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) 2013, Joyent, Inc. All rights reserved.
25 * Copyright (c) 2012 by Delphix. All rights reserved.
26 */
27
28 /*
29 * DTrace - Dynamic Tracing for Solaris
30 *
31 * This is the implementation of the Solaris Dynamic Tracing framework
32 * (DTrace). The user-visible interface to DTrace is described at length in
33 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
34 * library, the in-kernel DTrace framework, and the DTrace providers are
35 * described in the block comments in the <sys/dtrace.h> header file. The
36 * internal architecture of DTrace is described in the block comments in the
37 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
38 * implementation very much assume mastery of all of these sources; if one has
39 * an unanswered question about the implementation, one should consult them
40 * first.
41 *
42 * The functions here are ordered roughly as follows:
43 *
44 * - Probe context functions
45 * - Probe hashing functions
46 * - Non-probe context utility functions
47 * - Matching functions
48 * - Provider-to-Framework API functions
49 * - Probe management functions
50 * - DIF object functions
51 * - Format functions
52 * - Predicate functions
53 * - ECB functions
54 * - Buffer functions
55 * - Enabling functions
56 * - DOF functions
57 * - Anonymous enabling functions
58 * - Consumer state functions
59 * - Helper functions
60 * - Hook functions
61 * - Driver cookbook functions
62 *
63 * Each group of functions begins with a block comment labelled the "DTrace
64 * [Group] Functions", allowing one to find each block by searching forward
65 * on capital-f functions.
66 */
67 #include <sys/errno.h>
68 #include <sys/stat.h>
69 #include <sys/modctl.h>
70 #include <sys/conf.h>
71 #include <sys/systm.h>
72 #include <sys/ddi.h>
73 #include <sys/sunddi.h>
74 #include <sys/cpuvar.h>
75 #include <sys/kmem.h>
76 #include <sys/strsubr.h>
77 #include <sys/sysmacros.h>
78 #include <sys/dtrace_impl.h>
79 #include <sys/atomic.h>
80 #include <sys/cmn_err.h>
81 #include <sys/mutex_impl.h>
82 #include <sys/rwlock_impl.h>
83 #include <sys/ctf_api.h>
84 #include <sys/panic.h>
85 #include <sys/priv_impl.h>
86 #include <sys/policy.h>
87 #include <sys/cred_impl.h>
88 #include <sys/procfs_isa.h>
89 #include <sys/taskq.h>
90 #include <sys/mkdev.h>
91 #include <sys/kdi.h>
92 #include <sys/zone.h>
93 #include <sys/socket.h>
94 #include <netinet/in.h>
95 #include "strtolctype.h"
96
97 /*
98 * DTrace Tunable Variables
99 *
100 * The following variables may be tuned by adding a line to /etc/system that
101 * includes both the name of the DTrace module ("dtrace") and the name of the
102 * variable. For example:
103 *
104 * set dtrace:dtrace_destructive_disallow = 1
105 *
106 * In general, the only variables that one should be tuning this way are those
107 * that affect system-wide DTrace behavior, and for which the default behavior
108 * is undesirable. Most of these variables are tunable on a per-consumer
109 * basis using DTrace options, and need not be tuned on a system-wide basis.
110 * When tuning these variables, avoid pathological values; while some attempt
111 * is made to verify the integrity of these variables, they are not considered
112 * part of the supported interface to DTrace, and they are therefore not
113 * checked comprehensively. Further, these variables should not be tuned
114 * dynamically via "mdb -kw" or other means; they should only be tuned via
115 * /etc/system.
116 */
117 int dtrace_destructive_disallow = 0;
118 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
119 size_t dtrace_difo_maxsize = (256 * 1024);
120 dtrace_optval_t dtrace_dof_maxsize = (256 * 1024);
121 size_t dtrace_global_maxsize = (16 * 1024);
122 size_t dtrace_actions_max = (16 * 1024);
123 size_t dtrace_retain_max = 1024;
124 dtrace_optval_t dtrace_helper_actions_max = 1024;
125 dtrace_optval_t dtrace_helper_providers_max = 32;
126 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
127 size_t dtrace_strsize_default = 256;
128 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
129 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
130 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
131 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
132 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
133 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
134 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
135 dtrace_optval_t dtrace_nspec_default = 1;
136 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
137 dtrace_optval_t dtrace_stackframes_default = 20;
138 dtrace_optval_t dtrace_ustackframes_default = 20;
139 dtrace_optval_t dtrace_jstackframes_default = 50;
140 dtrace_optval_t dtrace_jstackstrsize_default = 512;
141 int dtrace_msgdsize_max = 128;
142 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
143 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
144 int dtrace_devdepth_max = 32;
145 int dtrace_err_verbose;
146 hrtime_t dtrace_deadman_interval = NANOSEC;
147 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
148 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
149 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
150
151 /*
152 * DTrace External Variables
153 *
154 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
155 * available to DTrace consumers via the backtick (`) syntax. One of these,
156 * dtrace_zero, is made deliberately so: it is provided as a source of
157 * well-known, zero-filled memory. While this variable is not documented,
158 * it is used by some translators as an implementation detail.
159 */
160 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
161
162 /*
163 * DTrace Internal Variables
164 */
165 static dev_info_t *dtrace_devi; /* device info */
166 static vmem_t *dtrace_arena; /* probe ID arena */
167 static vmem_t *dtrace_minor; /* minor number arena */
168 static taskq_t *dtrace_taskq; /* task queue */
169 static dtrace_probe_t **dtrace_probes; /* array of all probes */
170 static int dtrace_nprobes; /* number of probes */
171 static dtrace_provider_t *dtrace_provider; /* provider list */
172 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
173 static int dtrace_opens; /* number of opens */
174 static int dtrace_helpers; /* number of helpers */
175 static int dtrace_getf; /* number of unpriv getf()s */
176 static void *dtrace_softstate; /* softstate pointer */
177 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
178 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
179 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
180 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
181 static int dtrace_toxranges; /* number of toxic ranges */
182 static int dtrace_toxranges_max; /* size of toxic range array */
183 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
184 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
185 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
186 static kthread_t *dtrace_panicked; /* panicking thread */
187 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
188 static dtrace_genid_t dtrace_probegen; /* current probe generation */
189 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
190 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
191 static dtrace_genid_t dtrace_retained_gen; /* current retained enab gen */
192 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
193 static int dtrace_dynvar_failclean; /* dynvars failed to clean */
194
195 /*
196 * DTrace Locking
197 * DTrace is protected by three (relatively coarse-grained) locks:
198 *
199 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
200 * including enabling state, probes, ECBs, consumer state, helper state,
201 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
202 * probe context is lock-free -- synchronization is handled via the
203 * dtrace_sync() cross call mechanism.
204 *
205 * (2) dtrace_provider_lock is required when manipulating provider state, or
206 * when provider state must be held constant.
207 *
208 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
209 * when meta provider state must be held constant.
210 *
211 * The lock ordering between these three locks is dtrace_meta_lock before
212 * dtrace_provider_lock before dtrace_lock. (In particular, there are
213 * several places where dtrace_provider_lock is held by the framework as it
214 * calls into the providers -- which then call back into the framework,
215 * grabbing dtrace_lock.)
216 *
217 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
218 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
219 * role as a coarse-grained lock; it is acquired before both of these locks.
220 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
221 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
222 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
223 * acquired _between_ dtrace_provider_lock and dtrace_lock.
224 */
225 static kmutex_t dtrace_lock; /* probe state lock */
226 static kmutex_t dtrace_provider_lock; /* provider state lock */
227 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
228
229 /*
230 * DTrace Provider Variables
231 *
232 * These are the variables relating to DTrace as a provider (that is, the
233 * provider of the BEGIN, END, and ERROR probes).
234 */
235 static dtrace_pattr_t dtrace_provider_attr = {
236 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
237 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
238 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
239 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
240 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
241 };
242
243 static void
244 dtrace_nullop(void)
245 {}
246
247 static int
248 dtrace_enable_nullop(void)
249 {
250 return (0);
251 }
252
253 static dtrace_pops_t dtrace_provider_ops = {
254 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
255 (void (*)(void *, struct modctl *))dtrace_nullop,
256 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop,
257 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
258 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
259 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
260 NULL,
261 NULL,
262 NULL,
263 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
264 };
265
266 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
267 static dtrace_id_t dtrace_probeid_end; /* special END probe */
268 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
269
270 /*
271 * DTrace Helper Tracing Variables
272 */
273 uint32_t dtrace_helptrace_next = 0;
274 uint32_t dtrace_helptrace_nlocals;
275 char *dtrace_helptrace_buffer;
276 int dtrace_helptrace_bufsize = 512 * 1024;
277
278 #ifdef DEBUG
279 int dtrace_helptrace_enabled = 1;
280 #else
281 int dtrace_helptrace_enabled = 0;
282 #endif
283
284 /*
285 * DTrace Error Hashing
286 *
287 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
288 * table. This is very useful for checking coverage of tests that are
289 * expected to induce DIF or DOF processing errors, and may be useful for
290 * debugging problems in the DIF code generator or in DOF generation . The
291 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
292 */
293 #ifdef DEBUG
294 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
295 static const char *dtrace_errlast;
296 static kthread_t *dtrace_errthread;
297 static kmutex_t dtrace_errlock;
298 #endif
299
300 /*
301 * DTrace Macros and Constants
302 *
303 * These are various macros that are useful in various spots in the
304 * implementation, along with a few random constants that have no meaning
305 * outside of the implementation. There is no real structure to this cpp
306 * mishmash -- but is there ever?
307 */
308 #define DTRACE_HASHSTR(hash, probe) \
309 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
310
311 #define DTRACE_HASHNEXT(hash, probe) \
312 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
313
314 #define DTRACE_HASHPREV(hash, probe) \
315 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
316
317 #define DTRACE_HASHEQ(hash, lhs, rhs) \
318 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
319 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
320
321 #define DTRACE_AGGHASHSIZE_SLEW 17
322
323 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
324
325 /*
326 * The key for a thread-local variable consists of the lower 61 bits of the
327 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
328 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
329 * equal to a variable identifier. This is necessary (but not sufficient) to
330 * assure that global associative arrays never collide with thread-local
331 * variables. To guarantee that they cannot collide, we must also define the
332 * order for keying dynamic variables. That order is:
333 *
334 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
335 *
336 * Because the variable-key and the tls-key are in orthogonal spaces, there is
337 * no way for a global variable key signature to match a thread-local key
338 * signature.
339 */
340 #define DTRACE_TLS_THRKEY(where) { \
341 uint_t intr = 0; \
342 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
343 for (; actv; actv >>= 1) \
344 intr++; \
345 ASSERT(intr < (1 << 3)); \
346 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
347 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
348 }
349
350 #define DT_BSWAP_8(x) ((x) & 0xff)
351 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
352 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
353 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
354
355 #define DT_MASK_LO 0x00000000FFFFFFFFULL
356
357 #define DTRACE_STORE(type, tomax, offset, what) \
358 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
359
360 #ifndef __x86
361 #define DTRACE_ALIGNCHECK(addr, size, flags) \
362 if (addr & (size - 1)) { \
363 *flags |= CPU_DTRACE_BADALIGN; \
364 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
365 return (0); \
366 }
367 #else
368 #define DTRACE_ALIGNCHECK(addr, size, flags)
369 #endif
370
371 /*
372 * Test whether a range of memory starting at testaddr of size testsz falls
373 * within the range of memory described by addr, sz. We take care to avoid
374 * problems with overflow and underflow of the unsigned quantities, and
375 * disallow all negative sizes. Ranges of size 0 are allowed.
376 */
377 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
378 ((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
379 (testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
380 (testaddr) + (testsz) >= (testaddr))
381
382 /*
383 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
384 * alloc_sz on the righthand side of the comparison in order to avoid overflow
385 * or underflow in the comparison with it. This is simpler than the INRANGE
386 * check above, because we know that the dtms_scratch_ptr is valid in the
387 * range. Allocations of size zero are allowed.
388 */
389 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
390 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
391 (mstate)->dtms_scratch_ptr >= (alloc_sz))
392
393 #define DTRACE_LOADFUNC(bits) \
394 /*CSTYLED*/ \
395 uint##bits##_t \
396 dtrace_load##bits(uintptr_t addr) \
397 { \
398 size_t size = bits / NBBY; \
399 /*CSTYLED*/ \
400 uint##bits##_t rval; \
401 int i; \
402 volatile uint16_t *flags = (volatile uint16_t *) \
403 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; \
404 \
405 DTRACE_ALIGNCHECK(addr, size, flags); \
406 \
407 for (i = 0; i < dtrace_toxranges; i++) { \
408 if (addr >= dtrace_toxrange[i].dtt_limit) \
409 continue; \
410 \
411 if (addr + size <= dtrace_toxrange[i].dtt_base) \
412 continue; \
413 \
414 /* \
415 * This address falls within a toxic region; return 0. \
416 */ \
417 *flags |= CPU_DTRACE_BADADDR; \
418 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
419 return (0); \
420 } \
421 \
422 *flags |= CPU_DTRACE_NOFAULT; \
423 /*CSTYLED*/ \
424 rval = *((volatile uint##bits##_t *)addr); \
425 *flags &= ~CPU_DTRACE_NOFAULT; \
426 \
427 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
428 }
429
430 #ifdef _LP64
431 #define dtrace_loadptr dtrace_load64
432 #else
433 #define dtrace_loadptr dtrace_load32
434 #endif
435
436 #define DTRACE_DYNHASH_FREE 0
437 #define DTRACE_DYNHASH_SINK 1
438 #define DTRACE_DYNHASH_VALID 2
439
440 #define DTRACE_MATCH_FAIL -1
441 #define DTRACE_MATCH_NEXT 0
442 #define DTRACE_MATCH_DONE 1
443 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
444 #define DTRACE_STATE_ALIGN 64
445
446 #define DTRACE_FLAGS2FLT(flags) \
447 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
448 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
449 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
450 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
451 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
452 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
453 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
454 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
455 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
456 DTRACEFLT_UNKNOWN)
457
458 #define DTRACEACT_ISSTRING(act) \
459 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
460 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
461
462 static size_t dtrace_strlen(const char *, size_t);
463 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
464 static void dtrace_enabling_provide(dtrace_provider_t *);
465 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
466 static void dtrace_enabling_matchall(void);
467 static void dtrace_enabling_reap(void);
468 static dtrace_state_t *dtrace_anon_grab(void);
469 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
470 dtrace_state_t *, uint64_t, uint64_t);
471 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
472 static void dtrace_buffer_drop(dtrace_buffer_t *);
473 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
474 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
475 dtrace_state_t *, dtrace_mstate_t *);
476 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
477 dtrace_optval_t);
478 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
479 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
480 static int dtrace_priv_proc(dtrace_state_t *, dtrace_mstate_t *);
481 static void dtrace_getf_barrier(void);
482
483 /*
484 * DTrace Probe Context Functions
485 *
486 * These functions are called from probe context. Because probe context is
487 * any context in which C may be called, arbitrarily locks may be held,
488 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
489 * As a result, functions called from probe context may only call other DTrace
490 * support functions -- they may not interact at all with the system at large.
491 * (Note that the ASSERT macro is made probe-context safe by redefining it in
492 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
493 * loads are to be performed from probe context, they _must_ be in terms of
494 * the safe dtrace_load*() variants.
495 *
496 * Some functions in this block are not actually called from probe context;
497 * for these functions, there will be a comment above the function reading
498 * "Note: not called from probe context."
499 */
500 void
501 dtrace_panic(const char *format, ...)
502 {
503 va_list alist;
504
505 va_start(alist, format);
506 dtrace_vpanic(format, alist);
507 va_end(alist);
508 }
509
510 int
511 dtrace_assfail(const char *a, const char *f, int l)
512 {
513 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
514
515 /*
516 * We just need something here that even the most clever compiler
517 * cannot optimize away.
518 */
519 return (a[(uintptr_t)f]);
520 }
521
522 /*
523 * Atomically increment a specified error counter from probe context.
524 */
525 static void
526 dtrace_error(uint32_t *counter)
527 {
528 /*
529 * Most counters stored to in probe context are per-CPU counters.
530 * However, there are some error conditions that are sufficiently
531 * arcane that they don't merit per-CPU storage. If these counters
532 * are incremented concurrently on different CPUs, scalability will be
533 * adversely affected -- but we don't expect them to be white-hot in a
534 * correctly constructed enabling...
535 */
536 uint32_t oval, nval;
537
538 do {
539 oval = *counter;
540
541 if ((nval = oval + 1) == 0) {
542 /*
543 * If the counter would wrap, set it to 1 -- assuring
544 * that the counter is never zero when we have seen
545 * errors. (The counter must be 32-bits because we
546 * aren't guaranteed a 64-bit compare&swap operation.)
547 * To save this code both the infamy of being fingered
548 * by a priggish news story and the indignity of being
549 * the target of a neo-puritan witch trial, we're
550 * carefully avoiding any colorful description of the
551 * likelihood of this condition -- but suffice it to
552 * say that it is only slightly more likely than the
553 * overflow of predicate cache IDs, as discussed in
554 * dtrace_predicate_create().
555 */
556 nval = 1;
557 }
558 } while (dtrace_cas32(counter, oval, nval) != oval);
559 }
560
561 /*
562 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
563 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
564 */
565 DTRACE_LOADFUNC(8)
566 DTRACE_LOADFUNC(16)
567 DTRACE_LOADFUNC(32)
568 DTRACE_LOADFUNC(64)
569
570 static int
571 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
572 {
573 if (dest < mstate->dtms_scratch_base)
574 return (0);
575
576 if (dest + size < dest)
577 return (0);
578
579 if (dest + size > mstate->dtms_scratch_ptr)
580 return (0);
581
582 return (1);
583 }
584
585 static int
586 dtrace_canstore_statvar(uint64_t addr, size_t sz,
587 dtrace_statvar_t **svars, int nsvars)
588 {
589 int i;
590
591 for (i = 0; i < nsvars; i++) {
592 dtrace_statvar_t *svar = svars[i];
593
594 if (svar == NULL || svar->dtsv_size == 0)
595 continue;
596
597 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
598 return (1);
599 }
600
601 return (0);
602 }
603
604 /*
605 * Check to see if the address is within a memory region to which a store may
606 * be issued. This includes the DTrace scratch areas, and any DTrace variable
607 * region. The caller of dtrace_canstore() is responsible for performing any
608 * alignment checks that are needed before stores are actually executed.
609 */
610 static int
611 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
612 dtrace_vstate_t *vstate)
613 {
614 /*
615 * First, check to see if the address is in scratch space...
616 */
617 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
618 mstate->dtms_scratch_size))
619 return (1);
620
621 /*
622 * Now check to see if it's a dynamic variable. This check will pick
623 * up both thread-local variables and any global dynamically-allocated
624 * variables.
625 */
626 if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
627 vstate->dtvs_dynvars.dtds_size)) {
628 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
629 uintptr_t base = (uintptr_t)dstate->dtds_base +
630 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
631 uintptr_t chunkoffs;
632
633 /*
634 * Before we assume that we can store here, we need to make
635 * sure that it isn't in our metadata -- storing to our
636 * dynamic variable metadata would corrupt our state. For
637 * the range to not include any dynamic variable metadata,
638 * it must:
639 *
640 * (1) Start above the hash table that is at the base of
641 * the dynamic variable space
642 *
643 * (2) Have a starting chunk offset that is beyond the
644 * dtrace_dynvar_t that is at the base of every chunk
645 *
646 * (3) Not span a chunk boundary
647 *
648 */
649 if (addr < base)
650 return (0);
651
652 chunkoffs = (addr - base) % dstate->dtds_chunksize;
653
654 if (chunkoffs < sizeof (dtrace_dynvar_t))
655 return (0);
656
657 if (chunkoffs + sz > dstate->dtds_chunksize)
658 return (0);
659
660 return (1);
661 }
662
663 /*
664 * Finally, check the static local and global variables. These checks
665 * take the longest, so we perform them last.
666 */
667 if (dtrace_canstore_statvar(addr, sz,
668 vstate->dtvs_locals, vstate->dtvs_nlocals))
669 return (1);
670
671 if (dtrace_canstore_statvar(addr, sz,
672 vstate->dtvs_globals, vstate->dtvs_nglobals))
673 return (1);
674
675 return (0);
676 }
677
678
679 /*
680 * Convenience routine to check to see if the address is within a memory
681 * region in which a load may be issued given the user's privilege level;
682 * if not, it sets the appropriate error flags and loads 'addr' into the
683 * illegal value slot.
684 *
685 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
686 * appropriate memory access protection.
687 */
688 static int
689 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
690 dtrace_vstate_t *vstate)
691 {
692 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
693 file_t *fp;
694
695 /*
696 * If we hold the privilege to read from kernel memory, then
697 * everything is readable.
698 */
699 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
700 return (1);
701
702 /*
703 * You can obviously read that which you can store.
704 */
705 if (dtrace_canstore(addr, sz, mstate, vstate))
706 return (1);
707
708 /*
709 * We're allowed to read from our own string table.
710 */
711 if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
712 mstate->dtms_difo->dtdo_strlen))
713 return (1);
714
715 if (vstate->dtvs_state != NULL &&
716 dtrace_priv_proc(vstate->dtvs_state, mstate)) {
717 proc_t *p;
718
719 /*
720 * When we have privileges to the current process, there are
721 * several context-related kernel structures that are safe to
722 * read, even absent the privilege to read from kernel memory.
723 * These reads are safe because these structures contain only
724 * state that (1) we're permitted to read, (2) is harmless or
725 * (3) contains pointers to additional kernel state that we're
726 * not permitted to read (and as such, do not present an
727 * opportunity for privilege escalation). Finally (and
728 * critically), because of the nature of their relation with
729 * the current thread context, the memory associated with these
730 * structures cannot change over the duration of probe context,
731 * and it is therefore impossible for this memory to be
732 * deallocated and reallocated as something else while it's
733 * being operated upon.
734 */
735 if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t)))
736 return (1);
737
738 if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
739 sz, curthread->t_procp, sizeof (proc_t))) {
740 return (1);
741 }
742
743 if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
744 curthread->t_cred, sizeof (cred_t))) {
745 return (1);
746 }
747
748 if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
749 &(p->p_pidp->pid_id), sizeof (pid_t))) {
750 return (1);
751 }
752
753 if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
754 curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
755 return (1);
756 }
757 }
758
759 if ((fp = mstate->dtms_getf) != NULL) {
760 uintptr_t psz = sizeof (void *);
761 vnode_t *vp;
762 vnodeops_t *op;
763
764 /*
765 * When getf() returns a file_t, the enabling is implicitly
766 * granted the (transient) right to read the returned file_t
767 * as well as the v_path and v_op->vnop_name of the underlying
768 * vnode. These accesses are allowed after a successful
769 * getf() because the members that they refer to cannot change
770 * once set -- and the barrier logic in the kernel's closef()
771 * path assures that the file_t and its referenced vode_t
772 * cannot themselves be stale (that is, it impossible for
773 * either dtms_getf itself or its f_vnode member to reference
774 * freed memory).
775 */
776 if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t)))
777 return (1);
778
779 if ((vp = fp->f_vnode) != NULL) {
780 if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz))
781 return (1);
782
783 if (vp->v_path != NULL && DTRACE_INRANGE(addr, sz,
784 vp->v_path, strlen(vp->v_path) + 1)) {
785 return (1);
786 }
787
788 if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz))
789 return (1);
790
791 if ((op = vp->v_op) != NULL &&
792 DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
793 return (1);
794 }
795
796 if (op != NULL && op->vnop_name != NULL &&
797 DTRACE_INRANGE(addr, sz, op->vnop_name,
798 strlen(op->vnop_name) + 1)) {
799 return (1);
800 }
801 }
802 }
803
804 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
805 *illval = addr;
806 return (0);
807 }
808
809 /*
810 * Convenience routine to check to see if a given string is within a memory
811 * region in which a load may be issued given the user's privilege level;
812 * this exists so that we don't need to issue unnecessary dtrace_strlen()
813 * calls in the event that the user has all privileges.
814 */
815 static int
816 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
817 dtrace_vstate_t *vstate)
818 {
819 size_t strsz;
820
821 /*
822 * If we hold the privilege to read from kernel memory, then
823 * everything is readable.
824 */
825 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
826 return (1);
827
828 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
829 if (dtrace_canload(addr, strsz, mstate, vstate))
830 return (1);
831
832 return (0);
833 }
834
835 /*
836 * Convenience routine to check to see if a given variable is within a memory
837 * region in which a load may be issued given the user's privilege level.
838 */
839 static int
840 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
841 dtrace_vstate_t *vstate)
842 {
843 size_t sz;
844 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
845
846 /*
847 * If we hold the privilege to read from kernel memory, then
848 * everything is readable.
849 */
850 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
851 return (1);
852
853 if (type->dtdt_kind == DIF_TYPE_STRING)
854 sz = dtrace_strlen(src,
855 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
856 else
857 sz = type->dtdt_size;
858
859 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
860 }
861
862 /*
863 * Convert a string to a signed integer using safe loads.
864 *
865 * NOTE: This function uses various macros from strtolctype.h to manipulate
866 * digit values, etc -- these have all been checked to ensure they make
867 * no additional function calls.
868 */
869 static int64_t
870 dtrace_strtoll(char *input, int base, size_t limit)
871 {
872 uintptr_t pos = (uintptr_t)input;
873 int64_t val = 0;
874 int x;
875 boolean_t neg = B_FALSE;
876 char c, cc, ccc;
877 uintptr_t end = pos + limit;
878
879 /*
880 * Consume any whitespace preceding digits.
881 */
882 while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
883 pos++;
884
885 /*
886 * Handle an explicit sign if one is present.
887 */
888 if (c == '-' || c == '+') {
889 if (c == '-')
890 neg = B_TRUE;
891 c = dtrace_load8(++pos);
892 }
893
894 /*
895 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
896 * if present.
897 */
898 if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
899 cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
900 pos += 2;
901 c = ccc;
902 }
903
904 /*
905 * Read in contiguous digits until the first non-digit character.
906 */
907 for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
908 c = dtrace_load8(++pos))
909 val = val * base + x;
910
911 return (neg ? -val : val);
912 }
913
914 /*
915 * Compare two strings using safe loads.
916 */
917 static int
918 dtrace_strncmp(char *s1, char *s2, size_t limit)
919 {
920 uint8_t c1, c2;
921 volatile uint16_t *flags;
922
923 if (s1 == s2 || limit == 0)
924 return (0);
925
926 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
927
928 do {
929 if (s1 == NULL) {
930 c1 = '\0';
931 } else {
932 c1 = dtrace_load8((uintptr_t)s1++);
933 }
934
935 if (s2 == NULL) {
936 c2 = '\0';
937 } else {
938 c2 = dtrace_load8((uintptr_t)s2++);
939 }
940
941 if (c1 != c2)
942 return (c1 - c2);
943 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
944
945 return (0);
946 }
947
948 /*
949 * Compute strlen(s) for a string using safe memory accesses. The additional
950 * len parameter is used to specify a maximum length to ensure completion.
951 */
952 static size_t
953 dtrace_strlen(const char *s, size_t lim)
954 {
955 uint_t len;
956
957 for (len = 0; len != lim; len++) {
958 if (dtrace_load8((uintptr_t)s++) == '\0')
959 break;
960 }
961
962 return (len);
963 }
964
965 /*
966 * Check if an address falls within a toxic region.
967 */
968 static int
969 dtrace_istoxic(uintptr_t kaddr, size_t size)
970 {
971 uintptr_t taddr, tsize;
972 int i;
973
974 for (i = 0; i < dtrace_toxranges; i++) {
975 taddr = dtrace_toxrange[i].dtt_base;
976 tsize = dtrace_toxrange[i].dtt_limit - taddr;
977
978 if (kaddr - taddr < tsize) {
979 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
980 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
981 return (1);
982 }
983
984 if (taddr - kaddr < size) {
985 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
986 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
987 return (1);
988 }
989 }
990
991 return (0);
992 }
993
994 /*
995 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
996 * memory specified by the DIF program. The dst is assumed to be safe memory
997 * that we can store to directly because it is managed by DTrace. As with
998 * standard bcopy, overlapping copies are handled properly.
999 */
1000 static void
1001 dtrace_bcopy(const void *src, void *dst, size_t len)
1002 {
1003 if (len != 0) {
1004 uint8_t *s1 = dst;
1005 const uint8_t *s2 = src;
1006
1007 if (s1 <= s2) {
1008 do {
1009 *s1++ = dtrace_load8((uintptr_t)s2++);
1010 } while (--len != 0);
1011 } else {
1012 s2 += len;
1013 s1 += len;
1014
1015 do {
1016 *--s1 = dtrace_load8((uintptr_t)--s2);
1017 } while (--len != 0);
1018 }
1019 }
1020 }
1021
1022 /*
1023 * Copy src to dst using safe memory accesses, up to either the specified
1024 * length, or the point that a nul byte is encountered. The src is assumed to
1025 * be unsafe memory specified by the DIF program. The dst is assumed to be
1026 * safe memory that we can store to directly because it is managed by DTrace.
1027 * Unlike dtrace_bcopy(), overlapping regions are not handled.
1028 */
1029 static void
1030 dtrace_strcpy(const void *src, void *dst, size_t len)
1031 {
1032 if (len != 0) {
1033 uint8_t *s1 = dst, c;
1034 const uint8_t *s2 = src;
1035
1036 do {
1037 *s1++ = c = dtrace_load8((uintptr_t)s2++);
1038 } while (--len != 0 && c != '\0');
1039 }
1040 }
1041
1042 /*
1043 * Copy src to dst, deriving the size and type from the specified (BYREF)
1044 * variable type. The src is assumed to be unsafe memory specified by the DIF
1045 * program. The dst is assumed to be DTrace variable memory that is of the
1046 * specified type; we assume that we can store to directly.
1047 */
1048 static void
1049 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1050 {
1051 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1052
1053 if (type->dtdt_kind == DIF_TYPE_STRING) {
1054 dtrace_strcpy(src, dst, type->dtdt_size);
1055 } else {
1056 dtrace_bcopy(src, dst, type->dtdt_size);
1057 }
1058 }
1059
1060 /*
1061 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1062 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1063 * safe memory that we can access directly because it is managed by DTrace.
1064 */
1065 static int
1066 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1067 {
1068 volatile uint16_t *flags;
1069
1070 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1071
1072 if (s1 == s2)
1073 return (0);
1074
1075 if (s1 == NULL || s2 == NULL)
1076 return (1);
1077
1078 if (s1 != s2 && len != 0) {
1079 const uint8_t *ps1 = s1;
1080 const uint8_t *ps2 = s2;
1081
1082 do {
1083 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1084 return (1);
1085 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1086 }
1087 return (0);
1088 }
1089
1090 /*
1091 * Zero the specified region using a simple byte-by-byte loop. Note that this
1092 * is for safe DTrace-managed memory only.
1093 */
1094 static void
1095 dtrace_bzero(void *dst, size_t len)
1096 {
1097 uchar_t *cp;
1098
1099 for (cp = dst; len != 0; len--)
1100 *cp++ = 0;
1101 }
1102
1103 static void
1104 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1105 {
1106 uint64_t result[2];
1107
1108 result[0] = addend1[0] + addend2[0];
1109 result[1] = addend1[1] + addend2[1] +
1110 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1111
1112 sum[0] = result[0];
1113 sum[1] = result[1];
1114 }
1115
1116 /*
1117 * Shift the 128-bit value in a by b. If b is positive, shift left.
1118 * If b is negative, shift right.
1119 */
1120 static void
1121 dtrace_shift_128(uint64_t *a, int b)
1122 {
1123 uint64_t mask;
1124
1125 if (b == 0)
1126 return;
1127
1128 if (b < 0) {
1129 b = -b;
1130 if (b >= 64) {
1131 a[0] = a[1] >> (b - 64);
1132 a[1] = 0;
1133 } else {
1134 a[0] >>= b;
1135 mask = 1LL << (64 - b);
1136 mask -= 1;
1137 a[0] |= ((a[1] & mask) << (64 - b));
1138 a[1] >>= b;
1139 }
1140 } else {
1141 if (b >= 64) {
1142 a[1] = a[0] << (b - 64);
1143 a[0] = 0;
1144 } else {
1145 a[1] <<= b;
1146 mask = a[0] >> (64 - b);
1147 a[1] |= mask;
1148 a[0] <<= b;
1149 }
1150 }
1151 }
1152
1153 /*
1154 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1155 * use native multiplication on those, and then re-combine into the
1156 * resulting 128-bit value.
1157 *
1158 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1159 * hi1 * hi2 << 64 +
1160 * hi1 * lo2 << 32 +
1161 * hi2 * lo1 << 32 +
1162 * lo1 * lo2
1163 */
1164 static void
1165 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1166 {
1167 uint64_t hi1, hi2, lo1, lo2;
1168 uint64_t tmp[2];
1169
1170 hi1 = factor1 >> 32;
1171 hi2 = factor2 >> 32;
1172
1173 lo1 = factor1 & DT_MASK_LO;
1174 lo2 = factor2 & DT_MASK_LO;
1175
1176 product[0] = lo1 * lo2;
1177 product[1] = hi1 * hi2;
1178
1179 tmp[0] = hi1 * lo2;
1180 tmp[1] = 0;
1181 dtrace_shift_128(tmp, 32);
1182 dtrace_add_128(product, tmp, product);
1183
1184 tmp[0] = hi2 * lo1;
1185 tmp[1] = 0;
1186 dtrace_shift_128(tmp, 32);
1187 dtrace_add_128(product, tmp, product);
1188 }
1189
1190 /*
1191 * This privilege check should be used by actions and subroutines to
1192 * verify that the user credentials of the process that enabled the
1193 * invoking ECB match the target credentials
1194 */
1195 static int
1196 dtrace_priv_proc_common_user(dtrace_state_t *state)
1197 {
1198 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1199
1200 /*
1201 * We should always have a non-NULL state cred here, since if cred
1202 * is null (anonymous tracing), we fast-path bypass this routine.
1203 */
1204 ASSERT(s_cr != NULL);
1205
1206 if ((cr = CRED()) != NULL &&
1207 s_cr->cr_uid == cr->cr_uid &&
1208 s_cr->cr_uid == cr->cr_ruid &&
1209 s_cr->cr_uid == cr->cr_suid &&
1210 s_cr->cr_gid == cr->cr_gid &&
1211 s_cr->cr_gid == cr->cr_rgid &&
1212 s_cr->cr_gid == cr->cr_sgid)
1213 return (1);
1214
1215 return (0);
1216 }
1217
1218 /*
1219 * This privilege check should be used by actions and subroutines to
1220 * verify that the zone of the process that enabled the invoking ECB
1221 * matches the target credentials
1222 */
1223 static int
1224 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1225 {
1226 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1227
1228 /*
1229 * We should always have a non-NULL state cred here, since if cred
1230 * is null (anonymous tracing), we fast-path bypass this routine.
1231 */
1232 ASSERT(s_cr != NULL);
1233
1234 if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1235 return (1);
1236
1237 return (0);
1238 }
1239
1240 /*
1241 * This privilege check should be used by actions and subroutines to
1242 * verify that the process has not setuid or changed credentials.
1243 */
1244 static int
1245 dtrace_priv_proc_common_nocd()
1246 {
1247 proc_t *proc;
1248
1249 if ((proc = ttoproc(curthread)) != NULL &&
1250 !(proc->p_flag & SNOCD))
1251 return (1);
1252
1253 return (0);
1254 }
1255
1256 static int
1257 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1258 {
1259 int action = state->dts_cred.dcr_action;
1260
1261 if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1262 goto bad;
1263
1264 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1265 dtrace_priv_proc_common_zone(state) == 0)
1266 goto bad;
1267
1268 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1269 dtrace_priv_proc_common_user(state) == 0)
1270 goto bad;
1271
1272 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1273 dtrace_priv_proc_common_nocd() == 0)
1274 goto bad;
1275
1276 return (1);
1277
1278 bad:
1279 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1280
1281 return (0);
1282 }
1283
1284 static int
1285 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1286 {
1287 if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1288 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1289 return (1);
1290
1291 if (dtrace_priv_proc_common_zone(state) &&
1292 dtrace_priv_proc_common_user(state) &&
1293 dtrace_priv_proc_common_nocd())
1294 return (1);
1295 }
1296
1297 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1298
1299 return (0);
1300 }
1301
1302 static int
1303 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1304 {
1305 if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1306 (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1307 return (1);
1308
1309 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1310
1311 return (0);
1312 }
1313
1314 static int
1315 dtrace_priv_kernel(dtrace_state_t *state)
1316 {
1317 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1318 return (1);
1319
1320 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1321
1322 return (0);
1323 }
1324
1325 static int
1326 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1327 {
1328 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1329 return (1);
1330
1331 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1332
1333 return (0);
1334 }
1335
1336 /*
1337 * Determine if the dte_cond of the specified ECB allows for processing of
1338 * the current probe to continue. Note that this routine may allow continued
1339 * processing, but with access(es) stripped from the mstate's dtms_access
1340 * field.
1341 */
1342 static int
1343 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1344 dtrace_ecb_t *ecb)
1345 {
1346 dtrace_probe_t *probe = ecb->dte_probe;
1347 dtrace_provider_t *prov = probe->dtpr_provider;
1348 dtrace_pops_t *pops = &prov->dtpv_pops;
1349 int mode = DTRACE_MODE_NOPRIV_DROP;
1350
1351 ASSERT(ecb->dte_cond);
1352
1353 if (pops->dtps_mode != NULL) {
1354 mode = pops->dtps_mode(prov->dtpv_arg,
1355 probe->dtpr_id, probe->dtpr_arg);
1356
1357 ASSERT(mode & (DTRACE_MODE_USER | DTRACE_MODE_KERNEL));
1358 ASSERT(mode & (DTRACE_MODE_NOPRIV_RESTRICT |
1359 DTRACE_MODE_NOPRIV_DROP));
1360 }
1361
1362 /*
1363 * If the dte_cond bits indicate that this consumer is only allowed to
1364 * see user-mode firings of this probe, check that the probe was fired
1365 * while in a user context. If that's not the case, use the policy
1366 * specified by the provider to determine if we drop the probe or
1367 * merely restrict operation.
1368 */
1369 if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1370 ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1371
1372 if (!(mode & DTRACE_MODE_USER)) {
1373 if (mode & DTRACE_MODE_NOPRIV_DROP)
1374 return (0);
1375
1376 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1377 }
1378 }
1379
1380 /*
1381 * This is more subtle than it looks. We have to be absolutely certain
1382 * that CRED() isn't going to change out from under us so it's only
1383 * legit to examine that structure if we're in constrained situations.
1384 * Currently, the only times we'll this check is if a non-super-user
1385 * has enabled the profile or syscall providers -- providers that
1386 * allow visibility of all processes. For the profile case, the check
1387 * above will ensure that we're examining a user context.
1388 */
1389 if (ecb->dte_cond & DTRACE_COND_OWNER) {
1390 cred_t *cr;
1391 cred_t *s_cr = state->dts_cred.dcr_cred;
1392 proc_t *proc;
1393
1394 ASSERT(s_cr != NULL);
1395
1396 if ((cr = CRED()) == NULL ||
1397 s_cr->cr_uid != cr->cr_uid ||
1398 s_cr->cr_uid != cr->cr_ruid ||
1399 s_cr->cr_uid != cr->cr_suid ||
1400 s_cr->cr_gid != cr->cr_gid ||
1401 s_cr->cr_gid != cr->cr_rgid ||
1402 s_cr->cr_gid != cr->cr_sgid ||
1403 (proc = ttoproc(curthread)) == NULL ||
1404 (proc->p_flag & SNOCD)) {
1405 if (mode & DTRACE_MODE_NOPRIV_DROP)
1406 return (0);
1407
1408 mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1409 }
1410 }
1411
1412 /*
1413 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1414 * in our zone, check to see if our mode policy is to restrict rather
1415 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1416 * and DTRACE_ACCESS_ARGS
1417 */
1418 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1419 cred_t *cr;
1420 cred_t *s_cr = state->dts_cred.dcr_cred;
1421
1422 ASSERT(s_cr != NULL);
1423
1424 if ((cr = CRED()) == NULL ||
1425 s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1426 if (mode & DTRACE_MODE_NOPRIV_DROP)
1427 return (0);
1428
1429 mstate->dtms_access &=
1430 ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1431 }
1432 }
1433
1434 /*
1435 * By merits of being in this code path at all, we have limited
1436 * privileges. If the provider has indicated that limited privileges
1437 * are to denote restricted operation, strip off the ability to access
1438 * arguments.
1439 */
1440 if (mode & DTRACE_MODE_LIMITEDPRIV_RESTRICT)
1441 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1442
1443 return (1);
1444 }
1445
1446 /*
1447 * Note: not called from probe context. This function is called
1448 * asynchronously (and at a regular interval) from outside of probe context to
1449 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1450 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1451 */
1452 void
1453 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1454 {
1455 dtrace_dynvar_t *dirty;
1456 dtrace_dstate_percpu_t *dcpu;
1457 dtrace_dynvar_t **rinsep;
1458 int i, j, work = 0;
1459
1460 for (i = 0; i < NCPU; i++) {
1461 dcpu = &dstate->dtds_percpu[i];
1462 rinsep = &dcpu->dtdsc_rinsing;
1463
1464 /*
1465 * If the dirty list is NULL, there is no dirty work to do.
1466 */
1467 if (dcpu->dtdsc_dirty == NULL)
1468 continue;
1469
1470 if (dcpu->dtdsc_rinsing != NULL) {
1471 /*
1472 * If the rinsing list is non-NULL, then it is because
1473 * this CPU was selected to accept another CPU's
1474 * dirty list -- and since that time, dirty buffers
1475 * have accumulated. This is a highly unlikely
1476 * condition, but we choose to ignore the dirty
1477 * buffers -- they'll be picked up a future cleanse.
1478 */
1479 continue;
1480 }
1481
1482 if (dcpu->dtdsc_clean != NULL) {
1483 /*
1484 * If the clean list is non-NULL, then we're in a
1485 * situation where a CPU has done deallocations (we
1486 * have a non-NULL dirty list) but no allocations (we
1487 * also have a non-NULL clean list). We can't simply
1488 * move the dirty list into the clean list on this
1489 * CPU, yet we also don't want to allow this condition
1490 * to persist, lest a short clean list prevent a
1491 * massive dirty list from being cleaned (which in
1492 * turn could lead to otherwise avoidable dynamic
1493 * drops). To deal with this, we look for some CPU
1494 * with a NULL clean list, NULL dirty list, and NULL
1495 * rinsing list -- and then we borrow this CPU to
1496 * rinse our dirty list.
1497 */
1498 for (j = 0; j < NCPU; j++) {
1499 dtrace_dstate_percpu_t *rinser;
1500
1501 rinser = &dstate->dtds_percpu[j];
1502
1503 if (rinser->dtdsc_rinsing != NULL)
1504 continue;
1505
1506 if (rinser->dtdsc_dirty != NULL)
1507 continue;
1508
1509 if (rinser->dtdsc_clean != NULL)
1510 continue;
1511
1512 rinsep = &rinser->dtdsc_rinsing;
1513 break;
1514 }
1515
1516 if (j == NCPU) {
1517 /*
1518 * We were unable to find another CPU that
1519 * could accept this dirty list -- we are
1520 * therefore unable to clean it now.
1521 */
1522 dtrace_dynvar_failclean++;
1523 continue;
1524 }
1525 }
1526
1527 work = 1;
1528
1529 /*
1530 * Atomically move the dirty list aside.
1531 */
1532 do {
1533 dirty = dcpu->dtdsc_dirty;
1534
1535 /*
1536 * Before we zap the dirty list, set the rinsing list.
1537 * (This allows for a potential assertion in
1538 * dtrace_dynvar(): if a free dynamic variable appears
1539 * on a hash chain, either the dirty list or the
1540 * rinsing list for some CPU must be non-NULL.)
1541 */
1542 *rinsep = dirty;
1543 dtrace_membar_producer();
1544 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1545 dirty, NULL) != dirty);
1546 }
1547
1548 if (!work) {
1549 /*
1550 * We have no work to do; we can simply return.
1551 */
1552 return;
1553 }
1554
1555 dtrace_sync();
1556
1557 for (i = 0; i < NCPU; i++) {
1558 dcpu = &dstate->dtds_percpu[i];
1559
1560 if (dcpu->dtdsc_rinsing == NULL)
1561 continue;
1562
1563 /*
1564 * We are now guaranteed that no hash chain contains a pointer
1565 * into this dirty list; we can make it clean.
1566 */
1567 ASSERT(dcpu->dtdsc_clean == NULL);
1568 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1569 dcpu->dtdsc_rinsing = NULL;
1570 }
1571
1572 /*
1573 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1574 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1575 * This prevents a race whereby a CPU incorrectly decides that
1576 * the state should be something other than DTRACE_DSTATE_CLEAN
1577 * after dtrace_dynvar_clean() has completed.
1578 */
1579 dtrace_sync();
1580
1581 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1582 }
1583
1584 /*
1585 * Depending on the value of the op parameter, this function looks-up,
1586 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1587 * allocation is requested, this function will return a pointer to a
1588 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1589 * variable can be allocated. If NULL is returned, the appropriate counter
1590 * will be incremented.
1591 */
1592 dtrace_dynvar_t *
1593 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1594 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1595 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1596 {
1597 uint64_t hashval = DTRACE_DYNHASH_VALID;
1598 dtrace_dynhash_t *hash = dstate->dtds_hash;
1599 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1600 processorid_t me = CPU->cpu_id, cpu = me;
1601 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1602 size_t bucket, ksize;
1603 size_t chunksize = dstate->dtds_chunksize;
1604 uintptr_t kdata, lock, nstate;
1605 uint_t i;
1606
1607 ASSERT(nkeys != 0);
1608
1609 /*
1610 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1611 * algorithm. For the by-value portions, we perform the algorithm in
1612 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1613 * bit, and seems to have only a minute effect on distribution. For
1614 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1615 * over each referenced byte. It's painful to do this, but it's much
1616 * better than pathological hash distribution. The efficacy of the
1617 * hashing algorithm (and a comparison with other algorithms) may be
1618 * found by running the ::dtrace_dynstat MDB dcmd.
1619 */
1620 for (i = 0; i < nkeys; i++) {
1621 if (key[i].dttk_size == 0) {
1622 uint64_t val = key[i].dttk_value;
1623
1624 hashval += (val >> 48) & 0xffff;
1625 hashval += (hashval << 10);
1626 hashval ^= (hashval >> 6);
1627
1628 hashval += (val >> 32) & 0xffff;
1629 hashval += (hashval << 10);
1630 hashval ^= (hashval >> 6);
1631
1632 hashval += (val >> 16) & 0xffff;
1633 hashval += (hashval << 10);
1634 hashval ^= (hashval >> 6);
1635
1636 hashval += val & 0xffff;
1637 hashval += (hashval << 10);
1638 hashval ^= (hashval >> 6);
1639 } else {
1640 /*
1641 * This is incredibly painful, but it beats the hell
1642 * out of the alternative.
1643 */
1644 uint64_t j, size = key[i].dttk_size;
1645 uintptr_t base = (uintptr_t)key[i].dttk_value;
1646
1647 if (!dtrace_canload(base, size, mstate, vstate))
1648 break;
1649
1650 for (j = 0; j < size; j++) {
1651 hashval += dtrace_load8(base + j);
1652 hashval += (hashval << 10);
1653 hashval ^= (hashval >> 6);
1654 }
1655 }
1656 }
1657
1658 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1659 return (NULL);
1660
1661 hashval += (hashval << 3);
1662 hashval ^= (hashval >> 11);
1663 hashval += (hashval << 15);
1664
1665 /*
1666 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1667 * comes out to be one of our two sentinel hash values. If this
1668 * actually happens, we set the hashval to be a value known to be a
1669 * non-sentinel value.
1670 */
1671 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1672 hashval = DTRACE_DYNHASH_VALID;
1673
1674 /*
1675 * Yes, it's painful to do a divide here. If the cycle count becomes
1676 * important here, tricks can be pulled to reduce it. (However, it's
1677 * critical that hash collisions be kept to an absolute minimum;
1678 * they're much more painful than a divide.) It's better to have a
1679 * solution that generates few collisions and still keeps things
1680 * relatively simple.
1681 */
1682 bucket = hashval % dstate->dtds_hashsize;
1683
1684 if (op == DTRACE_DYNVAR_DEALLOC) {
1685 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1686
1687 for (;;) {
1688 while ((lock = *lockp) & 1)
1689 continue;
1690
1691 if (dtrace_casptr((void *)lockp,
1692 (void *)lock, (void *)(lock + 1)) == (void *)lock)
1693 break;
1694 }
1695
1696 dtrace_membar_producer();
1697 }
1698
1699 top:
1700 prev = NULL;
1701 lock = hash[bucket].dtdh_lock;
1702
1703 dtrace_membar_consumer();
1704
1705 start = hash[bucket].dtdh_chain;
1706 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1707 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1708 op != DTRACE_DYNVAR_DEALLOC));
1709
1710 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1711 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1712 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1713
1714 if (dvar->dtdv_hashval != hashval) {
1715 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1716 /*
1717 * We've reached the sink, and therefore the
1718 * end of the hash chain; we can kick out of
1719 * the loop knowing that we have seen a valid
1720 * snapshot of state.
1721 */
1722 ASSERT(dvar->dtdv_next == NULL);
1723 ASSERT(dvar == &dtrace_dynhash_sink);
1724 break;
1725 }
1726
1727 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1728 /*
1729 * We've gone off the rails: somewhere along
1730 * the line, one of the members of this hash
1731 * chain was deleted. Note that we could also
1732 * detect this by simply letting this loop run
1733 * to completion, as we would eventually hit
1734 * the end of the dirty list. However, we
1735 * want to avoid running the length of the
1736 * dirty list unnecessarily (it might be quite
1737 * long), so we catch this as early as
1738 * possible by detecting the hash marker. In
1739 * this case, we simply set dvar to NULL and
1740 * break; the conditional after the loop will
1741 * send us back to top.
1742 */
1743 dvar = NULL;
1744 break;
1745 }
1746
1747 goto next;
1748 }
1749
1750 if (dtuple->dtt_nkeys != nkeys)
1751 goto next;
1752
1753 for (i = 0; i < nkeys; i++, dkey++) {
1754 if (dkey->dttk_size != key[i].dttk_size)
1755 goto next; /* size or type mismatch */
1756
1757 if (dkey->dttk_size != 0) {
1758 if (dtrace_bcmp(
1759 (void *)(uintptr_t)key[i].dttk_value,
1760 (void *)(uintptr_t)dkey->dttk_value,
1761 dkey->dttk_size))
1762 goto next;
1763 } else {
1764 if (dkey->dttk_value != key[i].dttk_value)
1765 goto next;
1766 }
1767 }
1768
1769 if (op != DTRACE_DYNVAR_DEALLOC)
1770 return (dvar);
1771
1772 ASSERT(dvar->dtdv_next == NULL ||
1773 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1774
1775 if (prev != NULL) {
1776 ASSERT(hash[bucket].dtdh_chain != dvar);
1777 ASSERT(start != dvar);
1778 ASSERT(prev->dtdv_next == dvar);
1779 prev->dtdv_next = dvar->dtdv_next;
1780 } else {
1781 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1782 start, dvar->dtdv_next) != start) {
1783 /*
1784 * We have failed to atomically swing the
1785 * hash table head pointer, presumably because
1786 * of a conflicting allocation on another CPU.
1787 * We need to reread the hash chain and try
1788 * again.
1789 */
1790 goto top;
1791 }
1792 }
1793
1794 dtrace_membar_producer();
1795
1796 /*
1797 * Now set the hash value to indicate that it's free.
1798 */
1799 ASSERT(hash[bucket].dtdh_chain != dvar);
1800 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1801
1802 dtrace_membar_producer();
1803
1804 /*
1805 * Set the next pointer to point at the dirty list, and
1806 * atomically swing the dirty pointer to the newly freed dvar.
1807 */
1808 do {
1809 next = dcpu->dtdsc_dirty;
1810 dvar->dtdv_next = next;
1811 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1812
1813 /*
1814 * Finally, unlock this hash bucket.
1815 */
1816 ASSERT(hash[bucket].dtdh_lock == lock);
1817 ASSERT(lock & 1);
1818 hash[bucket].dtdh_lock++;
1819
1820 return (NULL);
1821 next:
1822 prev = dvar;
1823 continue;
1824 }
1825
1826 if (dvar == NULL) {
1827 /*
1828 * If dvar is NULL, it is because we went off the rails:
1829 * one of the elements that we traversed in the hash chain
1830 * was deleted while we were traversing it. In this case,
1831 * we assert that we aren't doing a dealloc (deallocs lock
1832 * the hash bucket to prevent themselves from racing with
1833 * one another), and retry the hash chain traversal.
1834 */
1835 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1836 goto top;
1837 }
1838
1839 if (op != DTRACE_DYNVAR_ALLOC) {
1840 /*
1841 * If we are not to allocate a new variable, we want to
1842 * return NULL now. Before we return, check that the value
1843 * of the lock word hasn't changed. If it has, we may have
1844 * seen an inconsistent snapshot.
1845 */
1846 if (op == DTRACE_DYNVAR_NOALLOC) {
1847 if (hash[bucket].dtdh_lock != lock)
1848 goto top;
1849 } else {
1850 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1851 ASSERT(hash[bucket].dtdh_lock == lock);
1852 ASSERT(lock & 1);
1853 hash[bucket].dtdh_lock++;
1854 }
1855
1856 return (NULL);
1857 }
1858
1859 /*
1860 * We need to allocate a new dynamic variable. The size we need is the
1861 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1862 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1863 * the size of any referred-to data (dsize). We then round the final
1864 * size up to the chunksize for allocation.
1865 */
1866 for (ksize = 0, i = 0; i < nkeys; i++)
1867 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1868
1869 /*
1870 * This should be pretty much impossible, but could happen if, say,
1871 * strange DIF specified the tuple. Ideally, this should be an
1872 * assertion and not an error condition -- but that requires that the
1873 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1874 * bullet-proof. (That is, it must not be able to be fooled by
1875 * malicious DIF.) Given the lack of backwards branches in DIF,
1876 * solving this would presumably not amount to solving the Halting
1877 * Problem -- but it still seems awfully hard.
1878 */
1879 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1880 ksize + dsize > chunksize) {
1881 dcpu->dtdsc_drops++;
1882 return (NULL);
1883 }
1884
1885 nstate = DTRACE_DSTATE_EMPTY;
1886
1887 do {
1888 retry:
1889 free = dcpu->dtdsc_free;
1890
1891 if (free == NULL) {
1892 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1893 void *rval;
1894
1895 if (clean == NULL) {
1896 /*
1897 * We're out of dynamic variable space on
1898 * this CPU. Unless we have tried all CPUs,
1899 * we'll try to allocate from a different
1900 * CPU.
1901 */
1902 switch (dstate->dtds_state) {
1903 case DTRACE_DSTATE_CLEAN: {
1904 void *sp = &dstate->dtds_state;
1905
1906 if (++cpu >= NCPU)
1907 cpu = 0;
1908
1909 if (dcpu->dtdsc_dirty != NULL &&
1910 nstate == DTRACE_DSTATE_EMPTY)
1911 nstate = DTRACE_DSTATE_DIRTY;
1912
1913 if (dcpu->dtdsc_rinsing != NULL)
1914 nstate = DTRACE_DSTATE_RINSING;
1915
1916 dcpu = &dstate->dtds_percpu[cpu];
1917
1918 if (cpu != me)
1919 goto retry;
1920
1921 (void) dtrace_cas32(sp,
1922 DTRACE_DSTATE_CLEAN, nstate);
1923
1924 /*
1925 * To increment the correct bean
1926 * counter, take another lap.
