);
12354 }
12355
12356 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12357 continue; /* just ignore non-loadable sections */
12358
12359 if (sec->dofs_align & (sec->dofs_align - 1)) {
12360 dtrace_dof_error(dof, "bad section alignment");
12361 return (-1);
12362 }
12363
12364 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12365 dtrace_dof_error(dof, "misaligned section");
12366 return (-1);
12367 }
12368
12369 if (sec->dofs_offset > len || sec->dofs_size > len ||
12370 sec->dofs_offset + sec->dofs_size > len) {
12371 dtrace_dof_error(dof, "corrupt section header");
12372 return (-1);
12373 }
12374
12375 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12376 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12377 dtrace_dof_error(dof, "non-terminating string table");
12378 return (-1);
12379 }
12380 }
12381
12382 /*
12383 * Take a second pass through the sections and locate and perform any
12384 * relocations that are present. We do this after the first pass to
12385 * be sure that all sections have had their headers validated.
12386 */
12387 for (i = 0; i < dof->dofh_secnum; i++) {
12388 dof_sec_t *sec = (dof_sec_t *)(daddr +
12389 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12390
12391 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12392 continue; /* skip sections that are not loadable */
12393
12394 switch (sec->dofs_type) {
12395 case DOF_SECT_URELHDR:
12396 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12397 return (-1);
12398 break;
12399 }
12400 }
12401
12402 if ((enab = *enabp) == NULL)
12403 enab = *enabp = dtrace_enabling_create(vstate);
12404
12405 for (i = 0; i < dof->dofh_secnum; i++) {
12406 dof_sec_t *sec = (dof_sec_t *)(daddr +
12407 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12408
12409 if (sec->dofs_type != DOF_SECT_ECBDESC)
12410 continue;
12411
12412 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12413 dtrace_enabling_destroy(enab);
12414 *enabp = NULL;
12415 return (-1);
12416 }
12417
12418 dtrace_enabling_add(enab, ep);
12419 }
12420
12421 return (0);
12422 }
12423
12424 /*
12425 * Process DOF for any options. This routine assumes that the DOF has been
12426 * at least processed by dtrace_dof_slurp().
12427 */
12428 static int
12429 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12430 {
12431 int i, rval;
12432 uint32_t entsize;
12433 size_t offs;
12434 dof_optdesc_t *desc;
12435
12436 for (i = 0; i < dof->dofh_secnum; i++) {
12437 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12438 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12439
12440 if (sec->dofs_type != DOF_SECT_OPTDESC)
12441 continue;
12442
12443 if (sec->dofs_align != sizeof (uint64_t)) {
12444 dtrace_dof_error(dof, "bad alignment in "
12445 "option description");
12446 return (EINVAL);
12447 }
12448
12449 if ((entsize = sec->dofs_entsize) == 0) {
12450 dtrace_dof_error(dof, "zeroed option entry size");
12451 return (EINVAL);
12452 }
12453
12454 if (entsize < sizeof (dof_optdesc_t)) {
12455 dtrace_dof_error(dof, "bad option entry size");
12456 return (EINVAL);
12457 }
12458
12459 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12460 desc = (dof_optdesc_t *)((uintptr_t)dof +
12461 (uintptr_t)sec->dofs_offset + offs);
12462
12463 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12464 dtrace_dof_error(dof, "non-zero option string");
12465 return (EINVAL);
12466 }
12467
12468 if (desc->dofo_value == DTRACEOPT_UNSET) {
12469 dtrace_dof_error(dof, "unset option");
12470 return (EINVAL);
12471 }
12472
12473 if ((rval = dtrace_state_option(state,
12474 desc->dofo_option, desc->dofo_value)) != 0) {
12475 dtrace_dof_error(dof, "rejected option");
12476 return (rval);
12477 }
12478 }
12479 }
12480
12481 return (0);
12482 }
12483
12484 /*
12485 * DTrace Consumer State Functions
12486 */
12487 int
12488 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12489 {
12490 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12491 void *base;
12492 uintptr_t limit;
12493 dtrace_dynvar_t *dvar, *next, *start;
12494 int i;
12495
12496 ASSERT(MUTEX_HELD(&dtrace_lock));
12497 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12498
12499 bzero(dstate, sizeof (dtrace_dstate_t));
12500
12501 if ((dstate->dtds_chunksize = chunksize) == 0)
12502 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12503
12504 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12505 size = min;
12506
12507 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12508 return (ENOMEM);
12509
12510 dstate->dtds_size = size;
12511 dstate->dtds_base = base;
12512 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12513 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12514
12515 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12516
12517 if (hashsize != 1 && (hashsize & 1))
12518 hashsize--;
12519
12520 dstate->dtds_hashsize = hashsize;
12521 dstate->dtds_hash = dstate->dtds_base;
12522
12523 /*
12524 * Set all of our hash buckets to point to the single sink, and (if
12525 * it hasn't already been set), set the sink's hash value to be the
12526 * sink sentinel value. The sink is needed for dynamic variable
12527 * lookups to know that they have iterated over an entire, valid hash
12528 * chain.
12529 */
12530 for (i = 0; i < hashsize; i++)
12531 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12532
12533 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12534 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12535
12536 /*
12537 * Determine number of active CPUs. Divide free list evenly among
12538 * active CPUs.
12539 */
12540 start = (dtrace_dynvar_t *)
12541 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12542 limit = (uintptr_t)base + size;
12543
12544 maxper = (limit - (uintptr_t)start) / NCPU;
12545 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12546
12547 for (i = 0; i < NCPU; i++) {
12548 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12549
12550 /*
12551 * If we don't even have enough chunks to make it once through
12552 * NCPUs, we're just going to allocate everything to the first
12553 * CPU. And if we're on the last CPU, we're going to allocate
12554 * whatever is left over. In either case, we set the limit to
12555 * be the limit of the dynamic variable space.
12556 */
12557 if (maxper == 0 || i == NCPU - 1) {
12558 limit = (uintptr_t)base + size;
12559 start = NULL;
12560 } else {
12561 limit = (uintptr_t)start + maxper;
12562 start = (dtrace_dynvar_t *)limit;
12563 }
12564
12565 ASSERT(limit <= (uintptr_t)base + size);
12566
12567 for (;;) {
12568 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12569 dstate->dtds_chunksize);
12570
12571 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12572 break;
12573
12574 dvar->dtdv_next = next;
12575 dvar = next;
12576 }
12577
12578 if (maxper == 0)
12579 break;
12580 }
12581
12582 return (0);
12583 }
12584
12585 void
12586 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12587 {
12588 ASSERT(MUTEX_HELD(&cpu_lock));
12589
12590 if (dstate->dtds_base == NULL)
12591 return;
12592
12593 kmem_free(dstate->dtds_base, dstate->dtds_size);
12594 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12595 }
12596
12597 static void
12598 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12599 {
12600 /*
12601 * Logical XOR, where are you?
12602 */
12603 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12604
12605 if (vstate->dtvs_nglobals > 0) {
12606 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12607 sizeof (dtrace_statvar_t *));
12608 }
12609
12610 if (vstate->dtvs_ntlocals > 0) {
12611 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12612 sizeof (dtrace_difv_t));
12613 }
12614
12615 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
12616
12617 if (vstate->dtvs_nlocals > 0) {
12618 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
12619 sizeof (dtrace_statvar_t *));
12620 }
12621 }
12622
12623 static void
12624 dtrace_state_clean(dtrace_state_t *state)
12625 {
12626 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12627 return;
12628
12629 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12630 dtrace_speculation_clean(state);
12631 }
12632
12633 static void
12634 dtrace_state_deadman(dtrace_state_t *state)
12635 {
12636 hrtime_t now;
12637
12638 dtrace_sync();
12639
12640 now = dtrace_gethrtime();
12641
12642 if (state != dtrace_anon.dta_state &&
12643 now - state->dts_laststatus >= dtrace_deadman_user)
12644 return;
12645
12646 /*
12647 * We must be sure that dts_alive never appears to be less than the
12648 * value upon entry to dtrace_state_deadman(), and because we lack a
12649 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12650 * store INT64_MAX to it, followed by a memory barrier, followed by
12651 * the new value. This assures that dts_alive never appears to be
12652 * less than its true value, regardless of the order in which the
12653 * stores to the underlying storage are issued.
12654 */
12655 state->dts_alive = INT64_MAX;
12656 dtrace_membar_producer();
12657 state->dts_alive = now;
12658 }
12659
12660 dtrace_state_t *
12661 dtrace_state_create(dev_t *devp, cred_t *cr)
12662 {
12663 minor_t minor;
12664 major_t major;
12665 char c[30];
12666 dtrace_state_t *state;
12667 dtrace_optval_t *opt;
12668 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12669
12670 ASSERT(MUTEX_HELD(&dtrace_lock));
12671 ASSERT(MUTEX_HELD(&cpu_lock));
12672
12673 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12674 VM_BESTFIT | VM_SLEEP);
12675
12676 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12677 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12678 return (NULL);
12679 }
12680
12681 state = ddi_get_soft_state(dtrace_softstate, minor);
12682 state->dts_epid = DTRACE_EPIDNONE + 1;
12683
12684 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
12685 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12686 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12687
12688 if (devp != NULL) {
12689 major = getemajor(*devp);
12690 } else {
12691 major = ddi_driver_major(dtrace_devi);
12692 }
12693
12694 state->dts_dev = makedevice(major, minor);
12695
12696 if (devp != NULL)
12697 *devp = state->dts_dev;
12698
12699 /*
12700 * We allocate NCPU buffers. On the one hand, this can be quite
12701 * a bit of memory per instance (nearly 36K on a Starcat). On the
12702 * other hand, it saves an additional memory reference in the probe
12703 * path.
12704 */
12705 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12706 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12707 state->dts_cleaner = CYCLIC_NONE;
12708 state->dts_deadman = CYCLIC_NONE;
12709 state->dts_vstate.dtvs_state = state;
12710
12711 for (i = 0; i < DTRACEOPT_MAX; i++)
12712 state->dts_options[i] = DTRACEOPT_UNSET;
12713
12714 /*
12715 * Set the default options.
12716 */
12717 opt = state->dts_options;
12718 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12719 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12720 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12721 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12722 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12723 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12724 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12725 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12726 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12727 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12728 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12729 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12730 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12731 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12732
12733 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12734
12735 /*
12736 * Depending on the user credentials, we set flag bits which alter probe
12737 * visibility or the amount of destructiveness allowed. In the case of
12738 * actual anonymous tracing, or the possession of all privileges, all of
12739 * the normal checks are bypassed.
12740 */
12741 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
12742 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
12743 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
12744 } else {
12745 /*
12746 * Set up the credentials for this instantiation. We take a
12747 * hold on the credential to prevent it from disappearing on
12748 * us; this in turn prevents the zone_t referenced by this
12749 * credential from disappearing. This means that we can
12750 * examine the credential and the zone from probe context.
12751 */
12752 crhold(cr);
12753 state->dts_cred.dcr_cred = cr;
12754
12755 /*
12756 * CRA_PROC means "we have *some* privilege for dtrace" and
12757 * unlocks the use of variables like pid, zonename, etc.
12758 */
12759 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
12760 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12761 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
12762 }
12763
12764 /*
12765 * dtrace_user allows use of syscall and profile providers.
12766 * If the user also has proc_owner and/or proc_zone, we
12767 * extend the scope to include additional visibility and
12768 * destructive power.
12769 */
12770 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
12771 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
12772 state->dts_cred.dcr_visible |=
12773 DTRACE_CRV_ALLPROC;
12774
12775 state->dts_cred.dcr_action |=
12776 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12777 }
12778
12779 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
12780 state->dts_cred.dcr_visible |=
12781 DTRACE_CRV_ALLZONE;
12782
12783 state->dts_cred.dcr_action |=
12784 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12785 }
12786
12787 /*
12788 * If we have all privs in whatever zone this is,
12789 * we can do destructive things to processes which
12790 * have altered credentials.
12791 */
12792 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12793 cr->cr_zone->zone_privset)) {
12794 state->dts_cred.dcr_action |=
12795 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12796 }
12797 }
12798
12799 /*
12800 * Holding the dtrace_kernel privilege also implies that
12801 * the user has the dtrace_user privilege from a visibility
12802 * perspective. But without further privileges, some
12803 * destructive actions are not available.
12804 */
12805 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
12806 /*
12807 * Make all probes in all zones visible. However,
12808 * this doesn't mean that all actions become available
12809 * to all zones.
12810 */
12811 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
12812 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
12813
12814 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
12815 DTRACE_CRA_PROC;
12816 /*
12817 * Holding proc_owner means that destructive actions
12818 * for *this* zone are allowed.
12819 */
12820 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12821 state->dts_cred.dcr_action |=
12822 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12823
12824 /*
12825 * Holding proc_zone means that destructive actions
12826 * for this user/group ID in all zones is allowed.
12827 */
12828 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12829 state->dts_cred.dcr_action |=
12830 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12831
12832 /*
12833 * If we have all privs in whatever zone this is,
12834 * we can do destructive things to processes which
12835 * have altered credentials.
12836 */
12837 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12838 cr->cr_zone->zone_privset)) {
12839 state->dts_cred.dcr_action |=
12840 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12841 }
12842 }
12843
12844 /*
12845 * Holding the dtrace_proc privilege gives control over fasttrap
12846 * and pid providers. We need to grant wider destructive
12847 * privileges in the event that the user has proc_owner and/or
12848 * proc_zone.
12849 */
12850 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12851 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12852 state->dts_cred.dcr_action |=
12853 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12854
12855 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12856 state->dts_cred.dcr_action |=
12857 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12858 }
12859 }
12860
12861 return (state);
12862 }
12863
12864 static int
12865 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
12866 {
12867 dtrace_optval_t *opt = state->dts_options, size;
12868 processorid_t cpu;
12869 int flags = 0, rval, factor, divisor = 1;
12870
12871 ASSERT(MUTEX_HELD(&dtrace_lock));
12872 ASSERT(MUTEX_HELD(&cpu_lock));
12873 ASSERT(which < DTRACEOPT_MAX);
12874 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
12875 (state == dtrace_anon.dta_state &&
12876 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
12877
12878 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
12879 return (0);
12880
12881 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
12882 cpu = opt[DTRACEOPT_CPU];
12883
12884 if (which == DTRACEOPT_SPECSIZE)
12885 flags |= DTRACEBUF_NOSWITCH;
12886
12887 if (which == DTRACEOPT_BUFSIZE) {
12888 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
12889 flags |= DTRACEBUF_RING;
12890
12891 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
12892 flags |= DTRACEBUF_FILL;
12893
12894 if (state != dtrace_anon.dta_state ||
12895 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
12896 flags |= DTRACEBUF_INACTIVE;
12897 }
12898
12899 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
12900 /*
12901 * The size must be 8-byte aligned. If the size is not 8-byte
12902 * aligned, drop it down by the difference.
12903 */
12904 if (size & (sizeof (uint64_t) - 1))
12905 size -= size & (sizeof (uint64_t) - 1);
12906
12907 if (size < state->dts_reserve) {
12908 /*
12909 * Buffers always must be large enough to accommodate
12910 * their prereserved space. We return E2BIG instead
12911 * of ENOMEM in this case to allow for user-level
12912 * software to differentiate the cases.
12913 */
12914 return (E2BIG);
12915 }
12916
12917 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
12918
12919 if (rval != ENOMEM) {
12920 opt[which] = size;
12921 return (rval);
12922 }
12923
12924 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
12925 return (rval);
12926
12927 for (divisor = 2; divisor < factor; divisor <<= 1)
12928 continue;
12929 }
12930
12931 return (ENOMEM);
12932 }
12933
12934 static int
12935 dtrace_state_buffers(dtrace_state_t *state)
12936 {
12937 dtrace_speculation_t *spec = state->dts_speculations;
12938 int rval, i;
12939
12940 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
12941 DTRACEOPT_BUFSIZE)) != 0)
12942 return (rval);
12943
12944 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
12945 DTRACEOPT_AGGSIZE)) != 0)
12946 return (rval);
12947
12948 for (i = 0; i < state->dts_nspeculations; i++) {
12949 if ((rval = dtrace_state_buffer(state,
12950 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
12951 return (rval);
12952 }
12953
12954 return (0);
12955 }
12956
12957 static void
12958 dtrace_state_prereserve(dtrace_state_t *state)
12959 {
12960 dtrace_ecb_t *ecb;
12961 dtrace_probe_t *probe;
12962
12963 state->dts_reserve = 0;
12964
12965 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
12966 return;
12967
12968 /*
12969 * If our buffer policy is a "fill" buffer policy, we need to set the
12970 * prereserved space to be the space required by the END probes.
12971 */
12972 probe = dtrace_probes[dtrace_probeid_end - 1];
12973 ASSERT(probe != NULL);
12974
12975 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
12976 if (ecb->dte_state != state)
12977 continue;
12978
12979 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
12980 }
12981 }
12982
12983 static int
12984 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
12985 {
12986 dtrace_optval_t *opt = state->dts_options, sz, nspec;
12987 dtrace_speculation_t *spec;
12988 dtrace_buffer_t *buf;
12989 cyc_handler_t hdlr;
12990 cyc_time_t when;
12991 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
12992 dtrace_icookie_t cookie;
12993
12994 mutex_enter(&cpu_lock);
12995 mutex_enter(&dtrace_lock);
12996
12997 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
12998 rval = EBUSY;
12999 goto out;
13000 }
13001
13002 /*
13003 * Before we can perform any checks, we must prime all of the
13004 * retained enablings that correspond to this state.
13005 */
13006 dtrace_enabling_prime(state);
13007
13008 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13009 rval = EACCES;
13010 goto out;
13011 }
13012
13013 dtrace_state_prereserve(state);
13014
13015 /*
13016 * Now we want to do is try to allocate our speculations.
13017 * We do not automatically resize the number of speculations; if
13018 * this fails, we will fail the operation.
13019 */
13020 nspec = opt[DTRACEOPT_NSPEC];
13021 ASSERT(nspec != DTRACEOPT_UNSET);
13022
13023 if (nspec > INT_MAX) {
13024 rval = ENOMEM;
13025 goto out;
13026 }
13027
13028 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
13029 KM_NOSLEEP | KM_NORMALPRI);
13030
13031 if (spec == NULL) {
13032 rval = ENOMEM;
13033 goto out;
13034 }
13035
13036 state->dts_speculations = spec;
13037 state->dts_nspeculations = (int)nspec;
13038
13039 for (i = 0; i < nspec; i++) {
13040 if ((buf = kmem_zalloc(bufsize,
13041 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
13042 rval = ENOMEM;
13043 goto err;
13044 }
13045
13046 spec[i].dtsp_buffer = buf;
13047 }
13048
13049 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13050 if (dtrace_anon.dta_state == NULL) {
13051 rval = ENOENT;
13052 goto out;
13053 }
13054
13055 if (state->dts_necbs != 0) {
13056 rval = EALREADY;
13057 goto out;
13058 }
13059
13060 state->dts_anon = dtrace_anon_grab();
13061 ASSERT(state->dts_anon != NULL);
13062 state = state->dts_anon;
13063
13064 /*
13065 * We want "grabanon" to be set in the grabbed state, so we'll
13066 * copy that option value from the grabbing state into the
13067 * grabbed state.
13068 */
13069 state->dts_options[DTRACEOPT_GRABANON] =
13070 opt[DTRACEOPT_GRABANON];
13071
13072 *cpu = dtrace_anon.dta_beganon;
13073
13074 /*
13075 * If the anonymous state is active (as it almost certainly
13076 * is if the anonymous enabling ultimately matched anything),
13077 * we don't allow any further option processing -- but we
13078 * don't return failure.
13079 */
13080 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13081 goto out;
13082 }
13083
13084 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13085 opt[DTRACEOPT_AGGSIZE] != 0) {
13086 if (state->dts_aggregations == NULL) {
13087 /*
13088 * We're not going to create an aggregation buffer
13089 * because we don't have any ECBs that contain
13090 * aggregations -- set this option to 0.
13091 */
13092 opt[DTRACEOPT_AGGSIZE] = 0;
13093 } else {
13094 /*
13095 * If we have an aggregation buffer, we must also have
13096 * a buffer to use as scratch.
13097 */
13098 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13099 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13100 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13101 }
13102 }
13103 }
13104
13105 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13106 opt[DTRACEOPT_SPECSIZE] != 0) {
13107 if (!state->dts_speculates) {
13108 /*
13109 * We're not going to create speculation buffers
13110 * because we don't have any ECBs that actually
13111 * speculate -- set the speculation size to 0.
13112 */
13113 opt[DTRACEOPT_SPECSIZE] = 0;
13114 }
13115 }
13116
13117 /*
13118 * The bare minimum size for any buffer that we're actually going to
13119 * do anything to is sizeof (uint64_t).
13120 */
13121 sz = sizeof (uint64_t);
13122
13123 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13124 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13125 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13126 /*
13127 * A buffer size has been explicitly set to 0 (or to a size
13128 * that will be adjusted to 0) and we need the space -- we
13129 * need to return failure. We return ENOSPC to differentiate
13130 * it from failing to allocate a buffer due to failure to meet
13131 * the reserve (for which we return E2BIG).
13132 */
13133 rval = ENOSPC;
13134 goto out;
13135 }
13136
13137 if ((rval = dtrace_state_buffers(state)) != 0)
13138 goto err;
13139
13140 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13141 sz = dtrace_dstate_defsize;
13142
13143 do {
13144 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13145
13146 if (rval == 0)
13147 break;
13148
13149 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13150 goto err;
13151 } while (sz >>= 1);
13152
13153 opt[DTRACEOPT_DYNVARSIZE] = sz;
13154
13155 if (rval != 0)
13156 goto err;
13157
13158 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13159 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13160
13161 if (opt[DTRACEOPT_CLEANRATE] == 0)
13162 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13163
13164 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13165 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13166
13167 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13168 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13169
13170 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13171 hdlr.cyh_arg = state;
13172 hdlr.cyh_level = CY_LOW_LEVEL;
13173
13174 when.cyt_when = 0;
13175 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13176
13177 state->dts_cleaner = cyclic_add(&hdlr, &when);
13178
13179 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13180 hdlr.cyh_arg = state;
13181 hdlr.cyh_level = CY_LOW_LEVEL;
13182
13183 when.cyt_when = 0;
13184 when.cyt_interval = dtrace_deadman_interval;
13185
13186 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13187 state->dts_deadman = cyclic_add(&hdlr, &when);
13188
13189 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13190
13191 if (state->dts_getf != 0 &&
13192 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13193 /*
13194 * We don't have kernel privs but we have at least one call
13195 * to getf(); we need to bump our zone's count, and (if
13196 * this is the first enabling to have an unprivileged call
13197 * to getf()) we need to hook into closef().
13198 */
13199 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
13200
13201 if (dtrace_getf++ == 0) {
13202 ASSERT(dtrace_closef == NULL);
13203 dtrace_closef = dtrace_getf_barrier;
13204 }
13205 }
13206
13207 /*
13208 * Now it's time to actually fire the BEGIN probe. We need to disable
13209 * interrupts here both to record the CPU on which we fired the BEGIN
13210 * probe (the data from this CPU will be processed first at user
13211 * level) and to manually activate the buffer for this CPU.
13212 */
13213 cookie = dtrace_interrupt_disable();
13214 *cpu = CPU->cpu_id;
13215 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13216 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13217
13218 dtrace_probe(dtrace_probeid_begin,
13219 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13220 dtrace_interrupt_enable(cookie);
13221 /*
13222 * We may have had an exit action from a BEGIN probe; only change our
13223 * state to ACTIVE if we're still in WARMUP.
13224 */
13225 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13226 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13227
13228 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13229 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13230
13231 /*
13232 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13233 * want each CPU to transition its principal buffer out of the
13234 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13235 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13236 * atomically transition from processing none of a state's ECBs to
13237 * processing all of them.
13238 */
13239 dtrace_xcall(DTRACE_CPUALL,
13240 (dtrace_xcall_t)dtrace_buffer_activate, state);
13241 goto out;
13242
13243 err:
13244 dtrace_buffer_free(state->dts_buffer);
13245 dtrace_buffer_free(state->dts_aggbuffer);
13246
13247 if ((nspec = state->dts_nspeculations) == 0) {
13248 ASSERT(state->dts_speculations == NULL);
13249 goto out;
13250 }
13251
13252 spec = state->dts_speculations;
13253 ASSERT(spec != NULL);
13254
13255 for (i = 0; i < state->dts_nspeculations; i++) {
13256 if ((buf = spec[i].dtsp_buffer) == NULL)
13257 break;
13258
13259 dtrace_buffer_free(buf);
13260 kmem_free(buf, bufsize);
13261 }
13262
13263 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13264 state->dts_nspeculations = 0;
13265 state->dts_speculations = NULL;
13266
13267 out:
13268 mutex_exit(&dtrace_lock);
13269 mutex_exit(&cpu_lock);
13270
13271 return (rval);
13272 }
13273
13274 static int
13275 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13276 {
13277 dtrace_icookie_t cookie;
13278
13279 ASSERT(MUTEX_HELD(&dtrace_lock));
13280
13281 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13282 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13283 return (EINVAL);
13284
13285 /*
13286 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13287 * to be sure that every CPU has seen it. See below for the details
13288 * on why this is done.
13289 */
13290 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13291 dtrace_sync();
13292
13293 /*
13294 * By this point, it is impossible for any CPU to be still processing
13295 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13296 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13297 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13298 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13299 * iff we're in the END probe.
13300 */
13301 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13302 dtrace_sync();
13303 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13304
13305 /*
13306 * Finally, we can release the reserve and call the END probe. We
13307 * disable interrupts across calling the END probe to allow us to
13308 * return the CPU on which we actually called the END probe. This
13309 * allows user-land to be sure that this CPU's principal buffer is
13310 * processed last.
13311 */
13312 state->dts_reserve = 0;
13313
13314 cookie = dtrace_interrupt_disable();
13315 *cpu = CPU->cpu_id;
13316 dtrace_probe(dtrace_probeid_end,
13317 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13318 dtrace_interrupt_enable(cookie);
13319
13320 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13321 dtrace_sync();
13322
13323 if (state->dts_getf != 0 &&
13324 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13325 /*
13326 * We don't have kernel privs but we have at least one call
13327 * to getf(); we need to lower our zone's count, and (if
13328 * this is the last enabling to have an unprivileged call
13329 * to getf()) we need to clear the closef() hook.
13330 */
13331 ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
13332 ASSERT(dtrace_closef == dtrace_getf_barrier);
13333 ASSERT(dtrace_getf > 0);
13334
13335 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
13336
13337 if (--dtrace_getf == 0)
13338 dtrace_closef = NULL;
13339 }
13340
13341 return (0);
13342 }
13343
13344 static int
13345 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13346 dtrace_optval_t val)
13347 {
13348 ASSERT(MUTEX_HELD(&dtrace_lock));
13349
13350 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13351 return (EBUSY);
13352
13353 if (option >= DTRACEOPT_MAX)
13354 return (EINVAL);
13355
13356 if (option != DTRACEOPT_CPU && val < 0)
13357 return (EINVAL);
13358
13359 switch (option) {
13360 case DTRACEOPT_DESTRUCTIVE:
13361 if (dtrace_destructive_disallow)
13362 return (EACCES);
13363
13364 state->dts_cred.dcr_destructive = 1;
13365 break;
13366
13367 case DTRACEOPT_BUFSIZE:
13368 case DTRACEOPT_DYNVARSIZE:
13369 case DTRACEOPT_AGGSIZE:
13370 case DTRACEOPT_SPECSIZE:
13371 case DTRACEOPT_STRSIZE:
13372 if (val < 0)
13373 return (EINVAL);
13374
13375 if (val >= LONG_MAX) {
13376 /*
13377 * If this is an otherwise negative value, set it to
13378 * the highest multiple of 128m less than LONG_MAX.