1927 */
1928 goto retry;
1929 }
1930
1931 case DTRACE_DSTATE_DIRTY:
1932 dcpu->dtdsc_dirty_drops++;
1933 break;
1934
1935 case DTRACE_DSTATE_RINSING:
1936 dcpu->dtdsc_rinsing_drops++;
1937 break;
1938
1939 case DTRACE_DSTATE_EMPTY:
1940 dcpu->dtdsc_drops++;
1941 break;
1942 }
1943
1944 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1945 return (NULL);
1946 }
1947
1948 /*
1949 * The clean list appears to be non-empty. We want to
1950 * move the clean list to the free list; we start by
1951 * moving the clean pointer aside.
1952 */
1953 if (dtrace_casptr(&dcpu->dtdsc_clean,
1954 clean, NULL) != clean) {
1955 /*
1956 * We are in one of two situations:
1957 *
1958 * (a) The clean list was switched to the
1959 * free list by another CPU.
1960 *
1961 * (b) The clean list was added to by the
1962 * cleansing cyclic.
1963 *
1964 * In either of these situations, we can
1965 * just reattempt the free list allocation.
1966 */
1967 goto retry;
1968 }
1969
1970 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1971
1972 /*
1973 * Now we'll move the clean list to our free list.
1974 * It's impossible for this to fail: the only way
1975 * the free list can be updated is through this
1976 * code path, and only one CPU can own the clean list.
1977 * Thus, it would only be possible for this to fail if
1978 * this code were racing with dtrace_dynvar_clean().
1979 * (That is, if dtrace_dynvar_clean() updated the clean
1980 * list, and we ended up racing to update the free
1981 * list.) This race is prevented by the dtrace_sync()
1982 * in dtrace_dynvar_clean() -- which flushes the
1983 * owners of the clean lists out before resetting
1984 * the clean lists.
1985 */
1986 dcpu = &dstate->dtds_percpu[me];
1987 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1988 ASSERT(rval == NULL);
1989 goto retry;
1990 }
1991
1992 dvar = free;
1993 new_free = dvar->dtdv_next;
1994 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1995
1996 /*
1997 * We have now allocated a new chunk. We copy the tuple keys into the
1998 * tuple array and copy any referenced key data into the data space
1999 * following the tuple array. As we do this, we relocate dttk_value
2000 * in the final tuple to point to the key data address in the chunk.
2001 */
2002 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2003 dvar->dtdv_data = (void *)(kdata + ksize);
2004 dvar->dtdv_tuple.dtt_nkeys = nkeys;
2005
2006 for (i = 0; i < nkeys; i++) {
2007 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2008 size_t kesize = key[i].dttk_size;
2009
2010 if (kesize != 0) {
2011 dtrace_bcopy(
2012 (const void *)(uintptr_t)key[i].dttk_value,
2013 (void *)kdata, kesize);
2014 dkey->dttk_value = kdata;
2015 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2016 } else {
2017 dkey->dttk_value = key[i].dttk_value;
2018 }
2019
2020 dkey->dttk_size = kesize;
2021 }
2022
2023 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2024 dvar->dtdv_hashval = hashval;
2025 dvar->dtdv_next = start;
2026
2027 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2028 return (dvar);
2029
2030 /*
2031 * The cas has failed. Either another CPU is adding an element to
2032 * this hash chain, or another CPU is deleting an element from this
2033 * hash chain. The simplest way to deal with both of these cases
2034 * (though not necessarily the most efficient) is to free our
2035 * allocated block and tail-call ourselves. Note that the free is
2036 * to the dirty list and _not_ to the free list. This is to prevent
2037 * races with allocators, above.
2038 */
2039 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2040
2041 dtrace_membar_producer();
2042
2043 do {
2044 free = dcpu->dtdsc_dirty;
2045 dvar->dtdv_next = free;
2046 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2047
2048 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
2049 }
2050
2051 /*ARGSUSED*/
2052 static void
2053 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2054 {
2055 if ((int64_t)nval < (int64_t)*oval)
2056 *oval = nval;
2057 }
2058
2059 /*ARGSUSED*/
2060 static void
2061 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2062 {
2063 if ((int64_t)nval > (int64_t)*oval)
2064 *oval = nval;
2065 }
2066
2067 static void
2068 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2069 {
2070 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2071 int64_t val = (int64_t)nval;
2072
2073 if (val < 0) {
2074 for (i = 0; i < zero; i++) {
2075 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2076 quanta[i] += incr;
2077 return;
2078 }
2079 }
2080 } else {
2081 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2082 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2083 quanta[i - 1] += incr;
2084 return;
2085 }
2086 }
2087
2088 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2089 return;
2090 }
2091
2092 ASSERT(0);
2093 }
2094
2095 static void
2096 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2097 {
2098 uint64_t arg = *lquanta++;
2099 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2100 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2101 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2102 int32_t val = (int32_t)nval, level;
2103
2104 ASSERT(step != 0);
2105 ASSERT(levels != 0);
2106
2107 if (val < base) {
2108 /*
2109 * This is an underflow.
2110 */
2111 lquanta[0] += incr;
2112 return;
2113 }
2114
2115 level = (val - base) / step;
2116
2117 if (level < levels) {
2118 lquanta[level + 1] += incr;
2119 return;
2120 }
2121
2122 /*
2123 * This is an overflow.
2124 */
2125 lquanta[levels + 1] += incr;
2126 }
2127
2128 static int
2129 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2130 uint16_t high, uint16_t nsteps, int64_t value)
2131 {
2132 int64_t this = 1, last, next;
2133 int base = 1, order;
2134
2135 ASSERT(factor <= nsteps);
2136 ASSERT(nsteps % factor == 0);
2137
2138 for (order = 0; order < low; order++)
2139 this *= factor;
2140
2141 /*
2142 * If our value is less than our factor taken to the power of the
2143 * low order of magnitude, it goes into the zeroth bucket.
2144 */
2145 if (value < (last = this))
2146 return (0);
2147
2148 for (this *= factor; order <= high; order++) {
2149 int nbuckets = this > nsteps ? nsteps : this;
2150
2151 if ((next = this * factor) < this) {
2152 /*
2153 * We should not generally get log/linear quantizations
2154 * with a high magnitude that allows 64-bits to
2155 * overflow, but we nonetheless protect against this
2156 * by explicitly checking for overflow, and clamping
2157 * our value accordingly.
2158 */
2159 value = this - 1;
2160 }
2161
2162 if (value < this) {
2163 /*
2164 * If our value lies within this order of magnitude,
2165 * determine its position by taking the offset within
2166 * the order of magnitude, dividing by the bucket
2167 * width, and adding to our (accumulated) base.
2168 */
2169 return (base + (value - last) / (this / nbuckets));
2170 }
2171
2172 base += nbuckets - (nbuckets / factor);
2173 last = this;
2174 this = next;
2175 }
2176
2177 /*
2178 * Our value is greater than or equal to our factor taken to the
2179 * power of one plus the high magnitude -- return the top bucket.
2180 */
2181 return (base);
2182 }
2183
2184 static void
2185 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2186 {
2187 uint64_t arg = *llquanta++;
2188 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2189 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2190 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2191 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2192
2193 llquanta[dtrace_aggregate_llquantize_bucket(factor,
2194 low, high, nsteps, nval)] += incr;
2195 }
2196
2197 /*ARGSUSED*/
2198 static void
2199 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2200 {
2201 data[0]++;
2202 data[1] += nval;
2203 }
2204
2205 /*ARGSUSED*/
2206 static void
2207 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2208 {
2209 int64_t snval = (int64_t)nval;
2210 uint64_t tmp[2];
2211
2212 data[0]++;
2213 data[1] += nval;
2214
2215 /*
2216 * What we want to say here is:
2217 *
2218 * data[2] += nval * nval;
2219 *
2220 * But given that nval is 64-bit, we could easily overflow, so
2221 * we do this as 128-bit arithmetic.
2222 */
2223 if (snval < 0)
2224 snval = -snval;
2225
2226 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2227 dtrace_add_128(data + 2, tmp, data + 2);
2228 }
2229
2230 /*ARGSUSED*/
2231 static void
2232 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2233 {
2234 *oval = *oval + 1;
2235 }
2236
2237 /*ARGSUSED*/
2238 static void
2239 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2240 {
2241 *oval += nval;
2242 }
2243
2244 /*
2245 * Aggregate given the tuple in the principal data buffer, and the aggregating
2246 * action denoted by the specified dtrace_aggregation_t. The aggregation
2247 * buffer is specified as the buf parameter. This routine does not return
2248 * failure; if there is no space in the aggregation buffer, the data will be
2249 * dropped, and a corresponding counter incremented.
2250 */
2251 static void
2252 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2253 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2254 {
2255 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2256 uint32_t i, ndx, size, fsize;
2257 uint32_t align = sizeof (uint64_t) - 1;
2258 dtrace_aggbuffer_t *agb;
2259 dtrace_aggkey_t *key;
2260 uint32_t hashval = 0, limit, isstr;
2261 caddr_t tomax, data, kdata;
2262 dtrace_actkind_t action;
2263 dtrace_action_t *act;
2264 uintptr_t offs;
2265
2266 if (buf == NULL)
2267 return;
2268
2269 if (!agg->dtag_hasarg) {
2270 /*
2271 * Currently, only quantize() and lquantize() take additional
2272 * arguments, and they have the same semantics: an increment
2273 * value that defaults to 1 when not present. If additional
2274 * aggregating actions take arguments, the setting of the
2275 * default argument value will presumably have to become more
2276 * sophisticated...
2277 */
2278 arg = 1;
2279 }
2280
2281 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2282 size = rec->dtrd_offset - agg->dtag_base;
2283 fsize = size + rec->dtrd_size;
2284
2285 ASSERT(dbuf->dtb_tomax != NULL);
2286 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2287
2288 if ((tomax = buf->dtb_tomax) == NULL) {
2289 dtrace_buffer_drop(buf);
2290 return;
2291 }
2292
2293 /*
2294 * The metastructure is always at the bottom of the buffer.
2295 */
2296 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2297 sizeof (dtrace_aggbuffer_t));
2298
2299 if (buf->dtb_offset == 0) {
2300 /*
2301 * We just kludge up approximately 1/8th of the size to be
2302 * buckets. If this guess ends up being routinely
2303 * off-the-mark, we may need to dynamically readjust this
2304 * based on past performance.
2305 */
2306 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2307
2308 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2309 (uintptr_t)tomax || hashsize == 0) {
2310 /*
2311 * We've been given a ludicrously small buffer;
2312 * increment our drop count and leave.
2313 */
2314 dtrace_buffer_drop(buf);
2315 return;
2316 }
2317
2318 /*
2319 * And now, a pathetic attempt to try to get a an odd (or
2320 * perchance, a prime) hash size for better hash distribution.
2321 */
2322 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2323 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2324
2325 agb->dtagb_hashsize = hashsize;
2326 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2327 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2328 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2329
2330 for (i = 0; i < agb->dtagb_hashsize; i++)
2331 agb->dtagb_hash[i] = NULL;
2332 }
2333
2334 ASSERT(agg->dtag_first != NULL);
2335 ASSERT(agg->dtag_first->dta_intuple);
2336
2337 /*
2338 * Calculate the hash value based on the key. Note that we _don't_
2339 * include the aggid in the hashing (but we will store it as part of
2340 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2341 * algorithm: a simple, quick algorithm that has no known funnels, and
2342 * gets good distribution in practice. The efficacy of the hashing
2343 * algorithm (and a comparison with other algorithms) may be found by
2344 * running the ::dtrace_aggstat MDB dcmd.
2345 */
2346 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2347 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2348 limit = i + act->dta_rec.dtrd_size;
2349 ASSERT(limit <= size);
2350 isstr = DTRACEACT_ISSTRING(act);
2351
2352 for (; i < limit; i++) {
2353 hashval += data[i];
2354 hashval += (hashval << 10);
2355 hashval ^= (hashval >> 6);
2356
2357 if (isstr && data[i] == '\0')
2358 break;
2359 }
2360 }
2361
2362 hashval += (hashval << 3);
2363 hashval ^= (hashval >> 11);
2364 hashval += (hashval << 15);
2365
2366 /*
2367 * Yes, the divide here is expensive -- but it's generally the least
2368 * of the performance issues given the amount of data that we iterate
2369 * over to compute hash values, compare data, etc.
2370 */
2371 ndx = hashval % agb->dtagb_hashsize;
2372
2373 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2374 ASSERT((caddr_t)key >= tomax);
2375 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2376
2377 if (hashval != key->dtak_hashval || key->dtak_size != size)
2378 continue;
2379
2380 kdata = key->dtak_data;
2381 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2382
2383 for (act = agg->dtag_first; act->dta_intuple;
2384 act = act->dta_next) {
2385 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2386 limit = i + act->dta_rec.dtrd_size;
2387 ASSERT(limit <= size);
2388 isstr = DTRACEACT_ISSTRING(act);
2389
2390 for (; i < limit; i++) {
2391 if (kdata[i] != data[i])
2392 goto next;
2393
2394 if (isstr && data[i] == '\0')
2395 break;
2396 }
2397 }
2398
2399 if (action != key->dtak_action) {
2400 /*
2401 * We are aggregating on the same value in the same
2402 * aggregation with two different aggregating actions.
2403 * (This should have been picked up in the compiler,
2404 * so we may be dealing with errant or devious DIF.)
2405 * This is an error condition; we indicate as much,
2406 * and return.
2407 */
2408 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2409 return;
2410 }
2411
2412 /*
2413 * This is a hit: we need to apply the aggregator to
2414 * the value at this key.
2415 */
2416 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2417 return;
2418 next:
2419 continue;
2420 }
2421
2422 /*
2423 * We didn't find it. We need to allocate some zero-filled space,
2424 * link it into the hash table appropriately, and apply the aggregator
2425 * to the (zero-filled) value.
2426 */
2427 offs = buf->dtb_offset;
2428 while (offs & (align - 1))
2429 offs += sizeof (uint32_t);
2430
2431 /*
2432 * If we don't have enough room to both allocate a new key _and_
2433 * its associated data, increment the drop count and return.
2434 */
2435 if ((uintptr_t)tomax + offs + fsize >
2436 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2437 dtrace_buffer_drop(buf);
2438 return;
2439 }
2440
2441 /*CONSTCOND*/
2442 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2443 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2444 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2445
2446 key->dtak_data = kdata = tomax + offs;
2447 buf->dtb_offset = offs + fsize;
2448
2449 /*
2450 * Now copy the data across.
2451 */
2452 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2453
2454 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2455 kdata[i] = data[i];
2456
2457 /*
2458 * Because strings are not zeroed out by default, we need to iterate
2459 * looking for actions that store strings, and we need to explicitly
2460 * pad these strings out with zeroes.
2461 */
2462 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2463 int nul;
2464
2465 if (!DTRACEACT_ISSTRING(act))
2466 continue;
2467
2468 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2469 limit = i + act->dta_rec.dtrd_size;
2470 ASSERT(limit <= size);
2471
2472 for (nul = 0; i < limit; i++) {
2473 if (nul) {
2474 kdata[i] = '\0';
2475 continue;
2476 }
2477
2478 if (data[i] != '\0')
2479 continue;
2480
2481 nul = 1;
2482 }
2483 }
2484
2485 for (i = size; i < fsize; i++)
2486 kdata[i] = 0;
2487
2488 key->dtak_hashval = hashval;
2489 key->dtak_size = size;
2490 key->dtak_action = action;
2491 key->dtak_next = agb->dtagb_hash[ndx];
2492 agb->dtagb_hash[ndx] = key;
2493
2494 /*
2495 * Finally, apply the aggregator.
2496 */
2497 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2498 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2499 }
2500
2501 /*
2502 * Given consumer state, this routine finds a speculation in the INACTIVE
2503 * state and transitions it into the ACTIVE state. If there is no speculation
2504 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2505 * incremented -- it is up to the caller to take appropriate action.
2506 */
2507 static int
2508 dtrace_speculation(dtrace_state_t *state)
2509 {
2510 int i = 0;
2511 dtrace_speculation_state_t current;
2512 uint32_t *stat = &state->dts_speculations_unavail, count;
2513
2514 while (i < state->dts_nspeculations) {
2515 dtrace_speculation_t *spec = &state->dts_speculations[i];
2516
2517 current = spec->dtsp_state;
2518
2519 if (current != DTRACESPEC_INACTIVE) {
2520 if (current == DTRACESPEC_COMMITTINGMANY ||
2521 current == DTRACESPEC_COMMITTING ||
2522 current == DTRACESPEC_DISCARDING)
2523 stat = &state->dts_speculations_busy;
2524 i++;
2525 continue;
2526 }
2527
2528 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2529 current, DTRACESPEC_ACTIVE) == current)
2530 return (i + 1);
2531 }
2532
2533 /*
2534 * We couldn't find a speculation. If we found as much as a single
2535 * busy speculation buffer, we'll attribute this failure as "busy"
2536 * instead of "unavail".
2537 */
2538 do {
2539 count = *stat;
2540 } while (dtrace_cas32(stat, count, count + 1) != count);
2541
2542 return (0);
2543 }
2544
2545 /*
2546 * This routine commits an active speculation. If the specified speculation
2547 * is not in a valid state to perform a commit(), this routine will silently do
2548 * nothing. The state of the specified speculation is transitioned according
2549 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2550 */
2551 static void
2552 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2553 dtrace_specid_t which)
2554 {
2555 dtrace_speculation_t *spec;
2556 dtrace_buffer_t *src, *dest;
2557 uintptr_t daddr, saddr, dlimit, slimit;
2558 dtrace_speculation_state_t current, new;
2559 intptr_t offs;
2560 uint64_t timestamp;
2561
2562 if (which == 0)
2563 return;
2564
2565 if (which > state->dts_nspeculations) {
2566 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2567 return;
2568 }
2569
2570 spec = &state->dts_speculations[which - 1];
2571 src = &spec->dtsp_buffer[cpu];
2572 dest = &state->dts_buffer[cpu];
2573
2574 do {
2575 current = spec->dtsp_state;
2576
2577 if (current == DTRACESPEC_COMMITTINGMANY)
2578 break;
2579
2580 switch (current) {
2581 case DTRACESPEC_INACTIVE:
2582 case DTRACESPEC_DISCARDING:
2583 return;
2584
2585 case DTRACESPEC_COMMITTING:
2586 /*
2587 * This is only possible if we are (a) commit()'ing
2588 * without having done a prior speculate() on this CPU
2589 * and (b) racing with another commit() on a different
2590 * CPU. There's nothing to do -- we just assert that
2591 * our offset is 0.
2592 */
2593 ASSERT(src->dtb_offset == 0);
2594 return;
2595
2596 case DTRACESPEC_ACTIVE:
2597 new = DTRACESPEC_COMMITTING;
2598 break;
2599
2600 case DTRACESPEC_ACTIVEONE:
2601 /*
2602 * This speculation is active on one CPU. If our
2603 * buffer offset is non-zero, we know that the one CPU
2604 * must be us. Otherwise, we are committing on a
2605 * different CPU from the speculate(), and we must
2606 * rely on being asynchronously cleaned.
2607 */
2608 if (src->dtb_offset != 0) {
2609 new = DTRACESPEC_COMMITTING;
2610 break;
2611 }
2612 /*FALLTHROUGH*/
2613
2614 case DTRACESPEC_ACTIVEMANY:
2615 new = DTRACESPEC_COMMITTINGMANY;
2616 break;
2617
2618 default:
2619 ASSERT(0);
2620 }
2621 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2622 current, new) != current);
2623
2624 /*
2625 * We have set the state to indicate that we are committing this
2626 * speculation. Now reserve the necessary space in the destination
2627 * buffer.
2628 */
2629 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2630 sizeof (uint64_t), state, NULL)) < 0) {
2631 dtrace_buffer_drop(dest);
2632 goto out;
2633 }
2634
2635 /*
2636 * We have sufficient space to copy the speculative buffer into the
2637 * primary buffer. First, modify the speculative buffer, filling
2638 * in the timestamp of all entries with the current time. The data
2639 * must have the commit() time rather than the time it was traced,
2640 * so that all entries in the primary buffer are in timestamp order.
2641 */
2642 timestamp = dtrace_gethrtime();
2643 saddr = (uintptr_t)src->dtb_tomax;
2644 slimit = saddr + src->dtb_offset;
2645 while (saddr < slimit) {
2646 size_t size;
2647 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2648
2649 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2650 saddr += sizeof (dtrace_epid_t);
2651 continue;
2652 }
2653 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2654 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2655
2656 ASSERT3U(saddr + size, <=, slimit);
2657 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2658 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2659
2660 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2661
2662 saddr += size;
2663 }
2664
2665 /*
2666 * Copy the buffer across. (Note that this is a
2667 * highly subobtimal bcopy(); in the unlikely event that this becomes
2668 * a serious performance issue, a high-performance DTrace-specific
2669 * bcopy() should obviously be invented.)
2670 */
2671 daddr = (uintptr_t)dest->dtb_tomax + offs;
2672 dlimit = daddr + src->dtb_offset;
2673 saddr = (uintptr_t)src->dtb_tomax;
2674
2675 /*
2676 * First, the aligned portion.
2677 */
2678 while (dlimit - daddr >= sizeof (uint64_t)) {
2679 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2680
2681 daddr += sizeof (uint64_t);
2682 saddr += sizeof (uint64_t);
2683 }
2684
2685 /*
2686 * Now any left-over bit...
2687 */
2688 while (dlimit - daddr)
2689 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2690
2691 /*
2692 * Finally, commit the reserved space in the destination buffer.
2693 */
2694 dest->dtb_offset = offs + src->dtb_offset;
2695
2696 out:
2697 /*
2698 * If we're lucky enough to be the only active CPU on this speculation
2699 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2700 */
2701 if (current == DTRACESPEC_ACTIVE ||
2702 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2703 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2704 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2705
2706 ASSERT(rval == DTRACESPEC_COMMITTING);
2707 }
2708
2709 src->dtb_offset = 0;
2710 src->dtb_xamot_drops += src->dtb_drops;
2711 src->dtb_drops = 0;
2712 }
2713
2714 /*
2715 * This routine discards an active speculation. If the specified speculation
2716 * is not in a valid state to perform a discard(), this routine will silently
2717 * do nothing. The state of the specified speculation is transitioned
2718 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2719 */
2720 static void
2721 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2722 dtrace_specid_t which)
2723 {
2724 dtrace_speculation_t *spec;
2725 dtrace_speculation_state_t current, new;
2726 dtrace_buffer_t *buf;
2727
2728 if (which == 0)
2729 return;
2730
2731 if (which > state->dts_nspeculations) {
2732 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2733 return;
2734 }
2735
2736 spec = &state->dts_speculations[which - 1];
2737 buf = &spec->dtsp_buffer[cpu];
2738
2739 do {
2740 current = spec->dtsp_state;
2741
2742 switch (current) {
2743 case DTRACESPEC_INACTIVE:
2744 case DTRACESPEC_COMMITTINGMANY:
2745 case DTRACESPEC_COMMITTING:
2746 case DTRACESPEC_DISCARDING:
2747 return;
2748
2749 case DTRACESPEC_ACTIVE:
2750 case DTRACESPEC_ACTIVEMANY:
2751 new = DTRACESPEC_DISCARDING;
2752 break;
2753
2754 case DTRACESPEC_ACTIVEONE:
2755 if (buf->dtb_offset != 0) {
2756 new = DTRACESPEC_INACTIVE;
2757 } else {
2758 new = DTRACESPEC_DISCARDING;
2759 }
2760 break;
2761
2762 default:
2763 ASSERT(0);
2764 }
2765 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2766 current, new) != current);
2767
2768 buf->dtb_offset = 0;
2769 buf->dtb_drops = 0;
2770 }
2771
2772 /*
2773 * Note: not called from probe context. This function is called
2774 * asynchronously from cross call context to clean any speculations that are
2775 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2776 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2777 * speculation.
2778 */
2779 static void
2780 dtrace_speculation_clean_here(dtrace_state_t *state)
2781 {
2782 dtrace_icookie_t cookie;
2783 processorid_t cpu = CPU->cpu_id;
2784 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2785 dtrace_specid_t i;
2786
2787 cookie = dtrace_interrupt_disable();
2788
2789 if (dest->dtb_tomax == NULL) {
2790 dtrace_interrupt_enable(cookie);
2791 return;
2792 }
2793
2794 for (i = 0; i < state->dts_nspeculations; i++) {
2795 dtrace_speculation_t *spec = &state->dts_speculations[i];
2796 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2797
2798 if (src->dtb_tomax == NULL)
2799 continue;
2800
2801 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2802 src->dtb_offset = 0;
2803 continue;
2804 }
2805
2806 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2807 continue;
2808
2809 if (src->dtb_offset == 0)
2810 continue;
2811
2812 dtrace_speculation_commit(state, cpu, i + 1);
2813 }
2814
2815 dtrace_interrupt_enable(cookie);
2816 }
2817
2818 /*
2819 * Note: not called from probe context. This function is called
2820 * asynchronously (and at a regular interval) to clean any speculations that
2821 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2822 * is work to be done, it cross calls all CPUs to perform that work;
2823 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2824 * INACTIVE state until they have been cleaned by all CPUs.
2825 */
2826 static void
2827 dtrace_speculation_clean(dtrace_state_t *state)
2828 {
2829 int work = 0, rv;
2830 dtrace_specid_t i;
2831
2832 for (i = 0; i < state->dts_nspeculations; i++) {
2833 dtrace_speculation_t *spec = &state->dts_speculations[i];
2834
2835 ASSERT(!spec->dtsp_cleaning);
2836
2837 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2838 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2839 continue;
2840
2841 work++;
2842 spec->dtsp_cleaning = 1;
2843 }
2844
2845 if (!work)
2846 return;
2847
2848 dtrace_xcall(DTRACE_CPUALL,
2849 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2850
2851 /*
2852 * We now know that all CPUs have committed or discarded their
2853 * speculation buffers, as appropriate. We can now set the state
2854 * to inactive.
2855 */
2856 for (i = 0; i < state->dts_nspeculations; i++) {
2857 dtrace_speculation_t *spec = &state->dts_speculations[i];
2858 dtrace_speculation_state_t current, new;
2859
2860 if (!spec->dtsp_cleaning)
2861 continue;
2862
2863 current = spec->dtsp_state;
2864 ASSERT(current == DTRACESPEC_DISCARDING ||
2865 current == DTRACESPEC_COMMITTINGMANY);
2866
2867 new = DTRACESPEC_INACTIVE;
2868
2869 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2870 ASSERT(rv == current);
2871 spec->dtsp_cleaning = 0;
2872 }
2873 }
2874
2875 /*
2876 * Called as part of a speculate() to get the speculative buffer associated
2877 * with a given speculation. Returns NULL if the specified speculation is not
2878 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2879 * the active CPU is not the specified CPU -- the speculation will be
2880 * atomically transitioned into the ACTIVEMANY state.
2881 */
2882 static dtrace_buffer_t *
2883 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2884 dtrace_specid_t which)
2885 {
2886 dtrace_speculation_t *spec;
2887 dtrace_speculation_state_t current, new;
2888 dtrace_buffer_t *buf;
2889
2890 if (which == 0)
2891 return (NULL);
2892
2893 if (which > state->dts_nspeculations) {
2894 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2895 return (NULL);
2896 }
2897
2898 spec = &state->dts_speculations[which - 1];
2899 buf = &spec->dtsp_buffer[cpuid];
2900
2901 do {
2902 current = spec->dtsp_state;
2903
2904 switch (current) {
2905 case DTRACESPEC_INACTIVE:
2906 case DTRACESPEC_COMMITTINGMANY:
2907 case DTRACESPEC_DISCARDING:
2908 return (NULL);
2909
2910 case DTRACESPEC_COMMITTING:
2911 ASSERT(buf->dtb_offset == 0);
2912 return (NULL);
2913
2914 case DTRACESPEC_ACTIVEONE:
2915 /*
2916 * This speculation is currently active on one CPU.
2917 * Check the offset in the buffer; if it's non-zero,
2918 * that CPU must be us (and we leave the state alone).
2919 * If it's zero, assume that we're starting on a new
2920 * CPU -- and change the state to indicate that the
2921 * speculation is active on more than one CPU.
2922 */
2923 if (buf->dtb_offset != 0)
2924 return (buf);
2925
2926 new = DTRACESPEC_ACTIVEMANY;
2927 break;
2928
2929 case DTRACESPEC_ACTIVEMANY:
2930 return (buf);
2931
2932 case DTRACESPEC_ACTIVE:
2933 new = DTRACESPEC_ACTIVEONE;
2934 break;
2935
2936 default:
2937 ASSERT(0);
2938 }
2939 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2940 current, new) != current);
2941
2942 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2943 return (buf);
2944 }
2945
2946 /*
2947 * Return a string. In the event that the user lacks the privilege to access
2948 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2949 * don't fail access checking.
2950 *
2951 * dtrace_dif_variable() uses this routine as a helper for various
2952 * builtin values such as 'execname' and 'probefunc.'
2953 */
2954 uintptr_t
2955 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2956 dtrace_mstate_t *mstate)
2957 {
2958 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2959 uintptr_t ret;
2960 size_t strsz;
2961
2962 /*
2963 * The easy case: this probe is allowed to read all of memory, so
2964 * we can just return this as a vanilla pointer.
2965 */
2966 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2967 return (addr);
2968
2969 /*
2970 * This is the tougher case: we copy the string in question from
2971 * kernel memory into scratch memory and return it that way: this
2972 * ensures that we won't trip up when access checking tests the
2973 * BYREF return value.
2974 */
2975 strsz = dtrace_strlen((char *)addr, size) + 1;
2976
2977 if (mstate->dtms_scratch_ptr + strsz >
2978 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2979 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2980 return (NULL);
2981 }
2982
2983 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2984 strsz);
2985 ret = mstate->dtms_scratch_ptr;
2986 mstate->dtms_scratch_ptr += strsz;
2987 return (ret);
2988 }
2989
2990 /*
2991 * This function implements the DIF emulator's variable lookups. The emulator
2992 * passes a reserved variable identifier and optional built-in array index.
2993 */
2994 static uint64_t
2995 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2996 uint64_t ndx)
2997 {
2998 /*
2999 * If we're accessing one of the uncached arguments, we'll turn this
3000 * into a reference in the args array.
3001 */
3002 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3003 ndx = v - DIF_VAR_ARG0;
3004 v = DIF_VAR_ARGS;
3005 }
3006
3007 switch (v) {
3008 case DIF_VAR_ARGS:
3009 if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
3010 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
3011 CPU_DTRACE_KPRIV;
3012 return (0);
3013 }
3014
3015 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3016 if (ndx >= sizeof (mstate->dtms_arg) /
3017 sizeof (mstate->dtms_arg[0])) {
3018 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3019 dtrace_provider_t *pv;
3020 uint64_t val;
3021
3022 pv = mstate->dtms_probe->dtpr_provider;
3023 if (pv->dtpv_pops.dtps_getargval != NULL)
3024 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3025 mstate->dtms_probe->dtpr_id,
3026 mstate->dtms_probe->dtpr_arg, ndx, aframes);
3027 else
3028 val = dtrace_getarg(ndx, aframes);
3029
3030 /*
3031 * This is regrettably required to keep the compiler
3032 * from tail-optimizing the call to dtrace_getarg().
3033 * The condition always evaluates to true, but the
3034 * compiler has no way of figuring that out a priori.
3035 * (None of this would be necessary if the compiler
3036 * could be relied upon to _always_ tail-optimize
3037 * the call to dtrace_getarg() -- but it can't.)
3038 */
3039 if (mstate->dtms_probe != NULL)
3040 return (val);
3041
3042 ASSERT(0);
3043 }
3044
3045 return (mstate->dtms_arg[ndx]);
3046
3047 case DIF_VAR_UREGS: {
3048 klwp_t *lwp;
3049
3050 if (!dtrace_priv_proc(state, mstate))
3051 return (0);
3052
3053 if ((lwp = curthread->t_lwp) == NULL) {
3054 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3055 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
3056 return (0);
3057 }
3058
3059 return (dtrace_getreg(lwp->lwp_regs, ndx));
3060 }
3061
3062 case DIF_VAR_VMREGS: {
3063 uint64_t rval;
3064
3065 if (!dtrace_priv_kernel(state))
3066 return (0);
3067
3068 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3069
3070 rval = dtrace_getvmreg(ndx,
3071 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
3072
3073 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3074
3075 return (rval);
3076 }
3077
3078 case DIF_VAR_CURTHREAD:
3079 if (!dtrace_priv_proc(state, mstate))
3080 return (0);
3081 return ((uint64_t)(uintptr_t)curthread);
3082
3083 case DIF_VAR_TIMESTAMP:
3084 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3085 mstate->dtms_timestamp = dtrace_gethrtime();
3086 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3087 }
3088 return (mstate->dtms_timestamp);
3089
3090 case DIF_VAR_VTIMESTAMP:
3091 ASSERT(dtrace_vtime_references != 0);
3092 return (curthread->t_dtrace_vtime);
3093
3094 case DIF_VAR_WALLTIMESTAMP:
3095 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3096 mstate->dtms_walltimestamp = dtrace_gethrestime();
3097 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3098 }
3099 return (mstate->dtms_walltimestamp);
3100
3101 case DIF_VAR_IPL:
3102 if (!dtrace_priv_kernel(state))
3103 return (0);
3104 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3105 mstate->dtms_ipl = dtrace_getipl();
3106 mstate->dtms_present |= DTRACE_MSTATE_IPL;
3107 }
3108 return (mstate->dtms_ipl);
3109
3110 case DIF_VAR_EPID:
3111 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3112 return (mstate->dtms_epid);
3113
3114 case DIF_VAR_ID:
3115 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3116 return (mstate->dtms_probe->dtpr_id);
3117
3118 case DIF_VAR_STACKDEPTH:
3119 if (!dtrace_priv_kernel(state))
3120 return (0);
3121 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3122 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3123
3124 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3125 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3126 }
3127 return (mstate->dtms_stackdepth);
3128
3129 case DIF_VAR_USTACKDEPTH:
3130 if (!dtrace_priv_proc(state, mstate))
3131 return (0);
3132 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3133 /*
3134 * See comment in DIF_VAR_PID.
3135 */
3136 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3137 CPU_ON_INTR(CPU)) {
3138 mstate->dtms_ustackdepth = 0;
3139 } else {
3140 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3141 mstate->dtms_ustackdepth =
3142 dtrace_getustackdepth();
3143 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3144 }
3145 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3146 }
3147 return (mstate->dtms_ustackdepth);
3148
3149 case DIF_VAR_CALLER:
3150 if (!dtrace_priv_kernel(state))
3151 return (0);
3152 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3153 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3154
3155 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3156 /*
3157 * If this is an unanchored probe, we are
3158 * required to go through the slow path:
3159 * dtrace_caller() only guarantees correct
3160 * results for anchored probes.
3161 */
3162 pc_t caller[2];
3163
3164 dtrace_getpcstack(caller, 2, aframes,
3165 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3166 mstate->dtms_caller = caller[1];
3167 } else if ((mstate->dtms_caller =
3168 dtrace_caller(aframes)) == -1) {
3169 /*
3170 * We have failed to do this the quick way;
3171 * we must resort to the slower approach of
3172 * calling dtrace_getpcstack().
3173 */
3174 pc_t caller;
3175
3176 dtrace_getpcstack(&caller, 1, aframes, NULL);
3177 mstate->dtms_caller = caller;
3178 }
3179
3180 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3181 }
3182 return (mstate->dtms_caller);
3183
3184 case DIF_VAR_UCALLER:
3185 if (!dtrace_priv_proc(state, mstate))
3186 return (0);
3187
3188 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3189 uint64_t ustack[3];
3190
3191 /*
3192 * dtrace_getupcstack() fills in the first uint64_t
3193 * with the current PID. The second uint64_t will
3194 * be the program counter at user-level. The third
3195 * uint64_t will contain the caller, which is what
3196 * we're after.
3197 */
3198 ustack[2] = NULL;
3199 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3200 dtrace_getupcstack(ustack, 3);
3201 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3202 mstate->dtms_ucaller = ustack[2];
3203 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3204 }
3205
3206 return (mstate->dtms_ucaller);
3207
3208 case DIF_VAR_PROBEPROV:
3209 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3210 return (dtrace_dif_varstr(
3211 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3212 state, mstate));
3213
3214 case DIF_VAR_PROBEMOD:
3215 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3216 return (dtrace_dif_varstr(
3217 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3218 state, mstate));
3219
3220 case DIF_VAR_PROBEFUNC:
3221 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3222 return (dtrace_dif_varstr(
3223 (uintptr_t)mstate->dtms_probe->dtpr_func,
3224 state, mstate));
3225
3226 case DIF_VAR_PROBENAME:
3227 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3228 return (dtrace_dif_varstr(
3229 (uintptr_t)mstate->dtms_probe->dtpr_name,
3230 state, mstate));
3231
3232 case DIF_VAR_PID:
3233 if (!dtrace_priv_proc(state, mstate))
3234 return (0);
3235
3236 /*
3237 * Note that we are assuming that an unanchored probe is
3238 * always due to a high-level interrupt. (And we're assuming
3239 * that there is only a single high level interrupt.)
3240 */
3241 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3242 return (pid0.pid_id);
3243
3244 /*
3245 * It is always safe to dereference one's own t_procp pointer:
3246 * it always points to a valid, allocated proc structure.
3247 * Further, it is always safe to dereference the p_pidp member
3248 * of one's own proc structure. (These are truisms becuase
3249 * threads and processes don't clean up their own state --
3250 * they leave that task to whomever reaps them.)
3251 */
3252 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3253
3254 case DIF_VAR_PPID:
3255 if (!dtrace_priv_proc(state, mstate))
3256 return (0);
3257
3258 /*
3259 * See comment in DIF_VAR_PID.
3260 */
3261 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3262 return (pid0.pid_id);
3263
3264 /*
3265 * It is always safe to dereference one's own t_procp pointer:
3266 * it always points to a valid, allocated proc structure.
3267 * (This is true because threads don't clean up their own
3268 * state -- they leave that task to whomever reaps them.)
3269 */
3270 return ((uint64_t)curthread->t_procp->p_ppid);
3271
3272 case DIF_VAR_TID:
3273 /*
3274 * See comment in DIF_VAR_PID.
3275 */
3276 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3277 return (0);
3278
3279 return ((uint64_t)curthread->t_tid);
3280
3281 case DIF_VAR_EXECNAME:
3282 if (!dtrace_priv_proc(state, mstate))
3283 return (0);
3284
3285 /*
3286 * See comment in DIF_VAR_PID.
3287 */
3288 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3289 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3290
3291 /*
3292 * It is always safe to dereference one's own t_procp pointer:
3293 * it always points to a valid, allocated proc structure.
3294 * (This is true because threads don't clean up their own
3295 * state -- they leave that task to whomever reaps them.)
3296 */
3297 return (dtrace_dif_varstr(
3298 (uintptr_t)curthread->t_procp->p_user.u_comm,
3299 state, mstate));
3300
3301 case DIF_VAR_ZONENAME:
3302 if (!dtrace_priv_proc(state, mstate))
3303 return (0);
3304
3305 /*
3306 * See comment in DIF_VAR_PID.
3307 */
3308 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3309 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3310
3311 /*
3312 * It is always safe to dereference one's own t_procp pointer:
3313 * it always points to a valid, allocated proc structure.
3314 * (This is true because threads don't clean up their own
3315 * state -- they leave that task to whomever reaps them.)
3316 */
3317 return (dtrace_dif_varstr(
3318 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3319 state, mstate));
3320
3321 case DIF_VAR_UID:
3322 if (!dtrace_priv_proc(state, mstate))
3323 return (0);
3324
3325 /*
3326 * See comment in DIF_VAR_PID.
3327 */
3328 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3329 return ((uint64_t)p0.p_cred->cr_uid);
3330
3331 /*
3332 * It is always safe to dereference one's own t_procp pointer:
3333 * it always points to a valid, allocated proc structure.
3334 * (This is true because threads don't clean up their own
3335 * state -- they leave that task to whomever reaps them.)
3336 *
3337 * Additionally, it is safe to dereference one's own process
3338 * credential, since this is never NULL after process birth.
3339 */
3340 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3341
3342 case DIF_VAR_GID:
3343 if (!dtrace_priv_proc(state, mstate))
3344 return (0);
3345
3346 /*
3347 * See comment in DIF_VAR_PID.
3348 */
3349 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3350 return ((uint64_t)p0.p_cred->cr_gid);
3351
3352 /*
3353 * It is always safe to dereference one's own t_procp pointer:
3354 * it always points to a valid, allocated proc structure.
3355 * (This is true because threads don't clean up their own
3356 * state -- they leave that task to whomever reaps them.)
3357 *
3358 * Additionally, it is safe to dereference one's own process
3359 * credential, since this is never NULL after process birth.
3360 */
3361 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3362
3363 case DIF_VAR_ERRNO: {
3364 klwp_t *lwp;
3365 if (!dtrace_priv_proc(state, mstate))
3366 return (0);
3367
3368 /*
3369 * See comment in DIF_VAR_PID.
3370 */
3371 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3372 return (0);
3373
3374 /*
3375 * It is always safe to dereference one's own t_lwp pointer in
3376 * the event that this pointer is non-NULL. (This is true
3377 * because threads and lwps don't clean up their own state --
3378 * they leave that task to whomever reaps them.)
3379 */
3380 if ((lwp = curthread->t_lwp) == NULL)
3381 return (0);
3382
3383 return ((uint64_t)lwp->lwp_errno);
3384 }
3385 default:
3386 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3387 return (0);
3388 }
3389 }
3390
3391
3392 typedef enum dtrace_json_state {
3393 DTRACE_JSON_REST = 1,
3394 DTRACE_JSON_OBJECT,
3395 DTRACE_JSON_STRING,
3396 DTRACE_JSON_STRING_ESCAPE,
3397 DTRACE_JSON_STRING_ESCAPE_UNICODE,
3398 DTRACE_JSON_COLON,
3399 DTRACE_JSON_COMMA,
3400 DTRACE_JSON_VALUE,
3401 DTRACE_JSON_IDENTIFIER,
3402 DTRACE_JSON_NUMBER,
3403 DTRACE_JSON_NUMBER_FRAC,
3404 DTRACE_JSON_NUMBER_EXP,
3405 DTRACE_JSON_COLLECT_OBJECT
3406 } dtrace_json_state_t;
3407
3408 /*
3409 * This function possesses just enough knowledge about JSON to extract a single
3410 * value from a JSON string and store it in the scratch buffer. It is able
3411 * to extract nested object values, and members of arrays by index.
3412 *
3413 * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3414 * be looked up as we descend into the object tree. e.g.
3415 *
3416 * foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3417 * with nelems = 5.
3418 *
3419 * The run time of this function must be bounded above by strsize to limit the
3420 * amount of work done in probe context. As such, it is implemented as a
3421 * simple state machine, reading one character at a time using safe loads
3422 * until we find the requested element, hit a parsing error or run off the
3423 * end of the object or string.
3424 *
3425 * As there is no way for a subroutine to return an error without interrupting
3426 * clause execution, we simply return NULL in the event of a missing key or any
3427 * other error condition. Each NULL return in this function is commented with
3428 * the error condition it represents -- parsing or otherwise.
3429 *
3430 * The set of states for the state machine closely matches the JSON
3431 * specification (http://json.org/). Briefly:
3432 *
3433 * DTRACE_JSON_REST:
3434 * Skip whitespace until we find either a top-level Object, moving
3435 * to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3436 *
3437 * DTRACE_JSON_OBJECT:
3438 * Locate the next key String in an Object. Sets a flag to denote
3439 * the next String as a key string and moves to DTRACE_JSON_STRING.
3440 *
3441 * DTRACE_JSON_COLON:
3442 * Skip whitespace until we find the colon that separates key Strings
3443 * from their values. Once found, move to DTRACE_JSON_VALUE.
3444 *
3445 * DTRACE_JSON_VALUE:
3446 * Detects the type of the next value (String, Number, Identifier, Object
3447 * or Array) and routes to the states that process that type. Here we also
3448 * deal with the element selector list if we are requested to traverse down
3449 * into the object tree.
3450 *
3451 * DTRACE_JSON_COMMA:
3452 * Skip whitespace until we find the comma that separates key-value pairs
3453 * in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3454 * (similarly DTRACE_JSON_VALUE). All following literal value processing
3455 * states return to this state at the end of their value, unless otherwise
3456 * noted.
3457 *
3458 * DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3459 * Processes a Number literal from the JSON, including any exponent
3460 * component that may be present. Numbers are returned as strings, which
3461 * may be passed to strtoll() if an integer is required.
3462 *
3463 * DTRACE_JSON_IDENTIFIER:
3464 * Processes a "true", "false" or "null" literal in the JSON.
3465 *
3466 * DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3467 * DTRACE_JSON_STRING_ESCAPE_UNICODE:
3468 * Processes a String literal from the JSON, whether the String denotes
3469 * a key, a value or part of a larger Object. Handles all escape sequences
3470 * present in the specification, including four-digit unicode characters,
3471 * but merely includes the escape sequence without converting it to the
3472 * actual escaped character. If the String is flagged as a key, we
3473 * move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3474 *
3475 * DTRACE_JSON_COLLECT_OBJECT:
3476 * This state collects an entire Object (or Array), correctly handling
3477 * embedded strings. If the full element selector list matches this nested
3478 * object, we return the Object in full as a string. If not, we use this
3479 * state to skip to the next value at this level and continue processing.
3480 *
3481 * NOTE: This function uses various macros from strtolctype.h to manipulate
3482 * digit values, etc -- these have all been checked to ensure they make
3483 * no additional function calls.
3484 */
3485 static char *
3486 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3487 char *dest)
3488 {
3489 dtrace_json_state_t state = DTRACE_JSON_REST;
3490 int64_t array_elem = INT64_MIN;
3491 int64_t array_pos = 0;
3492 uint8_t escape_unicount = 0;
3493 boolean_t string_is_key = B_FALSE;
3494 boolean_t collect_object = B_FALSE;
3495 boolean_t found_key = B_FALSE;
3496 boolean_t in_array = B_FALSE;
3497 uint32_t braces = 0, brackets = 0;
3498 char *elem = elemlist;
3499 char *dd = dest;
3500 uintptr_t cur;
3501
3502 for (cur = json; cur < json + size; cur++) {
3503 char cc = dtrace_load8(cur);
3504 if (cc == '\0')
3505 return (NULL);
3506
3507 switch (state) {
3508 case DTRACE_JSON_REST:
3509 if (isspace(cc))
3510 break;
3511
3512 if (cc == '{') {
3513 state = DTRACE_JSON_OBJECT;
3514 break;
3515 }
3516
3517 if (cc == '[') {
3518 in_array = B_TRUE;
3519 array_pos = 0;
3520 array_elem = dtrace_strtoll(elem, 10, size);
3521 found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3522 state = DTRACE_JSON_VALUE;
3523 break;
3524 }
3525
3526 /*
3527 * ERROR: expected to find a top-level object or array.
3528 */
3529 return (NULL);
3530 case DTRACE_JSON_OBJECT:
3531 if (isspace(cc))
3532 break;
3533
3534 if (cc == '"') {
3535 state = DTRACE_JSON_STRING;
3536 string_is_key = B_TRUE;
3537 break;
3538 }
3539
3540 /*
3541 * ERROR: either the object did not start with a key
3542 * string, or we've run off the end of the object
3543 * without finding the requested key.
3544 */
3545 return (NULL);
3546 case DTRACE_JSON_STRING:
3547 if (cc == '\\') {
3548 *dd++ = '\\';
3549 state = DTRACE_JSON_STRING_ESCAPE;
3550 break;
3551 }
3552
3553 if (cc == '"') {
3554 if (collect_object) {
3555 /*
3556 * We don't reset the dest here, as
3557 * the string is part of a larger
3558 * object being collected.
3559 */
3560 *dd++ = cc;
3561 collect_object = B_FALSE;
3562 state = DTRACE_JSON_COLLECT_OBJECT;
3563 break;
3564 }
3565 *dd = '\0';
3566 dd = dest; /* reset string buffer */
3567 if (string_is_key) {
3568 if (dtrace_strncmp(dest, elem,
3569 size) == 0)
3570 found_key = B_TRUE;
3571 } else if (found_key) {
3572 if (nelems > 1) {
3573 /*
3574 * We expected an object, not
3575 * this string.
3576 */
3577 return (NULL);
3578 }
3579 return (dest);
3580 }
3581 state = string_is_key ? DTRACE_JSON_COLON :
3582 DTRACE_JSON_COMMA;
3583 string_is_key = B_FALSE;
3584 break;
3585 }
3586
3587 *dd++ = cc;
3588 break;
3589 case DTRACE_JSON_STRING_ESCAPE:
3590 *dd++ = cc;
3591 if (cc == 'u') {
3592 escape_unicount = 0;
3593 state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3594 } else {
3595 state = DTRACE_JSON_STRING;
3596 }
3597 break;
3598 case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3599 if (!isxdigit(cc)) {
3600 /*
3601 * ERROR: invalid unicode escape, expected
3602 * four valid hexidecimal digits.
3603 */
3604 return (NULL);
3605 }
3606
3607 *dd++ = cc;
3608 if (++escape_unicount == 4)
3609 state = DTRACE_JSON_STRING;
3610 break;
3611 case DTRACE_JSON_COLON:
3612 if (isspace(cc))
3613 break;
3614
3615 if (cc == ':') {
3616 state = DTRACE_JSON_VALUE;
3617 break;
3618 }
3619
3620 /*
3621 * ERROR: expected a colon.
3622 */
3623 return (NULL);
3624 case DTRACE_JSON_COMMA:
3625 if (isspace(cc))
3626 break;
3627
3628 if (cc == ',') {
3629 if (in_array) {
3630 state = DTRACE_JSON_VALUE;
3631 if (++array_pos == array_elem)
3632 found_key = B_TRUE;
3633 } else {
3634 state = DTRACE_JSON_OBJECT;
3635 }
3636 break;
3637 }
3638
3639 /*
3640 * ERROR: either we hit an unexpected character, or
3641 * we reached the end of the object or array without
3642 * finding the requested key.
3643 */
3644 return (NULL);
3645 case DTRACE_JSON_IDENTIFIER:
3646 if (islower(cc)) {
3647 *dd++ = cc;
3648 break;
3649 }
3650
3651 *dd = '\0';
3652 dd = dest; /* reset string buffer */
3653
3654 if (dtrace_strncmp(dest, "true", 5) == 0 ||
3655 dtrace_strncmp(dest, "false", 6) == 0 ||
3656 dtrace_strncmp(dest, "null", 5) == 0) {
3657 if (found_key) {
3658 if (nelems > 1) {
3659 /*
3660 * ERROR: We expected an object,
3661 * not this identifier.
3662 */
3663 return (NULL);
3664 }
3665 return (dest);
3666 } else {
3667 cur--;
3668 state = DTRACE_JSON_COMMA;
3669 break;
3670 }
3671 }
3672
3673 /*
3674 * ERROR: we did not recognise the identifier as one
3675 * of those in the JSON specification.
3676 */
3677 return (NULL);
3678 case DTRACE_JSON_NUMBER:
3679 if (cc == '.') {
3680 *dd++ = cc;
3681 state = DTRACE_JSON_NUMBER_FRAC;
3682 break;
3683 }
3684
3685 if (cc == 'x' || cc == 'X') {
3686 /*
3687 * ERROR: specification explicitly excludes
3688 * hexidecimal or octal numbers.
3689 */
3690 return (NULL);
3691 }
3692
3693 /* FALLTHRU */
3694 case DTRACE_JSON_NUMBER_FRAC:
3695 if (cc == 'e' || cc == 'E') {
3696 *dd++ = cc;
3697 state = DTRACE_JSON_NUMBER_EXP;
3698 break;
3699 }
3700
3701 if (cc == '+' || cc == '-') {
3702 /*
3703 * ERROR: expect sign as part of exponent only.
3704 */
3705 return (NULL);
3706 }
3707 /* FALLTHRU */
3708 case DTRACE_JSON_NUMBER_EXP:
3709 if (isdigit(cc) || cc == '+' || cc == '-') {
3710 *dd++ = cc;
3711 break;
3712 }
3713
3714 *dd = '\0';
3715 dd = dest; /* reset string buffer */
3716 if (found_key) {
3717 if (nelems > 1) {
3718 /*
3719 * ERROR: We expected an object, not
3720 * this number.
3721 */
3722 return (NULL);
3723 }
3724 return (dest);
3725 }
3726
3727 cur--;
3728 state = DTRACE_JSON_COMMA;
3729 break;
3730 case DTRACE_JSON_VALUE:
3731 if (isspace(cc))
3732 break;
3733
3734 if (cc == '{' || cc == '[') {
3735 if (nelems > 1 && found_key) {
3736 in_array = cc == '[' ? B_TRUE : B_FALSE;
3737 /*
3738 * If our element selector directs us
3739 * to descend into this nested object,
3740 * then move to the next selector
3741 * element in the list and restart the
3742 * state machine.
3743 */
3744 while (*elem != '\0')
3745 elem++;
3746 elem++; /* skip the inter-element NUL */
3747 nelems--;
3748 dd = dest;
3749 if (in_array) {
3750 state = DTRACE_JSON_VALUE;
3751 array_pos = 0;
3752 array_elem = dtrace_strtoll(
3753 elem, 10, size);
3754 found_key = array_elem == 0 ?
3755 B_TRUE : B_FALSE;
3756 } else {
3757 found_key = B_FALSE;
3758 state = DTRACE_JSON_OBJECT;
3759 }
3760 break;
3761 }
3762
3763 /*
3764 * Otherwise, we wish to either skip this
3765 * nested object or return it in full.
3766 */
3767 if (cc == '[')
3768 brackets = 1;
3769 else
3770 braces = 1;
3771 *dd++ = cc;
3772 state = DTRACE_JSON_COLLECT_OBJECT;
3773 break;
3774 }
3775
3776 if (cc == '"') {
3777 state = DTRACE_JSON_STRING;
3778 break;
3779 }
3780
3781 if (islower(cc)) {
3782 /*
3783 * Here we deal with true, false and null.
3784 */
3785 *dd++ = cc;
3786 state = DTRACE_JSON_IDENTIFIER;
3787 break;
3788 }
3789
3790 if (cc == '-' || isdigit(cc)) {
3791 *dd++ = cc;
3792 state = DTRACE_JSON_NUMBER;
3793 break;
3794 }
3795
3796 /*
3797 * ERROR: unexpected character at start of value.
3798 */
3799 return (NULL);
3800 case DTRACE_JSON_COLLECT_OBJECT:
3801 if (cc == '\0')
3802 /*
3803 * ERROR: unexpected end of input.
3804 */
3805 return (NULL);
3806
3807 *dd++ = cc;
3808 if (cc == '"') {
3809 collect_object = B_TRUE;
3810 state = DTRACE_JSON_STRING;
3811 break;
3812 }
3813
3814 if (cc == ']') {
3815 if (brackets-- == 0) {
3816 /*
3817 * ERROR: unbalanced brackets.
3818 */
3819 return (NULL);
3820 }
3821 } else if (cc == '}') {
3822 if (braces-- == 0) {
3823 /*
3824 * ERROR: unbalanced braces.
3825 */
3826 return (NULL);
3827 }
3828 } else if (cc == '{') {
3829 braces++;
3830 } else if (cc == '[') {
3831 brackets++;
3832 }
3833
3834 if (brackets == 0 && braces == 0) {
3835 if (found_key) {
3836 *dd = '\0';
3837 return (dest);
3838 }
3839 dd = dest; /* reset string buffer */
3840 state = DTRACE_JSON_COMMA;
3841 }
3842 break;
3843 }
3844 }
3845 return (NULL);
3846 }
3847
3848 /*
3849 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3850 * Notice that we don't bother validating the proper number of arguments or
3851 * their types in the tuple stack. This isn't needed because all argument
3852 * interpretation is safe because of our load safety -- the worst that can
3853 * happen is that a bogus program can obtain bogus results.
3854 */
3855 static void
3856 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3857 dtrace_key_t *tupregs, int nargs,
3858 dtrace_mstate_t *mstate, dtrace_state_t *state)
3859 {
3860 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
3861 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
3862 dtrace_vstate_t *vstate = &state->dts_vstate;
3863
3864 union {
3865 mutex_impl_t mi;
3866 uint64_t mx;
3867 } m;
3868
3869 union {
3870 krwlock_t ri;
3871 uintptr_t rw;
3872 } r;
3873
3874 switch (subr) {
3875 case DIF_SUBR_RAND:
3876 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3877 break;
3878
3879 case DIF_SUBR_MUTEX_OWNED:
3880 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3881 mstate, vstate)) {
3882 regs[rd] = NULL;
3883 break;
3884 }
3885
3886 m.mx = dtrace_load64(tupregs[0].dttk_value);
3887 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3888 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3889 else
3890 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3891 break;
3892
3893 case DIF_SUBR_MUTEX_OWNER:
3894 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3895 mstate, vstate)) {
3896 regs[rd] = NULL;
3897 break;
3898 }
3899
3900 m.mx = dtrace_load64(tupregs[0].dttk_value);
3901 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3902 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3903 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3904 else
3905 regs[rd] = 0;
3906 break;
3907
3908 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3909 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3910 mstate, vstate)) {
3911 regs[rd] = NULL;
3912 break;
3913 }
3914
3915 m.mx = dtrace_load64(tupregs[0].dttk_value);
3916 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3917 break;
3918
3919 case DIF_SUBR_MUTEX_TYPE_SPIN:
3920 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3921 mstate, vstate)) {
3922 regs[rd] = NULL;
3923 break;
3924 }
3925
3926 m.mx = dtrace_load64(tupregs[0].dttk_value);
3927 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3928 break;
3929
3930 case DIF_SUBR_RW_READ_HELD: {
3931 uintptr_t tmp;
3932
3933 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3934 mstate, vstate)) {
3935 regs[rd] = NULL;
3936 break;
3937 }
3938
3939 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3940 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3941 break;
3942 }
3943
3944 case DIF_SUBR_RW_WRITE_HELD:
3945 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3946 mstate, vstate)) {
3947 regs[rd] = NULL;
3948 break;
3949 }
3950
3951 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3952 regs[rd] = _RW_WRITE_HELD(&r.ri);
3953 break;
3954
3955 case DIF_SUBR_RW_ISWRITER:
3956 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3957 mstate, vstate)) {
3958 regs[rd] = NULL;
3959 break;
3960 }
3961
3962 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3963 regs[rd] = _RW_ISWRITER(&r.ri);
3964 break;
3965
3966 case DIF_SUBR_BCOPY: {
3967 /*
3968 * We need to be sure that the destination is in the scratch
3969 * region -- no other region is allowed.
3970 */
3971 uintptr_t src = tupregs[0].dttk_value;
3972 uintptr_t dest = tupregs[1].dttk_value;
3973 size_t size = tupregs[2].dttk_value;
3974
3975 if (!dtrace_inscratch(dest, size, mstate)) {
3976 *flags |= CPU_DTRACE_BADADDR;
3977 *illval = regs[rd];
3978 break;
3979 }
3980
3981 if (!dtrace_canload(src, size, mstate, vstate)) {
3982 regs[rd] = NULL;
3983 break;
3984 }
3985
3986 dtrace_bcopy((void *)src, (void *)dest, size);
3987 break;
3988 }
3989
3990 case DIF_SUBR_ALLOCA:
3991 case DIF_SUBR_COPYIN: {
3992 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3993 uint64_t size =
3994 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3995 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3996
3997 /*
3998 * This action doesn't require any credential checks since
3999 * probes will not activate in user contexts to which the
4000 * enabling user does not have permissions.
4001 */
4002
4003 /*
4004 * Rounding up the user allocation size could have overflowed
4005 * a large, bogus allocation (like -1ULL) to 0.
4006 */
4007 if (scratch_size < size ||
4008 !DTRACE_INSCRATCH(mstate, scratch_size)) {
4009 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4010 regs[rd] = NULL;
4011 break;
4012 }
4013
4014 if (subr == DIF_SUBR_COPYIN) {
4015 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4016 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4017 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4018 }
4019
4020 mstate->dtms_scratch_ptr += scratch_size;
4021 regs[rd] = dest;
4022 break;
4023 }
4024
4025 case DIF_SUBR_COPYINTO: {
4026 uint64_t size = tupregs[1].dttk_value;
4027 uintptr_t dest = tupregs[2].dttk_value;
4028
4029 /*
4030 * This action doesn't require any credential checks since
4031 * probes will not activate in user contexts to which the
4032 * enabling user does not have permissions.
4033 */
4034 if (!dtrace_inscratch(dest, size, mstate)) {
4035 *flags |= CPU_DTRACE_BADADDR;
4036 *illval = regs[rd];
4037 break;
4038 }
4039
4040 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4041 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4042 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4043 break;
4044 }
4045
4046 case DIF_SUBR_COPYINSTR: {
4047 uintptr_t dest = mstate->dtms_scratch_ptr;
4048 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4049
4050 if (nargs > 1 && tupregs[1].dttk_value < size)
4051 size = tupregs[1].dttk_value + 1;
4052
4053 /*
4054 * This action doesn't require any credential checks since
4055 * probes will not activate in user contexts to which the
4056 * enabling user does not have permissions.
4057 */
4058 if (!DTRACE_INSCRATCH(mstate, size)) {
4059 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4060 regs[rd] = NULL;
4061 break;
4062 }
4063
4064 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4065 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4066 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4067
4068 ((char *)dest)[size - 1] = '\0';
4069 mstate->dtms_scratch_ptr += size;
4070 regs[rd] = dest;
4071 break;
4072 }
4073
4074 case DIF_SUBR_MSGSIZE:
4075 case DIF_SUBR_MSGDSIZE: {
4076 uintptr_t baddr = tupregs[0].dttk_value, daddr;
4077 uintptr_t wptr, rptr;
4078 size_t count = 0;
4079 int cont = 0;
4080
4081 while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4082
4083 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4084 vstate)) {
4085 regs[rd] = NULL;
4086 break;
4087 }
4088
4089 wptr = dtrace_loadptr(baddr +
4090 offsetof(mblk_t, b_wptr));
4091
4092 rptr = dtrace_loadptr(baddr +
4093 offsetof(mblk_t, b_rptr));
4094
4095 if (wptr < rptr) {
4096 *flags |= CPU_DTRACE_BADADDR;
4097 *illval = tupregs[0].dttk_value;
4098 break;
4099 }
4100
4101 daddr = dtrace_loadptr(baddr +
4102 offsetof(mblk_t, b_datap));
4103
4104 baddr = dtrace_loadptr(baddr +
4105 offsetof(mblk_t, b_cont));
4106
4107 /*
4108 * We want to prevent against denial-of-service here,
4109 * so we're only going to search the list for
4110 * dtrace_msgdsize_max mblks.
4111 */
4112 if (cont++ > dtrace_msgdsize_max) {
4113 *flags |= CPU_DTRACE_ILLOP;
4114 break;
4115 }
4116
4117 if (subr == DIF_SUBR_MSGDSIZE) {
4118 if (dtrace_load8(daddr +
4119 offsetof(dblk_t, db_type)) != M_DATA)
4120 continue;
4121 }
4122
4123 count += wptr - rptr;
4124 }
4125
4126 if (!(*flags & CPU_DTRACE_FAULT))
4127 regs[rd] = count;
4128
4129 break;
4130 }
4131
4132 case DIF_SUBR_PROGENYOF: {
4133 pid_t pid = tupregs[0].dttk_value;
4134 proc_t *p;
4135 int rval = 0;
4136
4137 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4138
4139 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4140 if (p->p_pidp->pid_id == pid) {
4141 rval = 1;
4142 break;
4143 }
4144 }
4145
4146 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4147
4148 regs[rd] = rval;
4149 break;
4150 }
4151
4152 case DIF_SUBR_SPECULATION:
4153 regs[rd] = dtrace_speculation(state);
4154 break;
4155
4156 case DIF_SUBR_COPYOUT: {
4157 uintptr_t kaddr = tupregs[0].dttk_value;
4158 uintptr_t uaddr = tupregs[1].dttk_value;
4159 uint64_t size = tupregs[2].dttk_value;
4160
4161 if (!dtrace_destructive_disallow &&
4162 dtrace_priv_proc_control(state, mstate) &&
4163 !dtrace_istoxic(kaddr, size)) {
4164 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4165 dtrace_copyout(kaddr, uaddr, size, flags);
4166 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4167 }
4168 break;
4169 }
4170
4171 case DIF_SUBR_COPYOUTSTR: {
4172 uintptr_t kaddr = tupregs[0].dttk_value;
4173 uintptr_t uaddr = tupregs[1].dttk_value;
4174 uint64_t size = tupregs[2].dttk_value;
4175
4176 if (!dtrace_destructive_disallow &&
4177 dtrace_priv_proc_control(state, mstate) &&
4178 !dtrace_istoxic(kaddr, size)) {
4179 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4180 dtrace_copyoutstr(kaddr, uaddr, size, flags);
4181 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4182 }
4183 break;
4184 }
4185
4186 case DIF_SUBR_STRLEN: {
4187 size_t sz;
4188 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4189 sz = dtrace_strlen((char *)addr,
4190 state->dts_options[DTRACEOPT_STRSIZE]);
4191
4192 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
4193 regs[rd] = NULL;
4194 break;
4195 }
4196
4197 regs[rd] = sz;
4198
4199 break;
4200 }
4201
4202 case DIF_SUBR_STRCHR:
4203 case DIF_SUBR_STRRCHR: {
4204 /*
4205 * We're going to iterate over the string looking for the
4206 * specified character. We will iterate until we have reached
4207 * the string length or we have found the character. If this
4208 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4209 * of the specified character instead of the first.
4210 */
4211 uintptr_t saddr = tupregs[0].dttk_value;
4212 uintptr_t addr = tupregs[0].dttk_value;
4213 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
4214 char c, target = (char)tupregs[1].dttk_value;
4215
4216 for (regs[rd] = NULL; addr < limit; addr++) {
4217 if ((c = dtrace_load8(addr)) == target) {
4218 regs[rd] = addr;
4219
4220 if (subr == DIF_SUBR_STRCHR)
4221 break;
4222 }
4223
4224 if (c == '\0')
4225 break;
4226 }
4227
4228 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
4229 regs[rd] = NULL;
4230 break;
4231 }
4232
4233 break;
4234 }
4235
4236 case DIF_SUBR_STRSTR:
4237 case DIF_SUBR_INDEX:
4238 case DIF_SUBR_RINDEX: {
4239 /*
4240 * We're going to iterate over the string looking for the
4241 * specified string. We will iterate until we have reached
4242 * the string length or we have found the string. (Yes, this
4243 * is done in the most naive way possible -- but considering
4244 * that the string we're searching for is likely to be
4245 * relatively short, the complexity of Rabin-Karp or similar
4246 * hardly seems merited.)
4247 */
4248 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4249 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4250 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4251 size_t len = dtrace_strlen(addr, size);
4252 size_t sublen = dtrace_strlen(substr, size);
4253 char *limit = addr + len, *orig = addr;
4254 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4255 int inc = 1;
4256
4257 regs[rd] = notfound;
4258
4259 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4260 regs[rd] = NULL;
4261 break;
4262 }
4263
4264 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4265 vstate)) {
4266 regs[rd] = NULL;
4267 break;
4268 }
4269
4270 /*
4271 * strstr() and index()/rindex() have similar semantics if
4272 * both strings are the empty string: strstr() returns a
4273 * pointer to the (empty) string, and index() and rindex()
4274 * both return index 0 (regardless of any position argument).
4275 */
4276 if (sublen == 0 && len == 0) {
4277 if (subr == DIF_SUBR_STRSTR)
4278 regs[rd] = (uintptr_t)addr;
4279 else
4280 regs[rd] = 0;
4281 break;
4282 }
4283
4284 if (subr != DIF_SUBR_STRSTR) {
4285 if (subr == DIF_SUBR_RINDEX) {
4286 limit = orig - 1;
4287 addr += len;
4288 inc = -1;
4289 }
4290
4291 /*
4292 * Both index() and rindex() take an optional position
4293 * argument that denotes the starting position.
4294 */
4295 if (nargs == 3) {
4296 int64_t pos = (int64_t)tupregs[2].dttk_value;
4297
4298 /*
4299 * If the position argument to index() is
4300 * negative, Perl implicitly clamps it at
4301 * zero. This semantic is a little surprising
4302 * given the special meaning of negative
4303 * positions to similar Perl functions like
4304 * substr(), but it appears to reflect a
4305 * notion that index() can start from a
4306 * negative index and increment its way up to
4307 * the string. Given this notion, Perl's
4308 * rindex() is at least self-consistent in
4309 * that it implicitly clamps positions greater
4310 * than the string length to be the string
4311 * length. Where Perl completely loses
4312 * coherence, however, is when the specified
4313 * substring is the empty string (""). In
4314 * this case, even if the position is
4315 * negative, rindex() returns 0 -- and even if
4316 * the position is greater than the length,
4317 * index() returns the string length. These
4318 * semantics violate the notion that index()
4319 * should never return a value less than the
4320 * specified position and that rindex() should
4321 * never return a value greater than the
4322 * specified position. (One assumes that
4323 * these semantics are artifacts of Perl's
4324 * implementation and not the results of
4325 * deliberate design -- it beggars belief that
4326 * even Larry Wall could desire such oddness.)
4327 * While in the abstract one would wish for
4328 * consistent position semantics across
4329 * substr(), index() and rindex() -- or at the
4330 * very least self-consistent position
4331 * semantics for index() and rindex() -- we
4332 * instead opt to keep with the extant Perl
4333 * semantics, in all their broken glory. (Do
4334 * we have more desire to maintain Perl's
4335 * semantics than Perl does? Probably.)
4336 */
4337 if (subr == DIF_SUBR_RINDEX) {
4338 if (pos < 0) {
4339 if (sublen == 0)
4340 regs[rd] = 0;
4341 break;
4342 }
4343
4344 if (pos > len)
4345 pos = len;
4346 } else {
4347 if (pos < 0)
4348 pos = 0;
4349
4350 if (pos >= len) {
4351 if (sublen == 0)
4352 regs[rd] = len;
4353 break;
4354 }
4355 }
4356
4357 addr = orig + pos;
4358 }
4359 }
4360
4361 for (regs[rd] = notfound; addr != limit; addr += inc) {
4362 if (dtrace_strncmp(addr, substr, sublen) == 0) {
4363 if (subr != DIF_SUBR_STRSTR) {
4364 /*
4365 * As D index() and rindex() are
4366 * modeled on Perl (and not on awk),
4367 * we return a zero-based (and not a
4368 * one-based) index. (For you Perl
4369 * weenies: no, we're not going to add
4370 * $[ -- and shouldn't you be at a con
4371 * or something?)
4372 */
4373 regs[rd] = (uintptr_t)(addr - orig);
4374 break;
4375 }
4376
4377 ASSERT(subr == DIF_SUBR_STRSTR);
4378 regs[rd] = (uintptr_t)addr;
4379 break;
4380 }
4381 }
4382
4383 break;
4384 }
4385
4386 case DIF_SUBR_STRTOK: {
4387 uintptr_t addr = tupregs[0].dttk_value;
4388 uintptr_t tokaddr = tupregs[1].dttk_value;
4389 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4390 uintptr_t limit, toklimit = tokaddr + size;
4391 uint8_t c, tokmap[32]; /* 256 / 8 */
4392 char *dest = (char *)mstate->dtms_scratch_ptr;
4393 int i;
4394
4395 /*
4396 * Check both the token buffer and (later) the input buffer,
4397 * since both could be non-scratch addresses.
4398 */
4399 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
4400 regs[rd] = NULL;
4401 break;
4402 }
4403
4404 if (!DTRACE_INSCRATCH(mstate, size)) {
4405 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4406 regs[rd] = NULL;
4407 break;
4408 }
4409
4410 if (addr == NULL) {
4411 /*
4412 * If the address specified is NULL, we use our saved
4413 * strtok pointer from the mstate. Note that this
4414 * means that the saved strtok pointer is _only_
4415 * valid within multiple enablings of the same probe --
4416 * it behaves like an implicit clause-local variable.
4417 */
4418 addr = mstate->dtms_strtok;
4419 } else {
4420 /*
4421 * If the user-specified address is non-NULL we must
4422 * access check it. This is the only time we have
4423 * a chance to do so, since this address may reside
4424 * in the string table of this clause-- future calls
4425 * (when we fetch addr from mstate->dtms_strtok)
4426 * would fail this access check.
4427 */
4428 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
4429 regs[rd] = NULL;
4430 break;
4431 }
4432 }
4433
4434 /*
4435 * First, zero the token map, and then process the token
4436 * string -- setting a bit in the map for every character
4437 * found in the token string.
4438 */
4439 for (i = 0; i < sizeof (tokmap); i++)
4440 tokmap[i] = 0;
4441
4442 for (; tokaddr < toklimit; tokaddr++) {
4443 if ((c = dtrace_load8(tokaddr)) == '\0')
4444 break;
4445
4446 ASSERT((c >> 3) < sizeof (tokmap));
4447 tokmap[c >> 3] |= (1 << (c & 0x7));
4448 }
4449
4450 for (limit = addr + size; addr < limit; addr++) {
4451 /*
4452 * We're looking for a character that is _not_ contained
4453 * in the token string.
4454 */
4455 if ((c = dtrace_load8(addr)) == '\0')
4456 break;
4457
4458 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4459 break;
4460 }
4461
4462 if (c == '\0') {
4463 /*
4464 * We reached the end of the string without finding
4465 * any character that was not in the token string.
4466 * We return NULL in this case, and we set the saved
4467 * address to NULL as well.
4468 */
4469 regs[rd] = NULL;
4470 mstate->dtms_strtok = NULL;
4471 break;
4472 }
4473
4474 /*
4475 * From here on, we're copying into the destination string.
4476 */
4477 for (i = 0; addr < limit && i < size - 1; addr++) {
4478 if ((c = dtrace_load8(addr)) == '\0')
4479 break;
4480
4481 if (tokmap[c >> 3] & (1 << (c & 0x7)))
4482 break;
4483
4484 ASSERT(i < size);
4485 dest[i++] = c;
4486 }
4487
4488 ASSERT(i < size);
4489 dest[i] = '\0';
4490 regs[rd] = (uintptr_t)dest;
4491 mstate->dtms_scratch_ptr += size;
4492 mstate->dtms_strtok = addr;
4493 break;
4494 }
4495
4496 case DIF_SUBR_SUBSTR: {
4497 uintptr_t s = tupregs[0].dttk_value;
4498 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4499 char *d = (char *)mstate->dtms_scratch_ptr;
4500 int64_t index = (int64_t)tupregs[1].dttk_value;
4501 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4502 size_t len = dtrace_strlen((char *)s, size);
4503 int64_t i;
4504
4505 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4506 regs[rd] = NULL;
4507 break;
4508 }
4509
4510 if (!DTRACE_INSCRATCH(mstate, size)) {
4511 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4512 regs[rd] = NULL;
4513 break;
4514 }
4515
4516 if (nargs <= 2)
4517 remaining = (int64_t)size;
4518
4519 if (index < 0) {
4520 index += len;
4521
4522 if (index < 0 && index + remaining > 0) {
4523 remaining += index;
4524 index = 0;
4525 }
4526 }
4527
4528 if (index >= len || index < 0) {
4529 remaining = 0;
4530 } else if (remaining < 0) {
4531 remaining += len - index;
4532 } else if (index + remaining > size) {
4533 remaining = size - index;
4534 }
4535
4536 for (i = 0; i < remaining; i++) {
4537 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4538 break;
4539 }
4540
4541 d[i] = '\0';
4542
4543 mstate->dtms_scratch_ptr += size;
4544 regs[rd] = (uintptr_t)d;
4545 break;
4546 }
4547
4548 case DIF_SUBR_JSON: {
4549 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4550 uintptr_t json = tupregs[0].dttk_value;
4551 size_t jsonlen = dtrace_strlen((char *)json, size);
4552 uintptr_t elem = tupregs[1].dttk_value;
4553 size_t elemlen = dtrace_strlen((char *)elem, size);
4554
4555 char *dest = (char *)mstate->dtms_scratch_ptr;
4556 char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
4557 char *ee = elemlist;
4558 int nelems = 1;
4559 uintptr_t cur;
4560
4561 if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
4562 !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
4563 regs[rd] = NULL;
4564 break;
4565 }
4566
4567 if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
4568 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4569 regs[rd] = NULL;
4570 break;
4571 }
4572
4573 /*
4574 * Read the element selector and split it up into a packed list
4575 * of strings.
4576 */
4577 for (cur = elem; cur < elem + elemlen; cur++) {
4578 char cc = dtrace_load8(cur);
4579
4580 if (cur == elem && cc == '[') {
4581 /*
4582 * If the first element selector key is
4583 * actually an array index then ignore the
4584 * bracket.
4585 */
4586 continue;
4587 }
4588
4589 if (cc == ']')
4590 continue;
4591
4592 if (cc == '.' || cc == '[') {
4593 nelems++;
4594 cc = '\0';
4595 }
4596
4597 *ee++ = cc;
4598 }
4599 *ee++ = '\0';
4600
4601 if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
4602 nelems, dest)) != NULL)
4603 mstate->dtms_scratch_ptr += jsonlen + 1;
4604 break;
4605 }
4606
4607 case DIF_SUBR_TOUPPER:
4608 case DIF_SUBR_TOLOWER: {
4609 uintptr_t s = tupregs[0].dttk_value;
4610 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4611 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4612 size_t len = dtrace_strlen((char *)s, size);
4613 char lower, upper, convert;
4614 int64_t i;
4615
4616 if (subr == DIF_SUBR_TOUPPER) {
4617 lower = 'a';
4618 upper = 'z';
4619 convert = 'A';
4620 } else {
4621 lower = 'A';
4622 upper = 'Z';
4623 convert = 'a';
4624 }
4625
4626 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4627 regs[rd] = NULL;
4628 break;
4629 }
4630
4631 if (!DTRACE_INSCRATCH(mstate, size)) {
4632 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4633 regs[rd] = NULL;
4634 break;
4635 }
4636
4637 for (i = 0; i < size - 1; i++) {
4638 if ((c = dtrace_load8(s + i)) == '\0')
4639 break;
4640
4641 if (c >= lower && c <= upper)
4642 c = convert + (c - lower);
4643
4644 dest[i] = c;
4645 }
4646
4647 ASSERT(i < size);
4648 dest[i] = '\0';
4649 regs[rd] = (uintptr_t)dest;
4650 mstate->dtms_scratch_ptr += size;
4651 break;
4652 }
4653
4654 case DIF_SUBR_GETMAJOR:
4655 #ifdef _LP64
4656 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4657 #else
4658 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4659 #endif
4660 break;
4661
4662 case DIF_SUBR_GETMINOR:
4663 #ifdef _LP64
4664 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4665 #else
4666 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4667 #endif
4668 break;
4669
4670 case DIF_SUBR_DDI_PATHNAME: {
4671 /*
4672 * This one is a galactic mess. We are going to roughly
4673 * emulate ddi_pathname(), but it's made more complicated
4674 * by the fact that we (a) want to include the minor name and
4675 * (b) must proceed iteratively instead of recursively.
4676 */
4677 uintptr_t dest = mstate->dtms_scratch_ptr;
4678 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4679 char *start = (char *)dest, *end = start + size - 1;
4680 uintptr_t daddr = tupregs[0].dttk_value;
4681 int64_t minor = (int64_t)tupregs[1].dttk_value;
4682 char *s;
4683 int i, len, depth = 0;
4684
4685 /*
4686 * Due to all the pointer jumping we do and context we must
4687 * rely upon, we just mandate that the user must have kernel
4688 * read privileges to use this routine.
4689 */
4690 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4691 *flags |= CPU_DTRACE_KPRIV;
4692 *illval = daddr;
4693 regs[rd] = NULL;
4694 }
4695
4696 if (!DTRACE_INSCRATCH(mstate, size)) {
4697 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4698 regs[rd] = NULL;
4699 break;
4700 }
4701
4702 *end = '\0';
4703
4704 /*
4705 * We want to have a name for the minor. In order to do this,
4706 * we need to walk the minor list from the devinfo. We want
4707 * to be sure that we don't infinitely walk a circular list,
4708 * so we check for circularity by sending a scout pointer
4709 * ahead two elements for every element that we iterate over;
4710 * if the list is circular, these will ultimately point to the
4711 * same element. You may recognize this little trick as the
4712 * answer to a stupid interview question -- one that always
4713 * seems to be asked by those who had to have it laboriously
4714 * explained to them, and who can't even concisely describe
4715 * the conditions under which one would be forced to resort to
4716 * this technique. Needless to say, those conditions are
4717 * found here -- and probably only here. Is this the only use
4718 * of this infamous trick in shipping, production code? If it
4719 * isn't, it probably should be...
4720 */
4721 if (minor != -1) {
4722 uintptr_t maddr = dtrace_loadptr(daddr +
4723 offsetof(struct dev_info, devi_minor));
4724
4725 uintptr_t next = offsetof(struct ddi_minor_data, next);
4726 uintptr_t name = offsetof(struct ddi_minor_data,
4727 d_minor) + offsetof(struct ddi_minor, name);
4728 uintptr_t dev = offsetof(struct ddi_minor_data,
4729 d_minor) + offsetof(struct ddi_minor, dev);
4730 uintptr_t scout;
4731
4732 if (maddr != NULL)
4733 scout = dtrace_loadptr(maddr + next);
4734
4735 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4736 uint64_t m;
4737 #ifdef _LP64
4738 m = dtrace_load64(maddr + dev) & MAXMIN64;
4739 #else
4740 m = dtrace_load32(maddr + dev) & MAXMIN;
4741 #endif
4742 if (m != minor) {
4743 maddr = dtrace_loadptr(maddr + next);
4744
4745 if (scout == NULL)
4746 continue;
4747
4748 scout = dtrace_loadptr(scout + next);
4749
4750 if (scout == NULL)
4751 continue;
4752
4753 scout = dtrace_loadptr(scout + next);
4754
4755 if (scout == NULL)
4756 continue;
4757
4758 if (scout == maddr) {
4759 *flags |= CPU_DTRACE_ILLOP;
4760 break;
4761 }
4762
4763 continue;
4764 }
4765
4766 /*
4767 * We have the minor data. Now we need to
4768 * copy the minor's name into the end of the
4769 * pathname.
4770 */
4771 s = (char *)dtrace_loadptr(maddr + name);
4772 len = dtrace_strlen(s, size);
4773
4774 if (*flags & CPU_DTRACE_FAULT)
4775 break;
4776
4777 if (len != 0) {
4778 if ((end -= (len + 1)) < start)
4779 break;
4780
4781 *end = ':';
4782 }
4783
4784 for (i = 1; i <= len; i++)
4785 end[i] = dtrace_load8((uintptr_t)s++);
4786 break;
4787 }
4788 }
4789
4790 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4791 ddi_node_state_t devi_state;
4792
4793 devi_state = dtrace_load32(daddr +
4794 offsetof(struct dev_info, devi_node_state));
4795
4796 if (*flags & CPU_DTRACE_FAULT)
4797 break;
4798
4799 if (devi_state >= DS_INITIALIZED) {
4800 s = (char *)dtrace_loadptr(daddr +
4801 offsetof(struct dev_info, devi_addr));
4802 len = dtrace_strlen(s, size);
4803
4804 if (*flags & CPU_DTRACE_FAULT)
4805 break;
4806
4807 if (len != 0) {
4808 if ((end -= (len + 1)) < start)
4809 break;
4810
4811 *end = '@';
4812 }
4813
4814 for (i = 1; i <= len; i++)
4815 end[i] = dtrace_load8((uintptr_t)s++);
4816 }
4817
4818 /*
4819 * Now for the node name...
4820 */
4821 s = (char *)dtrace_loadptr(daddr +
4822 offsetof(struct dev_info, devi_node_name));
4823
4824 daddr = dtrace_loadptr(daddr +
4825 offsetof(struct dev_info, devi_parent));
4826
4827 /*
4828 * If our parent is NULL (that is, if we're the root
4829 * node), we're going to use the special path
4830 * "devices".
4831 */
4832 if (daddr == NULL)
4833 s = "devices";
4834
4835 len = dtrace_strlen(s, size);
4836 if (*flags & CPU_DTRACE_FAULT)
4837 break;
4838
4839 if ((end -= (len + 1)) < start)
4840 break;
4841
4842 for (i = 1; i <= len; i++)
4843 end[i] = dtrace_load8((uintptr_t)s++);
4844 *end = '/';
4845
4846 if (depth++ > dtrace_devdepth_max) {
4847 *flags |= CPU_DTRACE_ILLOP;
4848 break;
4849 }
4850 }
4851
4852 if (end < start)
4853 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4854
4855 if (daddr == NULL) {
4856 regs[rd] = (uintptr_t)end;
4857 mstate->dtms_scratch_ptr += size;
4858 }
4859
4860 break;
4861 }
4862
4863 case DIF_SUBR_STRJOIN: {
4864 char *d = (char *)mstate->dtms_scratch_ptr;
4865 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4866 uintptr_t s1 = tupregs[0].dttk_value;
4867 uintptr_t s2 = tupregs[1].dttk_value;
4868 int i = 0;
4869
4870 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4871 !dtrace_strcanload(s2, size, mstate, vstate)) {
4872 regs[rd] = NULL;
4873 break;
4874 }
4875
4876 if (!DTRACE_INSCRATCH(mstate, size)) {
4877 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4878 regs[rd] = NULL;
4879 break;
4880 }
4881
4882 for (;;) {
4883 if (i >= size) {
4884 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4885 regs[rd] = NULL;
4886 break;
4887 }
4888
4889 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4890 i--;
4891 break;
4892 }
4893 }
4894
4895 for (;;) {
4896 if (i >= size) {
4897 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4898 regs[rd] = NULL;
4899 break;
4900 }
4901
4902 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4903 break;
4904 }
4905
4906 if (i < size) {
4907 mstate->dtms_scratch_ptr += i;
4908 regs[rd] = (uintptr_t)d;
4909 }
4910
4911 break;
4912 }
4913
4914 case DIF_SUBR_STRTOLL: {
4915 uintptr_t s = tupregs[0].dttk_value;
4916 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4917 int base = 10;
4918
4919 if (nargs > 1) {
4920 if ((base = tupregs[1].dttk_value) <= 1 ||
4921 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4922 *flags |= CPU_DTRACE_ILLOP;
4923 break;
4924 }
4925 }
4926
4927 if (!dtrace_strcanload(s, size, mstate, vstate)) {
4928 regs[rd] = INT64_MIN;
4929 break;
4930 }
4931
4932 regs[rd] = dtrace_strtoll((char *)s, base, size);
4933 break;
4934 }
4935
4936 case DIF_SUBR_LLTOSTR: {
4937 int64_t i = (int64_t)tupregs[0].dttk_value;
4938 uint64_t val, digit;
4939 uint64_t size = 65; /* enough room for 2^64 in binary */
4940 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4941 int base = 10;
4942
4943 if (nargs > 1) {
4944 if ((base = tupregs[1].dttk_value) <= 1 ||
4945 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4946 *flags |= CPU_DTRACE_ILLOP;
4947 break;
4948 }
4949 }
4950
4951 val = (base == 10 && i < 0) ? i * -1 : i;
4952
4953 if (!DTRACE_INSCRATCH(mstate, size)) {
4954 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4955 regs[rd] = NULL;
4956 break;
4957 }
4958
4959 for (*end-- = '\0'; val; val /= base) {
4960 if ((digit = val % base) <= '9' - '0') {
4961 *end-- = '0' + digit;
4962 } else {
4963 *end-- = 'a' + (digit - ('9' - '0') - 1);
4964 }
4965 }
4966
4967 if (i == 0 && base == 16)
4968 *end-- = '0';
4969
4970 if (base == 16)
4971 *end-- = 'x';
4972
4973 if (i == 0 || base == 8 || base == 16)
4974 *end-- = '0';
4975
4976 if (i < 0 && base == 10)
4977 *end-- = '-';
4978
4979 regs[rd] = (uintptr_t)end + 1;
4980 mstate->dtms_scratch_ptr += size;
4981 break;
4982 }
4983
4984 case DIF_SUBR_HTONS:
4985 case DIF_SUBR_NTOHS:
4986 #ifdef _BIG_ENDIAN
4987 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4988 #else
4989 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4990 #endif
4991 break;
4992
4993
4994 case DIF_SUBR_HTONL:
4995 case DIF_SUBR_NTOHL:
4996 #ifdef _BIG_ENDIAN
4997 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4998 #else
4999 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5000 #endif
5001 break;
5002
5003
5004 case DIF_SUBR_HTONLL:
5005 case DIF_SUBR_NTOHLL:
5006 #ifdef _BIG_ENDIAN
5007 regs[rd] = (uint64_t)tupregs[0].dttk_value;
5008 #else
5009 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5010 #endif
5011 break;
5012
5013
5014 case DIF_SUBR_DIRNAME:
5015 case DIF_SUBR_BASENAME: {
5016 char *dest = (char *)mstate->dtms_scratch_ptr;
5017 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5018 uintptr_t src = tupregs[0].dttk_value;
5019 int i, j, len = dtrace_strlen((char *)src, size);
5020 int lastbase = -1, firstbase = -1, lastdir = -1;
5021 int start, end;
5022
5023 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5024 regs[rd] = NULL;
5025 break;
5026 }
5027
5028 if (!DTRACE_INSCRATCH(mstate, size)) {
5029 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5030 regs[rd] = NULL;
5031 break;
5032 }
5033
5034 /*
5035 * The basename and dirname for a zero-length string is
5036 * defined to be "."
5037 */
5038 if (len == 0) {
5039 len = 1;
5040 src = (uintptr_t)".";
5041 }
5042
5043 /*
5044 * Start from the back of the string, moving back toward the
5045 * front until we see a character that isn't a slash. That
5046 * character is the last character in the basename.
5047 */
5048 for (i = len - 1; i >= 0; i--) {
5049 if (dtrace_load8(src + i) != '/')
5050 break;
5051 }
5052
5053 if (i >= 0)
5054 lastbase = i;
5055
5056 /*
5057 * Starting from the last character in the basename, move
5058 * towards the front until we find a slash. The character
5059 * that we processed immediately before that is the first
5060 * character in the basename.
5061 */
5062 for (; i >= 0; i--) {
5063 if (dtrace_load8(src + i) == '/')
5064 break;
5065 }
5066
5067 if (i >= 0)
5068 firstbase = i + 1;
5069
5070 /*
5071 * Now keep going until we find a non-slash character. That
5072 * character is the last character in the dirname.
5073 */
5074 for (; i >= 0; i--) {
5075 if (dtrace_load8(src + i) != '/')
5076 break;
5077 }
5078
5079 if (i >= 0)
5080 lastdir = i;
5081
5082 ASSERT(!(lastbase == -1 && firstbase != -1));
5083 ASSERT(!(firstbase == -1 && lastdir != -1));
5084
5085 if (lastbase == -1) {
5086 /*
5087 * We didn't find a non-slash character. We know that
5088 * the length is non-zero, so the whole string must be
5089 * slashes. In either the dirname or the basename
5090 * case, we return '/'.
5091 */
5092 ASSERT(firstbase == -1);
5093 firstbase = lastbase = lastdir = 0;
5094 }
5095
5096 if (firstbase == -1) {
5097 /*
5098 * The entire string consists only of a basename
5099 * component. If we're looking for dirname, we need
5100 * to change our string to be just "."; if we're
5101 * looking for a basename, we'll just set the first
5102 * character of the basename to be 0.
5103 */
5104 if (subr == DIF_SUBR_DIRNAME) {
5105 ASSERT(lastdir == -1);
5106 src = (uintptr_t)".";
5107 lastdir = 0;
5108 } else {
5109 firstbase = 0;
5110 }
5111 }
5112
5113 if (subr == DIF_SUBR_DIRNAME) {
5114 if (lastdir == -1) {
5115 /*
5116 * We know that we have a slash in the name --
5117 * or lastdir would be set to 0, above. And
5118 * because lastdir is -1, we know that this
5119 * slash must be the first character. (That
5120 * is, the full string must be of the form
5121 * "/basename".) In this case, the last
5122 * character of the directory name is 0.
5123 */
5124 lastdir = 0;
5125 }
5126
5127 start = 0;
5128 end = lastdir;
5129 } else {
5130 ASSERT(subr == DIF_SUBR_BASENAME);
5131 ASSERT(firstbase != -1 && lastbase != -1);
5132 start = firstbase;
5133 end = lastbase;
5134 }
5135
5136 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5137 dest[j] = dtrace_load8(src + i);
5138
5139 dest[j] = '\0';
5140 regs[rd] = (uintptr_t)dest;
5141 mstate->dtms_scratch_ptr += size;
5142 break;
5143 }
5144
5145 case DIF_SUBR_GETF: {
5146 uintptr_t fd = tupregs[0].dttk_value;
5147 uf_info_t *finfo = &curthread->t_procp->p_user.u_finfo;
5148 file_t *fp;
5149
5150 if (!dtrace_priv_proc(state, mstate)) {
5151 regs[rd] = NULL;
5152 break;
5153 }
5154
5155 /*
5156 * This is safe because fi_nfiles only increases, and the
5157 * fi_list array is not freed when the array size doubles.
5158 * (See the comment in flist_grow() for details on the
5159 * management of the u_finfo structure.)
5160 */
5161 fp = fd < finfo->fi_nfiles ? finfo->fi_list[fd].uf_file : NULL;
5162
5163 mstate->dtms_getf = fp;
5164 regs[rd] = (uintptr_t)fp;
5165 break;
5166 }
5167
5168 case DIF_SUBR_CLEANPATH: {
5169 char *dest = (char *)mstate->dtms_scratch_ptr, c;
5170 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5171 uintptr_t src = tupregs[0].dttk_value;
5172 int i = 0, j = 0;
5173 zone_t *z;
5174
5175 if (!dtrace_strcanload(src, size, mstate, vstate)) {
5176 regs[rd] = NULL;
5177 break;
5178 }
5179
5180 if (!DTRACE_INSCRATCH(mstate, size)) {
5181 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5182 regs[rd] = NULL;
5183 break;
5184 }
5185
5186 /*
5187 * Move forward, loading each character.
5188 */
5189 do {
5190 c = dtrace_load8(src + i++);
5191 next:
5192 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
5193 break;
5194
5195 if (c != '/') {
5196 dest[j++] = c;
5197 continue;
5198 }
5199
5200 c = dtrace_load8(src + i++);
5201
5202 if (c == '/') {
5203 /*
5204 * We have two slashes -- we can just advance
5205 * to the next character.
5206 */
5207 goto next;
5208 }
5209
5210 if (c != '.') {
5211 /*
5212 * This is not "." and it's not ".." -- we can
5213 * just store the "/" and this character and
5214 * drive on.
5215 */
5216 dest[j++] = '/';
5217 dest[j++] = c;
5218 continue;
5219 }
5220
5221 c = dtrace_load8(src + i++);
5222
5223 if (c == '/') {
5224 /*
5225 * This is a "/./" component. We're not going
5226 * to store anything in the destination buffer;
5227 * we're just going to go to the next component.
5228 */
5229 goto next;
5230 }
5231
5232 if (c != '.') {
5233 /*
5234 * This is not ".." -- we can just store the
5235 * "/." and this character and continue
5236 * processing.
5237 */
5238 dest[j++] = '/';
5239 dest[j++] = '.';
5240 dest[j++] = c;
5241 continue;
5242 }
5243
5244 c = dtrace_load8(src + i++);
5245
5246 if (c != '/' && c != '\0') {
5247 /*
5248 * This is not ".." -- it's "..[mumble]".
5249 * We'll store the "/.." and this character
5250 * and continue processing.
5251 */
5252 dest[j++] = '/';
5253 dest[j++] = '.';
5254 dest[j++] = '.';
5255 dest[j++] = c;
5256 continue;
5257 }
5258
5259 /*
5260 * This is "/../" or "/..\0". We need to back up
5261 * our destination pointer until we find a "/".
5262 */
5263 i--;
5264 while (j != 0 && dest[--j] != '/')
5265 continue;
5266
5267 if (c == '\0')
5268 dest[++j] = '/';
5269 } while (c != '\0');
5270
5271 dest[j] = '\0';
5272
5273 if (mstate->dtms_getf != NULL &&
5274 !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5275 (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5276 /*
5277 * If we've done a getf() as a part of this ECB and we
5278 * don't have kernel access (and we're not in the global
5279 * zone), check if the path we cleaned up begins with
5280 * the zone's root path, and trim it off if so. Note
5281 * that this is an output cleanliness issue, not a
5282 * security issue: knowing one's zone root path does
5283 * not enable privilege escalation.
5284 */
5285 if (strstr(dest, z->zone_rootpath) == dest)
5286 dest += strlen(z->zone_rootpath) - 1;
5287 }
5288
5289 regs[rd] = (uintptr_t)dest;
5290 mstate->dtms_scratch_ptr += size;
5291 break;
5292 }
5293
5294 case DIF_SUBR_INET_NTOA:
5295 case DIF_SUBR_INET_NTOA6:
5296 case DIF_SUBR_INET_NTOP: {
5297 size_t size;
5298 int af, argi, i;
5299 char *base, *end;
5300
5301 if (subr == DIF_SUBR_INET_NTOP) {
5302 af = (int)tupregs[0].dttk_value;
5303 argi = 1;
5304 } else {
5305 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5306 argi = 0;
5307 }
5308
5309 if (af == AF_INET) {
5310 ipaddr_t ip4;
5311 uint8_t *ptr8, val;
5312
5313 /*
5314 * Safely load the IPv4 address.
5315 */
5316 ip4 = dtrace_load32(tupregs[argi].dttk_value);
5317
5318 /*
5319 * Check an IPv4 string will fit in scratch.
5320 */
5321 size = INET_ADDRSTRLEN;
5322 if (!DTRACE_INSCRATCH(mstate, size)) {
5323 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5324 regs[rd] = NULL;
5325 break;
5326 }
5327 base = (char *)mstate->dtms_scratch_ptr;
5328 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5329
5330 /*
5331 * Stringify as a dotted decimal quad.
5332 */
5333 *end-- = '\0';
5334 ptr8 = (uint8_t *)&ip4;
5335 for (i = 3; i >= 0; i--) {
5336 val = ptr8[i];
5337
5338 if (val == 0) {
5339 *end-- = '0';
5340 } else {
5341 for (; val; val /= 10) {
5342 *end-- = '0' + (val % 10);
5343 }
5344 }
5345
5346 if (i > 0)
5347 *end-- = '.';
5348 }
5349 ASSERT(end + 1 >= base);
5350
5351 } else if (af == AF_INET6) {
5352 struct in6_addr ip6;
5353 int firstzero, tryzero, numzero, v6end;
5354 uint16_t val;
5355 const char digits[] = "0123456789abcdef";
5356
5357 /*
5358 * Stringify using RFC 1884 convention 2 - 16 bit
5359 * hexadecimal values with a zero-run compression.
5360 * Lower case hexadecimal digits are used.
5361 * eg, fe80::214:4fff:fe0b:76c8.
5362 * The IPv4 embedded form is returned for inet_ntop,
5363 * just the IPv4 string is returned for inet_ntoa6.
5364 */
5365
5366 /*
5367 * Safely load the IPv6 address.
5368 */
5369 dtrace_bcopy(
5370 (void *)(uintptr_t)tupregs[argi].dttk_value,
5371 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5372
5373 /*
5374 * Check an IPv6 string will fit in scratch.
5375 */
5376 size = INET6_ADDRSTRLEN;
5377 if (!DTRACE_INSCRATCH(mstate, size)) {
5378 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5379 regs[rd] = NULL;
5380 break;
5381 }
5382 base = (char *)mstate->dtms_scratch_ptr;
5383 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5384 *end-- = '\0';
5385
5386 /*
5387 * Find the longest run of 16 bit zero values
5388 * for the single allowed zero compression - "::".
5389 */
5390 firstzero = -1;
5391 tryzero = -1;
5392 numzero = 1;
5393 for (i = 0; i < sizeof (struct in6_addr); i++) {
5394 if (ip6._S6_un._S6_u8[i] == 0 &&
5395 tryzero == -1 && i % 2 == 0) {
5396 tryzero = i;
5397 continue;
5398 }
5399
5400 if (tryzero != -1 &&
5401 (ip6._S6_un._S6_u8[i] != 0 ||
5402 i == sizeof (struct in6_addr) - 1)) {
5403
5404 if (i - tryzero <= numzero) {
5405 tryzero = -1;
5406 continue;
5407 }
5408
5409 firstzero = tryzero;
5410 numzero = i - i % 2 - tryzero;
5411 tryzero = -1;
5412
5413 if (ip6._S6_un._S6_u8[i] == 0 &&
5414 i == sizeof (struct in6_addr) - 1)
5415 numzero += 2;
5416 }
5417 }
5418 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5419
5420 /*
5421 * Check for an IPv4 embedded address.
5422 */
5423 v6end = sizeof (struct in6_addr) - 2;
5424 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5425 IN6_IS_ADDR_V4COMPAT(&ip6)) {
5426 for (i = sizeof (struct in6_addr) - 1;
5427 i >= DTRACE_V4MAPPED_OFFSET; i--) {
5428 ASSERT(end >= base);
5429
5430 val = ip6._S6_un._S6_u8[i];
5431
5432 if (val == 0) {
5433 *end-- = '0';
5434 } else {
5435 for (; val; val /= 10) {
5436 *end-- = '0' + val % 10;
5437 }
5438 }
5439
5440 if (i > DTRACE_V4MAPPED_OFFSET)
5441 *end-- = '.';
5442 }
5443
5444 if (subr == DIF_SUBR_INET_NTOA6)
5445 goto inetout;
5446
5447 /*
5448 * Set v6end to skip the IPv4 address that
5449 * we have already stringified.
5450 */
5451 v6end = 10;
5452 }
5453
5454 /*
5455 * Build the IPv6 string by working through the
5456 * address in reverse.
5457 */
5458 for (i = v6end; i >= 0; i -= 2) {
5459 ASSERT(end >= base);
5460
5461 if (i == firstzero + numzero - 2) {
5462 *end-- = ':';
5463 *end-- = ':';
5464 i -= numzero - 2;
5465 continue;
5466 }
5467
5468 if (i < 14 && i != firstzero - 2)
5469 *end-- = ':';
5470
5471 val = (ip6._S6_un._S6_u8[i] << 8) +
5472 ip6._S6_un._S6_u8[i + 1];
5473
5474 if (val == 0) {
5475 *end-- = '0';
5476 } else {
5477 for (; val; val /= 16) {
5478 *end-- = digits[val % 16];
5479 }
5480 }
5481 }
5482 ASSERT(end + 1 >= base);
5483
5484 } else {
5485 /*
5486 * The user didn't use AH_INET or AH_INET6.
5487 */
5488 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5489 regs[rd] = NULL;
5490 break;
5491 }
5492
5493 inetout: regs[rd] = (uintptr_t)end + 1;
5494 mstate->dtms_scratch_ptr += size;
5495 break;
5496 }
5497
5498 }
5499 }
5500
5501 /*
5502 * Emulate the execution of DTrace IR instructions specified by the given
5503 * DIF object. This function is deliberately void of assertions as all of
5504 * the necessary checks are handled by a call to dtrace_difo_validate().
5505 */
5506 static uint64_t
5507 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
5508 dtrace_vstate_t *vstate, dtrace_state_t *state)
5509 {
5510 const dif_instr_t *text = difo->dtdo_buf;
5511 const uint_t textlen = difo->dtdo_len;
5512 const char *strtab = difo->dtdo_strtab;
5513 const uint64_t *inttab = difo->dtdo_inttab;
5514
5515 uint64_t rval = 0;
5516 dtrace_statvar_t *svar;
5517 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5518 dtrace_difv_t *v;
5519 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5520 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
5521
5522 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5523 uint64_t regs[DIF_DIR_NREGS];
5524 uint64_t *tmp;
5525
5526 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5527 int64_t cc_r;
5528 uint_t pc = 0, id, opc;
5529 uint8_t ttop = 0;
5530 dif_instr_t instr;
5531 uint_t r1, r2, rd;
5532
5533 /*
5534 * We stash the current DIF object into the machine state: we need it
5535 * for subsequent access checking.
5536 */
5537 mstate->dtms_difo = difo;
5538
5539 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
5540
5541 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5542 opc = pc;
5543
5544 instr = text[pc++];
5545 r1 = DIF_INSTR_R1(instr);
5546 r2 = DIF_INSTR_R2(instr);
5547 rd = DIF_INSTR_RD(instr);
5548
5549 switch (DIF_INSTR_OP(instr)) {
5550 case DIF_OP_OR:
5551 regs[rd] = regs[r1] | regs[r2];
5552 break;
5553 case DIF_OP_XOR:
5554 regs[rd] = regs[r1] ^ regs[r2];
5555 break;
5556 case DIF_OP_AND:
5557 regs[rd] = regs[r1] & regs[r2];
5558 break;
5559 case DIF_OP_SLL:
5560 regs[rd] = regs[r1] << regs[r2];
5561 break;
5562 case DIF_OP_SRL:
5563 regs[rd] = regs[r1] >> regs[r2];
5564 break;
5565 case DIF_OP_SUB:
5566 regs[rd] = regs[r1] - regs[r2];
5567 break;
5568 case DIF_OP_ADD:
5569 regs[rd] = regs[r1] + regs[r2];
5570 break;
5571 case DIF_OP_MUL:
5572 regs[rd] = regs[r1] * regs[r2];
5573 break;
5574 case DIF_OP_SDIV:
5575 if (regs[r2] == 0) {
5576 regs[rd] = 0;
5577 *flags |= CPU_DTRACE_DIVZERO;
5578 } else {
5579 regs[rd] = (int64_t)regs[r1] /
5580 (int64_t)regs[r2];
5581 }
5582 break;
5583
5584 case DIF_OP_UDIV:
5585 if (regs[r2] == 0) {
5586 regs[rd] = 0;
5587 *flags |= CPU_DTRACE_DIVZERO;
5588 } else {
5589 regs[rd] = regs[r1] / regs[r2];
5590 }
5591 break;
5592
5593 case DIF_OP_SREM:
5594 if (regs[r2] == 0) {
5595 regs[rd] = 0;
5596 *flags |= CPU_DTRACE_DIVZERO;
5597 } else {
5598 regs[rd] = (int64_t)regs[r1] %
5599 (int64_t)regs[r2];
5600 }
5601 break;
5602
5603 case DIF_OP_UREM:
5604 if (regs[r2] == 0) {
5605 regs[rd] = 0;
5606 *flags |= CPU_DTRACE_DIVZERO;
5607 } else {
5608 regs[rd] = regs[r1] % regs[r2];
5609 }
5610 break;
5611
5612 case DIF_OP_NOT:
5613 regs[rd] = ~regs[r1];
5614 break;
5615 case DIF_OP_MOV:
5616 regs[rd] = regs[r1];
5617 break;
5618 case DIF_OP_CMP:
5619 cc_r = regs[r1] - regs[r2];
5620 cc_n = cc_r < 0;
5621 cc_z = cc_r == 0;
5622 cc_v = 0;
5623 cc_c = regs[r1] < regs[r2];
5624 break;
5625 case DIF_OP_TST:
5626 cc_n = cc_v = cc_c = 0;
5627 cc_z = regs[r1] == 0;
5628 break;
5629 case DIF_OP_BA:
5630 pc = DIF_INSTR_LABEL(instr);
5631 break;
5632 case DIF_OP_BE:
5633 if (cc_z)
5634 pc = DIF_INSTR_LABEL(instr);
5635 break;
5636 case DIF_OP_BNE:
5637 if (cc_z == 0)
5638 pc = DIF_INSTR_LABEL(instr);
5639 break;
5640 case DIF_OP_BG:
5641 if ((cc_z | (cc_n ^ cc_v)) == 0)
5642 pc = DIF_INSTR_LABEL(instr);
5643 break;
5644 case DIF_OP_BGU:
5645 if ((cc_c | cc_z) == 0)
5646 pc = DIF_INSTR_LABEL(instr);
5647 break;
5648 case DIF_OP_BGE:
5649 if ((cc_n ^ cc_v) == 0)
5650 pc = DIF_INSTR_LABEL(instr);
5651 break;
5652 case DIF_OP_BGEU:
5653 if (cc_c == 0)
5654 pc = DIF_INSTR_LABEL(instr);
5655 break;
5656 case DIF_OP_BL:
5657 if (cc_n ^ cc_v)
5658 pc = DIF_INSTR_LABEL(instr);
5659 break;
5660 case DIF_OP_BLU:
5661 if (cc_c)
5662 pc = DIF_INSTR_LABEL(instr);
5663 break;
5664 case DIF_OP_BLE:
5665 if (cc_z | (cc_n ^ cc_v))
5666 pc = DIF_INSTR_LABEL(instr);
5667 break;
5668 case DIF_OP_BLEU:
5669 if (cc_c | cc_z)
5670 pc = DIF_INSTR_LABEL(instr);
5671 break;
5672 case DIF_OP_RLDSB:
5673 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5674 break;
5675 /*FALLTHROUGH*/
5676 case DIF_OP_LDSB:
5677 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5678 break;
5679 case DIF_OP_RLDSH:
5680 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5681 break;
5682 /*FALLTHROUGH*/
5683 case DIF_OP_LDSH:
5684 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5685 break;
5686 case DIF_OP_RLDSW:
5687 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5688 break;
5689 /*FALLTHROUGH*/
5690 case DIF_OP_LDSW:
5691 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5692 break;
5693 case DIF_OP_RLDUB:
5694 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5695 break;
5696 /*FALLTHROUGH*/
5697 case DIF_OP_LDUB:
5698 regs[rd] = dtrace_load8(regs[r1]);
5699 break;
5700 case DIF_OP_RLDUH:
5701 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5702 break;
5703 /*FALLTHROUGH*/
5704 case DIF_OP_LDUH:
5705 regs[rd] = dtrace_load16(regs[r1]);
5706 break;
5707 case DIF_OP_RLDUW:
5708 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5709 break;
5710 /*FALLTHROUGH*/
5711 case DIF_OP_LDUW:
5712 regs[rd] = dtrace_load32(regs[r1]);
5713 break;
5714 case DIF_OP_RLDX:
5715 if (!dtrace_canload(regs[r1], 8, mstate, vstate))
5716 break;
5717 /*FALLTHROUGH*/
5718 case DIF_OP_LDX:
5719 regs[rd] = dtrace_load64(regs[r1]);
5720 break;
5721 case DIF_OP_ULDSB:
5722 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5723 regs[rd] = (int8_t)
5724 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5725 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5726 break;
5727 case DIF_OP_ULDSH:
5728 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5729 regs[rd] = (int16_t)
5730 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5731 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5732 break;
5733 case DIF_OP_ULDSW:
5734 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5735 regs[rd] = (int32_t)
5736 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5737 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5738 break;
5739 case DIF_OP_ULDUB:
5740 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5741 regs[rd] =
5742 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5743 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5744 break;
5745 case DIF_OP_ULDUH:
5746 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5747 regs[rd] =
5748 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5749 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5750 break;
5751 case DIF_OP_ULDUW:
5752 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5753 regs[rd] =
5754 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5755 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5756 break;
5757 case DIF_OP_ULDX:
5758 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5759 regs[rd] =
5760 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5761 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5762 break;
5763 case DIF_OP_RET:
5764 rval = regs[rd];
5765 pc = textlen;
5766 break;
5767 case DIF_OP_NOP:
5768 break;
5769 case DIF_OP_SETX:
5770 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5771 break;
5772 case DIF_OP_SETS:
5773 regs[rd] = (uint64_t)(uintptr_t)
5774 (strtab + DIF_INSTR_STRING(instr));
5775 break;
5776 case DIF_OP_SCMP: {
5777 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5778 uintptr_t s1 = regs[r1];
5779 uintptr_t s2 = regs[r2];
5780
5781 if (s1 != NULL &&
5782 !dtrace_strcanload(s1, sz, mstate, vstate))
5783 break;
5784 if (s2 != NULL &&
5785 !dtrace_strcanload(s2, sz, mstate, vstate))
5786 break;
5787
5788 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5789
5790 cc_n = cc_r < 0;
5791 cc_z = cc_r == 0;
5792 cc_v = cc_c = 0;
5793 break;
5794 }
5795 case DIF_OP_LDGA:
5796 regs[rd] = dtrace_dif_variable(mstate, state,
5797 r1, regs[r2]);
5798 break;
5799 case DIF_OP_LDGS:
5800 id = DIF_INSTR_VAR(instr);
5801
5802 if (id >= DIF_VAR_OTHER_UBASE) {
5803 uintptr_t a;
5804
5805 id -= DIF_VAR_OTHER_UBASE;
5806 svar = vstate->dtvs_globals[id];
5807 ASSERT(svar != NULL);
5808 v = &svar->dtsv_var;
5809
5810 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5811 regs[rd] = svar->dtsv_data;
5812 break;
5813 }
5814
5815 a = (uintptr_t)svar->dtsv_data;
5816
5817 if (*(uint8_t *)a == UINT8_MAX) {
5818 /*
5819 * If the 0th byte is set to UINT8_MAX
5820 * then this is to be treated as a
5821 * reference to a NULL variable.