13379 * Technically, we're adjusting the size without
13380 * regard to the buffer resizing policy, but in fact,
13381 * this has no effect -- if we set the buffer size to
13382 * ~LONG_MAX and the buffer policy is ultimately set to
13383 * be "manual", the buffer allocation is guaranteed to
13384 * fail, if only because the allocation requires two
13385 * buffers. (We set the the size to the highest
13386 * multiple of 128m because it ensures that the size
13387 * will remain a multiple of a megabyte when
13388 * repeatedly halved -- all the way down to 15m.)
13389 */
13390 val = LONG_MAX - (1 << 27) + 1;
13391 }
13392 }
13393
13394 state->dts_options[option] = val;
13395
13396 return (0);
13397 }
13398
13399 static void
13400 dtrace_state_destroy(dtrace_state_t *state)
13401 {
13402 dtrace_ecb_t *ecb;
13403 dtrace_vstate_t *vstate = &state->dts_vstate;
13404 minor_t minor = getminor(state->dts_dev);
13405 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13406 dtrace_speculation_t *spec = state->dts_speculations;
13407 int nspec = state->dts_nspeculations;
13408 uint32_t match;
13409
13410 ASSERT(MUTEX_HELD(&dtrace_lock));
13411 ASSERT(MUTEX_HELD(&cpu_lock));
13412
13413 /*
13414 * First, retract any retained enablings for this state.
13415 */
13416 dtrace_enabling_retract(state);
13417 ASSERT(state->dts_nretained == 0);
13418
13419 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13420 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13421 /*
13422 * We have managed to come into dtrace_state_destroy() on a
13423 * hot enabling -- almost certainly because of a disorderly
13424 * shutdown of a consumer. (That is, a consumer that is
13425 * exiting without having called dtrace_stop().) In this case,
13426 * we're going to set our activity to be KILLED, and then
13427 * issue a sync to be sure that everyone is out of probe
13428 * context before we start blowing away ECBs.
13429 */
13430 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13431 dtrace_sync();
13432 }
13433
13434 /*
13435 * Release the credential hold we took in dtrace_state_create().
13436 */
13437 if (state->dts_cred.dcr_cred != NULL)
13438 crfree(state->dts_cred.dcr_cred);
13439
13440 /*
13441 * Now we can safely disable and destroy any enabled probes. Because
13442 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13443 * (especially if they're all enabled), we take two passes through the
13444 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13445 * in the second we disable whatever is left over.
13446 */
13447 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13448 for (i = 0; i < state->dts_necbs; i++) {
13449 if ((ecb = state->dts_ecbs[i]) == NULL)
13450 continue;
13451
13452 if (match && ecb->dte_probe != NULL) {
13453 dtrace_probe_t *probe = ecb->dte_probe;
13454 dtrace_provider_t *prov = probe->dtpr_provider;
13455
13456 if (!(prov->dtpv_priv.dtpp_flags & match))
13457 continue;
13458 }
13459
13460 dtrace_ecb_disable(ecb);
13461 dtrace_ecb_destroy(ecb);
13462 }
13463
13464 if (!match)
13465 break;
13466 }
13467
13468 /*
13469 * Before we free the buffers, perform one more sync to assure that
13470 * every CPU is out of probe context.
13471 */
13472 dtrace_sync();
13473
13474 dtrace_buffer_free(state->dts_buffer);
13475 dtrace_buffer_free(state->dts_aggbuffer);
13476
13477 for (i = 0; i < nspec; i++)
13478 dtrace_buffer_free(spec[i].dtsp_buffer);
13479
13480 if (state->dts_cleaner != CYCLIC_NONE)
13481 cyclic_remove(state->dts_cleaner);
13482
13483 if (state->dts_deadman != CYCLIC_NONE)
13484 cyclic_remove(state->dts_deadman);
13485
13486 dtrace_dstate_fini(&vstate->dtvs_dynvars);
13487 dtrace_vstate_fini(vstate);
13488 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13489
13490 if (state->dts_aggregations != NULL) {
13491 #ifdef DEBUG
13492 for (i = 0; i < state->dts_naggregations; i++)
13493 ASSERT(state->dts_aggregations[i] == NULL);
13494 #endif
13495 ASSERT(state->dts_naggregations > 0);
13496 kmem_free(state->dts_aggregations,
13497 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13498 }
13499
13500 kmem_free(state->dts_buffer, bufsize);
13501 kmem_free(state->dts_aggbuffer, bufsize);
13502
13503 for (i = 0; i < nspec; i++)
13504 kmem_free(spec[i].dtsp_buffer, bufsize);
13505
13506 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13507
13508 dtrace_format_destroy(state);
13509
13510 vmem_destroy(state->dts_aggid_arena);
13511 ddi_soft_state_free(dtrace_softstate, minor);
13512 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13513 }
13514
13515 /*
13516 * DTrace Anonymous Enabling Functions
13517 */
13518 static dtrace_state_t *
13519 dtrace_anon_grab(void)
13520 {
13521 dtrace_state_t *state;
13522
13523 ASSERT(MUTEX_HELD(&dtrace_lock));
13524
13525 if ((state = dtrace_anon.dta_state) == NULL) {
13526 ASSERT(dtrace_anon.dta_enabling == NULL);
13527 return (NULL);
13528 }
13529
13530 ASSERT(dtrace_anon.dta_enabling != NULL);
13531 ASSERT(dtrace_retained != NULL);
13532
13533 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13534 dtrace_anon.dta_enabling = NULL;
13535 dtrace_anon.dta_state = NULL;
13536
13537 return (state);
13538 }
13539
13540 static void
13541 dtrace_anon_property(void)
13542 {
13543 int i, rv;
13544 dtrace_state_t *state;
13545 dof_hdr_t *dof;
13546 char c[32]; /* enough for "dof-data-" + digits */
13547
13548 ASSERT(MUTEX_HELD(&dtrace_lock));
13549 ASSERT(MUTEX_HELD(&cpu_lock));
13550
13551 for (i = 0; ; i++) {
13552 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
13553
13554 dtrace_err_verbose = 1;
13555
13556 if ((dof = dtrace_dof_property(c)) == NULL) {
13557 dtrace_err_verbose = 0;
13558 break;
13559 }
13560
13561 /*
13562 * We want to create anonymous state, so we need to transition
13563 * the kernel debugger to indicate that DTrace is active. If
13564 * this fails (e.g. because the debugger has modified text in
13565 * some way), we won't continue with the processing.
13566 */
13567 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13568 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
13569 "enabling ignored.");
13570 dtrace_dof_destroy(dof);
13571 break;
13572 }
13573
13574 /*
13575 * If we haven't allocated an anonymous state, we'll do so now.
13576 */
13577 if ((state = dtrace_anon.dta_state) == NULL) {
13578 state = dtrace_state_create(NULL, NULL);
13579 dtrace_anon.dta_state = state;
13580
13581 if (state == NULL) {
13582 /*
13583 * This basically shouldn't happen: the only
13584 * failure mode from dtrace_state_create() is a
13585 * failure of ddi_soft_state_zalloc() that
13586 * itself should never happen. Still, the
13587 * interface allows for a failure mode, and
13588 * we want to fail as gracefully as possible:
13589 * we'll emit an error message and cease
13590 * processing anonymous state in this case.
13591 */
13592 cmn_err(CE_WARN, "failed to create "
13593 "anonymous state");
13594 dtrace_dof_destroy(dof);
13595 break;
13596 }
13597 }
13598
13599 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
13600 &dtrace_anon.dta_enabling, 0, B_TRUE);
13601
13602 if (rv == 0)
13603 rv = dtrace_dof_options(dof, state);
13604
13605 dtrace_err_verbose = 0;
13606 dtrace_dof_destroy(dof);
13607
13608 if (rv != 0) {
13609 /*
13610 * This is malformed DOF; chuck any anonymous state
13611 * that we created.
13612 */
13613 ASSERT(dtrace_anon.dta_enabling == NULL);
13614 dtrace_state_destroy(state);
13615 dtrace_anon.dta_state = NULL;
13616 break;
13617 }
13618
13619 ASSERT(dtrace_anon.dta_enabling != NULL);
13620 }
13621
13622 if (dtrace_anon.dta_enabling != NULL) {
13623 int rval;
13624
13625 /*
13626 * dtrace_enabling_retain() can only fail because we are
13627 * trying to retain more enablings than are allowed -- but
13628 * we only have one anonymous enabling, and we are guaranteed
13629 * to be allowed at least one retained enabling; we assert
13630 * that dtrace_enabling_retain() returns success.
13631 */
13632 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
13633 ASSERT(rval == 0);
13634
13635 dtrace_enabling_dump(dtrace_anon.dta_enabling);
13636 }
13637 }
13638
13639 /*
13640 * DTrace Helper Functions
13641 */
13642 static void
13643 dtrace_helper_trace(dtrace_helper_action_t *helper,
13644 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
13645 {
13646 uint32_t size, next, nnext, i;
13647 dtrace_helptrace_t *ent;
13648 uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13649
13650 if (!dtrace_helptrace_enabled)
13651 return;
13652
13653 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
13654
13655 /*
13656 * What would a tracing framework be without its own tracing
13657 * framework? (Well, a hell of a lot simpler, for starters...)
13658 */
13659 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
13660 sizeof (uint64_t) - sizeof (uint64_t);
13661
13662 /*
13663 * Iterate until we can allocate a slot in the trace buffer.
13664 */
13665 do {
13666 next = dtrace_helptrace_next;
13667
13668 if (next + size < dtrace_helptrace_bufsize) {
13669 nnext = next + size;
13670 } else {
13671 nnext = size;
13672 }
13673 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
13674
13675 /*
13676 * We have our slot; fill it in.
13677 */
13678 if (nnext == size)
13679 next = 0;
13680
13681 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
13682 ent->dtht_helper = helper;
13683 ent->dtht_where = where;
13684 ent->dtht_nlocals = vstate->dtvs_nlocals;
13685
13686 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13687 mstate->dtms_fltoffs : -1;
13688 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13689 ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
13690
13691 for (i = 0; i < vstate->dtvs_nlocals; i++) {
13692 dtrace_statvar_t *svar;
13693
13694 if ((svar = vstate->dtvs_locals[i]) == NULL)
13695 continue;
13696
13697 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13698 ent->dtht_locals[i] =
13699 ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
13700 }
13701 }
13702
13703 static uint64_t
13704 dtrace_helper(int which, dtrace_mstate_t *mstate,
13705 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13706 {
13707 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13708 uint64_t sarg0 = mstate->dtms_arg[0];
13709 uint64_t sarg1 = mstate->dtms_arg[1];
13710 uint64_t rval;
13711 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13712 dtrace_helper_action_t *helper;
13713 dtrace_vstate_t *vstate;
13714 dtrace_difo_t *pred;
13715 int i, trace = dtrace_helptrace_enabled;
13716
13717 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13718
13719 if (helpers == NULL)
13720 return (0);
13721
13722 if ((helper = helpers->dthps_actions[which]) == NULL)
13723 return (0);
13724
13725 vstate = &helpers->dthps_vstate;
13726 mstate->dtms_arg[0] = arg0;
13727 mstate->dtms_arg[1] = arg1;
13728
13729 /*
13730 * Now iterate over each helper. If its predicate evaluates to 'true',
13731 * we'll call the corresponding actions. Note that the below calls
13732 * to dtrace_dif_emulate() may set faults in machine state. This is
13733 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
13734 * the stored DIF offset with its own (which is the desired behavior).
13735 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13736 * from machine state; this is okay, too.
13737 */
13738 for (; helper != NULL; helper = helper->dtha_next) {
13739 if ((pred = helper->dtha_predicate) != NULL) {
13740 if (trace)
13741 dtrace_helper_trace(helper, mstate, vstate, 0);
13742
13743 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
13744 goto next;
13745
13746 if (*flags & CPU_DTRACE_FAULT)
13747 goto err;
13748 }
13749
13750 for (i = 0; i < helper->dtha_nactions; i++) {
13751 if (trace)
13752 dtrace_helper_trace(helper,
13753 mstate, vstate, i + 1);
13754
13755 rval = dtrace_dif_emulate(helper->dtha_actions[i],
13756 mstate, vstate, state);
13757
13758 if (*flags & CPU_DTRACE_FAULT)
13759 goto err;
13760 }
13761
13762 next:
13763 if (trace)
13764 dtrace_helper_trace(helper, mstate, vstate,
13765 DTRACE_HELPTRACE_NEXT);
13766 }
13767
13768 if (trace)
13769 dtrace_helper_trace(helper, mstate, vstate,
13770 DTRACE_HELPTRACE_DONE);
13771
13772 /*
13773 * Restore the arg0 that we saved upon entry.
13774 */
13775 mstate->dtms_arg[0] = sarg0;
13776 mstate->dtms_arg[1] = sarg1;
13777
13778 return (rval);
13779
13780 err:
13781 if (trace)
13782 dtrace_helper_trace(helper, mstate, vstate,
13783 DTRACE_HELPTRACE_ERR);
13784
13785 /*
13786 * Restore the arg0 that we saved upon entry.
13787 */
13788 mstate->dtms_arg[0] = sarg0;
13789 mstate->dtms_arg[1] = sarg1;
13790
13791 return (NULL);
13792 }
13793
13794 static void
13795 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
13796 dtrace_vstate_t *vstate)
13797 {
13798 int i;
13799
13800 if (helper->dtha_predicate != NULL)
13801 dtrace_difo_release(helper->dtha_predicate, vstate);
13802
13803 for (i = 0; i < helper->dtha_nactions; i++) {
13804 ASSERT(helper->dtha_actions[i] != NULL);
13805 dtrace_difo_release(helper->dtha_actions[i], vstate);
13806 }
13807
13808 kmem_free(helper->dtha_actions,
13809 helper->dtha_nactions * sizeof (dtrace_difo_t *));
13810 kmem_free(helper, sizeof (dtrace_helper_action_t));
13811 }
13812
13813 static int
13814 dtrace_helper_destroygen(int gen)
13815 {
13816 proc_t *p = curproc;
13817 dtrace_helpers_t *help = p->p_dtrace_helpers;
13818 dtrace_vstate_t *vstate;
13819 int i;
13820
13821 ASSERT(MUTEX_HELD(&dtrace_lock));
13822
13823 if (help == NULL || gen > help->dthps_generation)
13824 return (EINVAL);
13825
13826 vstate = &help->dthps_vstate;
13827
13828 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13829 dtrace_helper_action_t *last = NULL, *h, *next;
13830
13831 for (h = help->dthps_actions[i]; h != NULL; h = next) {
13832 next = h->dtha_next;
13833
13834 if (h->dtha_generation == gen) {
13835 if (last != NULL) {
13836 last->dtha_next = next;
13837 } else {
13838 help->dthps_actions[i] = next;
13839 }
13840
13841 dtrace_helper_action_destroy(h, vstate);
13842 } else {
13843 last = h;
13844 }
13845 }
13846 }
13847
13848 /*
13849 * Interate until we've cleared out all helper providers with the
13850 * given generation number.