5822 */
5823 regs[rd] = NULL;
5824 } else {
5825 regs[rd] = a + sizeof (uint64_t);
5826 }
5827
5828 break;
5829 }
5830
5831 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5832 break;
5833
5834 case DIF_OP_STGS:
5835 id = DIF_INSTR_VAR(instr);
5836
5837 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5838 id -= DIF_VAR_OTHER_UBASE;
5839
5840 svar = vstate->dtvs_globals[id];
5841 ASSERT(svar != NULL);
5842 v = &svar->dtsv_var;
5843
5844 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5845 uintptr_t a = (uintptr_t)svar->dtsv_data;
5846
5847 ASSERT(a != NULL);
5848 ASSERT(svar->dtsv_size != 0);
5849
5850 if (regs[rd] == NULL) {
5851 *(uint8_t *)a = UINT8_MAX;
5852 break;
5853 } else {
5854 *(uint8_t *)a = 0;
5855 a += sizeof (uint64_t);
5856 }
5857 if (!dtrace_vcanload(
5858 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5859 mstate, vstate))
5860 break;
5861
5862 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5863 (void *)a, &v->dtdv_type);
5864 break;
5865 }
5866
5867 svar->dtsv_data = regs[rd];
5868 break;
5869
5870 case DIF_OP_LDTA:
5871 /*
5872 * There are no DTrace built-in thread-local arrays at
5873 * present. This opcode is saved for future work.
5874 */
5875 *flags |= CPU_DTRACE_ILLOP;
5876 regs[rd] = 0;
5877 break;
5878
5879 case DIF_OP_LDLS:
5880 id = DIF_INSTR_VAR(instr);
5881
5882 if (id < DIF_VAR_OTHER_UBASE) {
5883 /*
5884 * For now, this has no meaning.
5885 */
5886 regs[rd] = 0;
5887 break;
5888 }
5889
5890 id -= DIF_VAR_OTHER_UBASE;
5891
5892 ASSERT(id < vstate->dtvs_nlocals);
5893 ASSERT(vstate->dtvs_locals != NULL);
5894
5895 svar = vstate->dtvs_locals[id];
5896 ASSERT(svar != NULL);
5897 v = &svar->dtsv_var;
5898
5899 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5900 uintptr_t a = (uintptr_t)svar->dtsv_data;
5901 size_t sz = v->dtdv_type.dtdt_size;
5902
5903 sz += sizeof (uint64_t);
5904 ASSERT(svar->dtsv_size == NCPU * sz);
5905 a += CPU->cpu_id * sz;
5906
5907 if (*(uint8_t *)a == UINT8_MAX) {
5908 /*
5909 * If the 0th byte is set to UINT8_MAX
5910 * then this is to be treated as a
5911 * reference to a NULL variable.
5912 */
5913 regs[rd] = NULL;
5914 } else {
5915 regs[rd] = a + sizeof (uint64_t);
5916 }
5917
5918 break;
5919 }
5920
5921 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5922 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5923 regs[rd] = tmp[CPU->cpu_id];
5924 break;
5925
5926 case DIF_OP_STLS:
5927 id = DIF_INSTR_VAR(instr);
5928
5929 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5930 id -= DIF_VAR_OTHER_UBASE;
5931 ASSERT(id < vstate->dtvs_nlocals);
5932
5933 ASSERT(vstate->dtvs_locals != NULL);
5934 svar = vstate->dtvs_locals[id];
5935 ASSERT(svar != NULL);
5936 v = &svar->dtsv_var;
5937
5938 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5939 uintptr_t a = (uintptr_t)svar->dtsv_data;
5940 size_t sz = v->dtdv_type.dtdt_size;
5941
5942 sz += sizeof (uint64_t);
5943 ASSERT(svar->dtsv_size == NCPU * sz);
5944 a += CPU->cpu_id * sz;
5945
5946 if (regs[rd] == NULL) {
5947 *(uint8_t *)a = UINT8_MAX;
5948 break;
5949 } else {
5950 *(uint8_t *)a = 0;
5951 a += sizeof (uint64_t);
5952 }
5953
5954 if (!dtrace_vcanload(
5955 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5956 mstate, vstate))
5957 break;
5958
5959 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5960 (void *)a, &v->dtdv_type);
5961 break;
5962 }
5963
5964 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5965 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5966 tmp[CPU->cpu_id] = regs[rd];
5967 break;
5968
5969 case DIF_OP_LDTS: {
5970 dtrace_dynvar_t *dvar;
5971 dtrace_key_t *key;
5972
5973 id = DIF_INSTR_VAR(instr);
5974 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5975 id -= DIF_VAR_OTHER_UBASE;
5976 v = &vstate->dtvs_tlocals[id];
5977
5978 key = &tupregs[DIF_DTR_NREGS];
5979 key[0].dttk_value = (uint64_t)id;
5980 key[0].dttk_size = 0;
5981 DTRACE_TLS_THRKEY(key[1].dttk_value);
5982 key[1].dttk_size = 0;
5983
5984 dvar = dtrace_dynvar(dstate, 2, key,
5985 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5986 mstate, vstate);
5987
5988 if (dvar == NULL) {
5989 regs[rd] = 0;
5990 break;
5991 }
5992
5993 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5994 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5995 } else {
5996 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5997 }
5998
5999 break;
6000 }
6001
6002 case DIF_OP_STTS: {
6003 dtrace_dynvar_t *dvar;
6004 dtrace_key_t *key;
6005
6006 id = DIF_INSTR_VAR(instr);
6007 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6008 id -= DIF_VAR_OTHER_UBASE;
6009
6010 key = &tupregs[DIF_DTR_NREGS];
6011 key[0].dttk_value = (uint64_t)id;
6012 key[0].dttk_size = 0;
6013 DTRACE_TLS_THRKEY(key[1].dttk_value);
6014 key[1].dttk_size = 0;
6015 v = &vstate->dtvs_tlocals[id];
6016
6017 dvar = dtrace_dynvar(dstate, 2, key,
6018 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6019 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6020 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6021 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6022
6023 /*
6024 * Given that we're storing to thread-local data,
6025 * we need to flush our predicate cache.
6026 */
6027 curthread->t_predcache = NULL;
6028
6029 if (dvar == NULL)
6030 break;
6031
6032 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6033 if (!dtrace_vcanload(
6034 (void *)(uintptr_t)regs[rd],
6035 &v->dtdv_type, mstate, vstate))
6036 break;
6037
6038 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6039 dvar->dtdv_data, &v->dtdv_type);
6040 } else {
6041 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6042 }
6043
6044 break;
6045 }
6046
6047 case DIF_OP_SRA:
6048 regs[rd] = (int64_t)regs[r1] >> regs[r2];
6049 break;
6050
6051 case DIF_OP_CALL:
6052 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6053 regs, tupregs, ttop, mstate, state);
6054 break;
6055
6056 case DIF_OP_PUSHTR:
6057 if (ttop == DIF_DTR_NREGS) {
6058 *flags |= CPU_DTRACE_TUPOFLOW;
6059 break;
6060 }
6061
6062 if (r1 == DIF_TYPE_STRING) {
6063 /*
6064 * If this is a string type and the size is 0,
6065 * we'll use the system-wide default string
6066 * size. Note that we are _not_ looking at
6067 * the value of the DTRACEOPT_STRSIZE option;
6068 * had this been set, we would expect to have
6069 * a non-zero size value in the "pushtr".
6070 */
6071 tupregs[ttop].dttk_size =
6072 dtrace_strlen((char *)(uintptr_t)regs[rd],
6073 regs[r2] ? regs[r2] :
6074 dtrace_strsize_default) + 1;
6075 } else {
6076 tupregs[ttop].dttk_size = regs[r2];
6077 }
6078
6079 tupregs[ttop++].dttk_value = regs[rd];
6080 break;
6081
6082 case DIF_OP_PUSHTV:
6083 if (ttop == DIF_DTR_NREGS) {
6084 *flags |= CPU_DTRACE_TUPOFLOW;
6085 break;
6086 }
6087
6088 tupregs[ttop].dttk_value = regs[rd];
6089 tupregs[ttop++].dttk_size = 0;
6090 break;
6091
6092 case DIF_OP_POPTS:
6093 if (ttop != 0)
6094 ttop--;
6095 break;
6096
6097 case DIF_OP_FLUSHTS:
6098 ttop = 0;
6099 break;
6100
6101 case DIF_OP_LDGAA:
6102 case DIF_OP_LDTAA: {
6103 dtrace_dynvar_t *dvar;
6104 dtrace_key_t *key = tupregs;
6105 uint_t nkeys = ttop;
6106
6107 id = DIF_INSTR_VAR(instr);
6108 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6109 id -= DIF_VAR_OTHER_UBASE;
6110
6111 key[nkeys].dttk_value = (uint64_t)id;
6112 key[nkeys++].dttk_size = 0;
6113
6114 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6115 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6116 key[nkeys++].dttk_size = 0;
6117 v = &vstate->dtvs_tlocals[id];
6118 } else {
6119 v = &vstate->dtvs_globals[id]->dtsv_var;
6120 }
6121
6122 dvar = dtrace_dynvar(dstate, nkeys, key,
6123 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6124 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6125 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6126
6127 if (dvar == NULL) {
6128 regs[rd] = 0;
6129 break;
6130 }
6131
6132 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6133 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6134 } else {
6135 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6136 }
6137
6138 break;
6139 }
6140
6141 case DIF_OP_STGAA:
6142 case DIF_OP_STTAA: {
6143 dtrace_dynvar_t *dvar;
6144 dtrace_key_t *key = tupregs;
6145 uint_t nkeys = ttop;
6146
6147 id = DIF_INSTR_VAR(instr);
6148 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6149 id -= DIF_VAR_OTHER_UBASE;
6150
6151 key[nkeys].dttk_value = (uint64_t)id;
6152 key[nkeys++].dttk_size = 0;
6153
6154 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6155 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6156 key[nkeys++].dttk_size = 0;
6157 v = &vstate->dtvs_tlocals[id];
6158 } else {
6159 v = &vstate->dtvs_globals[id]->dtsv_var;
6160 }
6161
6162 dvar = dtrace_dynvar(dstate, nkeys, key,
6163 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6164 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6165 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6166 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6167
6168 if (dvar == NULL)
6169 break;
6170
6171 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6172 if (!dtrace_vcanload(
6173 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6174 mstate, vstate))
6175 break;
6176
6177 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6178 dvar->dtdv_data, &v->dtdv_type);
6179 } else {
6180 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6181 }
6182
6183 break;
6184 }
6185
6186 case DIF_OP_ALLOCS: {
6187 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6188 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6189
6190 /*
6191 * Rounding up the user allocation size could have
6192 * overflowed large, bogus allocations (like -1ULL) to
6193 * 0.
6194 */
6195 if (size < regs[r1] ||
6196 !DTRACE_INSCRATCH(mstate, size)) {
6197 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6198 regs[rd] = NULL;
6199 break;
6200 }
6201
6202 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6203 mstate->dtms_scratch_ptr += size;
6204 regs[rd] = ptr;
6205 break;
6206 }
6207
6208 case DIF_OP_COPYS:
6209 if (!dtrace_canstore(regs[rd], regs[r2],
6210 mstate, vstate)) {
6211 *flags |= CPU_DTRACE_BADADDR;
6212 *illval = regs[rd];
6213 break;
6214 }
6215
6216 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6217 break;
6218
6219 dtrace_bcopy((void *)(uintptr_t)regs[r1],
6220 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6221 break;
6222
6223 case DIF_OP_STB:
6224 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6225 *flags |= CPU_DTRACE_BADADDR;
6226 *illval = regs[rd];
6227 break;
6228 }
6229 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6230 break;
6231
6232 case DIF_OP_STH:
6233 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6234 *flags |= CPU_DTRACE_BADADDR;
6235 *illval = regs[rd];
6236 break;
6237 }
6238 if (regs[rd] & 1) {
6239 *flags |= CPU_DTRACE_BADALIGN;
6240 *illval = regs[rd];
6241 break;
6242 }
6243 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6244 break;
6245
6246 case DIF_OP_STW:
6247 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6248 *flags |= CPU_DTRACE_BADADDR;
6249 *illval = regs[rd];
6250 break;
6251 }
6252 if (regs[rd] & 3) {
6253 *flags |= CPU_DTRACE_BADALIGN;
6254 *illval = regs[rd];
6255 break;
6256 }
6257 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6258 break;
6259
6260 case DIF_OP_STX:
6261 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6262 *flags |= CPU_DTRACE_BADADDR;
6263 *illval = regs[rd];
6264 break;
6265 }
6266 if (regs[rd] & 7) {
6267 *flags |= CPU_DTRACE_BADALIGN;
6268 *illval = regs[rd];
6269 break;
6270 }
6271 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6272 break;
6273 }
6274 }
6275
6276 if (!(*flags & CPU_DTRACE_FAULT))
6277 return (rval);
6278
6279 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6280 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6281
6282 return (0);
6283 }
6284
6285 static void
6286 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6287 {
6288 dtrace_probe_t *probe = ecb->dte_probe;
6289 dtrace_provider_t *prov = probe->dtpr_provider;
6290 char c[DTRACE_FULLNAMELEN + 80], *str;
6291 char *msg = "dtrace: breakpoint action at probe ";
6292 char *ecbmsg = " (ecb ";
6293 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6294 uintptr_t val = (uintptr_t)ecb;
6295 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6296
6297 if (dtrace_destructive_disallow)
6298 return;
6299
6300 /*
6301 * It's impossible to be taking action on the NULL probe.
6302 */
6303 ASSERT(probe != NULL);
6304
6305 /*
6306 * This is a poor man's (destitute man's?) sprintf(): we want to
6307 * print the provider name, module name, function name and name of
6308 * the probe, along with the hex address of the ECB with the breakpoint
6309 * action -- all of which we must place in the character buffer by
6310 * hand.
6311 */
6312 while (*msg != '\0')
6313 c[i++] = *msg++;
6314
6315 for (str = prov->dtpv_name; *str != '\0'; str++)
6316 c[i++] = *str;
6317 c[i++] = ':';
6318
6319 for (str = probe->dtpr_mod; *str != '\0'; str++)
6320 c[i++] = *str;
6321 c[i++] = ':';
6322
6323 for (str = probe->dtpr_func; *str != '\0'; str++)
6324 c[i++] = *str;
6325 c[i++] = ':';
6326
6327 for (str = probe->dtpr_name; *str != '\0'; str++)
6328 c[i++] = *str;
6329
6330 while (*ecbmsg != '\0')
6331 c[i++] = *ecbmsg++;
6332
6333 while (shift >= 0) {
6334 mask = (uintptr_t)0xf << shift;
6335
6336 if (val >= ((uintptr_t)1 << shift))
6337 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6338 shift -= 4;
6339 }
6340
6341 c[i++] = ')';
6342 c[i] = '\0';
6343
6344 debug_enter(c);
6345 }
6346
6347 static void
6348 dtrace_action_panic(dtrace_ecb_t *ecb)
6349 {
6350 dtrace_probe_t *probe = ecb->dte_probe;
6351
6352 /*
6353 * It's impossible to be taking action on the NULL probe.
6354 */
6355 ASSERT(probe != NULL);
6356
6357 if (dtrace_destructive_disallow)
6358 return;
6359
6360 if (dtrace_panicked != NULL)
6361 return;
6362
6363 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6364 return;
6365
6366 /*
6367 * We won the right to panic. (We want to be sure that only one
6368 * thread calls panic() from dtrace_probe(), and that panic() is
6369 * called exactly once.)
6370 */
6371 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6372 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6373 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6374 }
6375
6376 static void
6377 dtrace_action_raise(uint64_t sig)
6378 {
6379 if (dtrace_destructive_disallow)
6380 return;
6381
6382 if (sig >= NSIG) {
6383 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6384 return;
6385 }
6386
6387 /*
6388 * raise() has a queue depth of 1 -- we ignore all subsequent
6389 * invocations of the raise() action.
6390 */
6391 if (curthread->t_dtrace_sig == 0)
6392 curthread->t_dtrace_sig = (uint8_t)sig;
6393
6394 curthread->t_sig_check = 1;
6395 aston(curthread);
6396 }
6397
6398 static void
6399 dtrace_action_stop(void)
6400 {
6401 if (dtrace_destructive_disallow)
6402 return;
6403
6404 if (!curthread->t_dtrace_stop) {
6405 curthread->t_dtrace_stop = 1;
6406 curthread->t_sig_check = 1;
6407 aston(curthread);
6408 }
6409 }
6410
6411 static void
6412 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
6413 {
6414 hrtime_t now;
6415 volatile uint16_t *flags;
6416 cpu_t *cpu = CPU;
6417
6418 if (dtrace_destructive_disallow)
6419 return;
6420
6421 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
6422
6423 now = dtrace_gethrtime();
6424
6425 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
6426 /*
6427 * We need to advance the mark to the current time.
6428 */
6429 cpu->cpu_dtrace_chillmark = now;
6430 cpu->cpu_dtrace_chilled = 0;
6431 }
6432
6433 /*
6434 * Now check to see if the requested chill time would take us over
6435 * the maximum amount of time allowed in the chill interval. (Or
6436 * worse, if the calculation itself induces overflow.)
6437 */
6438 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
6439 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
6440 *flags |= CPU_DTRACE_ILLOP;
6441 return;
6442 }
6443
6444 while (dtrace_gethrtime() - now < val)
6445 continue;
6446
6447 /*
6448 * Normally, we assure that the value of the variable "timestamp" does
6449 * not change within an ECB. The presence of chill() represents an
6450 * exception to this rule, however.
6451 */
6452 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
6453 cpu->cpu_dtrace_chilled += val;
6454 }
6455
6456 static void
6457 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
6458 uint64_t *buf, uint64_t arg)
6459 {
6460 int nframes = DTRACE_USTACK_NFRAMES(arg);
6461 int strsize = DTRACE_USTACK_STRSIZE(arg);
6462 uint64_t *pcs = &buf[1], *fps;
6463 char *str = (char *)&pcs[nframes];
6464 int size, offs = 0, i, j;
6465 uintptr_t old = mstate->dtms_scratch_ptr, saved;
6466 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6467 char *sym;
6468
6469 /*
6470 * Should be taking a faster path if string space has not been
6471 * allocated.
6472 */
6473 ASSERT(strsize != 0);
6474
6475 /*
6476 * We will first allocate some temporary space for the frame pointers.
6477 */
6478 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6479 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
6480 (nframes * sizeof (uint64_t));
6481
6482 if (!DTRACE_INSCRATCH(mstate, size)) {
6483 /*
6484 * Not enough room for our frame pointers -- need to indicate
6485 * that we ran out of scratch space.
6486 */
6487 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6488 return;
6489 }
6490
6491 mstate->dtms_scratch_ptr += size;
6492 saved = mstate->dtms_scratch_ptr;
6493
6494 /*
6495 * Now get a stack with both program counters and frame pointers.
6496 */
6497 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6498 dtrace_getufpstack(buf, fps, nframes + 1);
6499 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6500
6501 /*
6502 * If that faulted, we're cooked.
6503 */
6504 if (*flags & CPU_DTRACE_FAULT)
6505 goto out;
6506
6507 /*
6508 * Now we want to walk up the stack, calling the USTACK helper. For
6509 * each iteration, we restore the scratch pointer.
6510 */
6511 for (i = 0; i < nframes; i++) {
6512 mstate->dtms_scratch_ptr = saved;
6513
6514 if (offs >= strsize)
6515 break;
6516
6517 sym = (char *)(uintptr_t)dtrace_helper(
6518 DTRACE_HELPER_ACTION_USTACK,
6519 mstate, state, pcs[i], fps[i]);
6520
6521 /*
6522 * If we faulted while running the helper, we're going to
6523 * clear the fault and null out the corresponding string.
6524 */
6525 if (*flags & CPU_DTRACE_FAULT) {
6526 *flags &= ~CPU_DTRACE_FAULT;
6527 str[offs++] = '\0';
6528 continue;
6529 }
6530
6531 if (sym == NULL) {
6532 str[offs++] = '\0';
6533 continue;
6534 }
6535
6536 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6537
6538 /*
6539 * Now copy in the string that the helper returned to us.
6540 */
6541 for (j = 0; offs + j < strsize; j++) {
6542 if ((str[offs + j] = sym[j]) == '\0')
6543 break;
6544 }
6545
6546 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6547
6548 offs += j + 1;
6549 }
6550
6551 if (offs >= strsize) {
6552 /*
6553 * If we didn't have room for all of the strings, we don't
6554 * abort processing -- this needn't be a fatal error -- but we
6555 * still want to increment a counter (dts_stkstroverflows) to
6556 * allow this condition to be warned about. (If this is from
6557 * a jstack() action, it is easily tuned via jstackstrsize.)
6558 */
6559 dtrace_error(&state->dts_stkstroverflows);
6560 }
6561
6562 while (offs < strsize)
6563 str[offs++] = '\0';
6564
6565 out:
6566 mstate->dtms_scratch_ptr = old;
6567 }
6568
6569 static void
6570 dtrace_store_by_ref(dtrace_difo_t *dp, caddr_t tomax, size_t size,
6571 size_t *valoffsp, uint64_t *valp, uint64_t end, int intuple, int dtkind)
6572 {
6573 volatile uint16_t *flags;
6574 uint64_t val = *valp;
6575 size_t valoffs = *valoffsp;
6576
6577 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6578 ASSERT(dtkind == DIF_TF_BYREF || dtkind == DIF_TF_BYUREF);
6579
6580 /*
6581 * If this is a string, we're going to only load until we find the zero
6582 * byte -- after which we'll store zero bytes.
6583 */
6584 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
6585 char c = '\0' + 1;
6586 size_t s;
6587
6588 for (s = 0; s < size; s++) {
6589 if (c != '\0' && dtkind == DIF_TF_BYREF) {
6590 c = dtrace_load8(val++);
6591 } else if (c != '\0' && dtkind == DIF_TF_BYUREF) {
6592 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6593 c = dtrace_fuword8((void *)(uintptr_t)val++);
6594 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6595 if (*flags & CPU_DTRACE_FAULT)
6596 break;
6597 }
6598
6599 DTRACE_STORE(uint8_t, tomax, valoffs++, c);
6600
6601 if (c == '\0' && intuple)
6602 break;
6603 }
6604 } else {
6605 uint8_t c;
6606 while (valoffs < end) {
6607 if (dtkind == DIF_TF_BYREF) {
6608 c = dtrace_load8(val++);
6609 } else if (dtkind == DIF_TF_BYUREF) {
6610 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6611 c = dtrace_fuword8((void *)(uintptr_t)val++);
6612 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6613 if (*flags & CPU_DTRACE_FAULT)
6614 break;
6615 }
6616
6617 DTRACE_STORE(uint8_t, tomax,
6618 valoffs++, c);
6619 }
6620 }
6621
6622 *valp = val;
6623 *valoffsp = valoffs;
6624 }
6625
6626 /*
6627 * If you're looking for the epicenter of DTrace, you just found it. This
6628 * is the function called by the provider to fire a probe -- from which all
6629 * subsequent probe-context DTrace activity emanates.
6630 */
6631 void
6632 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6633 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6634 {
6635 processorid_t cpuid;
6636 dtrace_icookie_t cookie;
6637 dtrace_probe_t *probe;
6638 dtrace_mstate_t mstate;
6639 dtrace_ecb_t *ecb;
6640 dtrace_action_t *act;
6641 intptr_t offs;
6642 size_t size;
6643 int vtime, onintr;
6644 volatile uint16_t *flags;
6645 hrtime_t now, end;
6646
6647 /*
6648 * Kick out immediately if this CPU is still being born (in which case
6649 * curthread will be set to -1) or the current thread can't allow
6650 * probes in its current context.
6651 */
6652 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6653 return;
6654
6655 cookie = dtrace_interrupt_disable();
6656 probe = dtrace_probes[id - 1];
6657 cpuid = CPU->cpu_id;
6658 onintr = CPU_ON_INTR(CPU);
6659
6660 CPU->cpu_dtrace_probes++;
6661
6662 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6663 probe->dtpr_predcache == curthread->t_predcache) {
6664 /*
6665 * We have hit in the predicate cache; we know that
6666 * this predicate would evaluate to be false.
6667 */
6668 dtrace_interrupt_enable(cookie);
6669 return;
6670 }
6671
6672 if (panic_quiesce) {
6673 /*
6674 * We don't trace anything if we're panicking.
6675 */
6676 dtrace_interrupt_enable(cookie);
6677 return;
6678 }
6679
6680 now = dtrace_gethrtime();
6681 vtime = dtrace_vtime_references != 0;
6682
6683 if (vtime && curthread->t_dtrace_start)
6684 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6685
6686 mstate.dtms_difo = NULL;
6687 mstate.dtms_probe = probe;
6688 mstate.dtms_strtok = NULL;
6689 mstate.dtms_arg[0] = arg0;
6690 mstate.dtms_arg[1] = arg1;
6691 mstate.dtms_arg[2] = arg2;
6692 mstate.dtms_arg[3] = arg3;
6693 mstate.dtms_arg[4] = arg4;
6694
6695 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6696
6697 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6698 dtrace_predicate_t *pred = ecb->dte_predicate;
6699 dtrace_state_t *state = ecb->dte_state;
6700 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6701 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6702 dtrace_vstate_t *vstate = &state->dts_vstate;
6703 dtrace_provider_t *prov = probe->dtpr_provider;
6704 uint64_t tracememsize = 0;
6705 int committed = 0;
6706 caddr_t tomax;
6707
6708 /*
6709 * A little subtlety with the following (seemingly innocuous)
6710 * declaration of the automatic 'val': by looking at the
6711 * code, you might think that it could be declared in the
6712 * action processing loop, below. (That is, it's only used in
6713 * the action processing loop.) However, it must be declared
6714 * out of that scope because in the case of DIF expression
6715 * arguments to aggregating actions, one iteration of the
6716 * action loop will use the last iteration's value.
6717 */
6718 #ifdef lint
6719 uint64_t val = 0;
6720 #else
6721 uint64_t val;
6722 #endif
6723
6724 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6725 mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
6726 mstate.dtms_getf = NULL;
6727
6728 *flags &= ~CPU_DTRACE_ERROR;
6729
6730 if (prov == dtrace_provider) {
6731 /*
6732 * If dtrace itself is the provider of this probe,
6733 * we're only going to continue processing the ECB if
6734 * arg0 (the dtrace_state_t) is equal to the ECB's
6735 * creating state. (This prevents disjoint consumers
6736 * from seeing one another's metaprobes.)
6737 */
6738 if (arg0 != (uint64_t)(uintptr_t)state)
6739 continue;
6740 }
6741
6742 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6743 /*
6744 * We're not currently active. If our provider isn't
6745 * the dtrace pseudo provider, we're not interested.
6746 */
6747 if (prov != dtrace_provider)
6748 continue;
6749
6750 /*
6751 * Now we must further check if we are in the BEGIN
6752 * probe. If we are, we will only continue processing
6753 * if we're still in WARMUP -- if one BEGIN enabling
6754 * has invoked the exit() action, we don't want to
6755 * evaluate subsequent BEGIN enablings.
6756 */
6757 if (probe->dtpr_id == dtrace_probeid_begin &&
6758 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6759 ASSERT(state->dts_activity ==
6760 DTRACE_ACTIVITY_DRAINING);
6761 continue;
6762 }
6763 }
6764
6765 if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
6766 continue;
6767
6768 if (now - state->dts_alive > dtrace_deadman_timeout) {
6769 /*
6770 * We seem to be dead. Unless we (a) have kernel
6771 * destructive permissions (b) have explicitly enabled
6772 * destructive actions and (c) destructive actions have
6773 * not been disabled, we're going to transition into
6774 * the KILLED state, from which no further processing
6775 * on this state will be performed.
6776 */
6777 if (!dtrace_priv_kernel_destructive(state) ||
6778 !state->dts_cred.dcr_destructive ||
6779 dtrace_destructive_disallow) {
6780 void *activity = &state->dts_activity;
6781 dtrace_activity_t current;
6782
6783 do {
6784 current = state->dts_activity;
6785 } while (dtrace_cas32(activity, current,
6786 DTRACE_ACTIVITY_KILLED) != current);
6787
6788 continue;
6789 }
6790 }
6791
6792 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6793 ecb->dte_alignment, state, &mstate)) < 0)
6794 continue;
6795
6796 tomax = buf->dtb_tomax;
6797 ASSERT(tomax != NULL);
6798
6799 if (ecb->dte_size != 0) {
6800 dtrace_rechdr_t dtrh;
6801 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6802 mstate.dtms_timestamp = dtrace_gethrtime();
6803 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6804 }
6805 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6806 dtrh.dtrh_epid = ecb->dte_epid;
6807 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6808 mstate.dtms_timestamp);
6809 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6810 }
6811
6812 mstate.dtms_epid = ecb->dte_epid;
6813 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6814
6815 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6816 mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
6817
6818 if (pred != NULL) {
6819 dtrace_difo_t *dp = pred->dtp_difo;
6820 int rval;
6821
6822 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6823
6824 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6825 dtrace_cacheid_t cid = probe->dtpr_predcache;
6826
6827 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6828 /*
6829 * Update the predicate cache...
6830 */
6831 ASSERT(cid == pred->dtp_cacheid);
6832 curthread->t_predcache = cid;
6833 }
6834
6835 continue;
6836 }
6837 }
6838
6839 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6840 act != NULL; act = act->dta_next) {
6841 size_t valoffs;
6842 dtrace_difo_t *dp;
6843 dtrace_recdesc_t *rec = &act->dta_rec;
6844
6845 size = rec->dtrd_size;
6846 valoffs = offs + rec->dtrd_offset;
6847
6848 if (DTRACEACT_ISAGG(act->dta_kind)) {
6849 uint64_t v = 0xbad;
6850 dtrace_aggregation_t *agg;
6851
6852 agg = (dtrace_aggregation_t *)act;
6853
6854 if ((dp = act->dta_difo) != NULL)
6855 v = dtrace_dif_emulate(dp,
6856 &mstate, vstate, state);
6857
6858 if (*flags & CPU_DTRACE_ERROR)
6859 continue;
6860
6861 /*
6862 * Note that we always pass the expression
6863 * value from the previous iteration of the
6864 * action loop. This value will only be used
6865 * if there is an expression argument to the
6866 * aggregating action, denoted by the
6867 * dtag_hasarg field.
6868 */
6869 dtrace_aggregate(agg, buf,
6870 offs, aggbuf, v, val);
6871 continue;
6872 }
6873
6874 switch (act->dta_kind) {
6875 case DTRACEACT_STOP:
6876 if (dtrace_priv_proc_destructive(state,
6877 &mstate))
6878 dtrace_action_stop();
6879 continue;
6880
6881 case DTRACEACT_BREAKPOINT:
6882 if (dtrace_priv_kernel_destructive(state))
6883 dtrace_action_breakpoint(ecb);
6884 continue;
6885
6886 case DTRACEACT_PANIC:
6887 if (dtrace_priv_kernel_destructive(state))
6888 dtrace_action_panic(ecb);
6889 continue;
6890
6891 case DTRACEACT_STACK:
6892 if (!dtrace_priv_kernel(state))
6893 continue;
6894
6895 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6896 size / sizeof (pc_t), probe->dtpr_aframes,
6897 DTRACE_ANCHORED(probe) ? NULL :
6898 (uint32_t *)arg0);
6899
6900 continue;
6901
6902 case DTRACEACT_JSTACK:
6903 case DTRACEACT_USTACK:
6904 if (!dtrace_priv_proc(state, &mstate))
6905 continue;
6906
6907 /*
6908 * See comment in DIF_VAR_PID.
6909 */
6910 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6911 CPU_ON_INTR(CPU)) {
6912 int depth = DTRACE_USTACK_NFRAMES(
6913 rec->dtrd_arg) + 1;
6914
6915 dtrace_bzero((void *)(tomax + valoffs),
6916 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6917 + depth * sizeof (uint64_t));
6918
6919 continue;
6920 }
6921
6922 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6923 curproc->p_dtrace_helpers != NULL) {
6924 /*
6925 * This is the slow path -- we have
6926 * allocated string space, and we're
6927 * getting the stack of a process that
6928 * has helpers. Call into a separate
6929 * routine to perform this processing.
6930 */
6931 dtrace_action_ustack(&mstate, state,
6932 (uint64_t *)(tomax + valoffs),
6933 rec->dtrd_arg);
6934 continue;
6935 }
6936
6937 /*
6938 * Clear the string space, since there's no
6939 * helper to do it for us.
6940 */
6941 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
6942 int depth = DTRACE_USTACK_NFRAMES(
6943 rec->dtrd_arg);
6944 size_t strsize = DTRACE_USTACK_STRSIZE(
6945 rec->dtrd_arg);
6946 uint64_t *buf = (uint64_t *)(tomax +
6947 valoffs);
6948 void *strspace = &buf[depth + 1];
6949
6950 dtrace_bzero(strspace,
6951 MIN(depth, strsize));
6952 }
6953
6954 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6955 dtrace_getupcstack((uint64_t *)
6956 (tomax + valoffs),
6957 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6958 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6959 continue;
6960
6961 default:
6962 break;
6963 }
6964
6965 dp = act->dta_difo;
6966 ASSERT(dp != NULL);
6967
6968 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6969
6970 if (*flags & CPU_DTRACE_ERROR)
6971 continue;
6972
6973 switch (act->dta_kind) {
6974 case DTRACEACT_SPECULATE: {
6975 dtrace_rechdr_t *dtrh;
6976
6977 ASSERT(buf == &state->dts_buffer[cpuid]);
6978 buf = dtrace_speculation_buffer(state,
6979 cpuid, val);
6980
6981 if (buf == NULL) {
6982 *flags |= CPU_DTRACE_DROP;
6983 continue;
6984 }
6985
6986 offs = dtrace_buffer_reserve(buf,
6987 ecb->dte_needed, ecb->dte_alignment,
6988 state, NULL);
6989
6990 if (offs < 0) {
6991 *flags |= CPU_DTRACE_DROP;
6992 continue;
6993 }
6994
6995 tomax = buf->dtb_tomax;
6996 ASSERT(tomax != NULL);
6997
6998 if (ecb->dte_size == 0)
6999 continue;
7000
7001 ASSERT3U(ecb->dte_size, >=,
7002 sizeof (dtrace_rechdr_t));
7003 dtrh = ((void *)(tomax + offs));
7004 dtrh->dtrh_epid = ecb->dte_epid;
7005 /*
7006 * When the speculation is committed, all of
7007 * the records in the speculative buffer will
7008 * have their timestamps set to the commit
7009 * time. Until then, it is set to a sentinel
7010 * value, for debugability.
7011 */
7012 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
7013 continue;
7014 }
7015
7016 case DTRACEACT_CHILL:
7017 if (dtrace_priv_kernel_destructive(state))
7018 dtrace_action_chill(&mstate, val);
7019 continue;
7020
7021 case DTRACEACT_RAISE:
7022 if (dtrace_priv_proc_destructive(state,
7023 &mstate))
7024 dtrace_action_raise(val);
7025 continue;
7026
7027 case DTRACEACT_COMMIT:
7028 ASSERT(!committed);
7029
7030 /*
7031 * We need to commit our buffer state.
7032 */
7033 if (ecb->dte_size)
7034 buf->dtb_offset = offs + ecb->dte_size;
7035 buf = &state->dts_buffer[cpuid];
7036 dtrace_speculation_commit(state, cpuid, val);
7037 committed = 1;
7038 continue;
7039
7040 case DTRACEACT_DISCARD:
7041 dtrace_speculation_discard(state, cpuid, val);
7042 continue;
7043
7044 case DTRACEACT_DIFEXPR:
7045 case DTRACEACT_LIBACT:
7046 case DTRACEACT_PRINTF:
7047 case DTRACEACT_PRINTA:
7048 case DTRACEACT_SYSTEM:
7049 case DTRACEACT_FREOPEN:
7050 case DTRACEACT_TRACEMEM:
7051 break;
7052
7053 case DTRACEACT_TRACEMEM_DYNSIZE:
7054 tracememsize = val;
7055 break;
7056
7057 case DTRACEACT_SYM:
7058 case DTRACEACT_MOD:
7059 if (!dtrace_priv_kernel(state))
7060 continue;
7061 break;
7062
7063 case DTRACEACT_USYM:
7064 case DTRACEACT_UMOD:
7065 case DTRACEACT_UADDR: {
7066 struct pid *pid = curthread->t_procp->p_pidp;
7067
7068 if (!dtrace_priv_proc(state, &mstate))
7069 continue;
7070
7071 DTRACE_STORE(uint64_t, tomax,
7072 valoffs, (uint64_t)pid->pid_id);
7073 DTRACE_STORE(uint64_t, tomax,
7074 valoffs + sizeof (uint64_t), val);
7075
7076 continue;
7077 }
7078
7079 case DTRACEACT_EXIT: {
7080 /*
7081 * For the exit action, we are going to attempt
7082 * to atomically set our activity to be
7083 * draining. If this fails (either because
7084 * another CPU has beat us to the exit action,
7085 * or because our current activity is something
7086 * other than ACTIVE or WARMUP), we will
7087 * continue. This assures that the exit action
7088 * can be successfully recorded at most once
7089 * when we're in the ACTIVE state. If we're
7090 * encountering the exit() action while in
7091 * COOLDOWN, however, we want to honor the new
7092 * status code. (We know that we're the only
7093 * thread in COOLDOWN, so there is no race.)
7094 */
7095 void *activity = &state->dts_activity;
7096 dtrace_activity_t current = state->dts_activity;
7097
7098 if (current == DTRACE_ACTIVITY_COOLDOWN)
7099 break;
7100
7101 if (current != DTRACE_ACTIVITY_WARMUP)
7102 current = DTRACE_ACTIVITY_ACTIVE;
7103
7104 if (dtrace_cas32(activity, current,
7105 DTRACE_ACTIVITY_DRAINING) != current) {
7106 *flags |= CPU_DTRACE_DROP;
7107 continue;
7108 }
7109
7110 break;
7111 }
7112
7113 default:
7114 ASSERT(0);
7115 }
7116
7117 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ||
7118 dp->dtdo_rtype.dtdt_flags & DIF_TF_BYUREF) {
7119 uintptr_t end = valoffs + size;
7120
7121 if (tracememsize != 0 &&
7122 valoffs + tracememsize < end) {
7123 end = valoffs + tracememsize;
7124 tracememsize = 0;
7125 }
7126
7127 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF &&
7128 !dtrace_vcanload((void *)(uintptr_t)val,
7129 &dp->dtdo_rtype, &mstate, vstate))
7130 continue;
7131
7132 dtrace_store_by_ref(dp, tomax, size, &valoffs,
7133 &val, end, act->dta_intuple,
7134 dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ?
7135 DIF_TF_BYREF: DIF_TF_BYUREF);
7136 continue;
7137 }
7138
7139 switch (size) {
7140 case 0:
7141 break;
7142
7143 case sizeof (uint8_t):
7144 DTRACE_STORE(uint8_t, tomax, valoffs, val);
7145 break;
7146 case sizeof (uint16_t):
7147 DTRACE_STORE(uint16_t, tomax, valoffs, val);
7148 break;
7149 case sizeof (uint32_t):
7150 DTRACE_STORE(uint32_t, tomax, valoffs, val);
7151 break;
7152 case sizeof (uint64_t):
7153 DTRACE_STORE(uint64_t, tomax, valoffs, val);
7154 break;
7155 default:
7156 /*
7157 * Any other size should have been returned by
7158 * reference, not by value.
7159 */
7160 ASSERT(0);
7161 break;
7162 }
7163 }
7164
7165 if (*flags & CPU_DTRACE_DROP)
7166 continue;
7167
7168 if (*flags & CPU_DTRACE_FAULT) {
7169 int ndx;
7170 dtrace_action_t *err;
7171
7172 buf->dtb_errors++;
7173
7174 if (probe->dtpr_id == dtrace_probeid_error) {
7175 /*
7176 * There's nothing we can do -- we had an
7177 * error on the error probe. We bump an
7178 * error counter to at least indicate that
7179 * this condition happened.
7180 */
7181 dtrace_error(&state->dts_dblerrors);
7182 continue;
7183 }
7184
7185 if (vtime) {
7186 /*
7187 * Before recursing on dtrace_probe(), we
7188 * need to explicitly clear out our start
7189 * time to prevent it from being accumulated
7190 * into t_dtrace_vtime.
7191 */
7192 curthread->t_dtrace_start = 0;
7193 }
7194
7195 /*
7196 * Iterate over the actions to figure out which action
7197 * we were processing when we experienced the error.
7198 * Note that act points _past_ the faulting action; if
7199 * act is ecb->dte_action, the fault was in the
7200 * predicate, if it's ecb->dte_action->dta_next it's
7201 * in action #1, and so on.
7202 */
7203 for (err = ecb->dte_action, ndx = 0;
7204 err != act; err = err->dta_next, ndx++)
7205 continue;
7206
7207 dtrace_probe_error(state, ecb->dte_epid, ndx,
7208 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7209 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7210 cpu_core[cpuid].cpuc_dtrace_illval);
7211
7212 continue;
7213 }
7214
7215 if (!committed)
7216 buf->dtb_offset = offs + ecb->dte_size;
7217 }
7218
7219 end = dtrace_gethrtime();
7220 if (vtime)
7221 curthread->t_dtrace_start = end;
7222
7223 CPU->cpu_dtrace_nsec += end - now;
7224
7225 dtrace_interrupt_enable(cookie);
7226 }
7227
7228 /*
7229 * DTrace Probe Hashing Functions
7230 *
7231 * The functions in this section (and indeed, the functions in remaining
7232 * sections) are not _called_ from probe context. (Any exceptions to this are
7233 * marked with a "Note:".) Rather, they are called from elsewhere in the
7234 * DTrace framework to look-up probes in, add probes to and remove probes from
7235 * the DTrace probe hashes. (Each probe is hashed by each element of the
7236 * probe tuple -- allowing for fast lookups, regardless of what was
7237 * specified.)
7238 */
7239 static uint_t
7240 dtrace_hash_str(char *p)
7241 {
7242 unsigned int g;
7243 uint_t hval = 0;
7244
7245 while (*p) {
7246 hval = (hval << 4) + *p++;
7247 if ((g = (hval & 0xf0000000)) != 0)
7248 hval ^= g >> 24;
7249 hval &= ~g;
7250 }
7251 return (hval);
7252 }
7253
7254 static dtrace_hash_t *
7255 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
7256 {
7257 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
7258
7259 hash->dth_stroffs = stroffs;
7260 hash->dth_nextoffs = nextoffs;
7261 hash->dth_prevoffs = prevoffs;
7262
7263 hash->dth_size = 1;
7264 hash->dth_mask = hash->dth_size - 1;
7265
7266 hash->dth_tab = kmem_zalloc(hash->dth_size *
7267 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
7268
7269 return (hash);
7270 }
7271
7272 static void
7273 dtrace_hash_destroy(dtrace_hash_t *hash)
7274 {
7275 #ifdef DEBUG
7276 int i;
7277
7278 for (i = 0; i < hash->dth_size; i++)
7279 ASSERT(hash->dth_tab[i] == NULL);
7280 #endif
7281
7282 kmem_free(hash->dth_tab,
7283 hash->dth_size * sizeof (dtrace_hashbucket_t *));
7284 kmem_free(hash, sizeof (dtrace_hash_t));
7285 }
7286
7287 static void
7288 dtrace_hash_resize(dtrace_hash_t *hash)
7289 {
7290 int size = hash->dth_size, i, ndx;
7291 int new_size = hash->dth_size << 1;
7292 int new_mask = new_size - 1;
7293 dtrace_hashbucket_t **new_tab, *bucket, *next;
7294
7295 ASSERT((new_size & new_mask) == 0);
7296
7297 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
7298
7299 for (i = 0; i < size; i++) {
7300 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
7301 dtrace_probe_t *probe = bucket->dthb_chain;
7302
7303 ASSERT(probe != NULL);
7304 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
7305
7306 next = bucket->dthb_next;
7307 bucket->dthb_next = new_tab[ndx];
7308 new_tab[ndx] = bucket;
7309 }
7310 }
7311
7312 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
7313 hash->dth_tab = new_tab;
7314 hash->dth_size = new_size;
7315 hash->dth_mask = new_mask;
7316 }
7317
7318 static void
7319 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
7320 {
7321 int hashval = DTRACE_HASHSTR(hash, new);
7322 int ndx = hashval & hash->dth_mask;
7323 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7324 dtrace_probe_t **nextp, **prevp;
7325
7326 for (; bucket != NULL; bucket = bucket->dthb_next) {
7327 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
7328 goto add;
7329 }
7330
7331 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
7332 dtrace_hash_resize(hash);
7333 dtrace_hash_add(hash, new);
7334 return;
7335 }
7336
7337 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
7338 bucket->dthb_next = hash->dth_tab[ndx];
7339 hash->dth_tab[ndx] = bucket;
7340 hash->dth_nbuckets++;
7341
7342 add:
7343 nextp = DTRACE_HASHNEXT(hash, new);
7344 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
7345 *nextp = bucket->dthb_chain;
7346
7347 if (bucket->dthb_chain != NULL) {
7348 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
7349 ASSERT(*prevp == NULL);
7350 *prevp = new;
7351 }
7352
7353 bucket->dthb_chain = new;
7354 bucket->dthb_len++;
7355 }
7356
7357 static dtrace_probe_t *
7358 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
7359 {
7360 int hashval = DTRACE_HASHSTR(hash, template);
7361 int ndx = hashval & hash->dth_mask;
7362 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7363
7364 for (; bucket != NULL; bucket = bucket->dthb_next) {
7365 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7366 return (bucket->dthb_chain);
7367 }
7368
7369 return (NULL);
7370 }
7371
7372 static int
7373 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
7374 {
7375 int hashval = DTRACE_HASHSTR(hash, template);
7376 int ndx = hashval & hash->dth_mask;
7377 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7378
7379 for (; bucket != NULL; bucket = bucket->dthb_next) {
7380 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7381 return (bucket->dthb_len);
7382 }
7383
7384 return (NULL);
7385 }
7386
7387 static void
7388 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7389 {
7390 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7391 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7392
7393 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7394 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7395
7396 /*
7397 * Find the bucket that we're removing this probe from.
7398 */
7399 for (; bucket != NULL; bucket = bucket->dthb_next) {
7400 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7401 break;
7402 }
7403
7404 ASSERT(bucket != NULL);
7405
7406 if (*prevp == NULL) {
7407 if (*nextp == NULL) {
7408 /*
7409 * The removed probe was the only probe on this
7410 * bucket; we need to remove the bucket.
7411 */
7412 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7413
7414 ASSERT(bucket->dthb_chain == probe);
7415 ASSERT(b != NULL);
7416
7417 if (b == bucket) {
7418 hash->dth_tab[ndx] = bucket->dthb_next;
7419 } else {
7420 while (b->dthb_next != bucket)
7421 b = b->dthb_next;
7422 b->dthb_next = bucket->dthb_next;
7423 }
7424
7425 ASSERT(hash->dth_nbuckets > 0);
7426 hash->dth_nbuckets--;
7427 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7428 return;
7429 }
7430
7431 bucket->dthb_chain = *nextp;
7432 } else {
7433 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7434 }
7435
7436 if (*nextp != NULL)
7437 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7438 }
7439
7440 /*
7441 * DTrace Utility Functions
7442 *
7443 * These are random utility functions that are _not_ called from probe context.
7444 */
7445 static int
7446 dtrace_badattr(const dtrace_attribute_t *a)
7447 {
7448 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7449 a->dtat_data > DTRACE_STABILITY_MAX ||
7450 a->dtat_class > DTRACE_CLASS_MAX);
7451 }
7452
7453 /*
7454 * Return a duplicate copy of a string. If the specified string is NULL,
7455 * this function returns a zero-length string.
7456 */
7457 static char *
7458 dtrace_strdup(const char *str)
7459 {
7460 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7461
7462 if (str != NULL)
7463 (void) strcpy(new, str);
7464
7465 return (new);
7466 }
7467
7468 #define DTRACE_ISALPHA(c) \
7469 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7470
7471 static int
7472 dtrace_badname(const char *s)
7473 {
7474 char c;
7475
7476 if (s == NULL || (c = *s++) == '\0')
7477 return (0);
7478
7479 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7480 return (1);
7481
7482 while ((c = *s++) != '\0') {
7483 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7484 c != '-' && c != '_' && c != '.' && c != '`')
7485 return (1);
7486 }
7487
7488 return (0);
7489 }
7490
7491 static void
7492 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7493 {
7494 uint32_t priv;
7495
7496 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7497 /*
7498 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7499 */
7500 priv = DTRACE_PRIV_ALL;
7501 } else {
7502 *uidp = crgetuid(cr);
7503 *zoneidp = crgetzoneid(cr);
7504
7505 priv = 0;
7506 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7507 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7508 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7509 priv |= DTRACE_PRIV_USER;
7510 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7511 priv |= DTRACE_PRIV_PROC;
7512 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7513 priv |= DTRACE_PRIV_OWNER;
7514 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7515 priv |= DTRACE_PRIV_ZONEOWNER;
7516 }
7517
7518 *privp = priv;
7519 }
7520
7521 #ifdef DTRACE_ERRDEBUG
7522 static void
7523 dtrace_errdebug(const char *str)
7524 {
7525 int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
7526 int occupied = 0;
7527
7528 mutex_enter(&dtrace_errlock);
7529 dtrace_errlast = str;
7530 dtrace_errthread = curthread;
7531
7532 while (occupied++ < DTRACE_ERRHASHSZ) {
7533 if (dtrace_errhash[hval].dter_msg == str) {
7534 dtrace_errhash[hval].dter_count++;
7535 goto out;
7536 }
7537
7538 if (dtrace_errhash[hval].dter_msg != NULL) {
7539 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7540 continue;
7541 }
7542
7543 dtrace_errhash[hval].dter_msg = str;
7544 dtrace_errhash[hval].dter_count = 1;
7545 goto out;
7546 }
7547
7548 panic("dtrace: undersized error hash");
7549 out:
7550 mutex_exit(&dtrace_errlock);
7551 }
7552 #endif
7553
7554 /*
7555 * DTrace Matching Functions
7556 *
7557 * These functions are used to match groups of probes, given some elements of
7558 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7559 */
7560 static int
7561 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7562 zoneid_t zoneid)
7563 {
7564 if (priv != DTRACE_PRIV_ALL) {
7565 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7566 uint32_t match = priv & ppriv;
7567
7568 /*
7569 * No PRIV_DTRACE_* privileges...
7570 */
7571 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7572 DTRACE_PRIV_KERNEL)) == 0)
7573 return (0);
7574
7575 /*
7576 * No matching bits, but there were bits to match...
7577 */
7578 if (match == 0 && ppriv != 0)
7579 return (0);
7580
7581 /*
7582 * Need to have permissions to the process, but don't...
7583 */
7584 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7585 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7586 return (0);
7587 }
7588
7589 /*
7590 * Need to be in the same zone unless we possess the
7591 * privilege to examine all zones.
7592 */
7593 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7594 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7595 return (0);
7596 }
7597 }
7598
7599 return (1);
7600 }
7601
7602 /*
7603 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7604 * consists of input pattern strings and an ops-vector to evaluate them.
7605 * This function returns >0 for match, 0 for no match, and <0 for error.
7606 */
7607 static int
7608 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7609 uint32_t priv, uid_t uid, zoneid_t zoneid)
7610 {
7611 dtrace_provider_t *pvp = prp->dtpr_provider;
7612 int rv;
7613
7614 if (pvp->dtpv_defunct)
7615 return (0);
7616
7617 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7618 return (rv);
7619
7620 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7621 return (rv);
7622
7623 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7624 return (rv);
7625
7626 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7627 return (rv);
7628
7629 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7630 return (0);
7631
7632 return (rv);
7633 }
7634
7635 /*
7636 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7637 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7638 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7639 * In addition, all of the recursion cases except for '*' matching have been
7640 * unwound. For '*', we still implement recursive evaluation, but a depth
7641 * counter is maintained and matching is aborted if we recurse too deep.
7642 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7643 */
7644 static int
7645 dtrace_match_glob(const char *s, const char *p, int depth)
7646 {
7647 const char *olds;
7648 char s1, c;
7649 int gs;
7650
7651 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7652 return (-1);
7653
7654 if (s == NULL)
7655 s = ""; /* treat NULL as empty string */
7656
7657 top:
7658 olds = s;
7659 s1 = *s++;
7660
7661 if (p == NULL)
7662 return (0);
7663
7664 if ((c = *p++) == '\0')
7665 return (s1 == '\0');
7666
7667 switch (c) {
7668 case '[': {
7669 int ok = 0, notflag = 0;
7670 char lc = '\0';
7671
7672 if (s1 == '\0')
7673 return (0);
7674
7675 if (*p == '!') {
7676 notflag = 1;
7677 p++;
7678 }
7679
7680 if ((c = *p++) == '\0')
7681 return (0);
7682
7683 do {
7684 if (c == '-' && lc != '\0' && *p != ']') {
7685 if ((c = *p++) == '\0')
7686 return (0);
7687 if (c == '\\' && (c = *p++) == '\0')
7688 return (0);
7689
7690 if (notflag) {
7691 if (s1 < lc || s1 > c)
7692 ok++;
7693 else
7694 return (0);
7695 } else if (lc <= s1 && s1 <= c)
7696 ok++;
7697
7698 } else if (c == '\\' && (c = *p++) == '\0')
7699 return (0);
7700
7701 lc = c; /* save left-hand 'c' for next iteration */
7702
7703 if (notflag) {
7704 if (s1 != c)
7705 ok++;
7706 else
7707 return (0);
7708 } else if (s1 == c)
7709 ok++;
7710
7711 if ((c = *p++) == '\0')
7712 return (0);
7713
7714 } while (c != ']');
7715
7716 if (ok)
7717 goto top;
7718
7719 return (0);
7720 }
7721
7722 case '\\':
7723 if ((c = *p++) == '\0')
7724 return (0);
7725 /*FALLTHRU*/
7726
7727 default:
7728 if (c != s1)
7729 return (0);
7730 /*FALLTHRU*/
7731
7732 case '?':
7733 if (s1 != '\0')
7734 goto top;
7735 return (0);
7736
7737 case '*':
7738 while (*p == '*')
7739 p++; /* consecutive *'s are identical to a single one */
7740
7741 if (*p == '\0')
7742 return (1);
7743
7744 for (s = olds; *s != '\0'; s++) {
7745 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7746 return (gs);
7747 }
7748
7749 return (0);
7750 }
7751 }
7752
7753 /*ARGSUSED*/
7754 static int
7755 dtrace_match_string(const char *s, const char *p, int depth)
7756 {
7757 return (s != NULL && strcmp(s, p) == 0);
7758 }
7759
7760 /*ARGSUSED*/
7761 static int
7762 dtrace_match_nul(const char *s, const char *p, int depth)
7763 {
7764 return (1); /* always match the empty pattern */
7765 }
7766
7767 /*ARGSUSED*/
7768 static int
7769 dtrace_match_nonzero(const char *s, const char *p, int depth)
7770 {
7771 return (s != NULL && s[0] != '\0');
7772 }
7773
7774 static int
7775 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7776 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7777 {
7778 dtrace_probe_t template, *probe;
7779 dtrace_hash_t *hash = NULL;
7780 int len, rc, best = INT_MAX, nmatched = 0;
7781 dtrace_id_t i;
7782
7783 ASSERT(MUTEX_HELD(&dtrace_lock));
7784
7785 /*
7786 * If the probe ID is specified in the key, just lookup by ID and
7787 * invoke the match callback once if a matching probe is found.
7788 */
7789 if (pkp->dtpk_id != DTRACE_IDNONE) {
7790 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7791 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7792 if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
7793 return (DTRACE_MATCH_FAIL);
7794 nmatched++;
7795 }
7796 return (nmatched);
7797 }
7798
7799 template.dtpr_mod = (char *)pkp->dtpk_mod;
7800 template.dtpr_func = (char *)pkp->dtpk_func;
7801 template.dtpr_name = (char *)pkp->dtpk_name;
7802
7803 /*
7804 * We want to find the most distinct of the module name, function
7805 * name, and name. So for each one that is not a glob pattern or
7806 * empty string, we perform a lookup in the corresponding hash and
7807 * use the hash table with the fewest collisions to do our search.
7808 */
7809 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7810 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7811 best = len;
7812 hash = dtrace_bymod;
7813 }
7814
7815 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7816 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7817 best = len;
7818 hash = dtrace_byfunc;
7819 }
7820
7821 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7822 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7823 best = len;
7824 hash = dtrace_byname;
7825 }
7826
7827 /*
7828 * If we did not select a hash table, iterate over every probe and
7829 * invoke our callback for each one that matches our input probe key.
7830 */
7831 if (hash == NULL) {
7832 for (i = 0; i < dtrace_nprobes; i++) {
7833 if ((probe = dtrace_probes[i]) == NULL ||
7834 dtrace_match_probe(probe, pkp, priv, uid,
7835 zoneid) <= 0)
7836 continue;
7837
7838 nmatched++;
7839
7840 if ((rc = (*matched)(probe, arg)) !=
7841 DTRACE_MATCH_NEXT) {
7842 if (rc == DTRACE_MATCH_FAIL)
7843 return (DTRACE_MATCH_FAIL);
7844 break;
7845 }
7846 }
7847
7848 return (nmatched);
7849 }
7850
7851 /*
7852 * If we selected a hash table, iterate over each probe of the same key
7853 * name and invoke the callback for every probe that matches the other
7854 * attributes of our input probe key.
7855 */
7856 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7857 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7858
7859 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7860 continue;
7861
7862 nmatched++;
7863
7864 if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
7865 if (rc == DTRACE_MATCH_FAIL)
7866 return (DTRACE_MATCH_FAIL);
7867 break;
7868 }
7869 }
7870
7871 return (nmatched);
7872 }
7873
7874 /*
7875 * Return the function pointer dtrace_probecmp() should use to compare the
7876 * specified pattern with a string. For NULL or empty patterns, we select
7877 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7878 * For non-empty non-glob strings, we use dtrace_match_string().
7879 */
7880 static dtrace_probekey_f *
7881 dtrace_probekey_func(const char *p)
7882 {
7883 char c;
7884
7885 if (p == NULL || *p == '\0')
7886 return (&dtrace_match_nul);
7887
7888 while ((c = *p++) != '\0') {
7889 if (c == '[' || c == '?' || c == '*' || c == '\\')
7890 return (&dtrace_match_glob);
7891 }
7892
7893 return (&dtrace_match_string);
7894 }
7895
7896 /*
7897 * Build a probe comparison key for use with dtrace_match_probe() from the
7898 * given probe description. By convention, a null key only matches anchored
7899 * probes: if each field is the empty string, reset dtpk_fmatch to
7900 * dtrace_match_nonzero().
7901 */
7902 static void
7903 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7904 {
7905 pkp->dtpk_prov = pdp->dtpd_provider;
7906 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7907
7908 pkp->dtpk_mod = pdp->dtpd_mod;
7909 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7910
7911 pkp->dtpk_func = pdp->dtpd_func;
7912 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7913
7914 pkp->dtpk_name = pdp->dtpd_name;
7915 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7916
7917 pkp->dtpk_id = pdp->dtpd_id;
7918
7919 if (pkp->dtpk_id == DTRACE_IDNONE &&
7920 pkp->dtpk_pmatch == &dtrace_match_nul &&
7921 pkp->dtpk_mmatch == &dtrace_match_nul &&
7922 pkp->dtpk_fmatch == &dtrace_match_nul &&
7923 pkp->dtpk_nmatch == &dtrace_match_nul)
7924 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7925 }
7926
7927 /*
7928 * DTrace Provider-to-Framework API Functions
7929 *
7930 * These functions implement much of the Provider-to-Framework API, as
7931 * described in <sys/dtrace.h>. The parts of the API not in this section are
7932 * the functions in the API for probe management (found below), and
7933 * dtrace_probe() itself (found above).
7934 */
7935
7936 /*
7937 * Register the calling provider with the DTrace framework. This should
7938 * generally be called by DTrace providers in their attach(9E) entry point.
7939 */
7940 int
7941 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7942 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7943 {
7944 dtrace_provider_t *provider;
7945
7946 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7947 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7948 "arguments", name ? name : "<NULL>");
7949 return (EINVAL);
7950 }
7951
7952 if (name[0] == '\0' || dtrace_badname(name)) {
7953 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7954 "provider name", name);
7955 return (EINVAL);
7956 }
7957
7958 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7959 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7960 pops->dtps_destroy == NULL ||
7961 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7962 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7963 "provider ops", name);
7964 return (EINVAL);
7965 }
7966
7967 if (dtrace_badattr(&pap->dtpa_provider) ||
7968 dtrace_badattr(&pap->dtpa_mod) ||
7969 dtrace_badattr(&pap->dtpa_func) ||
7970 dtrace_badattr(&pap->dtpa_name) ||
7971 dtrace_badattr(&pap->dtpa_args)) {
7972 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7973 "provider attributes", name);
7974 return (EINVAL);
7975 }
7976
7977 if (priv & ~DTRACE_PRIV_ALL) {
7978 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7979 "privilege attributes", name);
7980 return (EINVAL);
7981 }
7982
7983 if ((priv & DTRACE_PRIV_KERNEL) &&
7984 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7985 pops->dtps_mode == NULL) {
7986 cmn_err(CE_WARN, "failed to register provider '%s': need "
7987 "dtps_mode() op for given privilege attributes", name);
7988 return (EINVAL);
7989 }
7990
7991 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7992 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7993 (void) strcpy(provider->dtpv_name, name);
7994
7995 provider->dtpv_attr = *pap;
7996 provider->dtpv_priv.dtpp_flags = priv;
7997 if (cr != NULL) {
7998 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7999 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
8000 }
8001 provider->dtpv_pops = *pops;
8002
8003 if (pops->dtps_provide == NULL) {
8004 ASSERT(pops->dtps_provide_module != NULL);
8005 provider->dtpv_pops.dtps_provide =
8006 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
8007 }
8008
8009 if (pops->dtps_provide_module == NULL) {
8010 ASSERT(pops->dtps_provide != NULL);
8011 provider->dtpv_pops.dtps_provide_module =
8012 (void (*)(void *, struct modctl *))dtrace_nullop;
8013 }
8014
8015 if (pops->dtps_suspend == NULL) {
8016 ASSERT(pops->dtps_resume == NULL);
8017 provider->dtpv_pops.dtps_suspend =
8018 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8019 provider->dtpv_pops.dtps_resume =
8020 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8021 }
8022
8023 provider->dtpv_arg = arg;
8024 *idp = (dtrace_provider_id_t)provider;
8025
8026 if (pops == &dtrace_provider_ops) {
8027 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8028 ASSERT(MUTEX_HELD(&dtrace_lock));
8029 ASSERT(dtrace_anon.dta_enabling == NULL);
8030
8031 /*
8032 * We make sure that the DTrace provider is at the head of
8033 * the provider chain.
8034 */
8035 provider->dtpv_next = dtrace_provider;
8036 dtrace_provider = provider;
8037 return (0);
8038 }
8039
8040 mutex_enter(&dtrace_provider_lock);
8041 mutex_enter(&dtrace_lock);
8042
8043 /*
8044 * If there is at least one provider registered, we'll add this
8045 * provider after the first provider.
8046 */
8047 if (dtrace_provider != NULL) {
8048 provider->dtpv_next = dtrace_provider->dtpv_next;
8049 dtrace_provider->dtpv_next = provider;
8050 } else {
8051 dtrace_provider = provider;
8052 }
8053
8054 if (dtrace_retained != NULL) {
8055 dtrace_enabling_provide(provider);
8056
8057 /*
8058 * Now we need to call dtrace_enabling_matchall() -- which
8059 * will acquire cpu_lock and dtrace_lock. We therefore need
8060 * to drop all of our locks before calling into it...
8061 */
8062 mutex_exit(&dtrace_lock);
8063 mutex_exit(&dtrace_provider_lock);
8064 dtrace_enabling_matchall();
8065
8066 return (0);
8067 }
8068
8069 mutex_exit(&dtrace_lock);
8070 mutex_exit(&dtrace_provider_lock);
8071
8072 return (0);
8073 }
8074
8075 /*
8076 * Unregister the specified provider from the DTrace framework. This should
8077 * generally be called by DTrace providers in their detach(9E) entry point.
8078 */
8079 int
8080 dtrace_unregister(dtrace_provider_id_t id)
8081 {
8082 dtrace_provider_t *old = (dtrace_provider_t *)id;
8083 dtrace_provider_t *prev = NULL;
8084 int i, self = 0, noreap = 0;
8085 dtrace_probe_t *probe, *first = NULL;
8086
8087 if (old->dtpv_pops.dtps_enable ==
8088 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
8089 /*
8090 * If DTrace itself is the provider, we're called with locks
8091 * already held.
8092 */
8093 ASSERT(old == dtrace_provider);
8094 ASSERT(dtrace_devi != NULL);
8095 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8096 ASSERT(MUTEX_HELD(&dtrace_lock));
8097 self = 1;
8098
8099 if (dtrace_provider->dtpv_next != NULL) {
8100 /*
8101 * There's another provider here; return failure.
8102 */
8103 return (EBUSY);
8104 }
8105 } else {
8106 mutex_enter(&dtrace_provider_lock);
8107 mutex_enter(&mod_lock);
8108 mutex_enter(&dtrace_lock);
8109 }
8110
8111 /*
8112 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8113 * probes, we refuse to let providers slither away, unless this
8114 * provider has already been explicitly invalidated.
8115 */
8116 if (!old->dtpv_defunct &&
8117 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8118 dtrace_anon.dta_state->dts_necbs > 0))) {
8119 if (!self) {
8120 mutex_exit(&dtrace_lock);
8121 mutex_exit(&mod_lock);
8122 mutex_exit(&dtrace_provider_lock);
8123 }
8124 return (EBUSY);
8125 }
8126
8127 /*
8128 * Attempt to destroy the probes associated with this provider.
8129 */
8130 for (i = 0; i < dtrace_nprobes; i++) {
8131 if ((probe = dtrace_probes[i]) == NULL)
8132 continue;
8133
8134 if (probe->dtpr_provider != old)
8135 continue;
8136
8137 if (probe->dtpr_ecb == NULL)
8138 continue;
8139
8140 /*
8141 * If we are trying to unregister a defunct provider, and the
8142 * provider was made defunct within the interval dictated by
8143 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8144 * attempt to reap our enablings. To denote that the provider
8145 * should reattempt to unregister itself at some point in the
8146 * future, we will return a differentiable error code (EAGAIN
8147 * instead of EBUSY) in this case.
8148 */
8149 if (dtrace_gethrtime() - old->dtpv_defunct >
8150 dtrace_unregister_defunct_reap)
8151 noreap = 1;
8152
8153 if (!self) {
8154 mutex_exit(&dtrace_lock);
8155 mutex_exit(&mod_lock);
8156 mutex_exit(&dtrace_provider_lock);
8157 }
8158
8159 if (noreap)
8160 return (EBUSY);
8161
8162 (void) taskq_dispatch(dtrace_taskq,
8163 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8164
8165 return (EAGAIN);
8166 }
8167
8168 /*
8169 * All of the probes for this provider are disabled; we can safely
8170 * remove all of them from their hash chains and from the probe array.
8171 */
8172 for (i = 0; i < dtrace_nprobes; i++) {
8173 if ((probe = dtrace_probes[i]) == NULL)
8174 continue;
8175
8176 if (probe->dtpr_provider != old)
8177 continue;
8178
8179 dtrace_probes[i] = NULL;
8180
8181 dtrace_hash_remove(dtrace_bymod, probe);
8182 dtrace_hash_remove(dtrace_byfunc, probe);
8183 dtrace_hash_remove(dtrace_byname, probe);
8184
8185 if (first == NULL) {
8186 first = probe;
8187 probe->dtpr_nextmod = NULL;
8188 } else {
8189 probe->dtpr_nextmod = first;
8190 first = probe;
8191 }
8192 }
8193
8194 /*
8195 * The provider's probes have been removed from the hash chains and
8196 * from the probe array. Now issue a dtrace_sync() to be sure that
8197 * everyone has cleared out from any probe array processing.
8198 */
8199 dtrace_sync();
8200
8201 for (probe = first; probe != NULL; probe = first) {
8202 first = probe->dtpr_nextmod;
8203
8204 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8205 probe->dtpr_arg);
8206 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8207 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8208 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8209 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8210 kmem_free(probe, sizeof (dtrace_probe_t));
8211 }
8212
8213 if ((prev = dtrace_provider) == old) {
8214 ASSERT(self || dtrace_devi == NULL);
8215 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8216 dtrace_provider = old->dtpv_next;
8217 } else {
8218 while (prev != NULL && prev->dtpv_next != old)
8219 prev = prev->dtpv_next;
8220
8221 if (prev == NULL) {
8222 panic("attempt to unregister non-existent "
8223 "dtrace provider %p\n", (void *)id);
8224 }
8225
8226 prev->dtpv_next = old->dtpv_next;
8227 }
8228
8229 if (!self) {
8230 mutex_exit(&dtrace_lock);
8231 mutex_exit(&mod_lock);
8232 mutex_exit(&dtrace_provider_lock);
8233 }
8234
8235 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
8236 kmem_free(old, sizeof (dtrace_provider_t));
8237
8238 return (0);
8239 }
8240
8241 /*
8242 * Invalidate the specified provider. All subsequent probe lookups for the
8243 * specified provider will fail, but its probes will not be removed.
8244 */
8245 void
8246 dtrace_invalidate(dtrace_provider_id_t id)
8247 {
8248 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
8249
8250 ASSERT(pvp->dtpv_pops.dtps_enable !=
8251 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8252
8253 mutex_enter(&dtrace_provider_lock);
8254 mutex_enter(&dtrace_lock);
8255
8256 pvp->dtpv_defunct = dtrace_gethrtime();
8257
8258 mutex_exit(&dtrace_lock);
8259 mutex_exit(&dtrace_provider_lock);
8260 }
8261
8262 /*
8263 * Indicate whether or not DTrace has attached.
8264 */
8265 int
8266 dtrace_attached(void)
8267 {
8268 /*
8269 * dtrace_provider will be non-NULL iff the DTrace driver has
8270 * attached. (It's non-NULL because DTrace is always itself a
8271 * provider.)
8272 */
8273 return (dtrace_provider != NULL);
8274 }
8275
8276 /*
8277 * Remove all the unenabled probes for the given provider. This function is
8278 * not unlike dtrace_unregister(), except that it doesn't remove the provider
8279 * -- just as many of its associated probes as it can.
8280 */
8281 int
8282 dtrace_condense(dtrace_provider_id_t id)
8283 {
8284 dtrace_provider_t *prov = (dtrace_provider_t *)id;
8285 int i;
8286 dtrace_probe_t *probe;
8287
8288 /*
8289 * Make sure this isn't the dtrace provider itself.
8290 */
8291 ASSERT(prov->dtpv_pops.dtps_enable !=
8292 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8293
8294 mutex_enter(&dtrace_provider_lock);
8295 mutex_enter(&dtrace_lock);
8296
8297 /*
8298 * Attempt to destroy the probes associated with this provider.
8299 */
8300 for (i = 0; i < dtrace_nprobes; i++) {
8301 if ((probe = dtrace_probes[i]) == NULL)
8302 continue;
8303
8304 if (probe->dtpr_provider != prov)
8305 continue;
8306
8307 if (probe->dtpr_ecb != NULL)
8308 continue;
8309
8310 dtrace_probes[i] = NULL;
8311
8312 dtrace_hash_remove(dtrace_bymod, probe);
8313 dtrace_hash_remove(dtrace_byfunc, probe);
8314 dtrace_hash_remove(dtrace_byname, probe);
8315
8316 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
8317 probe->dtpr_arg);
8318 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8319 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8320 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8321 kmem_free(probe, sizeof (dtrace_probe_t));
8322 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
8323 }
8324
8325 mutex_exit(&dtrace_lock);
8326 mutex_exit(&dtrace_provider_lock);
8327
8328 return (0);
8329 }
8330
8331 /*
8332 * DTrace Probe Management Functions
8333 *
8334 * The functions in this section perform the DTrace probe management,
8335 * including functions to create probes, look-up probes, and call into the
8336 * providers to request that probes be provided. Some of these functions are
8337 * in the Provider-to-Framework API; these functions can be identified by the
8338 * fact that they are not declared "static".
8339 */
8340
8341 /*
8342 * Create a probe with the specified module name, function name, and name.
8343 */
8344 dtrace_id_t
8345 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
8346 const char *func, const char *name, int aframes, void *arg)
8347 {
8348 dtrace_probe_t *probe, **probes;
8349 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
8350 dtrace_id_t id;
8351
8352 if (provider == dtrace_provider) {
8353 ASSERT(MUTEX_HELD(&dtrace_lock));
8354 } else {
8355 mutex_enter(&dtrace_lock);
8356 }
8357
8358 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
8359 VM_BESTFIT | VM_SLEEP);
8360 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8361
8362 probe->dtpr_id = id;
8363 probe->dtpr_gen = dtrace_probegen++;
8364 probe->dtpr_mod = dtrace_strdup(mod);
8365 probe->dtpr_func = dtrace_strdup(func);
8366 probe->dtpr_name = dtrace_strdup(name);
8367 probe->dtpr_arg = arg;
8368 probe->dtpr_aframes = aframes;
8369 probe->dtpr_provider = provider;
8370
8371 dtrace_hash_add(dtrace_bymod, probe);
8372 dtrace_hash_add(dtrace_byfunc, probe);
8373 dtrace_hash_add(dtrace_byname, probe);
8374
8375 if (id - 1 >= dtrace_nprobes) {
8376 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8377 size_t nsize = osize << 1;
8378
8379 if (nsize == 0) {
8380 ASSERT(osize == 0);
8381 ASSERT(dtrace_probes == NULL);
8382 nsize = sizeof (dtrace_probe_t *);
8383 }
8384
8385 probes = kmem_zalloc(nsize, KM_SLEEP);
8386
8387 if (dtrace_probes == NULL) {
8388 ASSERT(osize == 0);
8389 dtrace_probes = probes;
8390 dtrace_nprobes = 1;
8391 } else {
8392 dtrace_probe_t **oprobes = dtrace_probes;
8393
8394 bcopy(oprobes, probes, osize);
8395 dtrace_membar_producer();
8396 dtrace_probes = probes;
8397
8398 dtrace_sync();
8399
8400 /*
8401 * All CPUs are now seeing the new probes array; we can
8402 * safely free the old array.
8403 */
8404 kmem_free(oprobes, osize);
8405 dtrace_nprobes <<= 1;
8406 }
8407
8408 ASSERT(id - 1 < dtrace_nprobes);
8409 }
8410
8411 ASSERT(dtrace_probes[id - 1] == NULL);
8412 dtrace_probes[id - 1] = probe;
8413
8414 if (provider != dtrace_provider)
8415 mutex_exit(&dtrace_lock);
8416
8417 return (id);
8418 }
8419
8420 static dtrace_probe_t *
8421 dtrace_probe_lookup_id(dtrace_id_t id)
8422 {
8423 ASSERT(MUTEX_HELD(&dtrace_lock));
8424
8425 if (id == 0 || id > dtrace_nprobes)
8426 return (NULL);
8427
8428 return (dtrace_probes[id - 1]);
8429 }
8430
8431 static int
8432 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8433 {
8434 *((dtrace_id_t *)arg) = probe->dtpr_id;
8435
8436 return (DTRACE_MATCH_DONE);
8437 }
8438
8439 /*
8440 * Look up a probe based on provider and one or more of module name, function
8441 * name and probe name.
8442 */
8443 dtrace_id_t
8444 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
8445 const char *func, const char *name)
8446 {
8447 dtrace_probekey_t pkey;
8448 dtrace_id_t id;
8449 int match;
8450
8451 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8452 pkey.dtpk_pmatch = &dtrace_match_string;
8453 pkey.dtpk_mod = mod;
8454 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8455 pkey.dtpk_func = func;
8456 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8457 pkey.dtpk_name = name;
8458 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8459 pkey.dtpk_id = DTRACE_IDNONE;
8460
8461 mutex_enter(&dtrace_lock);
8462 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8463 dtrace_probe_lookup_match, &id);
8464 mutex_exit(&dtrace_lock);
8465
8466 ASSERT(match == 1 || match == 0);
8467 return (match ? id : 0);
8468 }
8469
8470 /*
8471 * Returns the probe argument associated with the specified probe.
8472 */
8473 void *
8474 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8475 {
8476 dtrace_probe_t *probe;
8477 void *rval = NULL;
8478
8479 mutex_enter(&dtrace_lock);
8480
8481 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8482 probe->dtpr_provider == (dtrace_provider_t *)id)
8483 rval = probe->dtpr_arg;
8484
8485 mutex_exit(&dtrace_lock);
8486
8487 return (rval);
8488 }
8489
8490 /*
8491 * Copy a probe into a probe description.
8492 */
8493 static void
8494 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8495 {
8496 bzero(pdp, sizeof (dtrace_probedesc_t));
8497 pdp->dtpd_id = prp->dtpr_id;
8498
8499 (void) strncpy(pdp->dtpd_provider,
8500 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8501
8502 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8503 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8504 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8505 }
8506
8507 /*
8508 * Called to indicate that a probe -- or probes -- should be provided by a
8509 * specfied provider. If the specified description is NULL, the provider will
8510 * be told to provide all of its probes. (This is done whenever a new
8511 * consumer comes along, or whenever a retained enabling is to be matched.) If
8512 * the specified description is non-NULL, the provider is given the
8513 * opportunity to dynamically provide the specified probe, allowing providers
8514 * to support the creation of probes on-the-fly. (So-called _autocreated_
8515 * probes.) If the provider is NULL, the operations will be applied to all
8516 * providers; if the provider is non-NULL the operations will only be applied
8517 * to the specified provider. The dtrace_provider_lock must be held, and the
8518 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8519 * will need to grab the dtrace_lock when it reenters the framework through
8520 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8521 */
8522 static void
8523 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8524 {
8525 struct modctl *ctl;
8526 int all = 0;
8527
8528 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8529
8530 if (prv == NULL) {
8531 all = 1;
8532 prv = dtrace_provider;
8533 }
8534
8535 do {
8536 /*
8537 * First, call the blanket provide operation.
8538 */
8539 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8540
8541 /*
8542 * Now call the per-module provide operation. We will grab
8543 * mod_lock to prevent the list from being modified. Note
8544 * that this also prevents the mod_busy bits from changing.
8545 * (mod_busy can only be changed with mod_lock held.)
8546 */
8547 mutex_enter(&mod_lock);
8548
8549 ctl = &modules;
8550 do {
8551 if (ctl->mod_busy || ctl->mod_mp == NULL)
8552 continue;
8553
8554 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8555
8556 } while ((ctl = ctl->mod_next) != &modules);
8557
8558 mutex_exit(&mod_lock);
8559 } while (all && (prv = prv->dtpv_next) != NULL);
8560 }
8561
8562 /*
8563 * Iterate over each probe, and call the Framework-to-Provider API function
8564 * denoted by offs.
8565 */
8566 static void
8567 dtrace_probe_foreach(uintptr_t offs)
8568 {
8569 dtrace_provider_t *prov;
8570 void (*func)(void *, dtrace_id_t, void *);
8571 dtrace_probe_t *probe;
8572 dtrace_icookie_t cookie;
8573 int i;
8574
8575 /*
8576 * We disable interrupts to walk through the probe array. This is
8577 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8578 * won't see stale data.
8579 */
8580 cookie = dtrace_interrupt_disable();
8581
8582 for (i = 0; i < dtrace_nprobes; i++) {
8583 if ((probe = dtrace_probes[i]) == NULL)
8584 continue;
8585
8586 if (probe->dtpr_ecb == NULL) {
8587 /*
8588 * This probe isn't enabled -- don't call the function.
8589 */
8590 continue;
8591 }
8592
8593 prov = probe->dtpr_provider;
8594 func = *((void(**)(void *, dtrace_id_t, void *))
8595 ((uintptr_t)&prov->dtpv_pops + offs));
8596
8597 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8598 }
8599
8600 dtrace_interrupt_enable(cookie);
8601 }
8602
8603 static int
8604 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8605 {
8606 dtrace_probekey_t pkey;
8607 uint32_t priv;
8608 uid_t uid;
8609 zoneid_t zoneid;
8610
8611 ASSERT(MUTEX_HELD(&dtrace_lock));
8612 dtrace_ecb_create_cache = NULL;
8613
8614 if (desc == NULL) {
8615 /*
8616 * If we're passed a NULL description, we're being asked to
8617 * create an ECB with a NULL probe.
8618 */
8619 (void) dtrace_ecb_create_enable(NULL, enab);
8620 return (0);
8621 }
8622
8623 dtrace_probekey(desc, &pkey);
8624 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8625 &priv, &uid, &zoneid);
8626
8627 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8628 enab));
8629 }
8630
8631 /*
8632 * DTrace Helper Provider Functions
8633 */
8634 static void
8635 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8636 {
8637 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8638 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8639 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8640 }
8641
8642 static void
8643 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8644 const dof_provider_t *dofprov, char *strtab)
8645 {
8646 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8647 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8648 dofprov->dofpv_provattr);
8649 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8650 dofprov->dofpv_modattr);
8651 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8652 dofprov->dofpv_funcattr);
8653 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8654 dofprov->dofpv_nameattr);
8655 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8656 dofprov->dofpv_argsattr);
8657 }
8658
8659 static void
8660 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8661 {
8662 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8663 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8664 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8665 dof_provider_t *provider;
8666 dof_probe_t *probe;
8667 uint32_t *off, *enoff;
8668 uint8_t *arg;
8669 char *strtab;
8670 uint_t i, nprobes;
8671 dtrace_helper_provdesc_t dhpv;
8672 dtrace_helper_probedesc_t dhpb;
8673 dtrace_meta_t *meta = dtrace_meta_pid;
8674 dtrace_mops_t *mops = &meta->dtm_mops;
8675 void *parg;
8676
8677 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8678 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8679 provider->dofpv_strtab * dof->dofh_secsize);
8680 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8681 provider->dofpv_probes * dof->dofh_secsize);
8682 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8683 provider->dofpv_prargs * dof->dofh_secsize);
8684 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8685 provider->dofpv_proffs * dof->dofh_secsize);
8686
8687 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8688 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8689 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8690 enoff = NULL;
8691
8692 /*
8693 * See dtrace_helper_provider_validate().
8694 */
8695 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8696 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8697 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8698 provider->dofpv_prenoffs * dof->dofh_secsize);
8699 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8700 }
8701
8702 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8703
8704 /*
8705 * Create the provider.
8706 */
8707 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8708
8709 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8710 return;
8711
8712 meta->dtm_count++;
8713
8714 /*
8715 * Create the probes.
8716 */
8717 for (i = 0; i < nprobes; i++) {
8718 probe = (dof_probe_t *)(uintptr_t)(daddr +
8719 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8720
8721 dhpb.dthpb_mod = dhp->dofhp_mod;
8722 dhpb.dthpb_func = strtab + probe->dofpr_func;
8723 dhpb.dthpb_name = strtab + probe->dofpr_name;
8724 dhpb.dthpb_base = probe->dofpr_addr;
8725 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8726 dhpb.dthpb_noffs = probe->dofpr_noffs;
8727 if (enoff != NULL) {
8728 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8729 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8730 } else {
8731 dhpb.dthpb_enoffs = NULL;
8732 dhpb.dthpb_nenoffs = 0;
8733 }
8734 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8735 dhpb.dthpb_nargc = probe->dofpr_nargc;
8736 dhpb.dthpb_xargc = probe->dofpr_xargc;
8737 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8738 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8739
8740 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8741 }
8742 }
8743
8744 static void
8745 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8746 {
8747 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8748 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8749 int i;
8750
8751 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8752
8753 for (i = 0; i < dof->dofh_secnum; i++) {
8754 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8755 dof->dofh_secoff + i * dof->dofh_secsize);
8756
8757 if (sec->dofs_type != DOF_SECT_PROVIDER)
8758 continue;
8759
8760 dtrace_helper_provide_one(dhp, sec, pid);
8761 }
8762
8763 /*
8764 * We may have just created probes, so we must now rematch against
8765 * any retained enablings. Note that this call will acquire both
8766 * cpu_lock and dtrace_lock; the fact that we are holding
8767 * dtrace_meta_lock now is what defines the ordering with respect to
8768 * these three locks.
8769 */
8770 dtrace_enabling_matchall();
8771 }
8772
8773 static void
8774 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8775 {
8776 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8777 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8778 dof_sec_t *str_sec;
8779 dof_provider_t *provider;
8780 char *strtab;
8781 dtrace_helper_provdesc_t dhpv;
8782 dtrace_meta_t *meta = dtrace_meta_pid;
8783 dtrace_mops_t *mops = &meta->dtm_mops;
8784
8785 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8786 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8787 provider->dofpv_strtab * dof->dofh_secsize);
8788
8789 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8790
8791 /*
8792 * Create the provider.
8793 */
8794 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8795
8796 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8797
8798 meta->dtm_count--;
8799 }
8800
8801 static void
8802 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8803 {
8804 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8805 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8806 int i;
8807
8808 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8809
8810 for (i = 0; i < dof->dofh_secnum; i++) {
8811 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8812 dof->dofh_secoff + i * dof->dofh_secsize);
8813
8814 if (sec->dofs_type != DOF_SECT_PROVIDER)
8815 continue;
8816
8817 dtrace_helper_provider_remove_one(dhp, sec, pid);
8818 }
8819 }
8820
8821 /*
8822 * DTrace Meta Provider-to-Framework API Functions
8823 *
8824 * These functions implement the Meta Provider-to-Framework API, as described
8825 * in <sys/dtrace.h>.
8826 */
8827 int
8828 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8829 dtrace_meta_provider_id_t *idp)
8830 {
8831 dtrace_meta_t *meta;
8832 dtrace_helpers_t *help, *next;
8833 int i;
8834
8835 *idp = DTRACE_METAPROVNONE;
8836
8837 /*
8838 * We strictly don't need the name, but we hold onto it for
8839 * debuggability. All hail error queues!
8840 */
8841 if (name == NULL) {
8842 cmn_err(CE_WARN, "failed to register meta-provider: "
8843 "invalid name");
8844 return (EINVAL);
8845 }
8846
8847 if (mops == NULL ||
8848 mops->dtms_create_probe == NULL ||
8849 mops->dtms_provide_pid == NULL ||
8850 mops->dtms_remove_pid == NULL) {
8851 cmn_err(CE_WARN, "failed to register meta-register %s: "
8852 "invalid ops", name);
8853 return (EINVAL);
8854 }
8855
8856 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8857 meta->dtm_mops = *mops;
8858 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8859 (void) strcpy(meta->dtm_name, name);
8860 meta->dtm_arg = arg;
8861
8862 mutex_enter(&dtrace_meta_lock);
8863 mutex_enter(&dtrace_lock);
8864
8865 if (dtrace_meta_pid != NULL) {
8866 mutex_exit(&dtrace_lock);
8867 mutex_exit(&dtrace_meta_lock);
8868 cmn_err(CE_WARN, "failed to register meta-register %s: "
8869 "user-land meta-provider exists", name);
8870 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8871 kmem_free(meta, sizeof (dtrace_meta_t));
8872 return (EINVAL);
8873 }
8874
8875 dtrace_meta_pid = meta;
8876 *idp = (dtrace_meta_provider_id_t)meta;
8877
8878 /*
8879 * If there are providers and probes ready to go, pass them
8880 * off to the new meta provider now.
8881 */
8882
8883 help = dtrace_deferred_pid;
8884 dtrace_deferred_pid = NULL;
8885
8886 mutex_exit(&dtrace_lock);
8887
8888 while (help != NULL) {
8889 for (i = 0; i < help->dthps_nprovs; i++) {
8890 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8891 help->dthps_pid);
8892 }
8893
8894 next = help->dthps_next;
8895 help->dthps_next = NULL;
8896 help->dthps_prev = NULL;
8897 help->dthps_deferred = 0;
8898 help = next;
8899 }
8900
8901 mutex_exit(&dtrace_meta_lock);
8902
8903 return (0);
8904 }
8905
8906 int
8907 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8908 {
8909 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8910
8911 mutex_enter(&dtrace_meta_lock);
8912 mutex_enter(&dtrace_lock);
8913
8914 if (old == dtrace_meta_pid) {
8915 pp = &dtrace_meta_pid;
8916 } else {
8917 panic("attempt to unregister non-existent "
8918 "dtrace meta-provider %p\n", (void *)old);
8919 }
8920
8921 if (old->dtm_count != 0) {
8922 mutex_exit(&dtrace_lock);
8923 mutex_exit(&dtrace_meta_lock);
8924 return (EBUSY);
8925 }
8926
8927 *pp = NULL;
8928
8929 mutex_exit(&dtrace_lock);
8930 mutex_exit(&dtrace_meta_lock);
8931
8932 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8933 kmem_free(old, sizeof (dtrace_meta_t));
8934
8935 return (0);
8936 }
8937
8938
8939 /*
8940 * DTrace DIF Object Functions
8941 */
8942 static int
8943 dtrace_difo_err(uint_t pc, const char *format, ...)
8944 {
8945 if (dtrace_err_verbose) {
8946 va_list alist;
8947
8948 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8949 va_start(alist, format);
8950 (void) vuprintf(format, alist);
8951 va_end(alist);
8952 }
8953
8954 #ifdef DTRACE_ERRDEBUG
8955 dtrace_errdebug(format);
8956 #endif
8957 return (1);
8958 }
8959
8960 /*
8961 * Validate a DTrace DIF object by checking the IR instructions. The following
8962 * rules are currently enforced by dtrace_difo_validate():
8963 *
8964 * 1. Each instruction must have a valid opcode
8965 * 2. Each register, string, variable, or subroutine reference must be valid
8966 * 3. No instruction can modify register %r0 (must be zero)
8967 * 4. All instruction reserved bits must be set to zero
8968 * 5. The last instruction must be a "ret" instruction
8969 * 6. All branch targets must reference a valid instruction _after_ the branch
8970 */
8971 static int
8972 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8973 cred_t *cr)
8974 {
8975 int err = 0, i;
8976 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8977 int kcheckload;
8978 uint_t pc;
8979
8980 kcheckload = cr == NULL ||
8981 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8982
8983 dp->dtdo_destructive = 0;
8984
8985 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8986 dif_instr_t instr = dp->dtdo_buf[pc];
8987
8988 uint_t r1 = DIF_INSTR_R1(instr);
8989 uint_t r2 = DIF_INSTR_R2(instr);
8990 uint_t rd = DIF_INSTR_RD(instr);
8991 uint_t rs = DIF_INSTR_RS(instr);
8992 uint_t label = DIF_INSTR_LABEL(instr);
8993 uint_t v = DIF_INSTR_VAR(instr);
8994 uint_t subr = DIF_INSTR_SUBR(instr);
8995 uint_t type = DIF_INSTR_TYPE(instr);
8996 uint_t op = DIF_INSTR_OP(instr);
8997
8998 switch (op) {
8999 case DIF_OP_OR:
9000 case DIF_OP_XOR:
9001 case DIF_OP_AND:
9002 case DIF_OP_SLL:
9003 case DIF_OP_SRL:
9004 case DIF_OP_SRA:
9005 case DIF_OP_SUB:
9006 case DIF_OP_ADD:
9007 case DIF_OP_MUL:
9008 case DIF_OP_SDIV:
9009 case DIF_OP_UDIV:
9010 case DIF_OP_SREM:
9011 case DIF_OP_UREM:
9012 case DIF_OP_COPYS:
9013 if (r1 >= nregs)
9014 err += efunc(pc, "invalid register %u\n", r1);
9015 if (r2 >= nregs)
9016 err += efunc(pc, "invalid register %u\n", r2);
9017 if (rd >= nregs)
9018 err += efunc(pc, "invalid register %u\n", rd);
9019 if (rd == 0)
9020 err += efunc(pc, "cannot write to %r0\n");
9021 break;
9022 case DIF_OP_NOT:
9023 case DIF_OP_MOV:
9024 case DIF_OP_ALLOCS:
9025 if (r1 >= nregs)
9026 err += efunc(pc, "invalid register %u\n", r1);
9027 if (r2 != 0)
9028 err += efunc(pc, "non-zero reserved bits\n");
9029 if (rd >= nregs)
9030 err += efunc(pc, "invalid register %u\n", rd);
9031 if (rd == 0)
9032 err += efunc(pc, "cannot write to %r0\n");
9033 break;
9034 case DIF_OP_LDSB:
9035 case DIF_OP_LDSH:
9036 case DIF_OP_LDSW:
9037 case DIF_OP_LDUB:
9038 case DIF_OP_LDUH:
9039 case DIF_OP_LDUW:
9040 case DIF_OP_LDX:
9041 if (r1 >= nregs)
9042 err += efunc(pc, "invalid register %u\n", r1);
9043 if (r2 != 0)
9044 err += efunc(pc, "non-zero reserved bits\n");
9045 if (rd >= nregs)
9046 err += efunc(pc, "invalid register %u\n", rd);
9047 if (rd == 0)
9048 err += efunc(pc, "cannot write to %r0\n");
9049 if (kcheckload)
9050 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9051 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9052 break;
9053 case DIF_OP_RLDSB:
9054 case DIF_OP_RLDSH:
9055 case DIF_OP_RLDSW:
9056 case DIF_OP_RLDUB:
9057 case DIF_OP_RLDUH:
9058 case DIF_OP_RLDUW:
9059 case DIF_OP_RLDX:
9060 if (r1 >= nregs)
9061 err += efunc(pc, "invalid register %u\n", r1);
9062 if (r2 != 0)
9063 err += efunc(pc, "non-zero reserved bits\n");
9064 if (rd >= nregs)
9065 err += efunc(pc, "invalid register %u\n", rd);
9066 if (rd == 0)
9067 err += efunc(pc, "cannot write to %r0\n");
9068 break;
9069 case DIF_OP_ULDSB:
9070 case DIF_OP_ULDSH:
9071 case DIF_OP_ULDSW:
9072 case DIF_OP_ULDUB:
9073 case DIF_OP_ULDUH:
9074 case DIF_OP_ULDUW:
9075 case DIF_OP_ULDX:
9076 if (r1 >= nregs)
9077 err += efunc(pc, "invalid register %u\n", r1);
9078 if (r2 != 0)
9079 err += efunc(pc, "non-zero reserved bits\n");
9080 if (rd >= nregs)
9081 err += efunc(pc, "invalid register %u\n", rd);
9082 if (rd == 0)
9083 err += efunc(pc, "cannot write to %r0\n");
9084 break;
9085 case DIF_OP_STB:
9086 case DIF_OP_STH:
9087 case DIF_OP_STW:
9088 case DIF_OP_STX:
9089 if (r1 >= nregs)
9090 err += efunc(pc, "invalid register %u\n", r1);
9091 if (r2 != 0)
9092 err += efunc(pc, "non-zero reserved bits\n");
9093 if (rd >= nregs)
9094 err += efunc(pc, "invalid register %u\n", rd);
9095 if (rd == 0)
9096 err += efunc(pc, "cannot write to 0 address\n");
9097 break;
9098 case DIF_OP_CMP:
9099 case DIF_OP_SCMP:
9100 if (r1 >= nregs)
9101 err += efunc(pc, "invalid register %u\n", r1);
9102 if (r2 >= nregs)
9103 err += efunc(pc, "invalid register %u\n", r2);
9104 if (rd != 0)
9105 err += efunc(pc, "non-zero reserved bits\n");
9106 break;
9107 case DIF_OP_TST:
9108 if (r1 >= nregs)
9109 err += efunc(pc, "invalid register %u\n", r1);
9110 if (r2 != 0 || rd != 0)
9111 err += efunc(pc, "non-zero reserved bits\n");
9112 break;
9113 case DIF_OP_BA:
9114 case DIF_OP_BE:
9115 case DIF_OP_BNE:
9116 case DIF_OP_BG:
9117 case DIF_OP_BGU:
9118 case DIF_OP_BGE:
9119 case DIF_OP_BGEU:
9120 case DIF_OP_BL:
9121 case DIF_OP_BLU:
9122 case DIF_OP_BLE:
9123 case DIF_OP_BLEU:
9124 if (label >= dp->dtdo_len) {
9125 err += efunc(pc, "invalid branch target %u\n",
9126 label);
9127 }
9128 if (label <= pc) {
9129 err += efunc(pc, "backward branch to %u\n",
9130 label);
9131 }
9132 break;
9133 case DIF_OP_RET:
9134 if (r1 != 0 || r2 != 0)
9135 err += efunc(pc, "non-zero reserved bits\n");
9136 if (rd >= nregs)
9137 err += efunc(pc, "invalid register %u\n", rd);
9138 break;
9139 case DIF_OP_NOP:
9140 case DIF_OP_POPTS:
9141 case DIF_OP_FLUSHTS:
9142 if (r1 != 0 || r2 != 0 || rd != 0)
9143 err += efunc(pc, "non-zero reserved bits\n");
9144 break;
9145 case DIF_OP_SETX:
9146 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9147 err += efunc(pc, "invalid integer ref %u\n",
9148 DIF_INSTR_INTEGER(instr));
9149 }
9150 if (rd >= nregs)
9151 err += efunc(pc, "invalid register %u\n", rd);
9152 if (rd == 0)
9153 err += efunc(pc, "cannot write to %r0\n");
9154 break;
9155 case DIF_OP_SETS:
9156 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9157 err += efunc(pc, "invalid string ref %u\n",
9158 DIF_INSTR_STRING(instr));
9159 }
9160 if (rd >= nregs)
9161 err += efunc(pc, "invalid register %u\n", rd);
9162 if (rd == 0)
9163 err += efunc(pc, "cannot write to %r0\n");
9164 break;
9165 case DIF_OP_LDGA:
9166 case DIF_OP_LDTA:
9167 if (r1 > DIF_VAR_ARRAY_MAX)
9168 err += efunc(pc, "invalid array %u\n", r1);
9169 if (r2 >= nregs)
9170 err += efunc(pc, "invalid register %u\n", r2);
9171 if (rd >= nregs)
9172 err += efunc(pc, "invalid register %u\n", rd);
9173 if (rd == 0)
9174 err += efunc(pc, "cannot write to %r0\n");
9175 break;
9176 case DIF_OP_LDGS:
9177 case DIF_OP_LDTS:
9178 case DIF_OP_LDLS:
9179 case DIF_OP_LDGAA:
9180 case DIF_OP_LDTAA:
9181 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9182 err += efunc(pc, "invalid variable %u\n", v);
9183 if (rd >= nregs)
9184 err += efunc(pc, "invalid register %u\n", rd);
9185 if (rd == 0)
9186 err += efunc(pc, "cannot write to %r0\n");
9187 break;
9188 case DIF_OP_STGS:
9189 case DIF_OP_STTS:
9190 case DIF_OP_STLS:
9191 case DIF_OP_STGAA:
9192 case DIF_OP_STTAA:
9193 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9194 err += efunc(pc, "invalid variable %u\n", v);
9195 if (rs >= nregs)
9196 err += efunc(pc, "invalid register %u\n", rd);
9197 break;
9198 case DIF_OP_CALL:
9199 if (subr > DIF_SUBR_MAX)
9200 err += efunc(pc, "invalid subr %u\n", subr);
9201 if (rd >= nregs)
9202 err += efunc(pc, "invalid register %u\n", rd);
9203 if (rd == 0)
9204 err += efunc(pc, "cannot write to %r0\n");
9205
9206 if (subr == DIF_SUBR_COPYOUT ||
9207 subr == DIF_SUBR_COPYOUTSTR) {
9208 dp->dtdo_destructive = 1;
9209 }
9210
9211 if (subr == DIF_SUBR_GETF) {
9212 /*
9213 * If we have a getf() we need to record that
9214 * in our state. Note that our state can be
9215 * NULL if this is a helper -- but in that
9216 * case, the call to getf() is itself illegal,
9217 * and will be caught (slightly later) when
9218 * the helper is validated.
9219 */
9220 if (vstate->dtvs_state != NULL)
9221 vstate->dtvs_state->dts_getf++;
9222 }
9223
9224 break;
9225 case DIF_OP_PUSHTR:
9226 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
9227 err += efunc(pc, "invalid ref type %u\n", type);
9228 if (r2 >= nregs)
9229 err += efunc(pc, "invalid register %u\n", r2);
9230 if (rs >= nregs)
9231 err += efunc(pc, "invalid register %u\n", rs);
9232 break;
9233 case DIF_OP_PUSHTV:
9234 if (type != DIF_TYPE_CTF)
9235 err += efunc(pc, "invalid val type %u\n", type);
9236 if (r2 >= nregs)
9237 err += efunc(pc, "invalid register %u\n", r2);
9238 if (rs >= nregs)
9239 err += efunc(pc, "invalid register %u\n", rs);
9240 break;
9241 default:
9242 err += efunc(pc, "invalid opcode %u\n",
9243 DIF_INSTR_OP(instr));
9244 }
9245 }
9246
9247 if (dp->dtdo_len != 0 &&
9248 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
9249 err += efunc(dp->dtdo_len - 1,
9250 "expected 'ret' as last DIF instruction\n");
9251 }
9252
9253 if (!(dp->dtdo_rtype.dtdt_flags & (DIF_TF_BYREF | DIF_TF_BYUREF))) {
9254 /*
9255 * If we're not returning by reference, the size must be either
9256 * 0 or the size of one of the base types.
9257 */
9258 switch (dp->dtdo_rtype.dtdt_size) {
9259 case 0:
9260 case sizeof (uint8_t):
9261 case sizeof (uint16_t):
9262 case sizeof (uint32_t):
9263 case sizeof (uint64_t):
9264 break;
9265
9266 default:
9267 err += efunc(dp->dtdo_len - 1, "bad return size\n");
9268 }
9269 }
9270
9271 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
9272 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
9273 dtrace_diftype_t *vt, *et;
9274 uint_t id, ndx;
9275
9276 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
9277 v->dtdv_scope != DIFV_SCOPE_THREAD &&
9278 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
9279 err += efunc(i, "unrecognized variable scope %d\n",
9280 v->dtdv_scope);
9281 break;
9282 }
9283
9284 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
9285 v->dtdv_kind != DIFV_KIND_SCALAR) {
9286 err += efunc(i, "unrecognized variable type %d\n",
9287 v->dtdv_kind);
9288 break;
9289 }
9290
9291 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
9292 err += efunc(i, "%d exceeds variable id limit\n", id);
9293 break;
9294 }
9295
9296 if (id < DIF_VAR_OTHER_UBASE)
9297 continue;
9298
9299 /*
9300 * For user-defined variables, we need to check that this
9301 * definition is identical to any previous definition that we
9302 * encountered.
9303 */
9304 ndx = id - DIF_VAR_OTHER_UBASE;
9305
9306 switch (v->dtdv_scope) {
9307 case DIFV_SCOPE_GLOBAL:
9308 if (ndx < vstate->dtvs_nglobals) {
9309 dtrace_statvar_t *svar;
9310
9311 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
9312 existing = &svar->dtsv_var;
9313 }
9314
9315 break;
9316
9317 case DIFV_SCOPE_THREAD:
9318 if (ndx < vstate->dtvs_ntlocals)
9319 existing = &vstate->dtvs_tlocals[ndx];
9320 break;
9321
9322 case DIFV_SCOPE_LOCAL:
9323 if (ndx < vstate->dtvs_nlocals) {
9324 dtrace_statvar_t *svar;
9325
9326 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
9327 existing = &svar->dtsv_var;
9328 }
9329
9330 break;
9331 }
9332
9333 vt = &v->dtdv_type;
9334
9335 if (vt->dtdt_flags & DIF_TF_BYREF) {
9336 if (vt->dtdt_size == 0) {
9337 err += efunc(i, "zero-sized variable\n");
9338 break;
9339 }
9340
9341 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
9342 vt->dtdt_size > dtrace_global_maxsize) {
9343 err += efunc(i, "oversized by-ref global\n");
9344 break;
9345 }
9346 }
9347
9348 if (existing == NULL || existing->dtdv_id == 0)
9349 continue;
9350
9351 ASSERT(existing->dtdv_id == v->dtdv_id);
9352 ASSERT(existing->dtdv_scope == v->dtdv_scope);
9353
9354 if (existing->dtdv_kind != v->dtdv_kind)
9355 err += efunc(i, "%d changed variable kind\n", id);
9356
9357 et = &existing->dtdv_type;
9358
9359 if (vt->dtdt_flags != et->dtdt_flags) {
9360 err += efunc(i, "%d changed variable type flags\n", id);
9361 break;
9362 }
9363
9364 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
9365 err += efunc(i, "%d changed variable type size\n", id);
9366 break;
9367 }
9368 }
9369
9370 return (err);
9371 }
9372
9373 /*
9374 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
9375 * are much more constrained than normal DIFOs. Specifically, they may
9376 * not:
9377 *
9378 * 1. Make calls to subroutines other than copyin(), copyinstr() or
9379 * miscellaneous string routines
9380 * 2. Access DTrace variables other than the args[] array, and the
9381 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9382 * 3. Have thread-local variables.
9383 * 4. Have dynamic variables.
9384 */
9385 static int
9386 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9387 {
9388 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9389 int err = 0;
9390 uint_t pc;
9391
9392 for (pc = 0; pc < dp->dtdo_len; pc++) {
9393 dif_instr_t instr = dp->dtdo_buf[pc];
9394
9395 uint_t v = DIF_INSTR_VAR(instr);
9396 uint_t subr = DIF_INSTR_SUBR(instr);
9397 uint_t op = DIF_INSTR_OP(instr);
9398
9399 switch (op) {
9400 case DIF_OP_OR:
9401 case DIF_OP_XOR:
9402 case DIF_OP_AND:
9403 case DIF_OP_SLL:
9404 case DIF_OP_SRL:
9405 case DIF_OP_SRA:
9406 case DIF_OP_SUB:
9407 case DIF_OP_ADD:
9408 case DIF_OP_MUL:
9409 case DIF_OP_SDIV:
9410 case DIF_OP_UDIV:
9411 case DIF_OP_SREM:
9412 case DIF_OP_UREM:
9413 case DIF_OP_COPYS:
9414 case DIF_OP_NOT:
9415 case DIF_OP_MOV:
9416 case DIF_OP_RLDSB:
9417 case DIF_OP_RLDSH:
9418 case DIF_OP_RLDSW:
9419 case DIF_OP_RLDUB:
9420 case DIF_OP_RLDUH:
9421 case DIF_OP_RLDUW:
9422 case DIF_OP_RLDX:
9423 case DIF_OP_ULDSB:
9424 case DIF_OP_ULDSH:
9425 case DIF_OP_ULDSW:
9426 case DIF_OP_ULDUB:
9427 case DIF_OP_ULDUH:
9428 case DIF_OP_ULDUW:
9429 case DIF_OP_ULDX:
9430 case DIF_OP_STB:
9431 case DIF_OP_STH:
9432 case DIF_OP_STW:
9433 case DIF_OP_STX:
9434 case DIF_OP_ALLOCS:
9435 case DIF_OP_CMP:
9436 case DIF_OP_SCMP:
9437 case DIF_OP_TST:
9438 case DIF_OP_BA:
9439 case DIF_OP_BE:
9440 case DIF_OP_BNE:
9441 case DIF_OP_BG:
9442 case DIF_OP_BGU:
9443 case DIF_OP_BGE:
9444 case DIF_OP_BGEU:
9445 case DIF_OP_BL:
9446 case DIF_OP_BLU:
9447 case DIF_OP_BLE:
9448 case DIF_OP_BLEU:
9449 case DIF_OP_RET:
9450 case DIF_OP_NOP:
9451 case DIF_OP_POPTS:
9452 case DIF_OP_FLUSHTS:
9453 case DIF_OP_SETX:
9454 case DIF_OP_SETS:
9455 case DIF_OP_LDGA:
9456 case DIF_OP_LDLS:
9457 case DIF_OP_STGS:
9458 case DIF_OP_STLS:
9459 case DIF_OP_PUSHTR:
9460 case DIF_OP_PUSHTV:
9461 break;
9462
9463 case DIF_OP_LDGS:
9464 if (v >= DIF_VAR_OTHER_UBASE)
9465 break;
9466
9467 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9468 break;
9469
9470 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9471 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9472 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9473 v == DIF_VAR_UID || v == DIF_VAR_GID)
9474 break;
9475
9476 err += efunc(pc, "illegal variable %u\n", v);
9477 break;
9478
9479 case DIF_OP_LDTA:
9480 case DIF_OP_LDTS:
9481 case DIF_OP_LDGAA:
9482 case DIF_OP_LDTAA:
9483 err += efunc(pc, "illegal dynamic variable load\n");
9484 break;
9485
9486 case DIF_OP_STTS:
9487 case DIF_OP_STGAA:
9488 case DIF_OP_STTAA:
9489 err += efunc(pc, "illegal dynamic variable store\n");
9490 break;
9491
9492 case DIF_OP_CALL:
9493 if (subr == DIF_SUBR_ALLOCA ||
9494 subr == DIF_SUBR_BCOPY ||
9495 subr == DIF_SUBR_COPYIN ||
9496 subr == DIF_SUBR_COPYINTO ||
9497 subr == DIF_SUBR_COPYINSTR ||
9498 subr == DIF_SUBR_INDEX ||
9499 subr == DIF_SUBR_INET_NTOA ||
9500 subr == DIF_SUBR_INET_NTOA6 ||
9501 subr == DIF_SUBR_INET_NTOP ||
9502 subr == DIF_SUBR_JSON ||
9503 subr == DIF_SUBR_LLTOSTR ||
9504 subr == DIF_SUBR_STRTOLL ||
9505 subr == DIF_SUBR_RINDEX ||
9506 subr == DIF_SUBR_STRCHR ||
9507 subr == DIF_SUBR_STRJOIN ||
9508 subr == DIF_SUBR_STRRCHR ||
9509 subr == DIF_SUBR_STRSTR ||
9510 subr == DIF_SUBR_HTONS ||
9511 subr == DIF_SUBR_HTONL ||
9512 subr == DIF_SUBR_HTONLL ||
9513 subr == DIF_SUBR_NTOHS ||
9514 subr == DIF_SUBR_NTOHL ||
9515 subr == DIF_SUBR_NTOHLL)
9516 break;
9517
9518 err += efunc(pc, "invalid subr %u\n", subr);
9519 break;
9520
9521 default:
9522 err += efunc(pc, "invalid opcode %u\n",
9523 DIF_INSTR_OP(instr));
9524 }
9525 }
9526
9527 return (err);
9528 }
9529
9530 /*
9531 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9532 * basis; 0 if not.