13851 */
13852 for (;;) {
13853 dtrace_helper_provider_t *prov;
13854
13855 /*
13856 * Look for a helper provider with the right generation. We
13857 * have to start back at the beginning of the list each time
13858 * because we drop dtrace_lock. It's unlikely that we'll make
13859 * more than two passes.
13860 */
13861 for (i = 0; i < help->dthps_nprovs; i++) {
13862 prov = help->dthps_provs[i];
13863
13864 if (prov->dthp_generation == gen)
13865 break;
13866 }
13867
13868 /*
13869 * If there were no matches, we're done.
13870 */
13871 if (i == help->dthps_nprovs)
13872 break;
13873
13874 /*
13875 * Move the last helper provider into this slot.
13876 */
13877 help->dthps_nprovs--;
13878 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
13879 help->dthps_provs[help->dthps_nprovs] = NULL;
13880
13881 mutex_exit(&dtrace_lock);
13882
13883 /*
13884 * If we have a meta provider, remove this helper provider.
13885 */
13886 mutex_enter(&dtrace_meta_lock);
13887 if (dtrace_meta_pid != NULL) {
13888 ASSERT(dtrace_deferred_pid == NULL);
13889 dtrace_helper_provider_remove(&prov->dthp_prov,
13890 p->p_pid);
13891 }
13892 mutex_exit(&dtrace_meta_lock);
13893
13894 dtrace_helper_provider_destroy(prov);
13895
13896 mutex_enter(&dtrace_lock);
13897 }
13898
13899 return (0);
13900 }
13901
13902 static int
13903 dtrace_helper_validate(dtrace_helper_action_t *helper)
13904 {
13905 int err = 0, i;
13906 dtrace_difo_t *dp;
13907
13908 if ((dp = helper->dtha_predicate) != NULL)
13909 err += dtrace_difo_validate_helper(dp);
13910
13911 for (i = 0; i < helper->dtha_nactions; i++)
13912 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
13913
13914 return (err == 0);
13915 }
13916
13917 static int
13918 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
13919 {
13920 dtrace_helpers_t *help;
13921 dtrace_helper_action_t *helper, *last;
13922 dtrace_actdesc_t *act;
13923 dtrace_vstate_t *vstate;
13924 dtrace_predicate_t *pred;
13925 int count = 0, nactions = 0, i;
13926
13927 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
13928 return (EINVAL);
13929
13930 help = curproc->p_dtrace_helpers;
13931 last = help->dthps_actions[which];
13932 vstate = &help->dthps_vstate;
13933
13934 for (count = 0; last != NULL; last = last->dtha_next) {
13935 count++;
13936 if (last->dtha_next == NULL)
13937 break;
13938 }
13939
13940 /*
13941 * If we already have dtrace_helper_actions_max helper actions for this
13942 * helper action type, we'll refuse to add a new one.
13943 */
13944 if (count >= dtrace_helper_actions_max)
13945 return (ENOSPC);
13946
13947 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
13948 helper->dtha_generation = help->dthps_generation;
13949
13950 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
13951 ASSERT(pred->dtp_difo != NULL);
13952 dtrace_difo_hold(pred->dtp_difo);
13953 helper->dtha_predicate = pred->dtp_difo;
13954 }
13955
13956 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
13957 if (act->dtad_kind != DTRACEACT_DIFEXPR)
13958 goto err;
13959
13960 if (act->dtad_difo == NULL)
13961 goto err;
13962
13963 nactions++;
13964 }
13965
13966 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
13967 (helper->dtha_nactions = nactions), KM_SLEEP);
13968
13969 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
13970 dtrace_difo_hold(act->dtad_difo);
13971 helper->dtha_actions[i++] = act->dtad_difo;
13972 }
13973
13974 if (!dtrace_helper_validate(helper))
13975 goto err;
13976
13977 if (last == NULL) {
13978 help->dthps_actions[which] = helper;
13979 } else {
13980 last->dtha_next = helper;
13981 }
13982
13983 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
13984 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
13985 dtrace_helptrace_next = 0;
13986 }
13987
13988 return (0);
13989 err:
13990 dtrace_helper_action_destroy(helper, vstate);
13991 return (EINVAL);
13992 }
13993
13994 static void
13995 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
13996 dof_helper_t *dofhp)
13997 {
13998 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
13999
14000 mutex_enter(&dtrace_meta_lock);
14001 mutex_enter(&dtrace_lock);
14002
14003 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14004 /*
14005 * If the dtrace module is loaded but not attached, or if
14006 * there aren't isn't a meta provider registered to deal with
14007 * these provider descriptions, we need to postpone creating
14008 * the actual providers until later.
14009 */
14010
14011 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14012 dtrace_deferred_pid != help) {
14013 help->dthps_deferred = 1;
14014 help->dthps_pid = p->p_pid;
14015 help->dthps_next = dtrace_deferred_pid;
14016 help->dthps_prev = NULL;
14017 if (dtrace_deferred_pid != NULL)
14018 dtrace_deferred_pid->dthps_prev = help;
14019 dtrace_deferred_pid = help;
14020 }
14021
14022 mutex_exit(&dtrace_lock);
14023
14024 } else if (dofhp != NULL) {
14025 /*
14026 * If the dtrace module is loaded and we have a particular
14027 * helper provider description, pass that off to the
14028 * meta provider.
14029 */
14030
14031 mutex_exit(&dtrace_lock);
14032
14033 dtrace_helper_provide(dofhp, p->p_pid);
14034
14035 } else {
14036 /*
14037 * Otherwise, just pass all the helper provider descriptions
14038 * off to the meta provider.
14039 */
14040
14041 int i;
14042 mutex_exit(&dtrace_lock);
14043
14044 for (i = 0; i < help->dthps_nprovs; i++) {
14045 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14046 p->p_pid);
14047 }
14048 }
14049
14050 mutex_exit(&dtrace_meta_lock);
14051 }
14052
14053 static int
14054 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14055 {
14056 dtrace_helpers_t *help;
14057 dtrace_helper_provider_t *hprov, **tmp_provs;
14058 uint_t tmp_maxprovs, i;
14059
14060 ASSERT(MUTEX_HELD(&dtrace_lock));
14061
14062 help = curproc->p_dtrace_helpers;
14063 ASSERT(help != NULL);
14064
14065 /*
14066 * If we already have dtrace_helper_providers_max helper providers,
14067 * we're refuse to add a new one.
14068 */
14069 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14070 return (ENOSPC);
14071
14072 /*
14073 * Check to make sure this isn't a duplicate.
14074 */
14075 for (i = 0; i < help->dthps_nprovs; i++) {
14076 if (dofhp->dofhp_addr ==
14077 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14078 return (EALREADY);
14079 }
14080
14081 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14082 hprov->dthp_prov = *dofhp;
14083 hprov->dthp_ref = 1;
14084 hprov->dthp_generation = gen;
14085
14086 /*
14087 * Allocate a bigger table for helper providers if it's already full.
14088 */
14089 if (help->dthps_maxprovs == help->dthps_nprovs) {
14090 tmp_maxprovs = help->dthps_maxprovs;
14091 tmp_provs = help->dthps_provs;
14092
14093 if (help->dthps_maxprovs == 0)
14094 help->dthps_maxprovs = 2;
14095 else
14096 help->dthps_maxprovs *= 2;
14097 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14098 help->dthps_maxprovs = dtrace_helper_providers_max;
14099
14100 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14101
14102 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14103 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14104
14105 if (tmp_provs != NULL) {
14106 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14107 sizeof (dtrace_helper_provider_t *));
14108 kmem_free(tmp_provs, tmp_maxprovs *
14109 sizeof (dtrace_helper_provider_t *));
14110 }
14111 }
14112
14113 help->dthps_provs[help->dthps_nprovs] = hprov;
14114 help->dthps_nprovs++;
14115
14116 return (0);
14117 }
14118
14119 static void
14120 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14121 {
14122 mutex_enter(&dtrace_lock);
14123
14124 if (--hprov->dthp_ref == 0) {
14125 dof_hdr_t *dof;
14126 mutex_exit(&dtrace_lock);
14127 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14128 dtrace_dof_destroy(dof);
14129 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14130 } else {
14131 mutex_exit(&dtrace_lock);
14132 }
14133 }
14134
14135 static int
14136 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14137 {
14138 uintptr_t daddr = (uintptr_t)dof;
14139 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14140 dof_provider_t *provider;
14141 dof_probe_t *probe;
14142 uint8_t *arg;
14143 char *strtab, *typestr;
14144 dof_stridx_t typeidx;
14145 size_t typesz;
14146 uint_t nprobes, j, k;
14147
14148 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14149
14150 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14151 dtrace_dof_error(dof, "misaligned section offset");
14152 return (-1);
14153 }
14154
14155 /*
14156 * The section needs to be large enough to contain the DOF provider
14157 * structure appropriate for the given version.
14158 */
14159 if (sec->dofs_size <
14160 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14161 offsetof(dof_provider_t, dofpv_prenoffs) :
14162 sizeof (dof_provider_t))) {
14163 dtrace_dof_error(dof, "provider section too small");
14164 return (-1);
14165 }
14166
14167 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14168 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14169 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14170 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14171 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14172
14173 if (str_sec == NULL || prb_sec == NULL ||
14174 arg_sec == NULL || off_sec == NULL)
14175 return (-1);
14176
14177 enoff_sec = NULL;
14178
14179 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14180 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14181 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14182 provider->dofpv_prenoffs)) == NULL)
14183 return (-1);
14184
14185 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14186
14187 if (provider->dofpv_name >= str_sec->dofs_size ||
14188 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14189 dtrace_dof_error(dof, "invalid provider name");
14190 return (-1);
14191 }
14192
14193 if (prb_sec->dofs_entsize == 0 ||
14194 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14195 dtrace_dof_error(dof, "invalid entry size");
14196 return (-1);
14197 }
14198
14199 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14200 dtrace_dof_error(dof, "misaligned entry size");
14201 return (-1);
14202 }
14203
14204 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14205 dtrace_dof_error(dof, "invalid entry size");
14206 return (-1);
14207 }
14208
14209 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14210 dtrace_dof_error(dof, "misaligned section offset");
14211 return (-1);
14212 }
14213
14214 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14215 dtrace_dof_error(dof, "invalid entry size");
14216 return (-1);
14217 }
14218
14219 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14220
14221 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14222
14223 /*
14224 * Take a pass through the probes to check for errors.
14225 */
14226 for (j = 0; j < nprobes; j++) {
14227 probe = (dof_probe_t *)(uintptr_t)(daddr +
14228 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14229
14230 if (probe->dofpr_func >= str_sec->dofs_size) {
14231 dtrace_dof_error(dof, "invalid function name");
14232 return (-1);
14233 }
14234
14235 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14236 dtrace_dof_error(dof, "function name too long");
14237 return (-1);
14238 }
14239
14240 if (probe->dofpr_name >= str_sec->dofs_size ||
14241 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14242 dtrace_dof_error(dof, "invalid probe name");
14243 return (-1);
14244 }
14245
14246 /*
14247 * The offset count must not wrap the index, and the offsets
14248 * must also not overflow the section's data.
14249 */
14250 if (probe->dofpr_offidx + probe->dofpr_noffs <
14251 probe->dofpr_offidx ||
14252 (probe->dofpr_offidx + probe->dofpr_noffs) *
14253 off_sec->dofs_entsize > off_sec->dofs_size) {
14254 dtrace_dof_error(dof, "invalid probe offset");
14255 return (-1);
14256 }
14257
14258 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14259 /*
14260 * If there's no is-enabled offset section, make sure
14261 * there aren't any is-enabled offsets. Otherwise
14262 * perform the same checks as for probe offsets
14263 * (immediately above).
14264 */
14265 if (enoff_sec == NULL) {
14266 if (probe->dofpr_enoffidx != 0 ||
14267 probe->dofpr_nenoffs != 0) {
14268 dtrace_dof_error(dof, "is-enabled "
14269 "offsets with null section");
14270 return (-1);
14271 }
14272 } else if (probe->dofpr_enoffidx +
14273 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14274 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14275 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14276 dtrace_dof_error(dof, "invalid is-enabled "
14277 "offset");
14278 return (-1);
14279 }
14280
14281 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14282 dtrace_dof_error(dof, "zero probe and "
14283 "is-enabled offsets");
14284 return (-1);
14285 }
14286 } else if (probe->dofpr_noffs == 0) {
14287 dtrace_dof_error(dof, "zero probe offsets");
14288 return (-1);
14289 }
14290
14291 if (probe->dofpr_argidx + probe->dofpr_xargc <
14292 probe->dofpr_argidx ||
14293 (probe->dofpr_argidx + probe->dofpr_xargc) *
14294 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14295 dtrace_dof_error(dof, "invalid args");
14296 return (-1);
14297 }
14298
14299 typeidx = probe->dofpr_nargv;
14300 typestr = strtab + probe->dofpr_nargv;
14301 for (k = 0; k < probe->dofpr_nargc; k++) {
14302 if (typeidx >= str_sec->dofs_size) {
14303 dtrace_dof_error(dof, "bad "
14304 "native argument type");
14305 return (-1);
14306 }
14307
14308 typesz = strlen(typestr) + 1;
14309 if (typesz > DTRACE_ARGTYPELEN) {
14310 dtrace_dof_error(dof, "native "
14311 "argument type too long");
14312 return (-1);
14313 }
14314 typeidx += typesz;
14315 typestr += typesz;
14316 }
14317
14318 typeidx = probe->dofpr_xargv;
14319 typestr = strtab + probe->dofpr_xargv;
14320 for (k = 0; k < probe->dofpr_xargc; k++) {
14321 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14322 dtrace_dof_error(dof, "bad "
14323 "native argument index");
14324 return (-1);
14325 }
14326
14327 if (typeidx >= str_sec->dofs_size) {
14328 dtrace_dof_error(dof, "bad "
14329 "translated argument type");
14330 return (-1);
14331 }
14332
14333 typesz = strlen(typestr) + 1;
14334 if (typesz > DTRACE_ARGTYPELEN) {
14335 dtrace_dof_error(dof, "translated argument "
14336 "type too long");
14337 return (-1);
14338 }
14339
14340 typeidx += typesz;
14341 typestr += typesz;
14342 }
14343 }
14344
14345 return (0);
14346 }
14347
14348 static int
14349 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14350 {
14351 dtrace_helpers_t *help;
14352 dtrace_vstate_t *vstate;
14353 dtrace_enabling_t *enab = NULL;
14354 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14355 uintptr_t daddr = (uintptr_t)dof;
14356
14357 ASSERT(MUTEX_HELD(&dtrace_lock));
14358
14359 if ((help = curproc->p_dtrace_helpers) == NULL)
14360 help = dtrace_helpers_create(curproc);
14361
14362 vstate = &help->dthps_vstate;
14363
14364 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14365 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14366 dtrace_dof_destroy(dof);
14367 return (rv);
14368 }
14369
14370 /*
14371 * Look for helper providers and validate their descriptions.