9533 */
9534 static int
9535 dtrace_difo_cacheable(dtrace_difo_t *dp)
9536 {
9537 int i;
9538
9539 if (dp == NULL)
9540 return (0);
9541
9542 for (i = 0; i < dp->dtdo_varlen; i++) {
9543 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9544
9545 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9546 continue;
9547
9548 switch (v->dtdv_id) {
9549 case DIF_VAR_CURTHREAD:
9550 case DIF_VAR_PID:
9551 case DIF_VAR_TID:
9552 case DIF_VAR_EXECNAME:
9553 case DIF_VAR_ZONENAME:
9554 break;
9555
9556 default:
9557 return (0);
9558 }
9559 }
9560
9561 /*
9562 * This DIF object may be cacheable. Now we need to look for any
9563 * array loading instructions, any memory loading instructions, or
9564 * any stores to thread-local variables.
9565 */
9566 for (i = 0; i < dp->dtdo_len; i++) {
9567 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9568
9569 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9570 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9571 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9572 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9573 return (0);
9574 }
9575
9576 return (1);
9577 }
9578
9579 static void
9580 dtrace_difo_hold(dtrace_difo_t *dp)
9581 {
9582 int i;
9583
9584 ASSERT(MUTEX_HELD(&dtrace_lock));
9585
9586 dp->dtdo_refcnt++;
9587 ASSERT(dp->dtdo_refcnt != 0);
9588
9589 /*
9590 * We need to check this DIF object for references to the variable
9591 * DIF_VAR_VTIMESTAMP.
9592 */
9593 for (i = 0; i < dp->dtdo_varlen; i++) {
9594 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9595
9596 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9597 continue;
9598
9599 if (dtrace_vtime_references++ == 0)
9600 dtrace_vtime_enable();
9601 }
9602 }
9603
9604 /*
9605 * This routine calculates the dynamic variable chunksize for a given DIF
9606 * object. The calculation is not fool-proof, and can probably be tricked by
9607 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9608 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9609 * if a dynamic variable size exceeds the chunksize.
9610 */
9611 static void
9612 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9613 {
9614 uint64_t sval;
9615 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9616 const dif_instr_t *text = dp->dtdo_buf;
9617 uint_t pc, srd = 0;
9618 uint_t ttop = 0;
9619 size_t size, ksize;
9620 uint_t id, i;
9621
9622 for (pc = 0; pc < dp->dtdo_len; pc++) {
9623 dif_instr_t instr = text[pc];
9624 uint_t op = DIF_INSTR_OP(instr);
9625 uint_t rd = DIF_INSTR_RD(instr);
9626 uint_t r1 = DIF_INSTR_R1(instr);
9627 uint_t nkeys = 0;
9628 uchar_t scope;
9629
9630 dtrace_key_t *key = tupregs;
9631
9632 switch (op) {
9633 case DIF_OP_SETX:
9634 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9635 srd = rd;
9636 continue;
9637
9638 case DIF_OP_STTS:
9639 key = &tupregs[DIF_DTR_NREGS];
9640 key[0].dttk_size = 0;
9641 key[1].dttk_size = 0;
9642 nkeys = 2;
9643 scope = DIFV_SCOPE_THREAD;
9644 break;
9645
9646 case DIF_OP_STGAA:
9647 case DIF_OP_STTAA:
9648 nkeys = ttop;
9649
9650 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9651 key[nkeys++].dttk_size = 0;
9652
9653 key[nkeys++].dttk_size = 0;
9654
9655 if (op == DIF_OP_STTAA) {
9656 scope = DIFV_SCOPE_THREAD;
9657 } else {
9658 scope = DIFV_SCOPE_GLOBAL;
9659 }
9660
9661 break;
9662
9663 case DIF_OP_PUSHTR:
9664 if (ttop == DIF_DTR_NREGS)
9665 return;
9666
9667 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9668 /*
9669 * If the register for the size of the "pushtr"
9670 * is %r0 (or the value is 0) and the type is
9671 * a string, we'll use the system-wide default
9672 * string size.
9673 */
9674 tupregs[ttop++].dttk_size =
9675 dtrace_strsize_default;
9676 } else {
9677 if (srd == 0)
9678 return;
9679
9680 tupregs[ttop++].dttk_size = sval;
9681 }
9682
9683 break;
9684
9685 case DIF_OP_PUSHTV:
9686 if (ttop == DIF_DTR_NREGS)
9687 return;
9688
9689 tupregs[ttop++].dttk_size = 0;
9690 break;
9691
9692 case DIF_OP_FLUSHTS:
9693 ttop = 0;
9694 break;
9695
9696 case DIF_OP_POPTS:
9697 if (ttop != 0)
9698 ttop--;
9699 break;
9700 }
9701
9702 sval = 0;
9703 srd = 0;
9704
9705 if (nkeys == 0)
9706 continue;
9707
9708 /*
9709 * We have a dynamic variable allocation; calculate its size.
9710 */
9711 for (ksize = 0, i = 0; i < nkeys; i++)
9712 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9713
9714 size = sizeof (dtrace_dynvar_t);
9715 size += sizeof (dtrace_key_t) * (nkeys - 1);
9716 size += ksize;
9717
9718 /*
9719 * Now we need to determine the size of the stored data.
9720 */
9721 id = DIF_INSTR_VAR(instr);
9722
9723 for (i = 0; i < dp->dtdo_varlen; i++) {
9724 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9725
9726 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9727 size += v->dtdv_type.dtdt_size;
9728 break;
9729 }
9730 }
9731
9732 if (i == dp->dtdo_varlen)
9733 return;
9734
9735 /*
9736 * We have the size. If this is larger than the chunk size
9737 * for our dynamic variable state, reset the chunk size.
9738 */
9739 size = P2ROUNDUP(size, sizeof (uint64_t));
9740
9741 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9742 vstate->dtvs_dynvars.dtds_chunksize = size;
9743 }
9744 }
9745
9746 static void
9747 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9748 {
9749 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9750 uint_t id;
9751
9752 ASSERT(MUTEX_HELD(&dtrace_lock));
9753 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9754
9755 for (i = 0; i < dp->dtdo_varlen; i++) {
9756 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9757 dtrace_statvar_t *svar, ***svarp;
9758 size_t dsize = 0;
9759 uint8_t scope = v->dtdv_scope;
9760 int *np;
9761
9762 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9763 continue;
9764
9765 id -= DIF_VAR_OTHER_UBASE;
9766
9767 switch (scope) {
9768 case DIFV_SCOPE_THREAD:
9769 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9770 dtrace_difv_t *tlocals;
9771
9772 if ((ntlocals = (otlocals << 1)) == 0)
9773 ntlocals = 1;
9774
9775 osz = otlocals * sizeof (dtrace_difv_t);
9776 nsz = ntlocals * sizeof (dtrace_difv_t);
9777
9778 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9779
9780 if (osz != 0) {
9781 bcopy(vstate->dtvs_tlocals,
9782 tlocals, osz);
9783 kmem_free(vstate->dtvs_tlocals, osz);
9784 }
9785
9786 vstate->dtvs_tlocals = tlocals;
9787 vstate->dtvs_ntlocals = ntlocals;
9788 }
9789
9790 vstate->dtvs_tlocals[id] = *v;
9791 continue;
9792
9793 case DIFV_SCOPE_LOCAL:
9794 np = &vstate->dtvs_nlocals;
9795 svarp = &vstate->dtvs_locals;
9796
9797 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9798 dsize = NCPU * (v->dtdv_type.dtdt_size +
9799 sizeof (uint64_t));
9800 else
9801 dsize = NCPU * sizeof (uint64_t);
9802
9803 break;
9804
9805 case DIFV_SCOPE_GLOBAL:
9806 np = &vstate->dtvs_nglobals;
9807 svarp = &vstate->dtvs_globals;
9808
9809 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9810 dsize = v->dtdv_type.dtdt_size +
9811 sizeof (uint64_t);
9812
9813 break;
9814
9815 default:
9816 ASSERT(0);
9817 }
9818
9819 while (id >= (oldsvars = *np)) {
9820 dtrace_statvar_t **statics;
9821 int newsvars, oldsize, newsize;
9822
9823 if ((newsvars = (oldsvars << 1)) == 0)
9824 newsvars = 1;
9825
9826 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9827 newsize = newsvars * sizeof (dtrace_statvar_t *);
9828
9829 statics = kmem_zalloc(newsize, KM_SLEEP);
9830
9831 if (oldsize != 0) {
9832 bcopy(*svarp, statics, oldsize);
9833 kmem_free(*svarp, oldsize);
9834 }
9835
9836 *svarp = statics;
9837 *np = newsvars;
9838 }
9839
9840 if ((svar = (*svarp)[id]) == NULL) {
9841 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9842 svar->dtsv_var = *v;
9843
9844 if ((svar->dtsv_size = dsize) != 0) {
9845 svar->dtsv_data = (uint64_t)(uintptr_t)
9846 kmem_zalloc(dsize, KM_SLEEP);
9847 }
9848
9849 (*svarp)[id] = svar;
9850 }
9851
9852 svar->dtsv_refcnt++;
9853 }
9854
9855 dtrace_difo_chunksize(dp, vstate);
9856 dtrace_difo_hold(dp);
9857 }
9858
9859 static dtrace_difo_t *
9860 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9861 {
9862 dtrace_difo_t *new;
9863 size_t sz;
9864
9865 ASSERT(dp->dtdo_buf != NULL);
9866 ASSERT(dp->dtdo_refcnt != 0);
9867
9868 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9869
9870 ASSERT(dp->dtdo_buf != NULL);
9871 sz = dp->dtdo_len * sizeof (dif_instr_t);
9872 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9873 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9874 new->dtdo_len = dp->dtdo_len;
9875
9876 if (dp->dtdo_strtab != NULL) {
9877 ASSERT(dp->dtdo_strlen != 0);
9878 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9879 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9880 new->dtdo_strlen = dp->dtdo_strlen;
9881 }
9882
9883 if (dp->dtdo_inttab != NULL) {
9884 ASSERT(dp->dtdo_intlen != 0);
9885 sz = dp->dtdo_intlen * sizeof (uint64_t);
9886 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9887 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9888 new->dtdo_intlen = dp->dtdo_intlen;
9889 }
9890
9891 if (dp->dtdo_vartab != NULL) {
9892 ASSERT(dp->dtdo_varlen != 0);
9893 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9894 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9895 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9896 new->dtdo_varlen = dp->dtdo_varlen;
9897 }
9898
9899 dtrace_difo_init(new, vstate);
9900 return (new);
9901 }
9902
9903 static void
9904 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9905 {
9906 int i;
9907
9908 ASSERT(dp->dtdo_refcnt == 0);
9909
9910 for (i = 0; i < dp->dtdo_varlen; i++) {
9911 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9912 dtrace_statvar_t *svar, **svarp;
9913 uint_t id;
9914 uint8_t scope = v->dtdv_scope;
9915 int *np;
9916
9917 switch (scope) {
9918 case DIFV_SCOPE_THREAD:
9919 continue;
9920
9921 case DIFV_SCOPE_LOCAL:
9922 np = &vstate->dtvs_nlocals;
9923 svarp = vstate->dtvs_locals;
9924 break;
9925
9926 case DIFV_SCOPE_GLOBAL:
9927 np = &vstate->dtvs_nglobals;
9928 svarp = vstate->dtvs_globals;
9929 break;
9930
9931 default:
9932 ASSERT(0);
9933 }
9934
9935 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9936 continue;
9937
9938 id -= DIF_VAR_OTHER_UBASE;
9939 ASSERT(id < *np);
9940
9941 svar = svarp[id];
9942 ASSERT(svar != NULL);
9943 ASSERT(svar->dtsv_refcnt > 0);
9944
9945 if (--svar->dtsv_refcnt > 0)
9946 continue;
9947
9948 if (svar->dtsv_size != 0) {
9949 ASSERT(svar->dtsv_data != NULL);
9950 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9951 svar->dtsv_size);
9952 }
9953
9954 kmem_free(svar, sizeof (dtrace_statvar_t));
9955 svarp[id] = NULL;
9956 }
9957
9958 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9959 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9960 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9961 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9962
9963 kmem_free(dp, sizeof (dtrace_difo_t));
9964 }
9965
9966 static void
9967 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9968 {
9969 int i;
9970
9971 ASSERT(MUTEX_HELD(&dtrace_lock));
9972 ASSERT(dp->dtdo_refcnt != 0);
9973
9974 for (i = 0; i < dp->dtdo_varlen; i++) {
9975 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9976
9977 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9978 continue;
9979
9980 ASSERT(dtrace_vtime_references > 0);
9981 if (--dtrace_vtime_references == 0)
9982 dtrace_vtime_disable();
9983 }
9984
9985 if (--dp->dtdo_refcnt == 0)
9986 dtrace_difo_destroy(dp, vstate);
9987 }
9988
9989 /*
9990 * DTrace Format Functions
9991 */
9992 static uint16_t
9993 dtrace_format_add(dtrace_state_t *state, char *str)
9994 {
9995 char *fmt, **new;
9996 uint16_t ndx, len = strlen(str) + 1;
9997
9998 fmt = kmem_zalloc(len, KM_SLEEP);
9999 bcopy(str, fmt, len);
10000
10001 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
10002 if (state->dts_formats[ndx] == NULL) {
10003 state->dts_formats[ndx] = fmt;
10004 return (ndx + 1);
10005 }
10006 }
10007
10008 if (state->dts_nformats == USHRT_MAX) {
10009 /*
10010 * This is only likely if a denial-of-service attack is being
10011 * attempted. As such, it's okay to fail silently here.
10012 */
10013 kmem_free(fmt, len);
10014 return (0);
10015 }
10016
10017 /*
10018 * For simplicity, we always resize the formats array to be exactly the
10019 * number of formats.
10020 */
10021 ndx = state->dts_nformats++;
10022 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10023
10024 if (state->dts_formats != NULL) {
10025 ASSERT(ndx != 0);
10026 bcopy(state->dts_formats, new, ndx * sizeof (char *));
10027 kmem_free(state->dts_formats, ndx * sizeof (char *));
10028 }
10029
10030 state->dts_formats = new;
10031 state->dts_formats[ndx] = fmt;
10032
10033 return (ndx + 1);
10034 }
10035
10036 static void
10037 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10038 {
10039 char *fmt;
10040
10041 ASSERT(state->dts_formats != NULL);
10042 ASSERT(format <= state->dts_nformats);
10043 ASSERT(state->dts_formats[format - 1] != NULL);
10044
10045 fmt = state->dts_formats[format - 1];
10046 kmem_free(fmt, strlen(fmt) + 1);
10047 state->dts_formats[format - 1] = NULL;
10048 }
10049
10050 static void
10051 dtrace_format_destroy(dtrace_state_t *state)
10052 {
10053 int i;
10054
10055 if (state->dts_nformats == 0) {
10056 ASSERT(state->dts_formats == NULL);
10057 return;
10058 }
10059
10060 ASSERT(state->dts_formats != NULL);
10061
10062 for (i = 0; i < state->dts_nformats; i++) {
10063 char *fmt = state->dts_formats[i];
10064
10065 if (fmt == NULL)
10066 continue;
10067
10068 kmem_free(fmt, strlen(fmt) + 1);
10069 }
10070
10071 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10072 state->dts_nformats = 0;
10073 state->dts_formats = NULL;
10074 }
10075
10076 /*
10077 * DTrace Predicate Functions
10078 */
10079 static dtrace_predicate_t *
10080 dtrace_predicate_create(dtrace_difo_t *dp)
10081 {
10082 dtrace_predicate_t *pred;
10083
10084 ASSERT(MUTEX_HELD(&dtrace_lock));
10085 ASSERT(dp->dtdo_refcnt != 0);
10086
10087 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10088 pred->dtp_difo = dp;
10089 pred->dtp_refcnt = 1;
10090
10091 if (!dtrace_difo_cacheable(dp))
10092 return (pred);
10093
10094 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10095 /*
10096 * This is only theoretically possible -- we have had 2^32
10097 * cacheable predicates on this machine. We cannot allow any
10098 * more predicates to become cacheable: as unlikely as it is,
10099 * there may be a thread caching a (now stale) predicate cache
10100 * ID. (N.B.: the temptation is being successfully resisted to
10101 * have this cmn_err() "Holy shit -- we executed this code!")
10102 */
10103 return (pred);
10104 }
10105
10106 pred->dtp_cacheid = dtrace_predcache_id++;
10107
10108 return (pred);
10109 }
10110
10111 static void
10112 dtrace_predicate_hold(dtrace_predicate_t *pred)
10113 {
10114 ASSERT(MUTEX_HELD(&dtrace_lock));
10115 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10116 ASSERT(pred->dtp_refcnt > 0);
10117
10118 pred->dtp_refcnt++;
10119 }
10120
10121 static void
10122 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10123 {
10124 dtrace_difo_t *dp = pred->dtp_difo;
10125
10126 ASSERT(MUTEX_HELD(&dtrace_lock));
10127 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
10128 ASSERT(pred->dtp_refcnt > 0);
10129
10130 if (--pred->dtp_refcnt == 0) {
10131 dtrace_difo_release(pred->dtp_difo, vstate);
10132 kmem_free(pred, sizeof (dtrace_predicate_t));
10133 }
10134 }
10135
10136 /*
10137 * DTrace Action Description Functions
10138 */
10139 static dtrace_actdesc_t *
10140 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
10141 uint64_t uarg, uint64_t arg)
10142 {
10143 dtrace_actdesc_t *act;
10144
10145 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
10146 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
10147
10148 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
10149 act->dtad_kind = kind;
10150 act->dtad_ntuple = ntuple;
10151 act->dtad_uarg = uarg;
10152 act->dtad_arg = arg;
10153 act->dtad_refcnt = 1;
10154
10155 return (act);
10156 }
10157
10158 static void
10159 dtrace_actdesc_hold(dtrace_actdesc_t *act)
10160 {
10161 ASSERT(act->dtad_refcnt >= 1);
10162 act->dtad_refcnt++;
10163 }
10164
10165 static void
10166 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
10167 {
10168 dtrace_actkind_t kind = act->dtad_kind;
10169 dtrace_difo_t *dp;
10170
10171 ASSERT(act->dtad_refcnt >= 1);
10172
10173 if (--act->dtad_refcnt != 0)
10174 return;
10175
10176 if ((dp = act->dtad_difo) != NULL)
10177 dtrace_difo_release(dp, vstate);
10178
10179 if (DTRACEACT_ISPRINTFLIKE(kind)) {
10180 char *str = (char *)(uintptr_t)act->dtad_arg;
10181
10182 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
10183 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
10184
10185 if (str != NULL)
10186 kmem_free(str, strlen(str) + 1);
10187 }
10188
10189 kmem_free(act, sizeof (dtrace_actdesc_t));
10190 }
10191
10192 /*
10193 * DTrace ECB Functions
10194 */
10195 static dtrace_ecb_t *
10196 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
10197 {
10198 dtrace_ecb_t *ecb;
10199 dtrace_epid_t epid;
10200
10201 ASSERT(MUTEX_HELD(&dtrace_lock));
10202
10203 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
10204 ecb->dte_predicate = NULL;
10205 ecb->dte_probe = probe;
10206
10207 /*
10208 * The default size is the size of the default action: recording
10209 * the header.
10210 */
10211 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
10212 ecb->dte_alignment = sizeof (dtrace_epid_t);
10213
10214 epid = state->dts_epid++;
10215
10216 if (epid - 1 >= state->dts_necbs) {
10217 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
10218 int necbs = state->dts_necbs << 1;
10219
10220 ASSERT(epid == state->dts_necbs + 1);
10221
10222 if (necbs == 0) {
10223 ASSERT(oecbs == NULL);
10224 necbs = 1;
10225 }
10226
10227 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
10228
10229 if (oecbs != NULL)
10230 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
10231
10232 dtrace_membar_producer();
10233 state->dts_ecbs = ecbs;
10234
10235 if (oecbs != NULL) {
10236 /*
10237 * If this state is active, we must dtrace_sync()
10238 * before we can free the old dts_ecbs array: we're
10239 * coming in hot, and there may be active ring
10240 * buffer processing (which indexes into the dts_ecbs
10241 * array) on another CPU.
10242 */
10243 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
10244 dtrace_sync();
10245
10246 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
10247 }
10248
10249 dtrace_membar_producer();
10250 state->dts_necbs = necbs;
10251 }
10252
10253 ecb->dte_state = state;
10254
10255 ASSERT(state->dts_ecbs[epid - 1] == NULL);
10256 dtrace_membar_producer();
10257 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
10258
10259 return (ecb);
10260 }
10261
10262 static int
10263 dtrace_ecb_enable(dtrace_ecb_t *ecb)
10264 {
10265 dtrace_probe_t *probe = ecb->dte_probe;
10266
10267 ASSERT(MUTEX_HELD(&cpu_lock));
10268 ASSERT(MUTEX_HELD(&dtrace_lock));
10269 ASSERT(ecb->dte_next == NULL);
10270
10271 if (probe == NULL) {
10272 /*
10273 * This is the NULL probe -- there's nothing to do.
10274 */
10275 return (0);
10276 }
10277
10278 if (probe->dtpr_ecb == NULL) {
10279 dtrace_provider_t *prov = probe->dtpr_provider;
10280
10281 /*
10282 * We're the first ECB on this probe.
10283 */
10284 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
10285
10286 if (ecb->dte_predicate != NULL)
10287 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
10288
10289 return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
10290 probe->dtpr_id, probe->dtpr_arg));
10291 } else {
10292 /*
10293 * This probe is already active. Swing the last pointer to
10294 * point to the new ECB, and issue a dtrace_sync() to assure
10295 * that all CPUs have seen the change.
10296 */
10297 ASSERT(probe->dtpr_ecb_last != NULL);
10298 probe->dtpr_ecb_last->dte_next = ecb;
10299 probe->dtpr_ecb_last = ecb;
10300 probe->dtpr_predcache = 0;
10301
10302 dtrace_sync();
10303 return (0);
10304 }
10305 }
10306
10307 static void
10308 dtrace_ecb_resize(dtrace_ecb_t *ecb)
10309 {
10310 dtrace_action_t *act;
10311 uint32_t curneeded = UINT32_MAX;
10312 uint32_t aggbase = UINT32_MAX;
10313
10314 /*
10315 * If we record anything, we always record the dtrace_rechdr_t. (And
10316 * we always record it first.)
10317 */
10318 ecb->dte_size = sizeof (dtrace_rechdr_t);
10319 ecb->dte_alignment = sizeof (dtrace_epid_t);
10320
10321 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10322 dtrace_recdesc_t *rec = &act->dta_rec;
10323 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
10324
10325 ecb->dte_alignment = MAX(ecb->dte_alignment,
10326 rec->dtrd_alignment);
10327
10328 if (DTRACEACT_ISAGG(act->dta_kind)) {
10329 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10330
10331 ASSERT(rec->dtrd_size != 0);
10332 ASSERT(agg->dtag_first != NULL);
10333 ASSERT(act->dta_prev->dta_intuple);
10334 ASSERT(aggbase != UINT32_MAX);
10335 ASSERT(curneeded != UINT32_MAX);
10336
10337 agg->dtag_base = aggbase;
10338
10339 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10340 rec->dtrd_offset = curneeded;
10341 curneeded += rec->dtrd_size;
10342 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
10343
10344 aggbase = UINT32_MAX;
10345 curneeded = UINT32_MAX;
10346 } else if (act->dta_intuple) {
10347 if (curneeded == UINT32_MAX) {
10348 /*
10349 * This is the first record in a tuple. Align
10350 * curneeded to be at offset 4 in an 8-byte
10351 * aligned block.
10352 */
10353 ASSERT(act->dta_prev == NULL ||
10354 !act->dta_prev->dta_intuple);
10355 ASSERT3U(aggbase, ==, UINT32_MAX);
10356 curneeded = P2PHASEUP(ecb->dte_size,
10357 sizeof (uint64_t), sizeof (dtrace_aggid_t));
10358
10359 aggbase = curneeded - sizeof (dtrace_aggid_t);
10360 ASSERT(IS_P2ALIGNED(aggbase,
10361 sizeof (uint64_t)));
10362 }
10363 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10364 rec->dtrd_offset = curneeded;
10365 curneeded += rec->dtrd_size;
10366 } else {
10367 /* tuples must be followed by an aggregation */
10368 ASSERT(act->dta_prev == NULL ||
10369 !act->dta_prev->dta_intuple);
10370
10371 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10372 rec->dtrd_alignment);
10373 rec->dtrd_offset = ecb->dte_size;
10374 ecb->dte_size += rec->dtrd_size;
10375 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10376 }
10377 }
10378
10379 if ((act = ecb->dte_action) != NULL &&
10380 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10381 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10382 /*
10383 * If the size is still sizeof (dtrace_rechdr_t), then all
10384 * actions store no data; set the size to 0.
10385 */
10386 ecb->dte_size = 0;
10387 }
10388
10389 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10390 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10391 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10392 ecb->dte_needed);
10393 }
10394
10395 static dtrace_action_t *
10396 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10397 {
10398 dtrace_aggregation_t *agg;
10399 size_t size = sizeof (uint64_t);
10400 int ntuple = desc->dtad_ntuple;
10401 dtrace_action_t *act;
10402 dtrace_recdesc_t *frec;
10403 dtrace_aggid_t aggid;
10404 dtrace_state_t *state = ecb->dte_state;
10405
10406 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10407 agg->dtag_ecb = ecb;
10408
10409 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10410
10411 switch (desc->dtad_kind) {
10412 case DTRACEAGG_MIN:
10413 agg->dtag_initial = INT64_MAX;
10414 agg->dtag_aggregate = dtrace_aggregate_min;
10415 break;
10416
10417 case DTRACEAGG_MAX:
10418 agg->dtag_initial = INT64_MIN;
10419 agg->dtag_aggregate = dtrace_aggregate_max;
10420 break;
10421
10422 case DTRACEAGG_COUNT:
10423 agg->dtag_aggregate = dtrace_aggregate_count;
10424 break;
10425
10426 case DTRACEAGG_QUANTIZE:
10427 agg->dtag_aggregate = dtrace_aggregate_quantize;
10428 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10429 sizeof (uint64_t);
10430 break;
10431
10432 case DTRACEAGG_LQUANTIZE: {
10433 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10434 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10435
10436 agg->dtag_initial = desc->dtad_arg;
10437 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10438
10439 if (step == 0 || levels == 0)
10440 goto err;
10441
10442 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10443 break;
10444 }
10445
10446 case DTRACEAGG_LLQUANTIZE: {
10447 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10448 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10449 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10450 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10451 int64_t v;
10452
10453 agg->dtag_initial = desc->dtad_arg;
10454 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10455
10456 if (factor < 2 || low >= high || nsteps < factor)
10457 goto err;
10458
10459 /*
10460 * Now check that the number of steps evenly divides a power
10461 * of the factor. (This assures both integer bucket size and
10462 * linearity within each magnitude.)
10463 */
10464 for (v = factor; v < nsteps; v *= factor)
10465 continue;
10466
10467 if ((v % nsteps) || (nsteps % factor))
10468 goto err;
10469
10470 size = (dtrace_aggregate_llquantize_bucket(factor,
10471 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10472 break;
10473 }
10474
10475 case DTRACEAGG_AVG:
10476 agg->dtag_aggregate = dtrace_aggregate_avg;
10477 size = sizeof (uint64_t) * 2;
10478 break;
10479
10480 case DTRACEAGG_STDDEV:
10481 agg->dtag_aggregate = dtrace_aggregate_stddev;
10482 size = sizeof (uint64_t) * 4;
10483 break;
10484
10485 case DTRACEAGG_SUM:
10486 agg->dtag_aggregate = dtrace_aggregate_sum;
10487 break;
10488
10489 default:
10490 goto err;
10491 }
10492
10493 agg->dtag_action.dta_rec.dtrd_size = size;
10494
10495 if (ntuple == 0)
10496 goto err;
10497
10498 /*
10499 * We must make sure that we have enough actions for the n-tuple.
10500 */
10501 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10502 if (DTRACEACT_ISAGG(act->dta_kind))
10503 break;
10504
10505 if (--ntuple == 0) {
10506 /*
10507 * This is the action with which our n-tuple begins.
10508 */
10509 agg->dtag_first = act;
10510 goto success;
10511 }
10512 }
10513
10514 /*
10515 * This n-tuple is short by ntuple elements. Return failure.
10516 */
10517 ASSERT(ntuple != 0);
10518 err:
10519 kmem_free(agg, sizeof (dtrace_aggregation_t));
10520 return (NULL);
10521
10522 success:
10523 /*
10524 * If the last action in the tuple has a size of zero, it's actually
10525 * an expression argument for the aggregating action.
10526 */
10527 ASSERT(ecb->dte_action_last != NULL);
10528 act = ecb->dte_action_last;
10529
10530 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10531 ASSERT(act->dta_difo != NULL);
10532
10533 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10534 agg->dtag_hasarg = 1;
10535 }
10536
10537 /*
10538 * We need to allocate an id for this aggregation.
10539 */
10540 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10541 VM_BESTFIT | VM_SLEEP);
10542
10543 if (aggid - 1 >= state->dts_naggregations) {
10544 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10545 dtrace_aggregation_t **aggs;
10546 int naggs = state->dts_naggregations << 1;
10547 int onaggs = state->dts_naggregations;
10548
10549 ASSERT(aggid == state->dts_naggregations + 1);
10550
10551 if (naggs == 0) {
10552 ASSERT(oaggs == NULL);
10553 naggs = 1;
10554 }
10555
10556 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10557
10558 if (oaggs != NULL) {
10559 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10560 kmem_free(oaggs, onaggs * sizeof (*aggs));
10561 }
10562
10563 state->dts_aggregations = aggs;
10564 state->dts_naggregations = naggs;
10565 }
10566
10567 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10568 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10569
10570 frec = &agg->dtag_first->dta_rec;
10571 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10572 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10573
10574 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10575 ASSERT(!act->dta_intuple);
10576 act->dta_intuple = 1;
10577 }
10578
10579 return (&agg->dtag_action);
10580 }
10581
10582 static void
10583 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10584 {
10585 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10586 dtrace_state_t *state = ecb->dte_state;
10587 dtrace_aggid_t aggid = agg->dtag_id;
10588
10589 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10590 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10591
10592 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10593 state->dts_aggregations[aggid - 1] = NULL;
10594
10595 kmem_free(agg, sizeof (dtrace_aggregation_t));
10596 }
10597
10598 static int
10599 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10600 {
10601 dtrace_action_t *action, *last;
10602 dtrace_difo_t *dp = desc->dtad_difo;
10603 uint32_t size = 0, align = sizeof (uint8_t), mask;
10604 uint16_t format = 0;
10605 dtrace_recdesc_t *rec;
10606 dtrace_state_t *state = ecb->dte_state;
10607 dtrace_optval_t *opt = state->dts_options, nframes, strsize;
10608 uint64_t arg = desc->dtad_arg;
10609
10610 ASSERT(MUTEX_HELD(&dtrace_lock));
10611 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10612
10613 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10614 /*
10615 * If this is an aggregating action, there must be neither
10616 * a speculate nor a commit on the action chain.
10617 */
10618 dtrace_action_t *act;
10619
10620 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10621 if (act->dta_kind == DTRACEACT_COMMIT)
10622 return (EINVAL);
10623
10624 if (act->dta_kind == DTRACEACT_SPECULATE)
10625 return (EINVAL);
10626 }
10627
10628 action = dtrace_ecb_aggregation_create(ecb, desc);
10629
10630 if (action == NULL)
10631 return (EINVAL);
10632 } else {
10633 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10634 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10635 dp != NULL && dp->dtdo_destructive)) {
10636 state->dts_destructive = 1;
10637 }
10638
10639 switch (desc->dtad_kind) {
10640 case DTRACEACT_PRINTF:
10641 case DTRACEACT_PRINTA:
10642 case DTRACEACT_SYSTEM:
10643 case DTRACEACT_FREOPEN:
10644 case DTRACEACT_DIFEXPR:
10645 /*
10646 * We know that our arg is a string -- turn it into a
10647 * format.
10648 */
10649 if (arg == NULL) {
10650 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10651 desc->dtad_kind == DTRACEACT_DIFEXPR);
10652 format = 0;
10653 } else {
10654 ASSERT(arg != NULL);
10655 ASSERT(arg > KERNELBASE);
10656 format = dtrace_format_add(state,
10657 (char *)(uintptr_t)arg);
10658 }
10659
10660 /*FALLTHROUGH*/
10661 case DTRACEACT_LIBACT:
10662 case DTRACEACT_TRACEMEM:
10663 case DTRACEACT_TRACEMEM_DYNSIZE:
10664 if (dp == NULL)
10665 return (EINVAL);
10666
10667 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10668 break;
10669
10670 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10671 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10672 return (EINVAL);
10673
10674 size = opt[DTRACEOPT_STRSIZE];
10675 }
10676
10677 break;
10678
10679 case DTRACEACT_STACK:
10680 if ((nframes = arg) == 0) {
10681 nframes = opt[DTRACEOPT_STACKFRAMES];
10682 ASSERT(nframes > 0);
10683 arg = nframes;
10684 }
10685
10686 size = nframes * sizeof (pc_t);
10687 break;
10688
10689 case DTRACEACT_JSTACK:
10690 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10691 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10692
10693 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10694 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10695
10696 arg = DTRACE_USTACK_ARG(nframes, strsize);
10697
10698 /*FALLTHROUGH*/
10699 case DTRACEACT_USTACK:
10700 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10701 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10702 strsize = DTRACE_USTACK_STRSIZE(arg);
10703 nframes = opt[DTRACEOPT_USTACKFRAMES];
10704 ASSERT(nframes > 0);
10705 arg = DTRACE_USTACK_ARG(nframes, strsize);
10706 }
10707
10708 /*
10709 * Save a slot for the pid.
10710 */
10711 size = (nframes + 1) * sizeof (uint64_t);
10712 size += DTRACE_USTACK_STRSIZE(arg);
10713 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10714
10715 break;
10716
10717 case DTRACEACT_SYM:
10718 case DTRACEACT_MOD:
10719 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10720 sizeof (uint64_t)) ||
10721 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10722 return (EINVAL);
10723 break;
10724
10725 case DTRACEACT_USYM:
10726 case DTRACEACT_UMOD:
10727 case DTRACEACT_UADDR:
10728 if (dp == NULL ||
10729 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10730 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10731 return (EINVAL);
10732
10733 /*
10734 * We have a slot for the pid, plus a slot for the
10735 * argument. To keep things simple (aligned with
10736 * bitness-neutral sizing), we store each as a 64-bit
10737 * quantity.
10738 */
10739 size = 2 * sizeof (uint64_t);
10740 break;
10741
10742 case DTRACEACT_STOP:
10743 case DTRACEACT_BREAKPOINT:
10744 case DTRACEACT_PANIC:
10745 break;
10746
10747 case DTRACEACT_CHILL:
10748 case DTRACEACT_DISCARD:
10749 case DTRACEACT_RAISE:
10750 if (dp == NULL)
10751 return (EINVAL);
10752 break;
10753
10754 case DTRACEACT_EXIT:
10755 if (dp == NULL ||
10756 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10757 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10758 return (EINVAL);
10759 break;
10760
10761 case DTRACEACT_SPECULATE:
10762 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10763 return (EINVAL);
10764
10765 if (dp == NULL)
10766 return (EINVAL);
10767
10768 state->dts_speculates = 1;
10769 break;
10770
10771 case DTRACEACT_COMMIT: {
10772 dtrace_action_t *act = ecb->dte_action;
10773
10774 for (; act != NULL; act = act->dta_next) {
10775 if (act->dta_kind == DTRACEACT_COMMIT)
10776 return (EINVAL);
10777 }
10778
10779 if (dp == NULL)
10780 return (EINVAL);
10781 break;
10782 }
10783
10784 default:
10785 return (EINVAL);
10786 }
10787
10788 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10789 /*
10790 * If this is a data-storing action or a speculate,
10791 * we must be sure that there isn't a commit on the
10792 * action chain.
10793 */
10794 dtrace_action_t *act = ecb->dte_action;
10795
10796 for (; act != NULL; act = act->dta_next) {
10797 if (act->dta_kind == DTRACEACT_COMMIT)
10798 return (EINVAL);
10799 }
10800 }
10801
10802 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10803 action->dta_rec.dtrd_size = size;
10804 }
10805
10806 action->dta_refcnt = 1;
10807 rec = &action->dta_rec;
10808 size = rec->dtrd_size;
10809
10810 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10811 if (!(size & mask)) {
10812 align = mask + 1;
10813 break;
10814 }
10815 }
10816
10817 action->dta_kind = desc->dtad_kind;
10818
10819 if ((action->dta_difo = dp) != NULL)
10820 dtrace_difo_hold(dp);
10821
10822 rec->dtrd_action = action->dta_kind;
10823 rec->dtrd_arg = arg;
10824 rec->dtrd_uarg = desc->dtad_uarg;
10825 rec->dtrd_alignment = (uint16_t)align;
10826 rec->dtrd_format = format;
10827
10828 if ((last = ecb->dte_action_last) != NULL) {
10829 ASSERT(ecb->dte_action != NULL);
10830 action->dta_prev = last;
10831 last->dta_next = action;
10832 } else {
10833 ASSERT(ecb->dte_action == NULL);
10834 ecb->dte_action = action;
10835 }
10836
10837 ecb->dte_action_last = action;
10838
10839 return (0);
10840 }
10841
10842 static void
10843 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10844 {
10845 dtrace_action_t *act = ecb->dte_action, *next;
10846 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10847 dtrace_difo_t *dp;
10848 uint16_t format;
10849
10850 if (act != NULL && act->dta_refcnt > 1) {
10851 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10852 act->dta_refcnt--;
10853 } else {
10854 for (; act != NULL; act = next) {
10855 next = act->dta_next;
10856 ASSERT(next != NULL || act == ecb->dte_action_last);
10857 ASSERT(act->dta_refcnt == 1);
10858
10859 if ((format = act->dta_rec.dtrd_format) != 0)
10860 dtrace_format_remove(ecb->dte_state, format);
10861
10862 if ((dp = act->dta_difo) != NULL)
10863 dtrace_difo_release(dp, vstate);
10864
10865 if (DTRACEACT_ISAGG(act->dta_kind)) {
10866 dtrace_ecb_aggregation_destroy(ecb, act);
10867 } else {
10868 kmem_free(act, sizeof (dtrace_action_t));
10869 }
10870 }
10871 }
10872
10873 ecb->dte_action = NULL;
10874 ecb->dte_action_last = NULL;
10875 ecb->dte_size = 0;
10876 }
10877
10878 static void
10879 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10880 {
10881 /*
10882 * We disable the ECB by removing it from its probe.
10883 */
10884 dtrace_ecb_t *pecb, *prev = NULL;
10885 dtrace_probe_t *probe = ecb->dte_probe;
10886
10887 ASSERT(MUTEX_HELD(&dtrace_lock));
10888
10889 if (probe == NULL) {
10890 /*
10891 * This is the NULL probe; there is nothing to disable.
10892 */
10893 return;
10894 }
10895
10896 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10897 if (pecb == ecb)
10898 break;
10899 prev = pecb;
10900 }
10901
10902 ASSERT(pecb != NULL);
10903
10904 if (prev == NULL) {
10905 probe->dtpr_ecb = ecb->dte_next;
10906 } else {
10907 prev->dte_next = ecb->dte_next;
10908 }
10909
10910 if (ecb == probe->dtpr_ecb_last) {
10911 ASSERT(ecb->dte_next == NULL);
10912 probe->dtpr_ecb_last = prev;
10913 }
10914
10915 /*
10916 * The ECB has been disconnected from the probe; now sync to assure
10917 * that all CPUs have seen the change before returning.
10918 */
10919 dtrace_sync();
10920
10921 if (probe->dtpr_ecb == NULL) {
10922 /*
10923 * That was the last ECB on the probe; clear the predicate
10924 * cache ID for the probe, disable it and sync one more time
10925 * to assure that we'll never hit it again.
10926 */
10927 dtrace_provider_t *prov = probe->dtpr_provider;
10928
10929 ASSERT(ecb->dte_next == NULL);
10930 ASSERT(probe->dtpr_ecb_last == NULL);
10931 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10932 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10933 probe->dtpr_id, probe->dtpr_arg);
10934 dtrace_sync();
10935 } else {
10936 /*
10937 * There is at least one ECB remaining on the probe. If there
10938 * is _exactly_ one, set the probe's predicate cache ID to be
10939 * the predicate cache ID of the remaining ECB.
10940 */
10941 ASSERT(probe->dtpr_ecb_last != NULL);
10942 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10943
10944 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10945 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10946
10947 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10948
10949 if (p != NULL)
10950 probe->dtpr_predcache = p->dtp_cacheid;
10951 }
10952
10953 ecb->dte_next = NULL;
10954 }
10955 }
10956
10957 static void
10958 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10959 {
10960 dtrace_state_t *state = ecb->dte_state;
10961 dtrace_vstate_t *vstate = &state->dts_vstate;
10962 dtrace_predicate_t *pred;
10963 dtrace_epid_t epid = ecb->dte_epid;
10964
10965 ASSERT(MUTEX_HELD(&dtrace_lock));
10966 ASSERT(ecb->dte_next == NULL);
10967 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10968
10969 if ((pred = ecb->dte_predicate) != NULL)
10970 dtrace_predicate_release(pred, vstate);
10971
10972 dtrace_ecb_action_remove(ecb);
10973
10974 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10975 state->dts_ecbs[epid - 1] = NULL;
10976
10977 kmem_free(ecb, sizeof (dtrace_ecb_t));
10978 }
10979
10980 static dtrace_ecb_t *
10981 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10982 dtrace_enabling_t *enab)
10983 {
10984 dtrace_ecb_t *ecb;
10985 dtrace_predicate_t *pred;
10986 dtrace_actdesc_t *act;
10987 dtrace_provider_t *prov;
10988 dtrace_ecbdesc_t *desc = enab->dten_current;
10989
10990 ASSERT(MUTEX_HELD(&dtrace_lock));
10991 ASSERT(state != NULL);
10992
10993 ecb = dtrace_ecb_add(state, probe);
10994 ecb->dte_uarg = desc->dted_uarg;
10995
10996 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10997 dtrace_predicate_hold(pred);
10998 ecb->dte_predicate = pred;
10999 }
11000
11001 if (probe != NULL) {
11002 /*
11003 * If the provider shows more leg than the consumer is old
11004 * enough to see, we need to enable the appropriate implicit
11005 * predicate bits to prevent the ecb from activating at
11006 * revealing times.
11007 *
11008 * Providers specifying DTRACE_PRIV_USER at register time
11009 * are stating that they need the /proc-style privilege
11010 * model to be enforced, and this is what DTRACE_COND_OWNER
11011 * and DTRACE_COND_ZONEOWNER will then do at probe time.
11012 */
11013 prov = probe->dtpr_provider;
11014 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
11015 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11016 ecb->dte_cond |= DTRACE_COND_OWNER;
11017
11018 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11019 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11020 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11021
11022 /*
11023 * If the provider shows us kernel innards and the user
11024 * is lacking sufficient privilege, enable the
11025 * DTRACE_COND_USERMODE implicit predicate.
11026 */
11027 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11028 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11029 ecb->dte_cond |= DTRACE_COND_USERMODE;
11030 }
11031
11032 if (dtrace_ecb_create_cache != NULL) {
11033 /*
11034 * If we have a cached ecb, we'll use its action list instead
11035 * of creating our own (saving both time and space).
11036 */
11037 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11038 dtrace_action_t *act = cached->dte_action;
11039
11040 if (act != NULL) {
11041 ASSERT(act->dta_refcnt > 0);
11042 act->dta_refcnt++;
11043 ecb->dte_action = act;
11044 ecb->dte_action_last = cached->dte_action_last;
11045 ecb->dte_needed = cached->dte_needed;
11046 ecb->dte_size = cached->dte_size;
11047 ecb->dte_alignment = cached->dte_alignment;
11048 }
11049
11050 return (ecb);
11051 }
11052
11053 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11054 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11055 dtrace_ecb_destroy(ecb);
11056 return (NULL);
11057 }
11058 }
11059
11060 dtrace_ecb_resize(ecb);
11061
11062 return (dtrace_ecb_create_cache = ecb);
11063 }
11064
11065 static int
11066 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11067 {
11068 dtrace_ecb_t *ecb;
11069 dtrace_enabling_t *enab = arg;
11070 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11071
11072 ASSERT(state != NULL);
11073
11074 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11075 /*
11076 * This probe was created in a generation for which this
11077 * enabling has previously created ECBs; we don't want to
11078 * enable it again, so just kick out.
11079 */
11080 return (DTRACE_MATCH_NEXT);
11081 }
11082
11083 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11084 return (DTRACE_MATCH_DONE);
11085
11086 if (dtrace_ecb_enable(ecb) < 0)
11087 return (DTRACE_MATCH_FAIL);
11088
11089 return (DTRACE_MATCH_NEXT);
11090 }
11091
11092 static dtrace_ecb_t *
11093 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11094 {
11095 dtrace_ecb_t *ecb;
11096
11097 ASSERT(MUTEX_HELD(&dtrace_lock));
11098
11099 if (id == 0 || id > state->dts_necbs)
11100 return (NULL);
11101
11102 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
11103 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
11104
11105 return (state->dts_ecbs[id - 1]);
11106 }
11107
11108 static dtrace_aggregation_t *
11109 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
11110 {
11111 dtrace_aggregation_t *agg;
11112
11113 ASSERT(MUTEX_HELD(&dtrace_lock));
11114
11115 if (id == 0 || id > state->dts_naggregations)
11116 return (NULL);
11117
11118 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
11119 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
11120 agg->dtag_id == id);
11121
11122 return (state->dts_aggregations[id - 1]);
11123 }
11124
11125 /*
11126 * DTrace Buffer Functions
11127 *
11128 * The following functions manipulate DTrace buffers. Most of these functions
11129 * are called in the context of establishing or processing consumer state;
11130 * exceptions are explicitly noted.
11131 */
11132
11133 /*
11134 * Note: called from cross call context. This function switches the two
11135 * buffers on a given CPU. The atomicity of this operation is assured by
11136 * disabling interrupts while the actual switch takes place; the disabling of
11137 * interrupts serializes the execution with any execution of dtrace_probe() on
11138 * the same CPU.
11139 */
11140 static void
11141 dtrace_buffer_switch(dtrace_buffer_t *buf)
11142 {
11143 caddr_t tomax = buf->dtb_tomax;
11144 caddr_t xamot = buf->dtb_xamot;
11145 dtrace_icookie_t cookie;
11146 hrtime_t now;
11147
11148 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11149 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
11150
11151 cookie = dtrace_interrupt_disable();
11152 now = dtrace_gethrtime();
11153 buf->dtb_tomax = xamot;
11154 buf->dtb_xamot = tomax;
11155 buf->dtb_xamot_drops = buf->dtb_drops;
11156 buf->dtb_xamot_offset = buf->dtb_offset;
11157 buf->dtb_xamot_errors = buf->dtb_errors;
11158 buf->dtb_xamot_flags = buf->dtb_flags;
11159 buf->dtb_offset = 0;
11160 buf->dtb_drops = 0;
11161 buf->dtb_errors = 0;
11162 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
11163 buf->dtb_interval = now - buf->dtb_switched;
11164 buf->dtb_switched = now;
11165 dtrace_interrupt_enable(cookie);
11166 }
11167
11168 /*
11169 * Note: called from cross call context. This function activates a buffer
11170 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
11171 * is guaranteed by the disabling of interrupts.
11172 */
11173 static void
11174 dtrace_buffer_activate(dtrace_state_t *state)
11175 {
11176 dtrace_buffer_t *buf;
11177 dtrace_icookie_t cookie = dtrace_interrupt_disable();
11178
11179 buf = &state->dts_buffer[CPU->cpu_id];
11180
11181 if (buf->dtb_tomax != NULL) {
11182 /*
11183 * We might like to assert that the buffer is marked inactive,
11184 * but this isn't necessarily true: the buffer for the CPU
11185 * that processes the BEGIN probe has its buffer activated
11186 * manually. In this case, we take the (harmless) action
11187 * re-clearing the bit INACTIVE bit.
11188 */
11189 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
11190 }
11191
11192 dtrace_interrupt_enable(cookie);
11193 }
11194
11195 static int
11196 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
11197 processorid_t cpu, int *factor)
11198 {
11199 cpu_t *cp;
11200 dtrace_buffer_t *buf;
11201 int allocated = 0, desired = 0;
11202
11203 ASSERT(MUTEX_HELD(&cpu_lock));
11204 ASSERT(MUTEX_HELD(&dtrace_lock));
11205
11206 *factor = 1;
11207
11208 if (size > dtrace_nonroot_maxsize &&
11209 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
11210 return (EFBIG);
11211
11212 cp = cpu_list;
11213
11214 do {
11215 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11216 continue;
11217
11218 buf = &bufs[cp->cpu_id];
11219
11220 /*
11221 * If there is already a buffer allocated for this CPU, it
11222 * is only possible that this is a DR event. In this case,
11223 * the buffer size must match our specified size.
11224 */
11225 if (buf->dtb_tomax != NULL) {
11226 ASSERT(buf->dtb_size == size);
11227 continue;
11228 }
11229
11230 ASSERT(buf->dtb_xamot == NULL);
11231
11232 if ((buf->dtb_tomax = kmem_zalloc(size,
11233 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11234 goto err;
11235
11236 buf->dtb_size = size;
11237 buf->dtb_flags = flags;
11238 buf->dtb_offset = 0;
11239 buf->dtb_drops = 0;
11240
11241 if (flags & DTRACEBUF_NOSWITCH)
11242 continue;
11243
11244 if ((buf->dtb_xamot = kmem_zalloc(size,
11245 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11246 goto err;
11247 } while ((cp = cp->cpu_next) != cpu_list);
11248
11249 return (0);
11250
11251 err:
11252 cp = cpu_list;
11253
11254 do {
11255 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11256 continue;
11257
11258 buf = &bufs[cp->cpu_id];
11259 desired += 2;
11260
11261 if (buf->dtb_xamot != NULL) {
11262 ASSERT(buf->dtb_tomax != NULL);
11263 ASSERT(buf->dtb_size == size);
11264 kmem_free(buf->dtb_xamot, size);
11265 allocated++;
11266 }
11267
11268 if (buf->dtb_tomax != NULL) {
11269 ASSERT(buf->dtb_size == size);
11270 kmem_free(buf->dtb_tomax, size);
11271 allocated++;
11272 }
11273
11274 buf->dtb_tomax = NULL;
11275 buf->dtb_xamot = NULL;
11276 buf->dtb_size = 0;
11277 } while ((cp = cp->cpu_next) != cpu_list);
11278
11279 *factor = desired / (allocated > 0 ? allocated : 1);
11280
11281 return (ENOMEM);
11282 }
11283
11284 /*
11285 * Note: called from probe context. This function just increments the drop
11286 * count on a buffer. It has been made a function to allow for the
11287 * possibility of understanding the source of mysterious drop counts. (A
11288 * problem for which one may be particularly disappointed that DTrace cannot
11289 * be used to understand DTrace.)
11290 */
11291 static void
11292 dtrace_buffer_drop(dtrace_buffer_t *buf)
11293 {
11294 buf->dtb_drops++;
11295 }
11296
11297 /*
11298 * Note: called from probe context. This function is called to reserve space
11299 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
11300 * mstate. Returns the new offset in the buffer, or a negative value if an
11301 * error has occurred.
11302 */
11303 static intptr_t
11304 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11305 dtrace_state_t *state, dtrace_mstate_t *mstate)
11306 {
11307 intptr_t offs = buf->dtb_offset, soffs;
11308 intptr_t woffs;
11309 caddr_t tomax;
11310 size_t total;
11311
11312 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11313 return (-1);
11314
11315 if ((tomax = buf->dtb_tomax) == NULL) {
11316 dtrace_buffer_drop(buf);
11317 return (-1);
11318 }
11319
11320 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11321 while (offs & (align - 1)) {
11322 /*
11323 * Assert that our alignment is off by a number which
11324 * is itself sizeof (uint32_t) aligned.
11325 */
11326 ASSERT(!((align - (offs & (align - 1))) &
11327 (sizeof (uint32_t) - 1)));
11328 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11329 offs += sizeof (uint32_t);
11330 }
11331
11332 if ((soffs = offs + needed) > buf->dtb_size) {
11333 dtrace_buffer_drop(buf);
11334 return (-1);
11335 }
11336
11337 if (mstate == NULL)
11338 return (offs);
11339
11340 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11341 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11342 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11343
11344 return (offs);
11345 }
11346
11347 if (buf->dtb_flags & DTRACEBUF_FILL) {
11348 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11349 (buf->dtb_flags & DTRACEBUF_FULL))
11350 return (-1);
11351 goto out;
11352 }
11353
11354 total = needed + (offs & (align - 1));
11355
11356 /*
11357 * For a ring buffer, life is quite a bit more complicated. Before
11358 * we can store any padding, we need to adjust our wrapping offset.
11359 * (If we've never before wrapped or we're not about to, no adjustment
11360 * is required.)
11361 */
11362 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11363 offs + total > buf->dtb_size) {
11364 woffs = buf->dtb_xamot_offset;
11365
11366 if (offs + total > buf->dtb_size) {
11367 /*
11368 * We can't fit in the end of the buffer. First, a
11369 * sanity check that we can fit in the buffer at all.
11370 */
11371 if (total > buf->dtb_size) {
11372 dtrace_buffer_drop(buf);
11373 return (-1);
11374 }
11375
11376 /*
11377 * We're going to be storing at the top of the buffer,
11378 * so now we need to deal with the wrapped offset. We
11379 * only reset our wrapped offset to 0 if it is
11380 * currently greater than the current offset. If it
11381 * is less than the current offset, it is because a
11382 * previous allocation induced a wrap -- but the
11383 * allocation didn't subsequently take the space due
11384 * to an error or false predicate evaluation. In this
11385 * case, we'll just leave the wrapped offset alone: if
11386 * the wrapped offset hasn't been advanced far enough
11387 * for this allocation, it will be adjusted in the
11388 * lower loop.