14372 */
14373 if (dhp != NULL) {
14374 for (i = 0; i < dof->dofh_secnum; i++) {
14375 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14376 dof->dofh_secoff + i * dof->dofh_secsize);
14377
14378 if (sec->dofs_type != DOF_SECT_PROVIDER)
14379 continue;
14380
14381 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14382 dtrace_enabling_destroy(enab);
14383 dtrace_dof_destroy(dof);
14384 return (-1);
14385 }
14386
14387 nprovs++;
14388 }
14389 }
14390
14391 /*
14392 * Now we need to walk through the ECB descriptions in the enabling.
14393 */
14394 for (i = 0; i < enab->dten_ndesc; i++) {
14395 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14396 dtrace_probedesc_t *desc = &ep->dted_probe;
14397
14398 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14399 continue;
14400
14401 if (strcmp(desc->dtpd_mod, "helper") != 0)
14402 continue;
14403
14404 if (strcmp(desc->dtpd_func, "ustack") != 0)
14405 continue;
14406
14407 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14408 ep)) != 0) {
14409 /*
14410 * Adding this helper action failed -- we are now going
14411 * to rip out the entire generation and return failure.
14412 */
14413 (void) dtrace_helper_destroygen(help->dthps_generation);
14414 dtrace_enabling_destroy(enab);
14415 dtrace_dof_destroy(dof);
14416 return (-1);
14417 }
14418
14419 nhelpers++;
14420 }
14421
14422 if (nhelpers < enab->dten_ndesc)
14423 dtrace_dof_error(dof, "unmatched helpers");
14424
14425 gen = help->dthps_generation++;
14426 dtrace_enabling_destroy(enab);
14427
14428 if (dhp != NULL && nprovs > 0) {
14429 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14430 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14431 mutex_exit(&dtrace_lock);
14432 dtrace_helper_provider_register(curproc, help, dhp);
14433 mutex_enter(&dtrace_lock);
14434
14435 destroy = 0;
14436 }
14437 }
14438
14439 if (destroy)
14440 dtrace_dof_destroy(dof);
14441
14442 return (gen);
14443 }
14444
14445 static dtrace_helpers_t *
14446 dtrace_helpers_create(proc_t *p)
14447 {
14448 dtrace_helpers_t *help;
14449
14450 ASSERT(MUTEX_HELD(&dtrace_lock));
14451 ASSERT(p->p_dtrace_helpers == NULL);
14452
14453 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14454 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14455 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14456
14457 p->p_dtrace_helpers = help;
14458 dtrace_helpers++;
14459
14460 return (help);
14461 }
14462
14463 static void
14464 dtrace_helpers_destroy(void)
14465 {
14466 dtrace_helpers_t *help;
14467 dtrace_vstate_t *vstate;
14468 proc_t *p = curproc;
14469 int i;
14470
14471 mutex_enter(&dtrace_lock);
14472
14473 ASSERT(p->p_dtrace_helpers != NULL);
14474 ASSERT(dtrace_helpers > 0);
14475
14476 help = p->p_dtrace_helpers;
14477 vstate = &help->dthps_vstate;
14478
14479 /*
14480 * We're now going to lose the help from this process.
14481 */
14482 p->p_dtrace_helpers = NULL;
14483 dtrace_sync();
14484
14485 /*
14486 * Destory the helper actions.
14487 */
14488 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14489 dtrace_helper_action_t *h, *next;
14490
14491 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14492 next = h->dtha_next;
14493 dtrace_helper_action_destroy(h, vstate);
14494 h = next;
14495 }
14496 }
14497
14498 mutex_exit(&dtrace_lock);
14499
14500 /*
14501 * Destroy the helper providers.
14502 */
14503 if (help->dthps_maxprovs > 0) {
14504 mutex_enter(&dtrace_meta_lock);
14505 if (dtrace_meta_pid != NULL) {
14506 ASSERT(dtrace_deferred_pid == NULL);
14507
14508 for (i = 0; i < help->dthps_nprovs; i++) {
14509 dtrace_helper_provider_remove(
14510 &help->dthps_provs[i]->dthp_prov, p->p_pid);
14511 }
14512 } else {
14513 mutex_enter(&dtrace_lock);
14514 ASSERT(help->dthps_deferred == 0 ||
14515 help->dthps_next != NULL ||
14516 help->dthps_prev != NULL ||
14517 help == dtrace_deferred_pid);
14518
14519 /*
14520 * Remove the helper from the deferred list.
14521 */
14522 if (help->dthps_next != NULL)
14523 help->dthps_next->dthps_prev = help->dthps_prev;
14524 if (help->dthps_prev != NULL)
14525 help->dthps_prev->dthps_next = help->dthps_next;
14526 if (dtrace_deferred_pid == help) {
14527 dtrace_deferred_pid = help->dthps_next;
14528 ASSERT(help->dthps_prev == NULL);
14529 }
14530
14531 mutex_exit(&dtrace_lock);
14532 }
14533
14534 mutex_exit(&dtrace_meta_lock);
14535
14536 for (i = 0; i < help->dthps_nprovs; i++) {
14537 dtrace_helper_provider_destroy(help->dthps_provs[i]);
14538 }
14539
14540 kmem_free(help->dthps_provs, help->dthps_maxprovs *
14541 sizeof (dtrace_helper_provider_t *));
14542 }
14543
14544 mutex_enter(&dtrace_lock);
14545
14546 dtrace_vstate_fini(&help->dthps_vstate);
14547 kmem_free(help->dthps_actions,
14548 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
14549 kmem_free(help, sizeof (dtrace_helpers_t));
14550
14551 --dtrace_helpers;
14552 mutex_exit(&dtrace_lock);
14553 }
14554
14555 static void
14556 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
14557 {
14558 dtrace_helpers_t *help, *newhelp;
14559 dtrace_helper_action_t *helper, *new, *last;
14560 dtrace_difo_t *dp;
14561 dtrace_vstate_t *vstate;
14562 int i, j, sz, hasprovs = 0;
14563
14564 mutex_enter(&dtrace_lock);
14565 ASSERT(from->p_dtrace_helpers != NULL);
14566 ASSERT(dtrace_helpers > 0);
14567
14568 help = from->p_dtrace_helpers;
14569 newhelp = dtrace_helpers_create(to);
14570 ASSERT(to->p_dtrace_helpers != NULL);
14571
14572 newhelp->dthps_generation = help->dthps_generation;
14573 vstate = &newhelp->dthps_vstate;
14574
14575 /*
14576 * Duplicate the helper actions.
14577 */
14578 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14579 if ((helper = help->dthps_actions[i]) == NULL)
14580 continue;
14581
14582 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
14583 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
14584 KM_SLEEP);
14585 new->dtha_generation = helper->dtha_generation;
14586
14587 if ((dp = helper->dtha_predicate) != NULL) {
14588 dp = dtrace_difo_duplicate(dp, vstate);
14589 new->dtha_predicate = dp;
14590 }
14591
14592 new->dtha_nactions = helper->dtha_nactions;
14593 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
14594 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
14595
14596 for (j = 0; j < new->dtha_nactions; j++) {
14597 dtrace_difo_t *dp = helper->dtha_actions[j];
14598
14599 ASSERT(dp != NULL);
14600 dp = dtrace_difo_duplicate(dp, vstate);
14601 new->dtha_actions[j] = dp;
14602 }
14603
14604 if (last != NULL) {
14605 last->dtha_next = new;
14606 } else {
14607 newhelp->dthps_actions[i] = new;
14608 }
14609
14610 last = new;
14611 }
14612 }
14613
14614 /*
14615 * Duplicate the helper providers and register them with the
14616 * DTrace framework.
14617 */
14618 if (help->dthps_nprovs > 0) {
14619 newhelp->dthps_nprovs = help->dthps_nprovs;
14620 newhelp->dthps_maxprovs = help->dthps_nprovs;
14621 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
14622 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14623 for (i = 0; i < newhelp->dthps_nprovs; i++) {
14624 newhelp->dthps_provs[i] = help->dthps_provs[i];
14625 newhelp->dthps_provs[i]->dthp_ref++;
14626 }
14627
14628 hasprovs = 1;
14629 }
14630
14631 mutex_exit(&dtrace_lock);
14632
14633 if (hasprovs)
14634 dtrace_helper_provider_register(to, newhelp, NULL);
14635 }
14636
14637 /*
14638 * DTrace Hook Functions
14639 */
14640 static void
14641 dtrace_module_loaded(struct modctl *ctl)
14642 {
14643 dtrace_provider_t *prv;
14644
14645 mutex_enter(&dtrace_provider_lock);
14646 mutex_enter(&mod_lock);
14647
14648 ASSERT(ctl->mod_busy);
14649
14650 /*
14651 * We're going to call each providers per-module provide operation
14652 * specifying only this module.
14653 */
14654 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
14655 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
14656
14657 mutex_exit(&mod_lock);
14658 mutex_exit(&dtrace_provider_lock);
14659
14660 /*
14661 * If we have any retained enablings, we need to match against them.
14662 * Enabling probes requires that cpu_lock be held, and we cannot hold
14663 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
14664 * module. (In particular, this happens when loading scheduling
14665 * classes.) So if we have any retained enablings, we need to dispatch
14666 * our task queue to do the match for us.
14667 */
14668 mutex_enter(&dtrace_lock);
14669
14670 if (dtrace_retained == NULL) {
14671 mutex_exit(&dtrace_lock);
14672 return;
14673 }
14674
14675 (void) taskq_dispatch(dtrace_taskq,
14676 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
14677
14678 mutex_exit(&dtrace_lock);
14679
14680 /*
14681 * And now, for a little heuristic sleaze: in general, we want to
14682 * match modules as soon as they load. However, we cannot guarantee
14683 * this, because it would lead us to the lock ordering violation
14684 * outlined above. The common case, of course, is that cpu_lock is
14685 * _not_ held -- so we delay here for a clock tick, hoping that that's
14686 * long enough for the task queue to do its work. If it's not, it's
14687 * not a serious problem -- it just means that the module that we
14688 * just loaded may not be immediately instrumentable.
14689 */
14690 delay(1);
14691 }
14692
14693 static void
14694 dtrace_module_unloaded(struct modctl *ctl)
14695 {
14696 dtrace_probe_t template, *probe, *first, *next;
14697 dtrace_provider_t *prov;
14698
14699 template.dtpr_mod = ctl->mod_modname;
14700
14701 mutex_enter(&dtrace_provider_lock);
14702 mutex_enter(&mod_lock);
14703 mutex_enter(&dtrace_lock);
14704
14705 if (dtrace_bymod == NULL) {
14706 /*
14707 * The DTrace module is loaded (obviously) but not attached;
14708 * we don't have any work to do.
14709 */
14710 mutex_exit(&dtrace_provider_lock);
14711 mutex_exit(&mod_lock);
14712 mutex_exit(&dtrace_lock);
14713 return;
14714 }
14715
14716 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14717 probe != NULL; probe = probe->dtpr_nextmod) {
14718 if (probe->dtpr_ecb != NULL) {
14719 mutex_exit(&dtrace_provider_lock);
14720 mutex_exit(&mod_lock);
14721 mutex_exit(&dtrace_lock);
14722
14723 /*
14724 * This shouldn't _actually_ be possible -- we're
14725 * unloading a module that has an enabled probe in it.
14726 * (It's normally up to the provider to make sure that
14727 * this can't happen.) However, because dtps_enable()
14728 * doesn't have a failure mode, there can be an
14729 * enable/unload race. Upshot: we don't want to
14730 * assert, but we're not going to disable the
14731 * probe, either.
14732 */
14733 if (dtrace_err_verbose) {
14734 cmn_err(CE_WARN, "unloaded module '%s' had "
14735 "enabled probes", ctl->mod_modname);
14736 }
14737
14738 return;
14739 }
14740 }
14741
14742 probe = first;
14743
14744 for (first = NULL; probe != NULL; probe = next) {
14745 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
14746
14747 dtrace_probes[probe->dtpr_id - 1] = NULL;
14748
14749 next = probe->dtpr_nextmod;
14750 dtrace_hash_remove(dtrace_bymod, probe);
14751 dtrace_hash_remove(dtrace_byfunc, probe);
14752 dtrace_hash_remove(dtrace_byname, probe);
14753
14754 if (first == NULL) {
14755 first = probe;
14756 probe->dtpr_nextmod = NULL;
14757 } else {
14758 probe->dtpr_nextmod = first;
14759 first = probe;
14760 }
14761 }
14762
14763 /*
14764 * We've removed all of the module's probes from the hash chains and
14765 * from the probe array. Now issue a dtrace_sync() to be sure that
14766 * everyone has cleared out from any probe array processing.
14767 */
14768 dtrace_sync();
14769
14770 for (probe = first; probe != NULL; probe = first) {
14771 first = probe->dtpr_nextmod;
14772 prov = probe->dtpr_provider;
14773 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
14774 probe->dtpr_arg);
14775 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
14776 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
14777 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
14778 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
14779 kmem_free(probe, sizeof (dtrace_probe_t));
14780 }
14781
14782 mutex_exit(&dtrace_lock);
14783 mutex_exit(&mod_lock);
14784 mutex_exit(&dtrace_provider_lock);
14785 }
14786
14787 void
14788 dtrace_suspend(void)
14789 {
14790 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
14791 }
14792
14793 void
14794 dtrace_resume(void)
14795 {
14796 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
14797 }
14798
14799 static int
14800 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
14801 {
14802 ASSERT(MUTEX_HELD(&cpu_lock));
14803 mutex_enter(&dtrace_lock);
14804
14805 switch (what) {
14806 case CPU_CONFIG: {
14807 dtrace_state_t *state;
14808 dtrace_optval_t *opt, rs, c;
14809
14810 /*
14811 * For now, we only allocate a new buffer for anonymous state.
14812 */
14813 if ((state = dtrace_anon.dta_state) == NULL)
14814 break;
14815
14816 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14817 break;
14818
14819 opt = state->dts_options;
14820 c = opt[DTRACEOPT_CPU];
14821
14822 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
14823 break;
14824
14825 /*
14826 * Regardless of what the actual policy is, we're going to
14827 * temporarily set our resize policy to be manual. We're
14828 * also going to temporarily set our CPU option to denote
14829 * the newly configured CPU.
14830 */
14831 rs = opt[DTRACEOPT_BUFRESIZE];
14832 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
14833 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
14834
14835 (void) dtrace_state_buffers(state);
14836
14837 opt[DTRACEOPT_BUFRESIZE] = rs;
14838 opt[DTRACEOPT_CPU] = c;
14839
14840 break;
14841 }
14842
14843 case CPU_UNCONFIG:
14844 /*
14845 * We don't free the buffer in the CPU_UNCONFIG case. (The
14846 * buffer will be freed when the consumer exits.)
14847 */
14848 break;
14849
14850 default:
14851 break;
14852 }
14853
14854 mutex_exit(&dtrace_lock);
14855 return (0);
14856 }
14857
14858 static void
14859 dtrace_cpu_setup_initial(processorid_t cpu)
14860 {
14861 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
14862 }
14863
14864 static void
14865 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
14866 {
14867 if (dtrace_toxranges >= dtrace_toxranges_max) {
14868 int osize, nsize;
14869 dtrace_toxrange_t *range;
14870
14871 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14872
14873 if (osize == 0) {
14874 ASSERT(dtrace_toxrange == NULL);
14875 ASSERT(dtrace_toxranges_max == 0);
14876 dtrace_toxranges_max = 1;
14877 } else {
14878 dtrace_toxranges_max <<= 1;
14879 }
14880
14881 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14882 range = kmem_zalloc(nsize, KM_SLEEP);
14883
14884 if (dtrace_toxrange != NULL) {
14885 ASSERT(osize != 0);
14886 bcopy(dtrace_toxrange, range, osize);
14887 kmem_free(dtrace_toxrange, osize);
14888 }
14889
14890 dtrace_toxrange = range;
14891 }
14892
14893 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
14894 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
14895
14896 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
14897 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
14898 dtrace_toxranges++;
14899 }
14900
14901 static void
14902 dtrace_getf_barrier()
14903 {
14904 /*
14905 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
14906 * that contain calls to getf(), this routine will be called on every
14907 * closef() before either the underlying vnode is released or the
14908 * file_t itself is freed. By the time we are here, it is essential
14909 * that the file_t can no longer be accessed from a call to getf()
14910 * in probe context -- that assures that a dtrace_sync() can be used
14911 * to clear out any enablings referring to the old structures.
14912 */
14913 if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
14914 kcred->cr_zone->zone_dtrace_getf != 0)
14915 dtrace_sync();
14916 }
14917
14918 /*
14919 * DTrace Driver Cookbook Functions
14920 */
14921 /*ARGSUSED*/
14922 static int
14923 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
14924 {
14925 dtrace_provider_id_t id;
14926 dtrace_state_t *state = NULL;
14927 dtrace_enabling_t *enab;
14928
14929 mutex_enter(&cpu_lock);
14930 mutex_enter(&dtrace_provider_lock);
14931 mutex_enter(&dtrace_lock);
14932
14933 if (ddi_soft_state_init(&dtrace_softstate,
14934 sizeof (dtrace_state_t), 0) != 0) {
14935 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
14936 mutex_exit(&cpu_lock);
14937 mutex_exit(&dtrace_provider_lock);
14938 mutex_exit(&dtrace_lock);
14939 return (DDI_FAILURE);
14940 }
14941
14942 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
14943 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
14944 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
14945 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
14946 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
14947 ddi_remove_minor_node(devi, NULL);
14948 ddi_soft_state_fini(&dtrace_softstate);
14949 mutex_exit(&cpu_lock);
14950 mutex_exit(&dtrace_provider_lock);
14951 mutex_exit(&dtrace_lock);
14952 return (DDI_FAILURE);
14953 }
14954
14955 ddi_report_dev(devi);
14956 dtrace_devi = devi;
14957
14958 dtrace_modload = dtrace_module_loaded;
14959 dtrace_modunload = dtrace_module_unloaded;
14960 dtrace_cpu_init = dtrace_cpu_setup_initial;
14961 dtrace_helpers_cleanup = dtrace_helpers_destroy;
14962 dtrace_helpers_fork = dtrace_helpers_duplicate;
14963 dtrace_cpustart_init = dtrace_suspend;
14964 dtrace_cpustart_fini = dtrace_resume;
14965 dtrace_debugger_init = dtrace_suspend;
14966 dtrace_debugger_fini = dtrace_resume;
14967
14968 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
14969
14970 ASSERT(MUTEX_HELD(&cpu_lock));
14971
14972 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
14973 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14974 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
14975 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
14976 VM_SLEEP | VMC_IDENTIFIER);
14977 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
14978 1, INT_MAX, 0);
14979
14980 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
14981 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
14982 NULL, NULL, NULL, NULL, NULL, 0);
14983
14984 ASSERT(MUTEX_HELD(&cpu_lock));
14985 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
14986 offsetof(dtrace_probe_t, dtpr_nextmod),
14987 offsetof(dtrace_probe_t, dtpr_prevmod));
14988
14989 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
14990 offsetof(dtrace_probe_t, dtpr_nextfunc),
14991 offsetof(dtrace_probe_t, dtpr_prevfunc));
14992
14993 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
14994 offsetof(dtrace_probe_t, dtpr_nextname),
14995 offsetof(dtrace_probe_t, dtpr_prevname));
14996
14997 if (dtrace_retain_max < 1) {
14998 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
14999 "setting to 1", dtrace_retain_max);
15000 dtrace_retain_max = 1;
15001 }
15002
15003 /*
15004 * Now discover our toxic ranges.
15005 */
15006 dtrace_toxic_ranges(dtrace_toxrange_add);
15007
15008 /*
15009 * Before we register ourselves as a provider to our own framework,
15010 * we would like to assert that dtrace_provider is NULL -- but that's
15011 * not true if we were loaded as a dependency of a DTrace provider.
15012 * Once we've registered, we can assert that dtrace_provider is our
15013 * pseudo provider.
15014 */
15015 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15016 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15017
15018 ASSERT(dtrace_provider != NULL);
15019 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15020
15021 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15022 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15023 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15024 dtrace_provider, NULL, NULL, "END", 0, NULL);
15025 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15026 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15027
15028 dtrace_anon_property();
15029 mutex_exit(&cpu_lock);
15030
15031 /*
15032 * If DTrace helper tracing is enabled, we need to allocate the
15033 * trace buffer and initialize the values.
15034 */
15035 if (dtrace_helptrace_enabled) {
15036 ASSERT(dtrace_helptrace_buffer == NULL);
15037 dtrace_helptrace_buffer =
15038 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15039 dtrace_helptrace_next = 0;
15040 }
15041
15042 /*
15043 * If there are already providers, we must ask them to provide their
15044 * probes, and then match any anonymous enabling against them. Note
15045 * that there should be no other retained enablings at this time:
15046 * the only retained enablings at this time should be the anonymous
15047 * enabling.
15048 */
15049 if (dtrace_anon.dta_enabling != NULL) {
15050 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15051
15052 dtrace_enabling_provide(NULL);
15053 state = dtrace_anon.dta_state;
15054
15055 /*
15056 * We couldn't hold cpu_lock across the above call to
15057 * dtrace_enabling_provide(), but we must hold it to actually
15058 * enable the probes. We have to drop all of our locks, pick
15059 * up cpu_lock, and regain our locks before matching the
15060 * retained anonymous enabling.
15061 */
15062 mutex_exit(&dtrace_lock);
15063 mutex_exit(&dtrace_provider_lock);
15064
15065 mutex_enter(&cpu_lock);
15066 mutex_enter(&dtrace_provider_lock);
15067 mutex_enter(&dtrace_lock);
15068
15069 if ((enab = dtrace_anon.dta_enabling) != NULL)
15070 (void) dtrace_enabling_match(enab, NULL);
15071
15072 mutex_exit(&cpu_lock);
15073 }
15074
15075 mutex_exit(&dtrace_lock);
15076 mutex_exit(&dtrace_provider_lock);
15077
15078 if (state != NULL) {
15079 /*
15080 * If we created any anonymous state, set it going now.
15081 */
15082 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15083 }
15084
15085 return (DDI_SUCCESS);
15086 }
15087
15088 /*ARGSUSED*/
15089 static int
15090 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15091 {
15092 dtrace_state_t *state;
15093 uint32_t priv;
15094 uid_t uid;
15095 zoneid_t zoneid;
15096
15097 if (getminor(*devp) == DTRACEMNRN_HELPER)
15098 return (0);
15099
15100 /*
15101 * If this wasn't an open with the "helper" minor, then it must be
15102 * the "dtrace" minor.
15103 */
15104 if (getminor(*devp) != DTRACEMNRN_DTRACE)
15105 return (ENXIO);
15106
15107 /*
15108 * If no DTRACE_PRIV_* bits are set in the credential, then the
15109 * caller lacks sufficient permission to do anything with DTrace.
15110 */
15111 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15112 if (priv == DTRACE_PRIV_NONE)
15113 return (EACCES);
15114
15115 /*
15116 * Ask all providers to provide all their probes.
15117 */
15118 mutex_enter(&dtrace_provider_lock);
15119 dtrace_probe_provide(NULL, NULL);
15120 mutex_exit(&dtrace_provider_lock);
15121
15122 mutex_enter(&cpu_lock);
15123 mutex_enter(&dtrace_lock);
15124 dtrace_opens++;
15125 dtrace_membar_producer();
15126
15127 /*
15128 * If the kernel debugger is active (that is, if the kernel debugger
15129 * modified text in some way), we won't allow the open.
15130 */
15131 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15132 dtrace_opens--;
15133 mutex_exit(&cpu_lock);
15134 mutex_exit(&dtrace_lock);
15135 return (EBUSY);
15136 }
15137
15138 state = dtrace_state_create(devp, cred_p);
15139 mutex_exit(&cpu_lock);
15140
15141 if (state == NULL) {
15142 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15143 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15144 mutex_exit(&dtrace_lock);
15145 return (EAGAIN);
15146 }
15147
15148 mutex_exit(&dtrace_lock);
15149
15150 return (0);
15151 }
15152
15153 /*ARGSUSED*/
15154 static int
15155 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15156 {
15157 minor_t minor = getminor(dev);
15158 dtrace_state_t *state;
15159
15160 if (minor == DTRACEMNRN_HELPER)
15161 return (0);
15162
15163 state = ddi_get_soft_state(dtrace_softstate, minor);
15164
15165 mutex_enter(&cpu_lock);
15166 mutex_enter(&dtrace_lock);
15167
15168 if (state->dts_anon) {
15169 /*
15170 * There is anonymous state. Destroy that first.
15171 */
15172 ASSERT(dtrace_anon.dta_state == NULL);
15173 dtrace_state_destroy(state->dts_anon);
15174 }
15175
15176 dtrace_state_destroy(state);
15177 ASSERT(dtrace_opens > 0);
15178
15179 /*
15180 * Only relinquish control of the kernel debugger interface when there
15181 * are no consumers and no anonymous enablings.
15182 */
15183 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15184 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15185
15186 mutex_exit(&dtrace_lock);
15187 mutex_exit(&cpu_lock);
15188
15189 return (0);
15190 }
15191
15192 /*ARGSUSED*/
15193 static int
15194 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15195 {
15196 int rval;
15197 dof_helper_t help, *dhp = NULL;
15198
15199 switch (cmd) {
15200 case DTRACEHIOC_ADDDOF:
15201 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15202 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15203 return (EFAULT);
15204 }
15205
15206 dhp = &help;
15207 arg = (intptr_t)help.dofhp_dof;
15208 /*FALLTHROUGH*/
15209
15210 case DTRACEHIOC_ADD: {
15211 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15212
15213 if (dof == NULL)
15214 return (rval);
15215
15216 mutex_enter(&dtrace_lock);
15217
15218 /*
15219 * dtrace_helper_slurp() takes responsibility for the dof --
15220 * it may free it now or it may save it and free it later.