11389 */
11390 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11391 if (woffs >= offs)
11392 woffs = 0;
11393 } else {
11394 woffs = 0;
11395 }
11396
11397 /*
11398 * Now we know that we're going to be storing to the
11399 * top of the buffer and that there is room for us
11400 * there. We need to clear the buffer from the current
11401 * offset to the end (there may be old gunk there).
11402 */
11403 while (offs < buf->dtb_size)
11404 tomax[offs++] = 0;
11405
11406 /*
11407 * We need to set our offset to zero. And because we
11408 * are wrapping, we need to set the bit indicating as
11409 * much. We can also adjust our needed space back
11410 * down to the space required by the ECB -- we know
11411 * that the top of the buffer is aligned.
11412 */
11413 offs = 0;
11414 total = needed;
11415 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11416 } else {
11417 /*
11418 * There is room for us in the buffer, so we simply
11419 * need to check the wrapped offset.
11420 */
11421 if (woffs < offs) {
11422 /*
11423 * The wrapped offset is less than the offset.
11424 * This can happen if we allocated buffer space
11425 * that induced a wrap, but then we didn't
11426 * subsequently take the space due to an error
11427 * or false predicate evaluation. This is
11428 * okay; we know that _this_ allocation isn't
11429 * going to induce a wrap. We still can't
11430 * reset the wrapped offset to be zero,
11431 * however: the space may have been trashed in
11432 * the previous failed probe attempt. But at
11433 * least the wrapped offset doesn't need to
11434 * be adjusted at all...
11435 */
11436 goto out;
11437 }
11438 }
11439
11440 while (offs + total > woffs) {
11441 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11442 size_t size;
11443
11444 if (epid == DTRACE_EPIDNONE) {
11445 size = sizeof (uint32_t);
11446 } else {
11447 ASSERT3U(epid, <=, state->dts_necbs);
11448 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11449
11450 size = state->dts_ecbs[epid - 1]->dte_size;
11451 }
11452
11453 ASSERT(woffs + size <= buf->dtb_size);
11454 ASSERT(size != 0);
11455
11456 if (woffs + size == buf->dtb_size) {
11457 /*
11458 * We've reached the end of the buffer; we want
11459 * to set the wrapped offset to 0 and break
11460 * out. However, if the offs is 0, then we're
11461 * in a strange edge-condition: the amount of
11462 * space that we want to reserve plus the size
11463 * of the record that we're overwriting is
11464 * greater than the size of the buffer. This
11465 * is problematic because if we reserve the
11466 * space but subsequently don't consume it (due
11467 * to a failed predicate or error) the wrapped
11468 * offset will be 0 -- yet the EPID at offset 0
11469 * will not be committed. This situation is
11470 * relatively easy to deal with: if we're in
11471 * this case, the buffer is indistinguishable
11472 * from one that hasn't wrapped; we need only
11473 * finish the job by clearing the wrapped bit,
11474 * explicitly setting the offset to be 0, and
11475 * zero'ing out the old data in the buffer.
11476 */
11477 if (offs == 0) {
11478 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11479 buf->dtb_offset = 0;
11480 woffs = total;
11481
11482 while (woffs < buf->dtb_size)
11483 tomax[woffs++] = 0;
11484 }
11485
11486 woffs = 0;
11487 break;
11488 }
11489
11490 woffs += size;
11491 }
11492
11493 /*
11494 * We have a wrapped offset. It may be that the wrapped offset
11495 * has become zero -- that's okay.
11496 */
11497 buf->dtb_xamot_offset = woffs;
11498 }
11499
11500 out:
11501 /*
11502 * Now we can plow the buffer with any necessary padding.
11503 */
11504 while (offs & (align - 1)) {
11505 /*
11506 * Assert that our alignment is off by a number which
11507 * is itself sizeof (uint32_t) aligned.
11508 */
11509 ASSERT(!((align - (offs & (align - 1))) &
11510 (sizeof (uint32_t) - 1)));
11511 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11512 offs += sizeof (uint32_t);
11513 }
11514
11515 if (buf->dtb_flags & DTRACEBUF_FILL) {
11516 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11517 buf->dtb_flags |= DTRACEBUF_FULL;
11518 return (-1);
11519 }
11520 }
11521
11522 if (mstate == NULL)
11523 return (offs);
11524
11525 /*
11526 * For ring buffers and fill buffers, the scratch space is always
11527 * the inactive buffer.
11528 */
11529 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11530 mstate->dtms_scratch_size = buf->dtb_size;
11531 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11532
11533 return (offs);
11534 }
11535
11536 static void
11537 dtrace_buffer_polish(dtrace_buffer_t *buf)
11538 {
11539 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11540 ASSERT(MUTEX_HELD(&dtrace_lock));
11541
11542 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11543 return;
11544
11545 /*
11546 * We need to polish the ring buffer. There are three cases:
11547 *
11548 * - The first (and presumably most common) is that there is no gap
11549 * between the buffer offset and the wrapped offset. In this case,
11550 * there is nothing in the buffer that isn't valid data; we can
11551 * mark the buffer as polished and return.
11552 *
11553 * - The second (less common than the first but still more common
11554 * than the third) is that there is a gap between the buffer offset
11555 * and the wrapped offset, and the wrapped offset is larger than the
11556 * buffer offset. This can happen because of an alignment issue, or
11557 * can happen because of a call to dtrace_buffer_reserve() that
11558 * didn't subsequently consume the buffer space. In this case,
11559 * we need to zero the data from the buffer offset to the wrapped
11560 * offset.
11561 *
11562 * - The third (and least common) is that there is a gap between the
11563 * buffer offset and the wrapped offset, but the wrapped offset is
11564 * _less_ than the buffer offset. This can only happen because a
11565 * call to dtrace_buffer_reserve() induced a wrap, but the space
11566 * was not subsequently consumed. In this case, we need to zero the
11567 * space from the offset to the end of the buffer _and_ from the
11568 * top of the buffer to the wrapped offset.
11569 */
11570 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11571 bzero(buf->dtb_tomax + buf->dtb_offset,
11572 buf->dtb_xamot_offset - buf->dtb_offset);
11573 }
11574
11575 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11576 bzero(buf->dtb_tomax + buf->dtb_offset,
11577 buf->dtb_size - buf->dtb_offset);
11578 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11579 }
11580 }
11581
11582 /*
11583 * This routine determines if data generated at the specified time has likely
11584 * been entirely consumed at user-level. This routine is called to determine
11585 * if an ECB on a defunct probe (but for an active enabling) can be safely
11586 * disabled and destroyed.
11587 */
11588 static int
11589 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11590 {
11591 int i;
11592
11593 for (i = 0; i < NCPU; i++) {
11594 dtrace_buffer_t *buf = &bufs[i];
11595
11596 if (buf->dtb_size == 0)
11597 continue;
11598
11599 if (buf->dtb_flags & DTRACEBUF_RING)
11600 return (0);
11601
11602 if (!buf->dtb_switched && buf->dtb_offset != 0)
11603 return (0);
11604
11605 if (buf->dtb_switched - buf->dtb_interval < when)
11606 return (0);
11607 }
11608
11609 return (1);
11610 }
11611
11612 static void
11613 dtrace_buffer_free(dtrace_buffer_t *bufs)
11614 {
11615 int i;
11616
11617 for (i = 0; i < NCPU; i++) {
11618 dtrace_buffer_t *buf = &bufs[i];
11619
11620 if (buf->dtb_tomax == NULL) {
11621 ASSERT(buf->dtb_xamot == NULL);
11622 ASSERT(buf->dtb_size == 0);
11623 continue;
11624 }
11625
11626 if (buf->dtb_xamot != NULL) {
11627 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11628 kmem_free(buf->dtb_xamot, buf->dtb_size);
11629 }
11630
11631 kmem_free(buf->dtb_tomax, buf->dtb_size);
11632 buf->dtb_size = 0;
11633 buf->dtb_tomax = NULL;
11634 buf->dtb_xamot = NULL;
11635 }
11636 }
11637
11638 /*
11639 * DTrace Enabling Functions
11640 */
11641 static dtrace_enabling_t *
11642 dtrace_enabling_create(dtrace_vstate_t *vstate)
11643 {
11644 dtrace_enabling_t *enab;
11645
11646 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11647 enab->dten_vstate = vstate;
11648
11649 return (enab);
11650 }
11651
11652 static void
11653 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11654 {
11655 dtrace_ecbdesc_t **ndesc;
11656 size_t osize, nsize;
11657
11658 /*
11659 * We can't add to enablings after we've enabled them, or after we've
11660 * retained them.
11661 */
11662 ASSERT(enab->dten_probegen == 0);
11663 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11664
11665 if (enab->dten_ndesc < enab->dten_maxdesc) {
11666 enab->dten_desc[enab->dten_ndesc++] = ecb;
11667 return;
11668 }
11669
11670 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11671
11672 if (enab->dten_maxdesc == 0) {
11673 enab->dten_maxdesc = 1;
11674 } else {
11675 enab->dten_maxdesc <<= 1;
11676 }
11677
11678 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11679
11680 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11681 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11682 bcopy(enab->dten_desc, ndesc, osize);
11683 kmem_free(enab->dten_desc, osize);
11684
11685 enab->dten_desc = ndesc;
11686 enab->dten_desc[enab->dten_ndesc++] = ecb;
11687 }
11688
11689 static void
11690 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11691 dtrace_probedesc_t *pd)
11692 {
11693 dtrace_ecbdesc_t *new;
11694 dtrace_predicate_t *pred;
11695 dtrace_actdesc_t *act;
11696
11697 /*
11698 * We're going to create a new ECB description that matches the
11699 * specified ECB in every way, but has the specified probe description.
11700 */
11701 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11702
11703 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11704 dtrace_predicate_hold(pred);
11705
11706 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11707 dtrace_actdesc_hold(act);
11708
11709 new->dted_action = ecb->dted_action;
11710 new->dted_pred = ecb->dted_pred;
11711 new->dted_probe = *pd;
11712 new->dted_uarg = ecb->dted_uarg;
11713
11714 dtrace_enabling_add(enab, new);
11715 }
11716
11717 static void
11718 dtrace_enabling_dump(dtrace_enabling_t *enab)
11719 {
11720 int i;
11721
11722 for (i = 0; i < enab->dten_ndesc; i++) {
11723 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11724
11725 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11726 desc->dtpd_provider, desc->dtpd_mod,
11727 desc->dtpd_func, desc->dtpd_name);
11728 }
11729 }
11730
11731 static void
11732 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11733 {
11734 int i;
11735 dtrace_ecbdesc_t *ep;
11736 dtrace_vstate_t *vstate = enab->dten_vstate;
11737
11738 ASSERT(MUTEX_HELD(&dtrace_lock));
11739
11740 for (i = 0; i < enab->dten_ndesc; i++) {
11741 dtrace_actdesc_t *act, *next;
11742 dtrace_predicate_t *pred;
11743
11744 ep = enab->dten_desc[i];
11745
11746 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11747 dtrace_predicate_release(pred, vstate);
11748
11749 for (act = ep->dted_action; act != NULL; act = next) {
11750 next = act->dtad_next;
11751 dtrace_actdesc_release(act, vstate);
11752 }
11753
11754 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11755 }
11756
11757 kmem_free(enab->dten_desc,
11758 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11759
11760 /*
11761 * If this was a retained enabling, decrement the dts_nretained count
11762 * and take it off of the dtrace_retained list.
11763 */
11764 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11765 dtrace_retained == enab) {
11766 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11767 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11768 enab->dten_vstate->dtvs_state->dts_nretained--;
11769 dtrace_retained_gen++;
11770 }
11771
11772 if (enab->dten_prev == NULL) {
11773 if (dtrace_retained == enab) {
11774 dtrace_retained = enab->dten_next;
11775
11776 if (dtrace_retained != NULL)
11777 dtrace_retained->dten_prev = NULL;
11778 }
11779 } else {
11780 ASSERT(enab != dtrace_retained);
11781 ASSERT(dtrace_retained != NULL);
11782 enab->dten_prev->dten_next = enab->dten_next;
11783 }
11784
11785 if (enab->dten_next != NULL) {
11786 ASSERT(dtrace_retained != NULL);
11787 enab->dten_next->dten_prev = enab->dten_prev;
11788 }
11789
11790 kmem_free(enab, sizeof (dtrace_enabling_t));
11791 }
11792
11793 static int
11794 dtrace_enabling_retain(dtrace_enabling_t *enab)
11795 {
11796 dtrace_state_t *state;
11797
11798 ASSERT(MUTEX_HELD(&dtrace_lock));
11799 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11800 ASSERT(enab->dten_vstate != NULL);
11801
11802 state = enab->dten_vstate->dtvs_state;
11803 ASSERT(state != NULL);
11804
11805 /*
11806 * We only allow each state to retain dtrace_retain_max enablings.
11807 */
11808 if (state->dts_nretained >= dtrace_retain_max)
11809 return (ENOSPC);
11810
11811 state->dts_nretained++;
11812 dtrace_retained_gen++;
11813
11814 if (dtrace_retained == NULL) {
11815 dtrace_retained = enab;
11816 return (0);
11817 }
11818
11819 enab->dten_next = dtrace_retained;
11820 dtrace_retained->dten_prev = enab;
11821 dtrace_retained = enab;
11822
11823 return (0);
11824 }
11825
11826 static int
11827 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11828 dtrace_probedesc_t *create)
11829 {
11830 dtrace_enabling_t *new, *enab;
11831 int found = 0, err = ENOENT;
11832
11833 ASSERT(MUTEX_HELD(&dtrace_lock));
11834 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11835 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11836 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11837 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11838
11839 new = dtrace_enabling_create(&state->dts_vstate);
11840
11841 /*
11842 * Iterate over all retained enablings, looking for enablings that
11843 * match the specified state.
11844 */
11845 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11846 int i;
11847
11848 /*
11849 * dtvs_state can only be NULL for helper enablings -- and
11850 * helper enablings can't be retained.
11851 */
11852 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11853
11854 if (enab->dten_vstate->dtvs_state != state)
11855 continue;
11856
11857 /*
11858 * Now iterate over each probe description; we're looking for
11859 * an exact match to the specified probe description.
11860 */
11861 for (i = 0; i < enab->dten_ndesc; i++) {
11862 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11863 dtrace_probedesc_t *pd = &ep->dted_probe;
11864
11865 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11866 continue;
11867
11868 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11869 continue;
11870
11871 if (strcmp(pd->dtpd_func, match->dtpd_func))
11872 continue;
11873
11874 if (strcmp(pd->dtpd_name, match->dtpd_name))
11875 continue;
11876
11877 /*
11878 * We have a winning probe! Add it to our growing
11879 * enabling.
11880 */
11881 found = 1;
11882 dtrace_enabling_addlike(new, ep, create);
11883 }
11884 }
11885
11886 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11887 dtrace_enabling_destroy(new);
11888 return (err);
11889 }
11890
11891 return (0);
11892 }
11893
11894 static void
11895 dtrace_enabling_retract(dtrace_state_t *state)
11896 {
11897 dtrace_enabling_t *enab, *next;
11898
11899 ASSERT(MUTEX_HELD(&dtrace_lock));
11900
11901 /*
11902 * Iterate over all retained enablings, destroy the enablings retained
11903 * for the specified state.
11904 */
11905 for (enab = dtrace_retained; enab != NULL; enab = next) {
11906 next = enab->dten_next;
11907
11908 /*
11909 * dtvs_state can only be NULL for helper enablings -- and
11910 * helper enablings can't be retained.
11911 */
11912 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11913
11914 if (enab->dten_vstate->dtvs_state == state) {
11915 ASSERT(state->dts_nretained > 0);
11916 dtrace_enabling_destroy(enab);
11917 }
11918 }
11919
11920 ASSERT(state->dts_nretained == 0);
11921 }
11922
11923 static int
11924 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11925 {
11926 int i = 0;
11927 int total_matched = 0, matched = 0;
11928
11929 ASSERT(MUTEX_HELD(&cpu_lock));
11930 ASSERT(MUTEX_HELD(&dtrace_lock));
11931
11932 for (i = 0; i < enab->dten_ndesc; i++) {
11933 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11934
11935 enab->dten_current = ep;
11936 enab->dten_error = 0;
11937
11938 /*
11939 * If a provider failed to enable a probe then get out and
11940 * let the consumer know we failed.
11941 */
11942 if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
11943 return (EBUSY);
11944
11945 total_matched += matched;
11946
11947 if (enab->dten_error != 0) {
11948 /*
11949 * If we get an error half-way through enabling the
11950 * probes, we kick out -- perhaps with some number of
11951 * them enabled. Leaving enabled probes enabled may
11952 * be slightly confusing for user-level, but we expect
11953 * that no one will attempt to actually drive on in
11954 * the face of such errors. If this is an anonymous
11955 * enabling (indicated with a NULL nmatched pointer),
11956 * we cmn_err() a message. We aren't expecting to
11957 * get such an error -- such as it can exist at all,
11958 * it would be a result of corrupted DOF in the driver
11959 * properties.
11960 */
11961 if (nmatched == NULL) {
11962 cmn_err(CE_WARN, "dtrace_enabling_match() "
11963 "error on %p: %d", (void *)ep,
11964 enab->dten_error);
11965 }
11966
11967 return (enab->dten_error);
11968 }
11969 }
11970
11971 enab->dten_probegen = dtrace_probegen;
11972 if (nmatched != NULL)
11973 *nmatched = total_matched;
11974
11975 return (0);
11976 }
11977
11978 static void
11979 dtrace_enabling_matchall(void)
11980 {
11981 dtrace_enabling_t *enab;
11982
11983 mutex_enter(&cpu_lock);
11984 mutex_enter(&dtrace_lock);
11985
11986 /*
11987 * Iterate over all retained enablings to see if any probes match
11988 * against them. We only perform this operation on enablings for which
11989 * we have sufficient permissions by virtue of being in the global zone
11990 * or in the same zone as the DTrace client. Because we can be called
11991 * after dtrace_detach() has been called, we cannot assert that there
11992 * are retained enablings. We can safely load from dtrace_retained,
11993 * however: the taskq_destroy() at the end of dtrace_detach() will
11994 * block pending our completion.
11995 */
11996 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11997 dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
11998 cred_t *cr = dcr->dcr_cred;
11999 zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
12000
12001 if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
12002 (zone == GLOBAL_ZONEID || getzoneid() == zone)))
12003 (void) dtrace_enabling_match(enab, NULL);
12004 }
12005
12006 mutex_exit(&dtrace_lock);
12007 mutex_exit(&cpu_lock);
12008 }
12009
12010 /*
12011 * If an enabling is to be enabled without having matched probes (that is, if
12012 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
12013 * enabling must be _primed_ by creating an ECB for every ECB description.
12014 * This must be done to assure that we know the number of speculations, the
12015 * number of aggregations, the minimum buffer size needed, etc. before we
12016 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
12017 * enabling any probes, we create ECBs for every ECB decription, but with a
12018 * NULL probe -- which is exactly what this function does.
12019 */
12020 static void
12021 dtrace_enabling_prime(dtrace_state_t *state)
12022 {
12023 dtrace_enabling_t *enab;
12024 int i;
12025
12026 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12027 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12028
12029 if (enab->dten_vstate->dtvs_state != state)
12030 continue;
12031
12032 /*
12033 * We don't want to prime an enabling more than once, lest
12034 * we allow a malicious user to induce resource exhaustion.
12035 * (The ECBs that result from priming an enabling aren't
12036 * leaked -- but they also aren't deallocated until the
12037 * consumer state is destroyed.)
12038 */
12039 if (enab->dten_primed)
12040 continue;
12041
12042 for (i = 0; i < enab->dten_ndesc; i++) {
12043 enab->dten_current = enab->dten_desc[i];
12044 (void) dtrace_probe_enable(NULL, enab);
12045 }
12046
12047 enab->dten_primed = 1;
12048 }
12049 }
12050
12051 /*
12052 * Called to indicate that probes should be provided due to retained
12053 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
12054 * must take an initial lap through the enabling calling the dtps_provide()
12055 * entry point explicitly to allow for autocreated probes.
12056 */
12057 static void
12058 dtrace_enabling_provide(dtrace_provider_t *prv)
12059 {
12060 int i, all = 0;
12061 dtrace_probedesc_t desc;
12062 dtrace_genid_t gen;
12063
12064 ASSERT(MUTEX_HELD(&dtrace_lock));
12065 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
12066
12067 if (prv == NULL) {
12068 all = 1;
12069 prv = dtrace_provider;
12070 }
12071
12072 do {
12073 dtrace_enabling_t *enab;
12074 void *parg = prv->dtpv_arg;
12075
12076 retry:
12077 gen = dtrace_retained_gen;
12078 for (enab = dtrace_retained; enab != NULL;
12079 enab = enab->dten_next) {
12080 for (i = 0; i < enab->dten_ndesc; i++) {
12081 desc = enab->dten_desc[i]->dted_probe;
12082 mutex_exit(&dtrace_lock);
12083 prv->dtpv_pops.dtps_provide(parg, &desc);
12084 mutex_enter(&dtrace_lock);
12085 /*
12086 * Process the retained enablings again if
12087 * they have changed while we weren't holding
12088 * dtrace_lock.
12089 */
12090 if (gen != dtrace_retained_gen)
12091 goto retry;
12092 }
12093 }
12094 } while (all && (prv = prv->dtpv_next) != NULL);
12095
12096 mutex_exit(&dtrace_lock);
12097 dtrace_probe_provide(NULL, all ? NULL : prv);
12098 mutex_enter(&dtrace_lock);
12099 }
12100
12101 /*
12102 * Called to reap ECBs that are attached to probes from defunct providers.
12103 */
12104 static void
12105 dtrace_enabling_reap(void)
12106 {
12107 dtrace_provider_t *prov;
12108 dtrace_probe_t *probe;
12109 dtrace_ecb_t *ecb;
12110 hrtime_t when;
12111 int i;
12112
12113 mutex_enter(&cpu_lock);
12114 mutex_enter(&dtrace_lock);
12115
12116 for (i = 0; i < dtrace_nprobes; i++) {
12117 if ((probe = dtrace_probes[i]) == NULL)
12118 continue;
12119
12120 if (probe->dtpr_ecb == NULL)
12121 continue;
12122
12123 prov = probe->dtpr_provider;
12124
12125 if ((when = prov->dtpv_defunct) == 0)
12126 continue;
12127
12128 /*
12129 * We have ECBs on a defunct provider: we want to reap these
12130 * ECBs to allow the provider to unregister. The destruction
12131 * of these ECBs must be done carefully: if we destroy the ECB
12132 * and the consumer later wishes to consume an EPID that
12133 * corresponds to the destroyed ECB (and if the EPID metadata
12134 * has not been previously consumed), the consumer will abort
12135 * processing on the unknown EPID. To reduce (but not, sadly,
12136 * eliminate) the possibility of this, we will only destroy an
12137 * ECB for a defunct provider if, for the state that
12138 * corresponds to the ECB:
12139 *
12140 * (a) There is no speculative tracing (which can effectively
12141 * cache an EPID for an arbitrary amount of time).
12142 *
12143 * (b) The principal buffers have been switched twice since the
12144 * provider became defunct.
12145 *
12146 * (c) The aggregation buffers are of zero size or have been
12147 * switched twice since the provider became defunct.
12148 *
12149 * We use dts_speculates to determine (a) and call a function
12150 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
12151 * that as soon as we've been unable to destroy one of the ECBs
12152 * associated with the probe, we quit trying -- reaping is only
12153 * fruitful in as much as we can destroy all ECBs associated
12154 * with the defunct provider's probes.
12155 */
12156 while ((ecb = probe->dtpr_ecb) != NULL) {
12157 dtrace_state_t *state = ecb->dte_state;
12158 dtrace_buffer_t *buf = state->dts_buffer;
12159 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
12160
12161 if (state->dts_speculates)
12162 break;
12163
12164 if (!dtrace_buffer_consumed(buf, when))
12165 break;
12166
12167 if (!dtrace_buffer_consumed(aggbuf, when))
12168 break;
12169
12170 dtrace_ecb_disable(ecb);
12171 ASSERT(probe->dtpr_ecb != ecb);
12172 dtrace_ecb_destroy(ecb);
12173 }
12174 }
12175
12176 mutex_exit(&dtrace_lock);
12177 mutex_exit(&cpu_lock);
12178 }
12179
12180 /*
12181 * DTrace DOF Functions
12182 */
12183 /*ARGSUSED*/
12184 static void
12185 dtrace_dof_error(dof_hdr_t *dof, const char *str)
12186 {
12187 if (dtrace_err_verbose)
12188 cmn_err(CE_WARN, "failed to process DOF: %s", str);
12189
12190 #ifdef DTRACE_ERRDEBUG
12191 dtrace_errdebug(str);
12192 #endif
12193 }
12194
12195 /*
12196 * Create DOF out of a currently enabled state. Right now, we only create
12197 * DOF containing the run-time options -- but this could be expanded to create
12198 * complete DOF representing the enabled state.
12199 */
12200 static dof_hdr_t *
12201 dtrace_dof_create(dtrace_state_t *state)
12202 {
12203 dof_hdr_t *dof;
12204 dof_sec_t *sec;
12205 dof_optdesc_t *opt;
12206 int i, len = sizeof (dof_hdr_t) +
12207 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
12208 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12209
12210 ASSERT(MUTEX_HELD(&dtrace_lock));
12211
12212 dof = kmem_zalloc(len, KM_SLEEP);
12213 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
12214 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
12215 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
12216 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
12217
12218 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
12219 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
12220 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
12221 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
12222 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
12223 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
12224
12225 dof->dofh_flags = 0;
12226 dof->dofh_hdrsize = sizeof (dof_hdr_t);
12227 dof->dofh_secsize = sizeof (dof_sec_t);
12228 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
12229 dof->dofh_secoff = sizeof (dof_hdr_t);
12230 dof->dofh_loadsz = len;
12231 dof->dofh_filesz = len;
12232 dof->dofh_pad = 0;
12233
12234 /*
12235 * Fill in the option section header...
12236 */
12237 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
12238 sec->dofs_type = DOF_SECT_OPTDESC;
12239 sec->dofs_align = sizeof (uint64_t);
12240 sec->dofs_flags = DOF_SECF_LOAD;
12241 sec->dofs_entsize = sizeof (dof_optdesc_t);
12242
12243 opt = (dof_optdesc_t *)((uintptr_t)sec +
12244 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
12245
12246 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
12247 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12248
12249 for (i = 0; i < DTRACEOPT_MAX; i++) {
12250 opt[i].dofo_option = i;
12251 opt[i].dofo_strtab = DOF_SECIDX_NONE;
12252 opt[i].dofo_value = state->dts_options[i];
12253 }
12254
12255 return (dof);
12256 }
12257
12258 static dof_hdr_t *
12259 dtrace_dof_copyin(uintptr_t uarg, int *errp)
12260 {
12261 dof_hdr_t hdr, *dof;
12262
12263 ASSERT(!MUTEX_HELD(&dtrace_lock));
12264
12265 /*
12266 * First, we're going to copyin() the sizeof (dof_hdr_t).
12267 */
12268 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
12269 dtrace_dof_error(NULL, "failed to copyin DOF header");
12270 *errp = EFAULT;
12271 return (NULL);
12272 }
12273
12274 /*
12275 * Now we'll allocate the entire DOF and copy it in -- provided
12276 * that the length isn't outrageous.
12277 */
12278 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
12279 dtrace_dof_error(&hdr, "load size exceeds maximum");
12280 *errp = E2BIG;
12281 return (NULL);
12282 }
12283
12284 if (hdr.dofh_loadsz < sizeof (hdr)) {
12285 dtrace_dof_error(&hdr, "invalid load size");
12286 *errp = EINVAL;
12287 return (NULL);
12288 }
12289
12290 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12291
12292 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
12293 dof->dofh_loadsz != hdr.dofh_loadsz) {
12294 kmem_free(dof, hdr.dofh_loadsz);
12295 *errp = EFAULT;
12296 return (NULL);
12297 }
12298
12299 return (dof);
12300 }
12301
12302 static dof_hdr_t *
12303 dtrace_dof_property(const char *name)
12304 {
12305 uchar_t *buf;
12306 uint64_t loadsz;
12307 unsigned int len, i;
12308 dof_hdr_t *dof;
12309
12310 /*
12311 * Unfortunately, array of values in .conf files are always (and
12312 * only) interpreted to be integer arrays. We must read our DOF
12313 * as an integer array, and then squeeze it into a byte array.
12314 */
12315 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12316 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12317 return (NULL);
12318
12319 for (i = 0; i < len; i++)
12320 buf[i] = (uchar_t)(((int *)buf)[i]);
12321
12322 if (len < sizeof (dof_hdr_t)) {
12323 ddi_prop_free(buf);
12324 dtrace_dof_error(NULL, "truncated header");
12325 return (NULL);
12326 }
12327
12328 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12329 ddi_prop_free(buf);
12330 dtrace_dof_error(NULL, "truncated DOF");
12331 return (NULL);
12332 }
12333
12334 if (loadsz >= dtrace_dof_maxsize) {
12335 ddi_prop_free(buf);
12336 dtrace_dof_error(NULL, "oversized DOF");
12337 return (NULL);
12338 }
12339
12340 dof = kmem_alloc(loadsz, KM_SLEEP);
12341 bcopy(buf, dof, loadsz);
12342 ddi_prop_free(buf);
12343
12344 return (dof);
12345 }
12346
12347 static void
12348 dtrace_dof_destroy(dof_hdr_t *dof)
12349 {
12350 kmem_free(dof, dof->dofh_loadsz);
12351 }
12352
12353 /*
12354 * Return the dof_sec_t pointer corresponding to a given section index. If the
12355 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12356 * a type other than DOF_SECT_NONE is specified, the header is checked against
12357 * this type and NULL is returned if the types do not match.
12358 */
12359 static dof_sec_t *
12360 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12361 {
12362 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12363 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12364
12365 if (i >= dof->dofh_secnum) {
12366 dtrace_dof_error(dof, "referenced section index is invalid");
12367 return (NULL);
12368 }
12369
12370 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12371 dtrace_dof_error(dof, "referenced section is not loadable");
12372 return (NULL);
12373 }
12374
12375 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12376 dtrace_dof_error(dof, "referenced section is the wrong type");
12377 return (NULL);
12378 }
12379
12380 return (sec);
12381 }
12382
12383 static dtrace_probedesc_t *
12384 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12385 {
12386 dof_probedesc_t *probe;
12387 dof_sec_t *strtab;
12388 uintptr_t daddr = (uintptr_t)dof;
12389 uintptr_t str;
12390 size_t size;
12391
12392 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12393 dtrace_dof_error(dof, "invalid probe section");
12394 return (NULL);
12395 }
12396
12397 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12398 dtrace_dof_error(dof, "bad alignment in probe description");
12399 return (NULL);
12400 }
12401
12402 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12403 dtrace_dof_error(dof, "truncated probe description");
12404 return (NULL);
12405 }
12406
12407 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12408 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12409
12410 if (strtab == NULL)
12411 return (NULL);
12412
12413 str = daddr + strtab->dofs_offset;
12414 size = strtab->dofs_size;
12415
12416 if (probe->dofp_provider >= strtab->dofs_size) {
12417 dtrace_dof_error(dof, "corrupt probe provider");
12418 return (NULL);
12419 }
12420
12421 (void) strncpy(desc->dtpd_provider,
12422 (char *)(str + probe->dofp_provider),
12423 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12424
12425 if (probe->dofp_mod >= strtab->dofs_size) {
12426 dtrace_dof_error(dof, "corrupt probe module");
12427 return (NULL);
12428 }
12429
12430 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12431 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12432
12433 if (probe->dofp_func >= strtab->dofs_size) {
12434 dtrace_dof_error(dof, "corrupt probe function");
12435 return (NULL);
12436 }
12437
12438 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12439 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12440
12441 if (probe->dofp_name >= strtab->dofs_size) {
12442 dtrace_dof_error(dof, "corrupt probe name");
12443 return (NULL);
12444 }
12445
12446 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12447 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12448
12449 return (desc);
12450 }
12451
12452 static dtrace_difo_t *
12453 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12454 cred_t *cr)
12455 {
12456 dtrace_difo_t *dp;
12457 size_t ttl = 0;
12458 dof_difohdr_t *dofd;
12459 uintptr_t daddr = (uintptr_t)dof;
12460 size_t max = dtrace_difo_maxsize;
12461 int i, l, n;
12462
12463 static const struct {
12464 int section;
12465 int bufoffs;
12466 int lenoffs;
12467 int entsize;
12468 int align;
12469 const char *msg;
12470 } difo[] = {
12471 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12472 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12473 sizeof (dif_instr_t), "multiple DIF sections" },
12474
12475 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12476 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12477 sizeof (uint64_t), "multiple integer tables" },
12478
12479 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12480 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12481 sizeof (char), "multiple string tables" },
12482
12483 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12484 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12485 sizeof (uint_t), "multiple variable tables" },
12486
12487 { DOF_SECT_NONE, 0, 0, 0, NULL }
12488 };
12489
12490 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12491 dtrace_dof_error(dof, "invalid DIFO header section");
12492 return (NULL);
12493 }
12494
12495 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12496 dtrace_dof_error(dof, "bad alignment in DIFO header");
12497 return (NULL);
12498 }
12499
12500 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12501 sec->dofs_size % sizeof (dof_secidx_t)) {
12502 dtrace_dof_error(dof, "bad size in DIFO header");
12503 return (NULL);
12504 }
12505
12506 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12507 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12508
12509 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12510 dp->dtdo_rtype = dofd->dofd_rtype;
12511
12512 for (l = 0; l < n; l++) {
12513 dof_sec_t *subsec;
12514 void **bufp;
12515 uint32_t *lenp;
12516
12517 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12518 dofd->dofd_links[l])) == NULL)
12519 goto err; /* invalid section link */
12520
12521 if (ttl + subsec->dofs_size > max) {
12522 dtrace_dof_error(dof, "exceeds maximum size");
12523 goto err;
12524 }
12525
12526 ttl += subsec->dofs_size;
12527
12528 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12529 if (subsec->dofs_type != difo[i].section)
12530 continue;
12531
12532 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12533 dtrace_dof_error(dof, "section not loaded");
12534 goto err;
12535 }
12536
12537 if (subsec->dofs_align != difo[i].align) {
12538 dtrace_dof_error(dof, "bad alignment");
12539 goto err;
12540 }
12541
12542 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12543 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12544
12545 if (*bufp != NULL) {
12546 dtrace_dof_error(dof, difo[i].msg);
12547 goto err;
12548 }
12549
12550 if (difo[i].entsize != subsec->dofs_entsize) {
12551 dtrace_dof_error(dof, "entry size mismatch");
12552 goto err;
12553 }
12554
12555 if (subsec->dofs_entsize != 0 &&
12556 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12557 dtrace_dof_error(dof, "corrupt entry size");
12558 goto err;
12559 }
12560
12561 *lenp = subsec->dofs_size;
12562 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12563 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12564 *bufp, subsec->dofs_size);
12565
12566 if (subsec->dofs_entsize != 0)
12567 *lenp /= subsec->dofs_entsize;
12568
12569 break;
12570 }
12571
12572 /*
12573 * If we encounter a loadable DIFO sub-section that is not
12574 * known to us, assume this is a broken program and fail.
12575 */
12576 if (difo[i].section == DOF_SECT_NONE &&
12577 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12578 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12579 goto err;
12580 }
12581 }
12582
12583 if (dp->dtdo_buf == NULL) {
12584 /*
12585 * We can't have a DIF object without DIF text.
12586 */
12587 dtrace_dof_error(dof, "missing DIF text");
12588 goto err;
12589 }
12590
12591 /*
12592 * Before we validate the DIF object, run through the variable table
12593 * looking for the strings -- if any of their size are under, we'll set
12594 * their size to be the system-wide default string size. Note that
12595 * this should _not_ happen if the "strsize" option has been set --
12596 * in this case, the compiler should have set the size to reflect the
12597 * setting of the option.
12598 */
12599 for (i = 0; i < dp->dtdo_varlen; i++) {
12600 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12601 dtrace_diftype_t *t = &v->dtdv_type;
12602
12603 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12604 continue;
12605
12606 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12607 t->dtdt_size = dtrace_strsize_default;
12608 }
12609
12610 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12611 goto err;
12612
12613 dtrace_difo_init(dp, vstate);
12614 return (dp);
12615
12616 err:
12617 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12618 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12619 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12620 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12621
12622 kmem_free(dp, sizeof (dtrace_difo_t));
12623 return (NULL);
12624 }
12625
12626 static dtrace_predicate_t *
12627 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12628 cred_t *cr)
12629 {
12630 dtrace_difo_t *dp;
12631
12632 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12633 return (NULL);
12634
12635 return (dtrace_predicate_create(dp));
12636 }
12637
12638 static dtrace_actdesc_t *
12639 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12640 cred_t *cr)
12641 {
12642 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12643 dof_actdesc_t *desc;
12644 dof_sec_t *difosec;
12645 size_t offs;
12646 uintptr_t daddr = (uintptr_t)dof;
12647 uint64_t arg;
12648 dtrace_actkind_t kind;
12649
12650 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12651 dtrace_dof_error(dof, "invalid action section");
12652 return (NULL);
12653 }
12654
12655 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12656 dtrace_dof_error(dof, "truncated action description");
12657 return (NULL);
12658 }
12659
12660 if (sec->dofs_align != sizeof (uint64_t)) {
12661 dtrace_dof_error(dof, "bad alignment in action description");
12662 return (NULL);
12663 }
12664
12665 if (sec->dofs_size < sec->dofs_entsize) {
12666 dtrace_dof_error(dof, "section entry size exceeds total size");
12667 return (NULL);
12668 }
12669
12670 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12671 dtrace_dof_error(dof, "bad entry size in action description");
12672 return (NULL);
12673 }
12674
12675 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12676 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12677 return (NULL);
12678 }
12679
12680 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12681 desc = (dof_actdesc_t *)(daddr +
12682 (uintptr_t)sec->dofs_offset + offs);
12683 kind = (dtrace_actkind_t)desc->dofa_kind;
12684
12685 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12686 (kind != DTRACEACT_PRINTA ||
12687 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12688 (kind == DTRACEACT_DIFEXPR &&
12689 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12690 dof_sec_t *strtab;
12691 char *str, *fmt;
12692 uint64_t i;
12693
12694 /*
12695 * The argument to these actions is an index into the
12696 * DOF string table. For printf()-like actions, this
12697 * is the format string. For print(), this is the
12698 * CTF type of the expression result.
12699 */
12700 if ((strtab = dtrace_dof_sect(dof,
12701 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12702 goto err;
12703
12704 str = (char *)((uintptr_t)dof +
12705 (uintptr_t)strtab->dofs_offset);
12706
12707 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12708 if (str[i] == '\0')
12709 break;
12710 }
12711
12712 if (i >= strtab->dofs_size) {
12713 dtrace_dof_error(dof, "bogus format string");
12714 goto err;
12715 }
12716
12717 if (i == desc->dofa_arg) {
12718 dtrace_dof_error(dof, "empty format string");
12719 goto err;
12720 }
12721
12722 i -= desc->dofa_arg;
12723 fmt = kmem_alloc(i + 1, KM_SLEEP);
12724 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12725 arg = (uint64_t)(uintptr_t)fmt;
12726 } else {
12727 if (kind == DTRACEACT_PRINTA) {
12728 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12729 arg = 0;
12730 } else {
12731 arg = desc->dofa_arg;
12732 }
12733 }
12734
12735 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12736 desc->dofa_uarg, arg);
12737
12738 if (last != NULL) {
12739 last->dtad_next = act;
12740 } else {
12741 first = act;
12742 }
12743
12744 last = act;
12745
12746 if (desc->dofa_difo == DOF_SECIDX_NONE)
12747 continue;
12748
12749 if ((difosec = dtrace_dof_sect(dof,
12750 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12751 goto err;
12752
12753 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12754
12755 if (act->dtad_difo == NULL)
12756 goto err;
12757 }
12758
12759 ASSERT(first != NULL);
12760 return (first);
12761
12762 err:
12763 for (act = first; act != NULL; act = next) {
12764 next = act->dtad_next;
12765 dtrace_actdesc_release(act, vstate);
12766 }
12767
12768 return (NULL);
12769 }
12770
12771 static dtrace_ecbdesc_t *
12772 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12773 cred_t *cr)
12774 {
12775 dtrace_ecbdesc_t *ep;
12776 dof_ecbdesc_t *ecb;
12777 dtrace_probedesc_t *desc;
12778 dtrace_predicate_t *pred = NULL;
12779
12780 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12781 dtrace_dof_error(dof, "truncated ECB description");
12782 return (NULL);
12783 }
12784
12785 if (sec->dofs_align != sizeof (uint64_t)) {
12786 dtrace_dof_error(dof, "bad alignment in ECB description");
12787 return (NULL);
12788 }
12789
12790 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12791 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12792
12793 if (sec == NULL)
12794 return (NULL);
12795
12796 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12797 ep->dted_uarg = ecb->dofe_uarg;
12798 desc = &ep->dted_probe;
12799
12800 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12801 goto err;
12802
12803 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12804 if ((sec = dtrace_dof_sect(dof,
12805 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12806 goto err;
12807
12808 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12809 goto err;
12810
12811 ep->dted_pred.dtpdd_predicate = pred;
12812 }
12813
12814 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12815 if ((sec = dtrace_dof_sect(dof,
12816 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12817 goto err;
12818
12819 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12820
12821 if (ep->dted_action == NULL)
12822 goto err;
12823 }
12824
12825 return (ep);
12826
12827 err:
12828 if (pred != NULL)
12829 dtrace_predicate_release(pred, vstate);
12830 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12831 return (NULL);
12832 }
12833
12834 /*
12835 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12836 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12837 * site of any user SETX relocations to account for load object base address.
12838 * In the future, if we need other relocations, this function can be extended.
12839 */
12840 static int
12841 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12842 {
12843 uintptr_t daddr = (uintptr_t)dof;
12844 dof_relohdr_t *dofr =
12845 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12846 dof_sec_t *ss, *rs, *ts;
12847 dof_relodesc_t *r;
12848 uint_t i, n;
12849
12850 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12851 sec->dofs_align != sizeof (dof_secidx_t)) {
12852 dtrace_dof_error(dof, "invalid relocation header");
12853 return (-1);
12854 }
12855
12856 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12857 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12858 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12859
12860 if (ss == NULL || rs == NULL || ts == NULL)
12861 return (-1); /* dtrace_dof_error() has been called already */
12862
12863 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12864 rs->dofs_align != sizeof (uint64_t)) {
12865 dtrace_dof_error(dof, "invalid relocation section");
12866 return (-1);
12867 }
12868
12869 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12870 n = rs->dofs_size / rs->dofs_entsize;
12871
12872 for (i = 0; i < n; i++) {
12873 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12874
12875 switch (r->dofr_type) {
12876 case DOF_RELO_NONE:
12877 break;
12878 case DOF_RELO_SETX:
12879 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12880 sizeof (uint64_t) > ts->dofs_size) {
12881 dtrace_dof_error(dof, "bad relocation offset");
12882 return (-1);
12883 }
12884
12885 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12886 dtrace_dof_error(dof, "misaligned setx relo");
12887 return (-1);
12888 }
12889
12890 *(uint64_t *)taddr += ubase;
12891 break;
12892 default:
12893 dtrace_dof_error(dof, "invalid relocation type");
12894 return (-1);
12895 }
12896
12897 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12898 }
12899
12900 return (0);
12901 }
12902
12903 /*
12904 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12905 * header: it should be at the front of a memory region that is at least
12906 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12907 * size. It need not be validated in any other way.
12908 */
12909 static int
12910 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12911 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12912 {
12913 uint64_t len = dof->dofh_loadsz, seclen;
12914 uintptr_t daddr = (uintptr_t)dof;
12915 dtrace_ecbdesc_t *ep;
12916 dtrace_enabling_t *enab;
12917 uint_t i;
12918
12919 ASSERT(MUTEX_HELD(&dtrace_lock));
12920 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12921
12922 /*
12923 * Check the DOF header identification bytes. In addition to checking
12924 * valid settings, we also verify that unused bits/bytes are zeroed so
12925 * we can use them later without fear of regressing existing binaries.
12926 */
12927 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12928 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12929 dtrace_dof_error(dof, "DOF magic string mismatch");
12930 return (-1);
12931 }
12932
12933 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12934 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12935 dtrace_dof_error(dof, "DOF has invalid data model");
12936 return (-1);
12937 }
12938
12939 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12940 dtrace_dof_error(dof, "DOF encoding mismatch");
12941 return (-1);
12942 }
12943
12944 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12945 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12946 dtrace_dof_error(dof, "DOF version mismatch");
12947 return (-1);
12948 }
12949
12950 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12951 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12952 return (-1);
12953 }
12954
12955 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12956 dtrace_dof_error(dof, "DOF uses too many integer registers");
12957 return (-1);
12958 }
12959
12960 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12961 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12962 return (-1);
12963 }
12964
12965 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12966 if (dof->dofh_ident[i] != 0) {
12967 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12968 return (-1);
12969 }
12970 }
12971
12972 if (dof->dofh_flags & ~DOF_FL_VALID) {
12973 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12974 return (-1);
12975 }
12976
12977 if (dof->dofh_secsize == 0) {
12978 dtrace_dof_error(dof, "zero section header size");
12979 return (-1);
12980 }
12981
12982 /*
12983 * Check that the section headers don't exceed the amount of DOF
12984 * data. Note that we cast the section size and number of sections
12985 * to uint64_t's to prevent possible overflow in the multiplication.
12986 */
12987 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12988
12989 if (dof->dofh_secoff > len || seclen > len ||
12990 dof->dofh_secoff + seclen > len) {
12991 dtrace_dof_error(dof, "truncated section headers");
12992 return (-1);
12993 }
12994
12995 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12996 dtrace_dof_error(dof, "misaligned section headers");
12997 return (-1);
12998 }
12999
13000 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
13001 dtrace_dof_error(dof, "misaligned section size");
13002 return (-1);
13003 }
13004
13005 /*
13006 * Take an initial pass through the section headers to be sure that
13007 * the headers don't have stray offsets. If the 'noprobes' flag is
13008 * set, do not permit sections relating to providers, probes, or args.
13009 */
13010 for (i = 0; i < dof->dofh_secnum; i++) {
13011 dof_sec_t *sec = (dof_sec_t *)(daddr +
13012 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13013
13014 if (noprobes) {
13015 switch (sec->dofs_type) {
13016 case DOF_SECT_PROVIDER:
13017 case DOF_SECT_PROBES:
13018 case DOF_SECT_PRARGS:
13019 case DOF_SECT_PROFFS:
13020 dtrace_dof_error(dof, "illegal sections "
13021 "for enabling");
13022 return (-1);
13023 }
13024 }
13025
13026 if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
13027 !(sec->dofs_flags & DOF_SECF_LOAD)) {
13028 dtrace_dof_error(dof, "loadable section with load "
13029 "flag unset");
13030 return (-1);
13031 }
13032
13033 if (!(sec->dofs_flags & DOF_SECF_LOAD))
13034 continue; /* just ignore non-loadable sections */
13035
13036 if (sec->dofs_align & (sec->dofs_align - 1)) {
13037 dtrace_dof_error(dof, "bad section alignment");
13038 return (-1);
13039 }
13040
13041 if (sec->dofs_offset & (sec->dofs_align - 1)) {
13042 dtrace_dof_error(dof, "misaligned section");
13043 return (-1);
13044 }
13045
13046 if (sec->dofs_offset > len || sec->dofs_size > len ||
13047 sec->dofs_offset + sec->dofs_size > len) {
13048 dtrace_dof_error(dof, "corrupt section header");
13049 return (-1);
13050 }
13051
13052 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
13053 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
13054 dtrace_dof_error(dof, "non-terminating string table");
13055 return (-1);
13056 }
13057 }
13058
13059 /*
13060 * Take a second pass through the sections and locate and perform any
13061 * relocations that are present. We do this after the first pass to
13062 * be sure that all sections have had their headers validated.
13063 */
13064 for (i = 0; i < dof->dofh_secnum; i++) {
13065 dof_sec_t *sec = (dof_sec_t *)(daddr +
13066 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13067
13068 if (!(sec->dofs_flags & DOF_SECF_LOAD))
13069 continue; /* skip sections that are not loadable */
13070
13071 switch (sec->dofs_type) {
13072 case DOF_SECT_URELHDR:
13073 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
13074 return (-1);
13075 break;
13076 }
13077 }
13078
13079 if ((enab = *enabp) == NULL)
13080 enab = *enabp = dtrace_enabling_create(vstate);
13081
13082 for (i = 0; i < dof->dofh_secnum; i++) {
13083 dof_sec_t *sec = (dof_sec_t *)(daddr +
13084 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13085
13086 if (sec->dofs_type != DOF_SECT_ECBDESC)
13087 continue;
13088
13089 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
13090 dtrace_enabling_destroy(enab);
13091 *enabp = NULL;
13092 return (-1);
13093 }
13094
13095 dtrace_enabling_add(enab, ep);
13096 }
13097
13098 return (0);
13099 }
13100
13101 /*
13102 * Process DOF for any options. This routine assumes that the DOF has been
13103 * at least processed by dtrace_dof_slurp().
13104 */
13105 static int
13106 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
13107 {
13108 int i, rval;
13109 uint32_t entsize;
13110 size_t offs;
13111 dof_optdesc_t *desc;
13112
13113 for (i = 0; i < dof->dofh_secnum; i++) {
13114 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
13115 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13116
13117 if (sec->dofs_type != DOF_SECT_OPTDESC)
13118 continue;
13119
13120 if (sec->dofs_align != sizeof (uint64_t)) {
13121 dtrace_dof_error(dof, "bad alignment in "
13122 "option description");
13123 return (EINVAL);
13124 }
13125
13126 if ((entsize = sec->dofs_entsize) == 0) {
13127 dtrace_dof_error(dof, "zeroed option entry size");
13128 return (EINVAL);
13129 }
13130
13131 if (entsize < sizeof (dof_optdesc_t)) {
13132 dtrace_dof_error(dof, "bad option entry size");
13133 return (EINVAL);
13134 }
13135
13136 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
13137 desc = (dof_optdesc_t *)((uintptr_t)dof +
13138 (uintptr_t)sec->dofs_offset + offs);
13139
13140 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
13141 dtrace_dof_error(dof, "non-zero option string");
13142 return (EINVAL);
13143 }
13144
13145 if (desc->dofo_value == DTRACEOPT_UNSET) {
13146 dtrace_dof_error(dof, "unset option");
13147 return (EINVAL);
13148 }
13149
13150 if ((rval = dtrace_state_option(state,
13151 desc->dofo_option, desc->dofo_value)) != 0) {
13152 dtrace_dof_error(dof, "rejected option");
13153 return (rval);
13154 }
13155 }
13156 }
13157
13158 return (0);
13159 }
13160
13161 /*
13162 * DTrace Consumer State Functions
13163 */
13164 int
13165 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
13166 {
13167 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
13168 void *base;
13169 uintptr_t limit;
13170 dtrace_dynvar_t *dvar, *next, *start;
13171 int i;
13172
13173 ASSERT(MUTEX_HELD(&dtrace_lock));
13174 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
13175
13176 bzero(dstate, sizeof (dtrace_dstate_t));
13177
13178 if ((dstate->dtds_chunksize = chunksize) == 0)
13179 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
13180
13181 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
13182 size = min;
13183
13184 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
13185 return (ENOMEM);
13186
13187 dstate->dtds_size = size;
13188 dstate->dtds_base = base;
13189 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
13190 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
13191
13192 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
13193
13194 if (hashsize != 1 && (hashsize & 1))
13195 hashsize--;
13196
13197 dstate->dtds_hashsize = hashsize;
13198 dstate->dtds_hash = dstate->dtds_base;
13199
13200 /*
13201 * Set all of our hash buckets to point to the single sink, and (if
13202 * it hasn't already been set), set the sink's hash value to be the
13203 * sink sentinel value. The sink is needed for dynamic variable
13204 * lookups to know that they have iterated over an entire, valid hash
13205 * chain.
13206 */
13207 for (i = 0; i < hashsize; i++)
13208 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
13209
13210 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
13211 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
13212
13213 /*
13214 * Determine number of active CPUs. Divide free list evenly among
13215 * active CPUs.
13216 */
13217 start = (dtrace_dynvar_t *)
13218 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
13219 limit = (uintptr_t)base + size;
13220
13221 maxper = (limit - (uintptr_t)start) / NCPU;
13222 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13223
13224 for (i = 0; i < NCPU; i++) {
13225 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13226
13227 /*
13228 * If we don't even have enough chunks to make it once through
13229 * NCPUs, we're just going to allocate everything to the first
13230 * CPU. And if we're on the last CPU, we're going to allocate
13231 * whatever is left over. In either case, we set the limit to
13232 * be the limit of the dynamic variable space.
13233 */
13234 if (maxper == 0 || i == NCPU - 1) {
13235 limit = (uintptr_t)base + size;
13236 start = NULL;
13237 } else {
13238 limit = (uintptr_t)start + maxper;
13239 start = (dtrace_dynvar_t *)limit;
13240 }
13241
13242 ASSERT(limit <= (uintptr_t)base + size);
13243
13244 for (;;) {
13245 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13246 dstate->dtds_chunksize);
13247
13248 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13249 break;
13250
13251 dvar->dtdv_next = next;
13252 dvar = next;
13253 }
13254
13255 if (maxper == 0)
13256 break;
13257 }
13258
13259 return (0);
13260 }
13261
13262 void
13263 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13264 {
13265 ASSERT(MUTEX_HELD(&cpu_lock));
13266
13267 if (dstate->dtds_base == NULL)
13268 return;
13269
13270 kmem_free(dstate->dtds_base, dstate->dtds_size);
13271 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13272 }
13273
13274 static void
13275 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13276 {
13277 /*
13278 * Logical XOR, where are you?
13279 */
13280 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13281
13282 if (vstate->dtvs_nglobals > 0) {
13283 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13284 sizeof (dtrace_statvar_t *));
13285 }
13286
13287 if (vstate->dtvs_ntlocals > 0) {
13288 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13289 sizeof (dtrace_difv_t));
13290 }
13291
13292 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13293
13294 if (vstate->dtvs_nlocals > 0) {
13295 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13296 sizeof (dtrace_statvar_t *));
13297 }
13298 }
13299
13300 static void
13301 dtrace_state_clean(dtrace_state_t *state)
13302 {
13303 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13304 return;
13305
13306 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13307 dtrace_speculation_clean(state);
13308 }
13309
13310 static void
13311 dtrace_state_deadman(dtrace_state_t *state)
13312 {
13313 hrtime_t now;
13314
13315 dtrace_sync();
13316
13317 now = dtrace_gethrtime();
13318
13319 if (state != dtrace_anon.dta_state &&
13320 now - state->dts_laststatus >= dtrace_deadman_user)
13321 return;
13322
13323 /*
13324 * We must be sure that dts_alive never appears to be less than the
13325 * value upon entry to dtrace_state_deadman(), and because we lack a
13326 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13327 * store INT64_MAX to it, followed by a memory barrier, followed by
13328 * the new value. This assures that dts_alive never appears to be
13329 * less than its true value, regardless of the order in which the
13330 * stores to the underlying storage are issued.