15221 */
15222 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15223 *rv = rval;
15224 rval = 0;
15225 } else {
15226 rval = EINVAL;
15227 }
15228
15229 mutex_exit(&dtrace_lock);
15230 return (rval);
15231 }
15232
15233 case DTRACEHIOC_REMOVE: {
15234 mutex_enter(&dtrace_lock);
15235 rval = dtrace_helper_destroygen(arg);
15236 mutex_exit(&dtrace_lock);
15237
15238 return (rval);
15239 }
15240
15241 default:
15242 break;
15243 }
15244
15245 return (ENOTTY);
15246 }
15247
15248 /*ARGSUSED*/
15249 static int
15250 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15251 {
15252 minor_t minor = getminor(dev);
15253 dtrace_state_t *state;
15254 int rval;
15255
15256 if (minor == DTRACEMNRN_HELPER)
15257 return (dtrace_ioctl_helper(cmd, arg, rv));
15258
15259 state = ddi_get_soft_state(dtrace_softstate, minor);
15260
15261 if (state->dts_anon) {
15262 ASSERT(dtrace_anon.dta_state == NULL);
15263 state = state->dts_anon;
15264 }
15265
15266 switch (cmd) {
15267 case DTRACEIOC_PROVIDER: {
15268 dtrace_providerdesc_t pvd;
15269 dtrace_provider_t *pvp;
15270
15271 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15272 return (EFAULT);
15273
15274 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15275 mutex_enter(&dtrace_provider_lock);
15276
15277 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15278 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15279 break;
15280 }
15281
15282 mutex_exit(&dtrace_provider_lock);
15283
15284 if (pvp == NULL)
15285 return (ESRCH);
15286
15287 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15288 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15289 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15290 return (EFAULT);
15291
15292 return (0);
15293 }
15294
15295 case DTRACEIOC_EPROBE: {
15296 dtrace_eprobedesc_t epdesc;
15297 dtrace_ecb_t *ecb;
15298 dtrace_action_t *act;
15299 void *buf;
15300 size_t size;
15301 uintptr_t dest;
15302 int nrecs;
15303
15304 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15305 return (EFAULT);
15306
15307 mutex_enter(&dtrace_lock);
15308
15309 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15310 mutex_exit(&dtrace_lock);
15311 return (EINVAL);
15312 }
15313
15314 if (ecb->dte_probe == NULL) {
15315 mutex_exit(&dtrace_lock);
15316 return (EINVAL);
15317 }
15318
15319 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15320 epdesc.dtepd_uarg = ecb->dte_uarg;
15321 epdesc.dtepd_size = ecb->dte_size;
15322
15323 nrecs = epdesc.dtepd_nrecs;
15324 epdesc.dtepd_nrecs = 0;
15325 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15326 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15327 continue;
15328
15329 epdesc.dtepd_nrecs++;
15330 }
15331
15332 /*
15333 * Now that we have the size, we need to allocate a temporary
15334 * buffer in which to store the complete description. We need
15335 * the temporary buffer to be able to drop dtrace_lock()
15336 * across the copyout(), below.
15337 */
15338 size = sizeof (dtrace_eprobedesc_t) +
15339 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15340
15341 buf = kmem_alloc(size, KM_SLEEP);
15342 dest = (uintptr_t)buf;
15343
15344 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15345 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15346
15347 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15348 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15349 continue;
15350
15351 if (nrecs-- == 0)
15352 break;
15353
15354 bcopy(&act->dta_rec, (void *)dest,
15355 sizeof (dtrace_recdesc_t));
15356 dest += sizeof (dtrace_recdesc_t);
15357 }
15358
15359 mutex_exit(&dtrace_lock);
15360
15361 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15362 kmem_free(buf, size);
15363 return (EFAULT);
15364 }
15365
15366 kmem_free(buf, size);
15367 return (0);
15368 }
15369
15370 case DTRACEIOC_AGGDESC: {
15371 dtrace_aggdesc_t aggdesc;
15372 dtrace_action_t *act;
15373 dtrace_aggregation_t *agg;
15374 int nrecs;
15375 uint32_t offs;
15376 dtrace_recdesc_t *lrec;
15377 void *buf;
15378 size_t size;
15379 uintptr_t dest;
15380
15381 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15382 return (EFAULT);
15383
15384 mutex_enter(&dtrace_lock);
15385
15386 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15387 mutex_exit(&dtrace_lock);
15388 return (EINVAL);
15389 }
15390
15391 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15392
15393 nrecs = aggdesc.dtagd_nrecs;
15394 aggdesc.dtagd_nrecs = 0;
15395
15396 offs = agg->dtag_base;
15397 lrec = &agg->dtag_action.dta_rec;
15398 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15399
15400 for (act = agg->dtag_first; ; act = act->dta_next) {
15401 ASSERT(act->dta_intuple ||
15402 DTRACEACT_ISAGG(act->dta_kind));
15403
15404 /*
15405 * If this action has a record size of zero, it
15406 * denotes an argument to the aggregating action.
15407 * Because the presence of this record doesn't (or
15408 * shouldn't) affect the way the data is interpreted,
15409 * we don't copy it out to save user-level the
15410 * confusion of dealing with a zero-length record.
15411 */
15412 if (act->dta_rec.dtrd_size == 0) {
15413 ASSERT(agg->dtag_hasarg);
15414 continue;
15415 }
15416
15417 aggdesc.dtagd_nrecs++;
15418
15419 if (act == &agg->dtag_action)
15420 break;
15421 }
15422
15423 /*
15424 * Now that we have the size, we need to allocate a temporary
15425 * buffer in which to store the complete description. We need
15426 * the temporary buffer to be able to drop dtrace_lock()
15427 * across the copyout(), below.
15428 */
15429 size = sizeof (dtrace_aggdesc_t) +
15430 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15431
15432 buf = kmem_alloc(size, KM_SLEEP);
15433 dest = (uintptr_t)buf;
15434
15435 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15436 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15437
15438 for (act = agg->dtag_first; ; act = act->dta_next) {
15439 dtrace_recdesc_t rec = act->dta_rec;
15440
15441 /*
15442 * See the comment in the above loop for why we pass
15443 * over zero-length records.
15444 */
15445 if (rec.dtrd_size == 0) {
15446 ASSERT(agg->dtag_hasarg);
15447 continue;
15448 }
15449
15450 if (nrecs-- == 0)
15451 break;
15452
15453 rec.dtrd_offset -= offs;
15454 bcopy(&rec, (void *)dest, sizeof (rec));
15455 dest += sizeof (dtrace_recdesc_t);
15456
15457 if (act == &agg->dtag_action)
15458 break;
15459 }
15460
15461 mutex_exit(&dtrace_lock);
15462
15463 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15464 kmem_free(buf, size);
15465 return (EFAULT);
15466 }
15467
15468 kmem_free(buf, size);
15469 return (0);
15470 }
15471
15472 case DTRACEIOC_ENABLE: {
15473 dof_hdr_t *dof;
15474 dtrace_enabling_t *enab = NULL;
15475 dtrace_vstate_t *vstate;
15476 int err = 0;
15477
15478 *rv = 0;
15479
15480 /*
15481 * If a NULL argument has been passed, we take this as our
15482 * cue to reevaluate our enablings.
15483 */
15484 if (arg == NULL) {
15485 dtrace_enabling_matchall();
15486
15487 return (0);
15488 }
15489
15490 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
15491 return (rval);
15492
15493 mutex_enter(&cpu_lock);
15494 mutex_enter(&dtrace_lock);
15495 vstate = &state->dts_vstate;
15496
15497 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
15498 mutex_exit(&dtrace_lock);
15499 mutex_exit(&cpu_lock);
15500 dtrace_dof_destroy(dof);
15501 return (EBUSY);
15502 }
15503
15504 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
15505 mutex_exit(&dtrace_lock);
15506 mutex_exit(&cpu_lock);
15507 dtrace_dof_destroy(dof);
15508 return (EINVAL);
15509 }
15510
15511 if ((rval = dtrace_dof_options(dof, state)) != 0) {
15512 dtrace_enabling_destroy(enab);
15513 mutex_exit(&dtrace_lock);
15514 mutex_exit(&cpu_lock);
15515 dtrace_dof_destroy(dof);
15516 return (rval);
15517 }
15518
15519 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
15520 err = dtrace_enabling_retain(enab);
15521 } else {
15522 dtrace_enabling_destroy(enab);
15523 }
15524
15525 mutex_exit(&cpu_lock);
15526 mutex_exit(&dtrace_lock);
15527 dtrace_dof_destroy(dof);
15528
15529 return (err);
15530 }
15531
15532 case DTRACEIOC_REPLICATE: {
15533 dtrace_repldesc_t desc;
15534 dtrace_probedesc_t *match = &desc.dtrpd_match;
15535 dtrace_probedesc_t *create = &desc.dtrpd_create;
15536 int err;
15537
15538 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15539 return (EFAULT);
15540
15541 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15542 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15543 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15544 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15545
15546 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15547 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15548 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15549 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15550
15551 mutex_enter(&dtrace_lock);
15552 err = dtrace_enabling_replicate(state, match, create);
15553 mutex_exit(&dtrace_lock);
15554
15555 return (err);
15556 }
15557
15558 case DTRACEIOC_PROBEMATCH:
15559 case DTRACEIOC_PROBES: {
15560 dtrace_probe_t *probe = NULL;
15561 dtrace_probedesc_t desc;
15562 dtrace_probekey_t pkey;
15563 dtrace_id_t i;
15564 int m = 0;
15565 uint32_t priv;
15566 uid_t uid;
15567 zoneid_t zoneid;
15568
15569 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15570 return (EFAULT);
15571
15572 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15573 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15574 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15575 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15576
15577 /*
15578 * Before we attempt to match this probe, we want to give
15579 * all providers the opportunity to provide it.
15580 */
15581 if (desc.dtpd_id == DTRACE_IDNONE) {
15582 mutex_enter(&dtrace_provider_lock);
15583 dtrace_probe_provide(&desc, NULL);
15584 mutex_exit(&dtrace_provider_lock);
15585 desc.dtpd_id++;
15586 }
15587
15588 if (cmd == DTRACEIOC_PROBEMATCH) {
15589 dtrace_probekey(&desc, &pkey);
15590 pkey.dtpk_id = DTRACE_IDNONE;
15591 }
15592
15593 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
15594
15595 mutex_enter(&dtrace_lock);
15596
15597 if (cmd == DTRACEIOC_PROBEMATCH) {
15598 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15599 if ((probe = dtrace_probes[i - 1]) != NULL &&
15600 (m = dtrace_match_probe(probe, &pkey,
15601 priv, uid, zoneid)) != 0)
15602 break;
15603 }
15604
15605 if (m < 0) {
15606 mutex_exit(&dtrace_lock);
15607 return (EINVAL);
15608 }
15609
15610 } else {
15611 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15612 if ((probe = dtrace_probes[i - 1]) != NULL &&
15613 dtrace_match_priv(probe, priv, uid, zoneid))
15614 break;
15615 }
15616 }
15617
15618 if (probe == NULL) {
15619 mutex_exit(&dtrace_lock);
15620 return (ESRCH);
15621 }
15622
15623 dtrace_probe_description(probe, &desc);
15624 mutex_exit(&dtrace_lock);
15625
15626 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15627 return (EFAULT);
15628
15629 return (0);
15630 }
15631
15632 case DTRACEIOC_PROBEARG: {
15633 dtrace_argdesc_t desc;
15634 dtrace_probe_t *probe;
15635 dtrace_provider_t *prov;
15636
15637 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15638 return (EFAULT);
15639
15640 if (desc.dtargd_id == DTRACE_IDNONE)
15641 return (EINVAL);
15642
15643 if (desc.dtargd_ndx == DTRACE_ARGNONE)
15644 return (EINVAL);
15645
15646 mutex_enter(&dtrace_provider_lock);
15647 mutex_enter(&mod_lock);
15648 mutex_enter(&dtrace_lock);
15649
15650 if (desc.dtargd_id > dtrace_nprobes) {
15651 mutex_exit(&dtrace_lock);
15652 mutex_exit(&mod_lock);
15653 mutex_exit(&dtrace_provider_lock);
15654 return (EINVAL);
15655 }
15656
15657 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
15658 mutex_exit(&dtrace_lock);
15659 mutex_exit(&mod_lock);
15660 mutex_exit(&dtrace_provider_lock);
15661 return (EINVAL);
15662 }
15663
15664 mutex_exit(&dtrace_lock);
15665
15666 prov = probe->dtpr_provider;
15667
15668 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
15669 /*
15670 * There isn't any typed information for this probe.
15671 * Set the argument number to DTRACE_ARGNONE.
15672 */
15673 desc.dtargd_ndx = DTRACE_ARGNONE;
15674 } else {
15675 desc.dtargd_native[0] = '\0';
15676 desc.dtargd_xlate[0] = '\0';
15677 desc.dtargd_mapping = desc.dtargd_ndx;
15678
15679 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
15680 probe->dtpr_id, probe->dtpr_arg, &desc);
15681 }
15682
15683 mutex_exit(&mod_lock);
15684 mutex_exit(&dtrace_provider_lock);
15685
15686 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15687 return (EFAULT);
15688
15689 return (0);
15690 }
15691
15692 case DTRACEIOC_GO: {
15693 processorid_t cpuid;
15694 rval = dtrace_state_go(state, &cpuid);
15695
15696 if (rval != 0)
15697 return (rval);
15698
15699 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15700 return (EFAULT);
15701
15702 return (0);
15703 }
15704
15705 case DTRACEIOC_STOP: {
15706 processorid_t cpuid;
15707
15708 mutex_enter(&dtrace_lock);
15709 rval = dtrace_state_stop(state, &cpuid);
15710 mutex_exit(&dtrace_lock);
15711
15712 if (rval != 0)
15713 return (rval);
15714
15715 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15716 return (EFAULT);
15717
15718 return (0);
15719 }
15720
15721 case DTRACEIOC_DOFGET: {
15722 dof_hdr_t hdr, *dof;
15723 uint64_t len;
15724
15725 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
15726 return (EFAULT);
15727
15728 mutex_enter(&dtrace_lock);
15729 dof = dtrace_dof_create(state);
15730 mutex_exit(&dtrace_lock);
15731
15732 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
15733 rval = copyout(dof, (void *)arg, len);
15734 dtrace_dof_destroy(dof);
15735
15736 return (rval == 0 ? 0 : EFAULT);
15737 }
15738
15739 case DTRACEIOC_AGGSNAP:
15740 case DTRACEIOC_BUFSNAP: {
15741 dtrace_bufdesc_t desc;
15742 caddr_t cached;
15743 dtrace_buffer_t *buf;
15744
15745 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15746 return (EFAULT);
15747
15748 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
15749 return (EINVAL);
15750
15751 mutex_enter(&dtrace_lock);
15752
15753 if (cmd == DTRACEIOC_BUFSNAP) {
15754 buf = &state->dts_buffer[desc.dtbd_cpu];
15755 } else {
15756 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
15757 }
15758
15759 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
15760 size_t sz = buf->dtb_offset;
15761
15762 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
15763 mutex_exit(&dtrace_lock);
15764 return (EBUSY);
15765 }
15766
15767 /*
15768 * If this buffer has already been consumed, we're
15769 * going to indicate that there's nothing left here
15770 * to consume.
15771 */
15772 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
15773 mutex_exit(&dtrace_lock);
15774
15775 desc.dtbd_size = 0;
15776 desc.dtbd_drops = 0;
15777 desc.dtbd_errors = 0;
15778 desc.dtbd_oldest = 0;
15779 sz = sizeof (desc);
15780
15781 if (copyout(&desc, (void *)arg, sz) != 0)
15782 return (EFAULT);
15783
15784 return (0);
15785 }
15786
15787 /*
15788 * If this is a ring buffer that has wrapped, we want
15789 * to copy the whole thing out.
15790 */
15791 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
15792 dtrace_buffer_polish(buf);
15793 sz = buf->dtb_size;
15794 }
15795
15796 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
15797 mutex_exit(&dtrace_lock);
15798 return (EFAULT);
15799 }
15800
15801 desc.dtbd_size = sz;
15802 desc.dtbd_drops = buf->dtb_drops;
15803 desc.dtbd_errors = buf->dtb_errors;
15804 desc.dtbd_oldest = buf->dtb_xamot_offset;
15805
15806 mutex_exit(&dtrace_lock);
15807
15808 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15809 return (EFAULT);
15810
15811 buf->dtb_flags |= DTRACEBUF_CONSUMED;
15812
15813 return (0);
15814 }
15815
15816 if (buf->dtb_tomax == NULL) {
15817 ASSERT(buf->dtb_xamot == NULL);
15818 mutex_exit(&dtrace_lock);
15819 return (ENOENT);
15820 }
15821
15822 cached = buf->dtb_tomax;
15823 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
15824
15825 dtrace_xcall(desc.dtbd_cpu,
15826 (dtrace_xcall_t)dtrace_buffer_switch, buf);
15827
15828 state->dts_errors += buf->dtb_xamot_errors;
15829
15830 /*
15831 * If the buffers did not actually switch, then the cross call
15832 * did not take place -- presumably because the given CPU is
15833 * not in the ready set. If this is the case, we'll return
15834 * ENOENT.
15835 */
15836 if (buf->dtb_tomax == cached) {
15837 ASSERT(buf->dtb_xamot != cached);
15838 mutex_exit(&dtrace_lock);
15839 return (ENOENT);
15840 }
15841
15842 ASSERT(cached == buf->dtb_xamot);
15843
15844 /*
15845 * We have our snapshot; now copy it out.
15846 */
15847 if (copyout(buf->dtb_xamot, desc.dtbd_data,
15848 buf->dtb_xamot_offset) != 0) {
15849 mutex_exit(&dtrace_lock);
15850 return (EFAULT);
15851 }
15852
15853 desc.dtbd_size = buf->dtb_xamot_offset;
15854 desc.dtbd_drops = buf->dtb_xamot_drops;
15855 desc.dtbd_errors = buf->dtb_xamot_errors;
15856 desc.dtbd_oldest = 0;
15857
15858 mutex_exit(&dtrace_lock);
15859
15860 /*
15861 * Finally, copy out the buffer description.
15862 */
15863 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15864 return (EFAULT);
15865
15866 return (0);
15867 }
15868
15869 case DTRACEIOC_CONF: {
15870 dtrace_conf_t conf;
15871
15872 bzero(&conf, sizeof (conf));
15873 conf.dtc_difversion = DIF_VERSION;
15874 conf.dtc_difintregs = DIF_DIR_NREGS;
15875 conf.dtc_diftupregs = DIF_DTR_NREGS;
15876 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
15877
15878 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
15879 return (EFAULT);
15880
15881 return (0);
15882 }
15883
15884 case DTRACEIOC_STATUS: {
15885 dtrace_status_t stat;
15886 dtrace_dstate_t *dstate;
15887 int i, j;
15888 uint64_t nerrs;
15889
15890 /*
15891 * See the comment in dtrace_state_deadman() for the reason
15892 * for setting dts_laststatus to INT64_MAX before setting
15893 * it to the correct value.
15894 */
15895 state->dts_laststatus = INT64_MAX;
15896 dtrace_membar_producer();
15897 state->dts_laststatus = dtrace_gethrtime();
15898
15899 bzero(&stat, sizeof (stat));
15900
15901 mutex_enter(&dtrace_lock);
15902
15903 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
15904 mutex_exit(&dtrace_lock);
15905 return (ENOENT);
15906 }
15907
15908 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
15909 stat.dtst_exiting = 1;
15910
15911 nerrs = state->dts_errors;
15912 dstate = &state->dts_vstate.dtvs_dynvars;
15913
15914 for (i = 0; i < NCPU; i++) {
15915 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
15916
15917 stat.dtst_dyndrops += dcpu->dtdsc_drops;
15918 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
15919 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
15920
15921 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
15922 stat.dtst_filled++;
15923
15924 nerrs += state->dts_buffer[i].dtb_errors;
15925
15926 for (j = 0; j < state->dts_nspeculations; j++) {
15927 dtrace_speculation_t *spec;
15928 dtrace_buffer_t *buf;
15929
15930 spec = &state->dts_speculations[j];
15931 buf = &spec->dtsp_buffer[i];
15932 stat.dtst_specdrops += buf->dtb_xamot_drops;
15933 }
15934 }
15935
15936 stat.dtst_specdrops_busy = state->dts_speculations_busy;
15937 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
15938 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
15939 stat.dtst_dblerrors = state->dts_dblerrors;
15940 stat.dtst_killed =
15941 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
15942 stat.dtst_errors = nerrs;
15943
15944 mutex_exit(&dtrace_lock);
15945
15946 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
15947 return (EFAULT);
15948
15949 return (0);
15950 }
15951
15952 case DTRACEIOC_FORMAT: {
15953 dtrace_fmtdesc_t fmt;
15954 char *str;
15955 int len;
15956
15957 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
15958 return (EFAULT);
15959
15960 mutex_enter(&dtrace_lock);
15961
15962 if (fmt.dtfd_format == 0 ||
15963 fmt.dtfd_format > state->dts_nformats) {
15964 mutex_exit(&dtrace_lock);
15965 return (EINVAL);
15966 }
15967
15968 /*
15969 * Format strings are allocated contiguously and they are
15970 * never freed; if a format index is less than the number
15971 * of formats, we can assert that the format map is non-NULL
15972 * and that the format for the specified index is non-NULL.
15973 */
15974 ASSERT(state->dts_formats != NULL);
15975 str = state->dts_formats[fmt.dtfd_format - 1];
15976 ASSERT(str != NULL);
15977
15978 len = strlen(str) + 1;
15979
15980 if (len > fmt.dtfd_length) {
15981 fmt.dtfd_length = len;
15982
15983 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
15984 mutex_exit(&dtrace_lock);
15985 return (EINVAL);
15986 }
15987 } else {
15988 if (copyout(str, fmt.dtfd_string, len) != 0) {
15989 mutex_exit(&dtrace_lock);
15990 return (EINVAL);
15991 }
15992 }
15993
15994 mutex_exit(&dtrace_lock);
15995 return (0);
15996 }
15997
15998 default:
15999 break;
16000 }
16001
16002 return (ENOTTY);
16003 }
16004
16005 /*ARGSUSED*/
16006 static int
16007 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16008 {
16009 dtrace_state_t *state;
16010
16011 switch (cmd) {
16012 case DDI_DETACH:
16013 break;
16014
16015 case DDI_SUSPEND:
16016 return (DDI_SUCCESS);
16017
16018 default:
16019 return (DDI_FAILURE);
16020 }
16021
16022 mutex_enter(&cpu_lock);
16023 mutex_enter(&dtrace_provider_lock);
16024 mutex_enter(&dtrace_lock);
16025
16026 ASSERT(dtrace_opens == 0);
16027
16028 if (dtrace_helpers > 0) {
16029 mutex_exit(&dtrace_provider_lock);
16030 mutex_exit(&dtrace_lock);
16031 mutex_exit(&cpu_lock);
16032 return (DDI_FAILURE);
16033 }
16034
16035 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16036 mutex_exit(&dtrace_provider_lock);
16037 mutex_exit(&dtrace_lock);
16038 mutex_exit(&cpu_lock);
16039 return (DDI_FAILURE);
16040 }
16041
16042 dtrace_provider = NULL;
16043
16044 if ((state = dtrace_anon_grab()) != NULL) {
16045 /*
16046 * If there were ECBs on this state, the provider should
16047 * have not been allowed to detach; assert that there is
16048 * none.
16049 */
16050 ASSERT(state->dts_necbs == 0);
16051 dtrace_state_destroy(state);
16052
16053 /*
16054 * If we're being detached with anonymous state, we need to
16055 * indicate to the kernel debugger that DTrace is now inactive.
16056 */
16057 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16058 }
16059
16060 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16061 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16062 dtrace_cpu_init = NULL;
16063 dtrace_helpers_cleanup = NULL;
16064 dtrace_helpers_fork = NULL;
16065 dtrace_cpustart_init = NULL;
16066 dtrace_cpustart_fini = NULL;
16067 dtrace_debugger_init = NULL;
16068 dtrace_debugger_fini = NULL;
16069 dtrace_modload = NULL;
16070 dtrace_modunload = NULL;
16071
16072 ASSERT(dtrace_getf == 0);
16073 ASSERT(dtrace_closef == NULL);
16074
16075 mutex_exit(&cpu_lock);
16076
16077 if (dtrace_helptrace_enabled) {
16078 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16079 dtrace_helptrace_buffer = NULL;
16080 }
16081
16082 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16083 dtrace_probes = NULL;
16084 dtrace_nprobes = 0;
16085
16086 dtrace_hash_destroy(dtrace_bymod);
16087 dtrace_hash_destroy(dtrace_byfunc);
16088 dtrace_hash_destroy(dtrace_byname);
16089 dtrace_bymod = NULL;
16090 dtrace_byfunc = NULL;
16091 dtrace_byname = NULL;
16092
16093 kmem_cache_destroy(dtrace_state_cache);
16094 vmem_destroy(dtrace_minor);
16095 vmem_destroy(dtrace_arena);
16096
16097 if (dtrace_toxrange != NULL) {
16098 kmem_free(dtrace_toxrange,
16099 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16100 dtrace_toxrange = NULL;
16101 dtrace_toxranges = 0;
16102 dtrace_toxranges_max = 0;
16103 }
16104
16105 ddi_remove_minor_node(dtrace_devi, NULL);
16106 dtrace_devi = NULL;
16107
16108 ddi_soft_state_fini(&dtrace_softstate);
16109
16110 ASSERT(dtrace_vtime_references == 0);
16111 ASSERT(dtrace_opens == 0);
16112 ASSERT(dtrace_retained == NULL);
16113
16114 mutex_exit(&dtrace_lock);
16115 mutex_exit(&dtrace_provider_lock);
16116
16117 /*
16118 * We don't destroy the task queue until after we have dropped our
16119 * locks (taskq_destroy() may block on running tasks). To prevent
16120 * attempting to do work after we have effectively detached but before
16121 * the task queue has been destroyed, all tasks dispatched via the
16122 * task queue must check that DTrace is still attached before
16123 * performing any operation.
16124 */
16125 taskq_destroy(dtrace_taskq);
16126 dtrace_taskq = NULL;
16127
16128 return (DDI_SUCCESS);
16129 }
16130
16131 /*ARGSUSED*/
16132 static int
16133 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16134 {
16135 int error;
16136
16137 switch (infocmd) {
16138 case DDI_INFO_DEVT2DEVINFO:
16139 *result = (void *)dtrace_devi;
16140 error = DDI_SUCCESS;
16141 break;
16142 case DDI_INFO_DEVT2INSTANCE:
16143 *result = (void *)0;
16144 error = DDI_SUCCESS;
16145 break;
16146 default:
16147 error = DDI_FAILURE;
16148 }
16149 return (error);
16150 }
16151
16152 static struct cb_ops dtrace_cb_ops = {
16153 dtrace_open, /* open */
16154 dtrace_close, /* close */
16155 nulldev, /* strategy */
16156 nulldev, /* print */
16157 nodev, /* dump */
16158 nodev, /* read */
16159 nodev, /* write */
16160 dtrace_ioctl, /* ioctl */
16161 nodev, /* devmap */
16162 nodev, /* mmap */
16163 nodev, /* segmap */
16164 nochpoll, /* poll */
16165 ddi_prop_op, /* cb_prop_op */
16166 0, /* streamtab */
16167 D_NEW | D_MP /* Driver compatibility flag */
16168 };
16169
16170 static struct dev_ops dtrace_ops = {
16171 DEVO_REV, /* devo_rev */
16172 0, /* refcnt */
16173 dtrace_info, /* get_dev_info */
16174 nulldev, /* identify */
16175 nulldev, /* probe */
16176 dtrace_attach, /* attach */
16177 dtrace_detach, /* detach */
16178 nodev, /* reset */
16179 &dtrace_cb_ops, /* driver operations */
16180 NULL, /* bus operations */
16181 nodev, /* dev power */
16182 ddi_quiesce_not_needed, /* quiesce */
16183 };
16184
16185 static struct modldrv modldrv = {
16186 &mod_driverops, /* module type (this is a pseudo driver) */
16187 "Dynamic Tracing", /* name of module */
16188 &dtrace_ops, /* driver ops */
16189 };
16190
16191 static struct modlinkage modlinkage = {
16192 MODREV_1,
16193 (void *)&modldrv,
16194 NULL
16195 };
16196
16197 int
16198 _init(void)
16199 {
16200 return (mod_install(&modlinkage));
16201 }
16202
16203 int
16204 _info(struct modinfo *modinfop)
16205 {
16206 return (mod_info(&modlinkage, modinfop));
16207 }
16208
16209 int
16210 _fini(void)
16211 {
16212 return (mod_remove(&modlinkage));
16213 }