13331 */
13332 state->dts_alive = INT64_MAX;
13333 dtrace_membar_producer();
13334 state->dts_alive = now;
13335 }
13336
13337 dtrace_state_t *
13338 dtrace_state_create(dev_t *devp, cred_t *cr)
13339 {
13340 minor_t minor;
13341 major_t major;
13342 char c[30];
13343 dtrace_state_t *state;
13344 dtrace_optval_t *opt;
13345 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13346
13347 ASSERT(MUTEX_HELD(&dtrace_lock));
13348 ASSERT(MUTEX_HELD(&cpu_lock));
13349
13350 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13351 VM_BESTFIT | VM_SLEEP);
13352
13353 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13354 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13355 return (NULL);
13356 }
13357
13358 state = ddi_get_soft_state(dtrace_softstate, minor);
13359 state->dts_epid = DTRACE_EPIDNONE + 1;
13360
13361 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
13362 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13363 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13364
13365 if (devp != NULL) {
13366 major = getemajor(*devp);
13367 } else {
13368 major = ddi_driver_major(dtrace_devi);
13369 }
13370
13371 state->dts_dev = makedevice(major, minor);
13372
13373 if (devp != NULL)
13374 *devp = state->dts_dev;
13375
13376 /*
13377 * We allocate NCPU buffers. On the one hand, this can be quite
13378 * a bit of memory per instance (nearly 36K on a Starcat). On the
13379 * other hand, it saves an additional memory reference in the probe
13380 * path.
13381 */
13382 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13383 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13384 state->dts_cleaner = CYCLIC_NONE;
13385 state->dts_deadman = CYCLIC_NONE;
13386 state->dts_vstate.dtvs_state = state;
13387
13388 for (i = 0; i < DTRACEOPT_MAX; i++)
13389 state->dts_options[i] = DTRACEOPT_UNSET;
13390
13391 /*
13392 * Set the default options.
13393 */
13394 opt = state->dts_options;
13395 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13396 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13397 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13398 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13399 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13400 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13401 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13402 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13403 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13404 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13405 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13406 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13407 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13408 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13409
13410 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13411
13412 /*
13413 * Depending on the user credentials, we set flag bits which alter probe
13414 * visibility or the amount of destructiveness allowed. In the case of
13415 * actual anonymous tracing, or the possession of all privileges, all of
13416 * the normal checks are bypassed.
13417 */
13418 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13419 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13420 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13421 } else {
13422 /*
13423 * Set up the credentials for this instantiation. We take a
13424 * hold on the credential to prevent it from disappearing on
13425 * us; this in turn prevents the zone_t referenced by this
13426 * credential from disappearing. This means that we can
13427 * examine the credential and the zone from probe context.
13428 */
13429 crhold(cr);
13430 state->dts_cred.dcr_cred = cr;
13431
13432 /*
13433 * CRA_PROC means "we have *some* privilege for dtrace" and
13434 * unlocks the use of variables like pid, zonename, etc.
13435 */
13436 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13437 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13438 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13439 }
13440
13441 /*
13442 * dtrace_user allows use of syscall and profile providers.
13443 * If the user also has proc_owner and/or proc_zone, we
13444 * extend the scope to include additional visibility and
13445 * destructive power.
13446 */
13447 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13448 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13449 state->dts_cred.dcr_visible |=
13450 DTRACE_CRV_ALLPROC;
13451
13452 state->dts_cred.dcr_action |=
13453 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13454 }
13455
13456 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13457 state->dts_cred.dcr_visible |=
13458 DTRACE_CRV_ALLZONE;
13459
13460 state->dts_cred.dcr_action |=
13461 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13462 }
13463
13464 /*
13465 * If we have all privs in whatever zone this is,
13466 * we can do destructive things to processes which
13467 * have altered credentials.
13468 */
13469 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13470 cr->cr_zone->zone_privset)) {
13471 state->dts_cred.dcr_action |=
13472 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13473 }
13474 }
13475
13476 /*
13477 * Holding the dtrace_kernel privilege also implies that
13478 * the user has the dtrace_user privilege from a visibility
13479 * perspective. But without further privileges, some
13480 * destructive actions are not available.
13481 */
13482 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13483 /*
13484 * Make all probes in all zones visible. However,
13485 * this doesn't mean that all actions become available
13486 * to all zones.
13487 */
13488 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13489 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13490
13491 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13492 DTRACE_CRA_PROC;
13493 /*
13494 * Holding proc_owner means that destructive actions
13495 * for *this* zone are allowed.
13496 */
13497 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13498 state->dts_cred.dcr_action |=
13499 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13500
13501 /*
13502 * Holding proc_zone means that destructive actions
13503 * for this user/group ID in all zones is allowed.
13504 */
13505 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13506 state->dts_cred.dcr_action |=
13507 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13508
13509 /*
13510 * If we have all privs in whatever zone this is,
13511 * we can do destructive things to processes which
13512 * have altered credentials.
13513 */
13514 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13515 cr->cr_zone->zone_privset)) {
13516 state->dts_cred.dcr_action |=
13517 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13518 }
13519 }
13520
13521 /*
13522 * Holding the dtrace_proc privilege gives control over fasttrap
13523 * and pid providers. We need to grant wider destructive
13524 * privileges in the event that the user has proc_owner and/or
13525 * proc_zone.
13526 */
13527 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13528 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13529 state->dts_cred.dcr_action |=
13530 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13531
13532 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13533 state->dts_cred.dcr_action |=
13534 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13535 }
13536 }
13537
13538 return (state);
13539 }
13540
13541 static int
13542 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13543 {
13544 dtrace_optval_t *opt = state->dts_options, size;
13545 processorid_t cpu;
13546 int flags = 0, rval, factor, divisor = 1;
13547
13548 ASSERT(MUTEX_HELD(&dtrace_lock));
13549 ASSERT(MUTEX_HELD(&cpu_lock));
13550 ASSERT(which < DTRACEOPT_MAX);
13551 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13552 (state == dtrace_anon.dta_state &&
13553 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13554
13555 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13556 return (0);
13557
13558 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13559 cpu = opt[DTRACEOPT_CPU];
13560
13561 if (which == DTRACEOPT_SPECSIZE)
13562 flags |= DTRACEBUF_NOSWITCH;
13563
13564 if (which == DTRACEOPT_BUFSIZE) {
13565 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13566 flags |= DTRACEBUF_RING;
13567
13568 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13569 flags |= DTRACEBUF_FILL;
13570
13571 if (state != dtrace_anon.dta_state ||
13572 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13573 flags |= DTRACEBUF_INACTIVE;
13574 }
13575
13576 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
13577 /*
13578 * The size must be 8-byte aligned. If the size is not 8-byte
13579 * aligned, drop it down by the difference.
13580 */
13581 if (size & (sizeof (uint64_t) - 1))
13582 size -= size & (sizeof (uint64_t) - 1);
13583
13584 if (size < state->dts_reserve) {
13585 /*
13586 * Buffers always must be large enough to accommodate
13587 * their prereserved space. We return E2BIG instead
13588 * of ENOMEM in this case to allow for user-level
13589 * software to differentiate the cases.
13590 */
13591 return (E2BIG);
13592 }
13593
13594 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
13595
13596 if (rval != ENOMEM) {
13597 opt[which] = size;
13598 return (rval);
13599 }
13600
13601 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13602 return (rval);
13603
13604 for (divisor = 2; divisor < factor; divisor <<= 1)
13605 continue;
13606 }
13607
13608 return (ENOMEM);
13609 }
13610
13611 static int
13612 dtrace_state_buffers(dtrace_state_t *state)
13613 {
13614 dtrace_speculation_t *spec = state->dts_speculations;
13615 int rval, i;
13616
13617 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13618 DTRACEOPT_BUFSIZE)) != 0)
13619 return (rval);
13620
13621 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13622 DTRACEOPT_AGGSIZE)) != 0)
13623 return (rval);
13624
13625 for (i = 0; i < state->dts_nspeculations; i++) {
13626 if ((rval = dtrace_state_buffer(state,
13627 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13628 return (rval);
13629 }
13630
13631 return (0);
13632 }
13633
13634 static void
13635 dtrace_state_prereserve(dtrace_state_t *state)
13636 {
13637 dtrace_ecb_t *ecb;
13638 dtrace_probe_t *probe;
13639
13640 state->dts_reserve = 0;
13641
13642 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13643 return;
13644
13645 /*
13646 * If our buffer policy is a "fill" buffer policy, we need to set the
13647 * prereserved space to be the space required by the END probes.
13648 */
13649 probe = dtrace_probes[dtrace_probeid_end - 1];
13650 ASSERT(probe != NULL);
13651
13652 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13653 if (ecb->dte_state != state)
13654 continue;
13655
13656 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13657 }
13658 }
13659
13660 static int
13661 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13662 {
13663 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13664 dtrace_speculation_t *spec;
13665 dtrace_buffer_t *buf;
13666 cyc_handler_t hdlr;
13667 cyc_time_t when;
13668 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13669 dtrace_icookie_t cookie;
13670
13671 mutex_enter(&cpu_lock);
13672 mutex_enter(&dtrace_lock);
13673
13674 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13675 rval = EBUSY;
13676 goto out;
13677 }
13678
13679 /*
13680 * Before we can perform any checks, we must prime all of the
13681 * retained enablings that correspond to this state.
13682 */
13683 dtrace_enabling_prime(state);
13684
13685 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13686 rval = EACCES;
13687 goto out;
13688 }
13689
13690 dtrace_state_prereserve(state);
13691
13692 /*
13693 * Now we want to do is try to allocate our speculations.
13694 * We do not automatically resize the number of speculations; if
13695 * this fails, we will fail the operation.
13696 */
13697 nspec = opt[DTRACEOPT_NSPEC];
13698 ASSERT(nspec != DTRACEOPT_UNSET);
13699
13700 if (nspec > INT_MAX) {
13701 rval = ENOMEM;
13702 goto out;
13703 }
13704
13705 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
13706 KM_NOSLEEP | KM_NORMALPRI);
13707
13708 if (spec == NULL) {
13709 rval = ENOMEM;
13710 goto out;
13711 }
13712
13713 state->dts_speculations = spec;
13714 state->dts_nspeculations = (int)nspec;
13715
13716 for (i = 0; i < nspec; i++) {
13717 if ((buf = kmem_zalloc(bufsize,
13718 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
13719 rval = ENOMEM;
13720 goto err;
13721 }
13722
13723 spec[i].dtsp_buffer = buf;
13724 }
13725
13726 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13727 if (dtrace_anon.dta_state == NULL) {
13728 rval = ENOENT;
13729 goto out;
13730 }
13731
13732 if (state->dts_necbs != 0) {
13733 rval = EALREADY;
13734 goto out;
13735 }
13736
13737 state->dts_anon = dtrace_anon_grab();
13738 ASSERT(state->dts_anon != NULL);
13739 state = state->dts_anon;
13740
13741 /*
13742 * We want "grabanon" to be set in the grabbed state, so we'll
13743 * copy that option value from the grabbing state into the
13744 * grabbed state.
13745 */
13746 state->dts_options[DTRACEOPT_GRABANON] =
13747 opt[DTRACEOPT_GRABANON];
13748
13749 *cpu = dtrace_anon.dta_beganon;
13750
13751 /*
13752 * If the anonymous state is active (as it almost certainly
13753 * is if the anonymous enabling ultimately matched anything),
13754 * we don't allow any further option processing -- but we
13755 * don't return failure.
13756 */
13757 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13758 goto out;
13759 }
13760
13761 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13762 opt[DTRACEOPT_AGGSIZE] != 0) {
13763 if (state->dts_aggregations == NULL) {
13764 /*
13765 * We're not going to create an aggregation buffer
13766 * because we don't have any ECBs that contain
13767 * aggregations -- set this option to 0.
13768 */
13769 opt[DTRACEOPT_AGGSIZE] = 0;
13770 } else {
13771 /*
13772 * If we have an aggregation buffer, we must also have
13773 * a buffer to use as scratch.
13774 */
13775 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13776 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13777 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13778 }
13779 }
13780 }
13781
13782 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13783 opt[DTRACEOPT_SPECSIZE] != 0) {
13784 if (!state->dts_speculates) {
13785 /*
13786 * We're not going to create speculation buffers
13787 * because we don't have any ECBs that actually
13788 * speculate -- set the speculation size to 0.
13789 */
13790 opt[DTRACEOPT_SPECSIZE] = 0;
13791 }
13792 }
13793
13794 /*
13795 * The bare minimum size for any buffer that we're actually going to
13796 * do anything to is sizeof (uint64_t).
13797 */
13798 sz = sizeof (uint64_t);
13799
13800 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13801 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13802 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13803 /*
13804 * A buffer size has been explicitly set to 0 (or to a size
13805 * that will be adjusted to 0) and we need the space -- we
13806 * need to return failure. We return ENOSPC to differentiate
13807 * it from failing to allocate a buffer due to failure to meet
13808 * the reserve (for which we return E2BIG).
13809 */
13810 rval = ENOSPC;
13811 goto out;
13812 }
13813
13814 if ((rval = dtrace_state_buffers(state)) != 0)
13815 goto err;
13816
13817 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13818 sz = dtrace_dstate_defsize;
13819
13820 do {
13821 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13822
13823 if (rval == 0)
13824 break;
13825
13826 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13827 goto err;
13828 } while (sz >>= 1);
13829
13830 opt[DTRACEOPT_DYNVARSIZE] = sz;
13831
13832 if (rval != 0)
13833 goto err;
13834
13835 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13836 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13837
13838 if (opt[DTRACEOPT_CLEANRATE] == 0)
13839 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13840
13841 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13842 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13843
13844 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13845 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13846
13847 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13848 hdlr.cyh_arg = state;
13849 hdlr.cyh_level = CY_LOW_LEVEL;
13850
13851 when.cyt_when = 0;
13852 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13853
13854 state->dts_cleaner = cyclic_add(&hdlr, &when);
13855
13856 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13857 hdlr.cyh_arg = state;
13858 hdlr.cyh_level = CY_LOW_LEVEL;
13859
13860 when.cyt_when = 0;
13861 when.cyt_interval = dtrace_deadman_interval;
13862
13863 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13864 state->dts_deadman = cyclic_add(&hdlr, &when);
13865
13866 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13867
13868 if (state->dts_getf != 0 &&
13869 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13870 /*
13871 * We don't have kernel privs but we have at least one call
13872 * to getf(); we need to bump our zone's count, and (if
13873 * this is the first enabling to have an unprivileged call
13874 * to getf()) we need to hook into closef().
13875 */
13876 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
13877
13878 if (dtrace_getf++ == 0) {
13879 ASSERT(dtrace_closef == NULL);
13880 dtrace_closef = dtrace_getf_barrier;
13881 }
13882 }
13883
13884 /*
13885 * Now it's time to actually fire the BEGIN probe. We need to disable
13886 * interrupts here both to record the CPU on which we fired the BEGIN
13887 * probe (the data from this CPU will be processed first at user
13888 * level) and to manually activate the buffer for this CPU.
13889 */
13890 cookie = dtrace_interrupt_disable();
13891 *cpu = CPU->cpu_id;
13892 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13893 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13894
13895 dtrace_probe(dtrace_probeid_begin,
13896 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13897 dtrace_interrupt_enable(cookie);
13898 /*
13899 * We may have had an exit action from a BEGIN probe; only change our
13900 * state to ACTIVE if we're still in WARMUP.
13901 */
13902 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13903 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13904
13905 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13906 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13907
13908 /*
13909 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13910 * want each CPU to transition its principal buffer out of the
13911 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13912 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13913 * atomically transition from processing none of a state's ECBs to
13914 * processing all of them.
13915 */
13916 dtrace_xcall(DTRACE_CPUALL,
13917 (dtrace_xcall_t)dtrace_buffer_activate, state);
13918 goto out;
13919
13920 err:
13921 dtrace_buffer_free(state->dts_buffer);
13922 dtrace_buffer_free(state->dts_aggbuffer);
13923
13924 if ((nspec = state->dts_nspeculations) == 0) {
13925 ASSERT(state->dts_speculations == NULL);
13926 goto out;
13927 }
13928
13929 spec = state->dts_speculations;
13930 ASSERT(spec != NULL);
13931
13932 for (i = 0; i < state->dts_nspeculations; i++) {
13933 if ((buf = spec[i].dtsp_buffer) == NULL)
13934 break;
13935
13936 dtrace_buffer_free(buf);
13937 kmem_free(buf, bufsize);
13938 }
13939
13940 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13941 state->dts_nspeculations = 0;
13942 state->dts_speculations = NULL;
13943
13944 out:
13945 mutex_exit(&dtrace_lock);
13946 mutex_exit(&cpu_lock);
13947
13948 return (rval);
13949 }
13950
13951 static int
13952 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13953 {
13954 dtrace_icookie_t cookie;
13955
13956 ASSERT(MUTEX_HELD(&dtrace_lock));
13957
13958 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13959 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13960 return (EINVAL);
13961
13962 /*
13963 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13964 * to be sure that every CPU has seen it. See below for the details
13965 * on why this is done.
13966 */
13967 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13968 dtrace_sync();
13969
13970 /*
13971 * By this point, it is impossible for any CPU to be still processing
13972 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13973 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13974 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13975 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13976 * iff we're in the END probe.
13977 */
13978 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13979 dtrace_sync();
13980 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13981
13982 /*
13983 * Finally, we can release the reserve and call the END probe. We
13984 * disable interrupts across calling the END probe to allow us to
13985 * return the CPU on which we actually called the END probe. This
13986 * allows user-land to be sure that this CPU's principal buffer is
13987 * processed last.
13988 */
13989 state->dts_reserve = 0;
13990
13991 cookie = dtrace_interrupt_disable();
13992 *cpu = CPU->cpu_id;
13993 dtrace_probe(dtrace_probeid_end,
13994 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13995 dtrace_interrupt_enable(cookie);
13996
13997 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13998 dtrace_sync();
13999
14000 if (state->dts_getf != 0 &&
14001 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14002 /*
14003 * We don't have kernel privs but we have at least one call
14004 * to getf(); we need to lower our zone's count, and (if
14005 * this is the last enabling to have an unprivileged call
14006 * to getf()) we need to clear the closef() hook.
14007 */
14008 ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
14009 ASSERT(dtrace_closef == dtrace_getf_barrier);
14010 ASSERT(dtrace_getf > 0);
14011
14012 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
14013
14014 if (--dtrace_getf == 0)
14015 dtrace_closef = NULL;
14016 }
14017
14018 return (0);
14019 }
14020
14021 static int
14022 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
14023 dtrace_optval_t val)
14024 {
14025 ASSERT(MUTEX_HELD(&dtrace_lock));
14026
14027 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14028 return (EBUSY);
14029
14030 if (option >= DTRACEOPT_MAX)
14031 return (EINVAL);
14032
14033 if (option != DTRACEOPT_CPU && val < 0)
14034 return (EINVAL);
14035
14036 switch (option) {
14037 case DTRACEOPT_DESTRUCTIVE:
14038 if (dtrace_destructive_disallow)
14039 return (EACCES);
14040
14041 state->dts_cred.dcr_destructive = 1;
14042 break;
14043
14044 case DTRACEOPT_BUFSIZE:
14045 case DTRACEOPT_DYNVARSIZE:
14046 case DTRACEOPT_AGGSIZE:
14047 case DTRACEOPT_SPECSIZE:
14048 case DTRACEOPT_STRSIZE:
14049 if (val < 0)
14050 return (EINVAL);
14051
14052 if (val >= LONG_MAX) {
14053 /*
14054 * If this is an otherwise negative value, set it to
14055 * the highest multiple of 128m less than LONG_MAX.
14056 * Technically, we're adjusting the size without
14057 * regard to the buffer resizing policy, but in fact,
14058 * this has no effect -- if we set the buffer size to
14059 * ~LONG_MAX and the buffer policy is ultimately set to
14060 * be "manual", the buffer allocation is guaranteed to
14061 * fail, if only because the allocation requires two
14062 * buffers. (We set the the size to the highest
14063 * multiple of 128m because it ensures that the size
14064 * will remain a multiple of a megabyte when
14065 * repeatedly halved -- all the way down to 15m.)
14066 */
14067 val = LONG_MAX - (1 << 27) + 1;
14068 }
14069 }
14070
14071 state->dts_options[option] = val;
14072
14073 return (0);
14074 }
14075
14076 static void
14077 dtrace_state_destroy(dtrace_state_t *state)
14078 {
14079 dtrace_ecb_t *ecb;
14080 dtrace_vstate_t *vstate = &state->dts_vstate;
14081 minor_t minor = getminor(state->dts_dev);
14082 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14083 dtrace_speculation_t *spec = state->dts_speculations;
14084 int nspec = state->dts_nspeculations;
14085 uint32_t match;
14086
14087 ASSERT(MUTEX_HELD(&dtrace_lock));
14088 ASSERT(MUTEX_HELD(&cpu_lock));
14089
14090 /*
14091 * First, retract any retained enablings for this state.
14092 */
14093 dtrace_enabling_retract(state);
14094 ASSERT(state->dts_nretained == 0);
14095
14096 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
14097 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
14098 /*
14099 * We have managed to come into dtrace_state_destroy() on a
14100 * hot enabling -- almost certainly because of a disorderly
14101 * shutdown of a consumer. (That is, a consumer that is
14102 * exiting without having called dtrace_stop().) In this case,
14103 * we're going to set our activity to be KILLED, and then
14104 * issue a sync to be sure that everyone is out of probe
14105 * context before we start blowing away ECBs.
14106 */
14107 state->dts_activity = DTRACE_ACTIVITY_KILLED;
14108 dtrace_sync();
14109 }
14110
14111 /*
14112 * Release the credential hold we took in dtrace_state_create().
14113 */
14114 if (state->dts_cred.dcr_cred != NULL)
14115 crfree(state->dts_cred.dcr_cred);
14116
14117 /*
14118 * Now we can safely disable and destroy any enabled probes. Because
14119 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
14120 * (especially if they're all enabled), we take two passes through the
14121 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
14122 * in the second we disable whatever is left over.
14123 */
14124 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
14125 for (i = 0; i < state->dts_necbs; i++) {
14126 if ((ecb = state->dts_ecbs[i]) == NULL)
14127 continue;
14128
14129 if (match && ecb->dte_probe != NULL) {
14130 dtrace_probe_t *probe = ecb->dte_probe;
14131 dtrace_provider_t *prov = probe->dtpr_provider;
14132
14133 if (!(prov->dtpv_priv.dtpp_flags & match))
14134 continue;
14135 }
14136
14137 dtrace_ecb_disable(ecb);
14138 dtrace_ecb_destroy(ecb);
14139 }
14140
14141 if (!match)
14142 break;
14143 }
14144
14145 /*
14146 * Before we free the buffers, perform one more sync to assure that
14147 * every CPU is out of probe context.
14148 */
14149 dtrace_sync();
14150
14151 dtrace_buffer_free(state->dts_buffer);
14152 dtrace_buffer_free(state->dts_aggbuffer);
14153
14154 for (i = 0; i < nspec; i++)
14155 dtrace_buffer_free(spec[i].dtsp_buffer);
14156
14157 if (state->dts_cleaner != CYCLIC_NONE)
14158 cyclic_remove(state->dts_cleaner);
14159
14160 if (state->dts_deadman != CYCLIC_NONE)
14161 cyclic_remove(state->dts_deadman);
14162
14163 dtrace_dstate_fini(&vstate->dtvs_dynvars);
14164 dtrace_vstate_fini(vstate);
14165 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14166
14167 if (state->dts_aggregations != NULL) {
14168 #ifdef DEBUG
14169 for (i = 0; i < state->dts_naggregations; i++)
14170 ASSERT(state->dts_aggregations[i] == NULL);
14171 #endif
14172 ASSERT(state->dts_naggregations > 0);
14173 kmem_free(state->dts_aggregations,
14174 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14175 }
14176
14177 kmem_free(state->dts_buffer, bufsize);
14178 kmem_free(state->dts_aggbuffer, bufsize);
14179
14180 for (i = 0; i < nspec; i++)
14181 kmem_free(spec[i].dtsp_buffer, bufsize);
14182
14183 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14184
14185 dtrace_format_destroy(state);
14186
14187 vmem_destroy(state->dts_aggid_arena);
14188 ddi_soft_state_free(dtrace_softstate, minor);
14189 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14190 }
14191
14192 /*
14193 * DTrace Anonymous Enabling Functions
14194 */
14195 static dtrace_state_t *
14196 dtrace_anon_grab(void)
14197 {
14198 dtrace_state_t *state;
14199
14200 ASSERT(MUTEX_HELD(&dtrace_lock));
14201
14202 if ((state = dtrace_anon.dta_state) == NULL) {
14203 ASSERT(dtrace_anon.dta_enabling == NULL);
14204 return (NULL);
14205 }
14206
14207 ASSERT(dtrace_anon.dta_enabling != NULL);
14208 ASSERT(dtrace_retained != NULL);
14209
14210 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14211 dtrace_anon.dta_enabling = NULL;
14212 dtrace_anon.dta_state = NULL;
14213
14214 return (state);
14215 }
14216
14217 static void
14218 dtrace_anon_property(void)
14219 {
14220 int i, rv;
14221 dtrace_state_t *state;
14222 dof_hdr_t *dof;
14223 char c[32]; /* enough for "dof-data-" + digits */
14224
14225 ASSERT(MUTEX_HELD(&dtrace_lock));
14226 ASSERT(MUTEX_HELD(&cpu_lock));
14227
14228 for (i = 0; ; i++) {
14229 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14230
14231 dtrace_err_verbose = 1;
14232
14233 if ((dof = dtrace_dof_property(c)) == NULL) {
14234 dtrace_err_verbose = 0;
14235 break;
14236 }
14237
14238 /*
14239 * We want to create anonymous state, so we need to transition
14240 * the kernel debugger to indicate that DTrace is active. If
14241 * this fails (e.g. because the debugger has modified text in
14242 * some way), we won't continue with the processing.
14243 */
14244 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14245 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14246 "enabling ignored.");
14247 dtrace_dof_destroy(dof);
14248 break;
14249 }
14250
14251 /*
14252 * If we haven't allocated an anonymous state, we'll do so now.
14253 */
14254 if ((state = dtrace_anon.dta_state) == NULL) {
14255 state = dtrace_state_create(NULL, NULL);
14256 dtrace_anon.dta_state = state;
14257
14258 if (state == NULL) {
14259 /*
14260 * This basically shouldn't happen: the only
14261 * failure mode from dtrace_state_create() is a
14262 * failure of ddi_soft_state_zalloc() that
14263 * itself should never happen. Still, the
14264 * interface allows for a failure mode, and
14265 * we want to fail as gracefully as possible:
14266 * we'll emit an error message and cease
14267 * processing anonymous state in this case.
14268 */
14269 cmn_err(CE_WARN, "failed to create "
14270 "anonymous state");
14271 dtrace_dof_destroy(dof);
14272 break;
14273 }
14274 }
14275
14276 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14277 &dtrace_anon.dta_enabling, 0, B_TRUE);
14278
14279 if (rv == 0)
14280 rv = dtrace_dof_options(dof, state);
14281
14282 dtrace_err_verbose = 0;
14283 dtrace_dof_destroy(dof);
14284
14285 if (rv != 0) {
14286 /*
14287 * This is malformed DOF; chuck any anonymous state
14288 * that we created.
14289 */
14290 ASSERT(dtrace_anon.dta_enabling == NULL);
14291 dtrace_state_destroy(state);
14292 dtrace_anon.dta_state = NULL;
14293 break;
14294 }
14295
14296 ASSERT(dtrace_anon.dta_enabling != NULL);
14297 }
14298
14299 if (dtrace_anon.dta_enabling != NULL) {
14300 int rval;
14301
14302 /*
14303 * dtrace_enabling_retain() can only fail because we are
14304 * trying to retain more enablings than are allowed -- but
14305 * we only have one anonymous enabling, and we are guaranteed
14306 * to be allowed at least one retained enabling; we assert
14307 * that dtrace_enabling_retain() returns success.
14308 */
14309 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14310 ASSERT(rval == 0);
14311
14312 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14313 }
14314 }
14315
14316 /*
14317 * DTrace Helper Functions
14318 */
14319 static void
14320 dtrace_helper_trace(dtrace_helper_action_t *helper,
14321 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14322 {
14323 uint32_t size, next, nnext, i;
14324 dtrace_helptrace_t *ent;
14325 uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14326
14327 if (!dtrace_helptrace_enabled)
14328 return;
14329
14330 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14331
14332 /*
14333 * What would a tracing framework be without its own tracing
14334 * framework? (Well, a hell of a lot simpler, for starters...)
14335 */
14336 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14337 sizeof (uint64_t) - sizeof (uint64_t);
14338
14339 /*
14340 * Iterate until we can allocate a slot in the trace buffer.
14341 */
14342 do {
14343 next = dtrace_helptrace_next;
14344
14345 if (next + size < dtrace_helptrace_bufsize) {
14346 nnext = next + size;
14347 } else {
14348 nnext = size;
14349 }
14350 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14351
14352 /*
14353 * We have our slot; fill it in.
14354 */
14355 if (nnext == size)
14356 next = 0;
14357
14358 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
14359 ent->dtht_helper = helper;
14360 ent->dtht_where = where;
14361 ent->dtht_nlocals = vstate->dtvs_nlocals;
14362
14363 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14364 mstate->dtms_fltoffs : -1;
14365 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14366 ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
14367
14368 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14369 dtrace_statvar_t *svar;
14370
14371 if ((svar = vstate->dtvs_locals[i]) == NULL)
14372 continue;
14373
14374 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14375 ent->dtht_locals[i] =
14376 ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
14377 }
14378 }
14379
14380 static uint64_t
14381 dtrace_helper(int which, dtrace_mstate_t *mstate,
14382 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14383 {
14384 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14385 uint64_t sarg0 = mstate->dtms_arg[0];
14386 uint64_t sarg1 = mstate->dtms_arg[1];
14387 uint64_t rval;
14388 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14389 dtrace_helper_action_t *helper;
14390 dtrace_vstate_t *vstate;
14391 dtrace_difo_t *pred;
14392 int i, trace = dtrace_helptrace_enabled;
14393
14394 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14395
14396 if (helpers == NULL)
14397 return (0);
14398
14399 if ((helper = helpers->dthps_actions[which]) == NULL)
14400 return (0);
14401
14402 vstate = &helpers->dthps_vstate;
14403 mstate->dtms_arg[0] = arg0;
14404 mstate->dtms_arg[1] = arg1;
14405
14406 /*
14407 * Now iterate over each helper. If its predicate evaluates to 'true',
14408 * we'll call the corresponding actions. Note that the below calls
14409 * to dtrace_dif_emulate() may set faults in machine state. This is
14410 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14411 * the stored DIF offset with its own (which is the desired behavior).
14412 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14413 * from machine state; this is okay, too.
14414 */
14415 for (; helper != NULL; helper = helper->dtha_next) {
14416 if ((pred = helper->dtha_predicate) != NULL) {
14417 if (trace)
14418 dtrace_helper_trace(helper, mstate, vstate, 0);
14419
14420 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14421 goto next;
14422
14423 if (*flags & CPU_DTRACE_FAULT)
14424 goto err;
14425 }
14426
14427 for (i = 0; i < helper->dtha_nactions; i++) {
14428 if (trace)
14429 dtrace_helper_trace(helper,
14430 mstate, vstate, i + 1);
14431
14432 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14433 mstate, vstate, state);
14434
14435 if (*flags & CPU_DTRACE_FAULT)
14436 goto err;
14437 }
14438
14439 next:
14440 if (trace)
14441 dtrace_helper_trace(helper, mstate, vstate,
14442 DTRACE_HELPTRACE_NEXT);
14443 }
14444
14445 if (trace)
14446 dtrace_helper_trace(helper, mstate, vstate,
14447 DTRACE_HELPTRACE_DONE);
14448
14449 /*
14450 * Restore the arg0 that we saved upon entry.
14451 */
14452 mstate->dtms_arg[0] = sarg0;
14453 mstate->dtms_arg[1] = sarg1;
14454
14455 return (rval);
14456
14457 err:
14458 if (trace)
14459 dtrace_helper_trace(helper, mstate, vstate,
14460 DTRACE_HELPTRACE_ERR);
14461
14462 /*
14463 * Restore the arg0 that we saved upon entry.
14464 */
14465 mstate->dtms_arg[0] = sarg0;
14466 mstate->dtms_arg[1] = sarg1;
14467
14468 return (NULL);
14469 }
14470
14471 static void
14472 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14473 dtrace_vstate_t *vstate)
14474 {
14475 int i;
14476
14477 if (helper->dtha_predicate != NULL)
14478 dtrace_difo_release(helper->dtha_predicate, vstate);
14479
14480 for (i = 0; i < helper->dtha_nactions; i++) {
14481 ASSERT(helper->dtha_actions[i] != NULL);
14482 dtrace_difo_release(helper->dtha_actions[i], vstate);
14483 }
14484
14485 kmem_free(helper->dtha_actions,
14486 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14487 kmem_free(helper, sizeof (dtrace_helper_action_t));
14488 }
14489
14490 static int
14491 dtrace_helper_destroygen(int gen)
14492 {
14493 proc_t *p = curproc;
14494 dtrace_helpers_t *help = p->p_dtrace_helpers;
14495 dtrace_vstate_t *vstate;
14496 int i;
14497
14498 ASSERT(MUTEX_HELD(&dtrace_lock));
14499
14500 if (help == NULL || gen > help->dthps_generation)
14501 return (EINVAL);
14502
14503 vstate = &help->dthps_vstate;
14504
14505 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14506 dtrace_helper_action_t *last = NULL, *h, *next;
14507
14508 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14509 next = h->dtha_next;
14510
14511 if (h->dtha_generation == gen) {
14512 if (last != NULL) {
14513 last->dtha_next = next;
14514 } else {
14515 help->dthps_actions[i] = next;
14516 }
14517
14518 dtrace_helper_action_destroy(h, vstate);
14519 } else {
14520 last = h;
14521 }
14522 }
14523 }
14524
14525 /*
14526 * Interate until we've cleared out all helper providers with the
14527 * given generation number.
14528 */
14529 for (;;) {
14530 dtrace_helper_provider_t *prov;
14531
14532 /*
14533 * Look for a helper provider with the right generation. We
14534 * have to start back at the beginning of the list each time
14535 * because we drop dtrace_lock. It's unlikely that we'll make
14536 * more than two passes.
14537 */
14538 for (i = 0; i < help->dthps_nprovs; i++) {
14539 prov = help->dthps_provs[i];
14540
14541 if (prov->dthp_generation == gen)
14542 break;
14543 }
14544
14545 /*
14546 * If there were no matches, we're done.
14547 */
14548 if (i == help->dthps_nprovs)
14549 break;
14550
14551 /*
14552 * Move the last helper provider into this slot.
14553 */
14554 help->dthps_nprovs--;
14555 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14556 help->dthps_provs[help->dthps_nprovs] = NULL;
14557
14558 mutex_exit(&dtrace_lock);
14559
14560 /*
14561 * If we have a meta provider, remove this helper provider.
14562 */
14563 mutex_enter(&dtrace_meta_lock);
14564 if (dtrace_meta_pid != NULL) {
14565 ASSERT(dtrace_deferred_pid == NULL);
14566 dtrace_helper_provider_remove(&prov->dthp_prov,
14567 p->p_pid);
14568 }
14569 mutex_exit(&dtrace_meta_lock);
14570
14571 dtrace_helper_provider_destroy(prov);
14572
14573 mutex_enter(&dtrace_lock);
14574 }
14575
14576 return (0);
14577 }
14578
14579 static int
14580 dtrace_helper_validate(dtrace_helper_action_t *helper)
14581 {
14582 int err = 0, i;
14583 dtrace_difo_t *dp;
14584
14585 if ((dp = helper->dtha_predicate) != NULL)
14586 err += dtrace_difo_validate_helper(dp);
14587
14588 for (i = 0; i < helper->dtha_nactions; i++)
14589 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14590
14591 return (err == 0);
14592 }
14593
14594 static int
14595 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14596 {
14597 dtrace_helpers_t *help;
14598 dtrace_helper_action_t *helper, *last;
14599 dtrace_actdesc_t *act;
14600 dtrace_vstate_t *vstate;
14601 dtrace_predicate_t *pred;
14602 int count = 0, nactions = 0, i;
14603
14604 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14605 return (EINVAL);
14606
14607 help = curproc->p_dtrace_helpers;
14608 last = help->dthps_actions[which];
14609 vstate = &help->dthps_vstate;
14610
14611 for (count = 0; last != NULL; last = last->dtha_next) {
14612 count++;
14613 if (last->dtha_next == NULL)
14614 break;
14615 }
14616
14617 /*
14618 * If we already have dtrace_helper_actions_max helper actions for this
14619 * helper action type, we'll refuse to add a new one.
14620 */
14621 if (count >= dtrace_helper_actions_max)
14622 return (ENOSPC);
14623
14624 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14625 helper->dtha_generation = help->dthps_generation;
14626
14627 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14628 ASSERT(pred->dtp_difo != NULL);
14629 dtrace_difo_hold(pred->dtp_difo);
14630 helper->dtha_predicate = pred->dtp_difo;
14631 }
14632
14633 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14634 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14635 goto err;
14636
14637 if (act->dtad_difo == NULL)
14638 goto err;
14639
14640 nactions++;
14641 }
14642
14643 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14644 (helper->dtha_nactions = nactions), KM_SLEEP);
14645
14646 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14647 dtrace_difo_hold(act->dtad_difo);
14648 helper->dtha_actions[i++] = act->dtad_difo;
14649 }
14650
14651 if (!dtrace_helper_validate(helper))
14652 goto err;
14653
14654 if (last == NULL) {
14655 help->dthps_actions[which] = helper;
14656 } else {
14657 last->dtha_next = helper;
14658 }
14659
14660 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14661 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14662 dtrace_helptrace_next = 0;
14663 }
14664
14665 return (0);
14666 err:
14667 dtrace_helper_action_destroy(helper, vstate);
14668 return (EINVAL);
14669 }
14670
14671 static void
14672 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14673 dof_helper_t *dofhp)
14674 {
14675 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14676
14677 mutex_enter(&dtrace_meta_lock);
14678 mutex_enter(&dtrace_lock);
14679
14680 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14681 /*
14682 * If the dtrace module is loaded but not attached, or if
14683 * there aren't isn't a meta provider registered to deal with
14684 * these provider descriptions, we need to postpone creating
14685 * the actual providers until later.
14686 */
14687
14688 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14689 dtrace_deferred_pid != help) {
14690 help->dthps_deferred = 1;
14691 help->dthps_pid = p->p_pid;
14692 help->dthps_next = dtrace_deferred_pid;
14693 help->dthps_prev = NULL;
14694 if (dtrace_deferred_pid != NULL)
14695 dtrace_deferred_pid->dthps_prev = help;
14696 dtrace_deferred_pid = help;
14697 }
14698
14699 mutex_exit(&dtrace_lock);
14700
14701 } else if (dofhp != NULL) {
14702 /*
14703 * If the dtrace module is loaded and we have a particular
14704 * helper provider description, pass that off to the
14705 * meta provider.
14706 */
14707
14708 mutex_exit(&dtrace_lock);
14709
14710 dtrace_helper_provide(dofhp, p->p_pid);
14711
14712 } else {
14713 /*
14714 * Otherwise, just pass all the helper provider descriptions
14715 * off to the meta provider.
14716 */
14717
14718 int i;
14719 mutex_exit(&dtrace_lock);
14720
14721 for (i = 0; i < help->dthps_nprovs; i++) {
14722 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14723 p->p_pid);
14724 }
14725 }
14726
14727 mutex_exit(&dtrace_meta_lock);
14728 }
14729
14730 static int
14731 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14732 {
14733 dtrace_helpers_t *help;
14734 dtrace_helper_provider_t *hprov, **tmp_provs;
14735 uint_t tmp_maxprovs, i;
14736
14737 ASSERT(MUTEX_HELD(&dtrace_lock));
14738
14739 help = curproc->p_dtrace_helpers;
14740 ASSERT(help != NULL);
14741
14742 /*
14743 * If we already have dtrace_helper_providers_max helper providers,
14744 * we're refuse to add a new one.
14745 */
14746 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14747 return (ENOSPC);
14748
14749 /*
14750 * Check to make sure this isn't a duplicate.
14751 */
14752 for (i = 0; i < help->dthps_nprovs; i++) {
14753 if (dofhp->dofhp_dof ==
14754 help->dthps_provs[i]->dthp_prov.dofhp_dof)
14755 return (EALREADY);
14756 }
14757
14758 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14759 hprov->dthp_prov = *dofhp;
14760 hprov->dthp_ref = 1;
14761 hprov->dthp_generation = gen;
14762
14763 /*
14764 * Allocate a bigger table for helper providers if it's already full.
14765 */
14766 if (help->dthps_maxprovs == help->dthps_nprovs) {
14767 tmp_maxprovs = help->dthps_maxprovs;
14768 tmp_provs = help->dthps_provs;
14769
14770 if (help->dthps_maxprovs == 0)
14771 help->dthps_maxprovs = 2;
14772 else
14773 help->dthps_maxprovs *= 2;
14774 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14775 help->dthps_maxprovs = dtrace_helper_providers_max;
14776
14777 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14778
14779 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14780 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14781
14782 if (tmp_provs != NULL) {
14783 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14784 sizeof (dtrace_helper_provider_t *));
14785 kmem_free(tmp_provs, tmp_maxprovs *
14786 sizeof (dtrace_helper_provider_t *));
14787 }
14788 }
14789
14790 help->dthps_provs[help->dthps_nprovs] = hprov;
14791 help->dthps_nprovs++;
14792
14793 return (0);
14794 }
14795
14796 static void
14797 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14798 {
14799 mutex_enter(&dtrace_lock);
14800
14801 if (--hprov->dthp_ref == 0) {
14802 dof_hdr_t *dof;
14803 mutex_exit(&dtrace_lock);
14804 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14805 dtrace_dof_destroy(dof);
14806 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14807 } else {
14808 mutex_exit(&dtrace_lock);
14809 }
14810 }
14811
14812 static int
14813 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14814 {
14815 uintptr_t daddr = (uintptr_t)dof;
14816 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14817 dof_provider_t *provider;
14818 dof_probe_t *probe;
14819 uint8_t *arg;
14820 char *strtab, *typestr;
14821 dof_stridx_t typeidx;
14822 size_t typesz;
14823 uint_t nprobes, j, k;
14824
14825 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14826
14827 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14828 dtrace_dof_error(dof, "misaligned section offset");
14829 return (-1);
14830 }
14831
14832 /*
14833 * The section needs to be large enough to contain the DOF provider
14834 * structure appropriate for the given version.
14835 */
14836 if (sec->dofs_size <
14837 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14838 offsetof(dof_provider_t, dofpv_prenoffs) :
14839 sizeof (dof_provider_t))) {
14840 dtrace_dof_error(dof, "provider section too small");
14841 return (-1);
14842 }
14843
14844 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14845 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14846 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14847 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14848 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14849
14850 if (str_sec == NULL || prb_sec == NULL ||
14851 arg_sec == NULL || off_sec == NULL)
14852 return (-1);
14853
14854 enoff_sec = NULL;
14855
14856 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14857 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14858 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14859 provider->dofpv_prenoffs)) == NULL)
14860 return (-1);
14861
14862 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14863
14864 if (provider->dofpv_name >= str_sec->dofs_size ||
14865 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14866 dtrace_dof_error(dof, "invalid provider name");
14867 return (-1);
14868 }
14869
14870 if (prb_sec->dofs_entsize == 0 ||
14871 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14872 dtrace_dof_error(dof, "invalid entry size");
14873 return (-1);
14874 }
14875
14876 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14877 dtrace_dof_error(dof, "misaligned entry size");
14878 return (-1);
14879 }
14880
14881 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14882 dtrace_dof_error(dof, "invalid entry size");
14883 return (-1);
14884 }
14885
14886 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14887 dtrace_dof_error(dof, "misaligned section offset");
14888 return (-1);
14889 }
14890
14891 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14892 dtrace_dof_error(dof, "invalid entry size");
14893 return (-1);
14894 }
14895
14896 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14897
14898 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14899
14900 /*
14901 * Take a pass through the probes to check for errors.
14902 */
14903 for (j = 0; j < nprobes; j++) {
14904 probe = (dof_probe_t *)(uintptr_t)(daddr +
14905 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14906
14907 if (probe->dofpr_func >= str_sec->dofs_size) {
14908 dtrace_dof_error(dof, "invalid function name");
14909 return (-1);
14910 }
14911
14912 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14913 dtrace_dof_error(dof, "function name too long");
14914 return (-1);
14915 }
14916
14917 if (probe->dofpr_name >= str_sec->dofs_size ||
14918 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14919 dtrace_dof_error(dof, "invalid probe name");
14920 return (-1);
14921 }
14922
14923 /*
14924 * The offset count must not wrap the index, and the offsets
14925 * must also not overflow the section's data.
14926 */
14927 if (probe->dofpr_offidx + probe->dofpr_noffs <
14928 probe->dofpr_offidx ||
14929 (probe->dofpr_offidx + probe->dofpr_noffs) *
14930 off_sec->dofs_entsize > off_sec->dofs_size) {
14931 dtrace_dof_error(dof, "invalid probe offset");
14932 return (-1);
14933 }
14934
14935 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14936 /*
14937 * If there's no is-enabled offset section, make sure
14938 * there aren't any is-enabled offsets. Otherwise
14939 * perform the same checks as for probe offsets
14940 * (immediately above).
14941 */
14942 if (enoff_sec == NULL) {
14943 if (probe->dofpr_enoffidx != 0 ||
14944 probe->dofpr_nenoffs != 0) {
14945 dtrace_dof_error(dof, "is-enabled "
14946 "offsets with null section");
14947 return (-1);
14948 }
14949 } else if (probe->dofpr_enoffidx +
14950 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14951 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14952 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14953 dtrace_dof_error(dof, "invalid is-enabled "
14954 "offset");
14955 return (-1);
14956 }
14957
14958 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14959 dtrace_dof_error(dof, "zero probe and "
14960 "is-enabled offsets");
14961 return (-1);
14962 }
14963 } else if (probe->dofpr_noffs == 0) {
14964 dtrace_dof_error(dof, "zero probe offsets");
14965 return (-1);
14966 }
14967
14968 if (probe->dofpr_argidx + probe->dofpr_xargc <
14969 probe->dofpr_argidx ||
14970 (probe->dofpr_argidx + probe->dofpr_xargc) *
14971 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14972 dtrace_dof_error(dof, "invalid args");
14973 return (-1);
14974 }
14975
14976 typeidx = probe->dofpr_nargv;
14977 typestr = strtab + probe->dofpr_nargv;
14978 for (k = 0; k < probe->dofpr_nargc; k++) {
14979 if (typeidx >= str_sec->dofs_size) {
14980 dtrace_dof_error(dof, "bad "
14981 "native argument type");
14982 return (-1);
14983 }
14984
14985 typesz = strlen(typestr) + 1;
14986 if (typesz > DTRACE_ARGTYPELEN) {
14987 dtrace_dof_error(dof, "native "
14988 "argument type too long");
14989 return (-1);
14990 }
14991 typeidx += typesz;
14992 typestr += typesz;
14993 }
14994
14995 typeidx = probe->dofpr_xargv;
14996 typestr = strtab + probe->dofpr_xargv;
14997 for (k = 0; k < probe->dofpr_xargc; k++) {
14998 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14999 dtrace_dof_error(dof, "bad "
15000 "native argument index");
15001 return (-1);
15002 }
15003
15004 if (typeidx >= str_sec->dofs_size) {
15005 dtrace_dof_error(dof, "bad "
15006 "translated argument type");
15007 return (-1);
15008 }
15009
15010 typesz = strlen(typestr) + 1;
15011 if (typesz > DTRACE_ARGTYPELEN) {
15012 dtrace_dof_error(dof, "translated argument "
15013 "type too long");
15014 return (-1);
15015 }
15016
15017 typeidx += typesz;
15018 typestr += typesz;
15019 }
15020 }
15021
15022 return (0);
15023 }
15024
15025 static int
15026 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
15027 {
15028 dtrace_helpers_t *help;
15029 dtrace_vstate_t *vstate;
15030 dtrace_enabling_t *enab = NULL;
15031 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
15032 uintptr_t daddr = (uintptr_t)dof;
15033
15034 ASSERT(MUTEX_HELD(&dtrace_lock));
15035
15036 if ((help = curproc->p_dtrace_helpers) == NULL)
15037 help = dtrace_helpers_create(curproc);
15038
15039 vstate = &help->dthps_vstate;
15040
15041 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
15042 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
15043 dtrace_dof_destroy(dof);
15044 return (rv);
15045 }
15046
15047 /*
15048 * Look for helper providers and validate their descriptions.
15049 */
15050 if (dhp != NULL) {
15051 for (i = 0; i < dof->dofh_secnum; i++) {
15052 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
15053 dof->dofh_secoff + i * dof->dofh_secsize);
15054
15055 if (sec->dofs_type != DOF_SECT_PROVIDER)
15056 continue;
15057
15058 if (dtrace_helper_provider_validate(dof, sec) != 0) {
15059 dtrace_enabling_destroy(enab);
15060 dtrace_dof_destroy(dof);
15061 return (-1);
15062 }
15063
15064 nprovs++;
15065 }
15066 }
15067
15068 /*
15069 * Now we need to walk through the ECB descriptions in the enabling.
15070 */
15071 for (i = 0; i < enab->dten_ndesc; i++) {
15072 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
15073 dtrace_probedesc_t *desc = &ep->dted_probe;
15074
15075 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
15076 continue;
15077
15078 if (strcmp(desc->dtpd_mod, "helper") != 0)
15079 continue;
15080
15081 if (strcmp(desc->dtpd_func, "ustack") != 0)
15082 continue;
15083
15084 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
15085 ep)) != 0) {
15086 /*
15087 * Adding this helper action failed -- we are now going
15088 * to rip out the entire generation and return failure.
15089 */
15090 (void) dtrace_helper_destroygen(help->dthps_generation);
15091 dtrace_enabling_destroy(enab);
15092 dtrace_dof_destroy(dof);
15093 return (-1);
15094 }
15095
15096 nhelpers++;
15097 }
15098
15099 if (nhelpers < enab->dten_ndesc)
15100 dtrace_dof_error(dof, "unmatched helpers");
15101
15102 gen = help->dthps_generation++;
15103 dtrace_enabling_destroy(enab);
15104
15105 if (dhp != NULL && nprovs > 0) {
15106 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
15107 if (dtrace_helper_provider_add(dhp, gen) == 0) {
15108 mutex_exit(&dtrace_lock);
15109 dtrace_helper_provider_register(curproc, help, dhp);
15110 mutex_enter(&dtrace_lock);
15111
15112 destroy = 0;
15113 }
15114 }
15115
15116 if (destroy)
15117 dtrace_dof_destroy(dof);
15118
15119 return (gen);
15120 }
15121
15122 static dtrace_helpers_t *
15123 dtrace_helpers_create(proc_t *p)
15124 {
15125 dtrace_helpers_t *help;
15126
15127 ASSERT(MUTEX_HELD(&dtrace_lock));
15128 ASSERT(p->p_dtrace_helpers == NULL);
15129
15130 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
15131 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
15132 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
15133
15134 p->p_dtrace_helpers = help;
15135 dtrace_helpers++;
15136
15137 return (help);
15138 }
15139
15140 static void
15141 dtrace_helpers_destroy(void)
15142 {
15143 dtrace_helpers_t *help;
15144 dtrace_vstate_t *vstate;
15145 proc_t *p = curproc;
15146 int i;
15147
15148 mutex_enter(&dtrace_lock);
15149
15150 ASSERT(p->p_dtrace_helpers != NULL);
15151 ASSERT(dtrace_helpers > 0);
15152
15153 help = p->p_dtrace_helpers;
15154 vstate = &help->dthps_vstate;
15155
15156 /*
15157 * We're now going to lose the help from this process.
15158 */
15159 p->p_dtrace_helpers = NULL;
15160 dtrace_sync();
15161
15162 /*
15163 * Destory the helper actions.
15164 */
15165 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15166 dtrace_helper_action_t *h, *next;
15167
15168 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15169 next = h->dtha_next;
15170 dtrace_helper_action_destroy(h, vstate);
15171 h = next;
15172 }
15173 }
15174
15175 mutex_exit(&dtrace_lock);
15176
15177 /*
15178 * Destroy the helper providers.
15179 */
15180 if (help->dthps_maxprovs > 0) {
15181 mutex_enter(&dtrace_meta_lock);
15182 if (dtrace_meta_pid != NULL) {
15183 ASSERT(dtrace_deferred_pid == NULL);
15184
15185 for (i = 0; i < help->dthps_nprovs; i++) {
15186 dtrace_helper_provider_remove(
15187 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15188 }
15189 } else {
15190 mutex_enter(&dtrace_lock);
15191 ASSERT(help->dthps_deferred == 0 ||
15192 help->dthps_next != NULL ||
15193 help->dthps_prev != NULL ||
15194 help == dtrace_deferred_pid);
15195
15196 /*
15197 * Remove the helper from the deferred list.
15198 */
15199 if (help->dthps_next != NULL)
15200 help->dthps_next->dthps_prev = help->dthps_prev;
15201 if (help->dthps_prev != NULL)
15202 help->dthps_prev->dthps_next = help->dthps_next;
15203 if (dtrace_deferred_pid == help) {
15204 dtrace_deferred_pid = help->dthps_next;
15205 ASSERT(help->dthps_prev == NULL);
15206 }
15207
15208 mutex_exit(&dtrace_lock);
15209 }
15210
15211 mutex_exit(&dtrace_meta_lock);
15212
15213 for (i = 0; i < help->dthps_nprovs; i++) {
15214 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15215 }
15216
15217 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15218 sizeof (dtrace_helper_provider_t *));
15219 }
15220
15221 mutex_enter(&dtrace_lock);
15222
15223 dtrace_vstate_fini(&help->dthps_vstate);
15224 kmem_free(help->dthps_actions,
15225 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15226 kmem_free(help, sizeof (dtrace_helpers_t));
15227
15228 --dtrace_helpers;
15229 mutex_exit(&dtrace_lock);
15230 }
15231
15232 static void
15233 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15234 {
15235 dtrace_helpers_t *help, *newhelp;
15236 dtrace_helper_action_t *helper, *new, *last;
15237 dtrace_difo_t *dp;
15238 dtrace_vstate_t *vstate;
15239 int i, j, sz, hasprovs = 0;
15240
15241 mutex_enter(&dtrace_lock);
15242 ASSERT(from->p_dtrace_helpers != NULL);
15243 ASSERT(dtrace_helpers > 0);
15244
15245 help = from->p_dtrace_helpers;
15246 newhelp = dtrace_helpers_create(to);
15247 ASSERT(to->p_dtrace_helpers != NULL);
15248
15249 newhelp->dthps_generation = help->dthps_generation;
15250 vstate = &newhelp->dthps_vstate;
15251
15252 /*
15253 * Duplicate the helper actions.
15254 */
15255 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15256 if ((helper = help->dthps_actions[i]) == NULL)
15257 continue;
15258
15259 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15260 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15261 KM_SLEEP);
15262 new->dtha_generation = helper->dtha_generation;
15263
15264 if ((dp = helper->dtha_predicate) != NULL) {
15265 dp = dtrace_difo_duplicate(dp, vstate);
15266 new->dtha_predicate = dp;
15267 }
15268
15269 new->dtha_nactions = helper->dtha_nactions;
15270 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15271 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15272
15273 for (j = 0; j < new->dtha_nactions; j++) {
15274 dtrace_difo_t *dp = helper->dtha_actions[j];
15275
15276 ASSERT(dp != NULL);
15277 dp = dtrace_difo_duplicate(dp, vstate);
15278 new->dtha_actions[j] = dp;
15279 }
15280
15281 if (last != NULL) {
15282 last->dtha_next = new;
15283 } else {
15284 newhelp->dthps_actions[i] = new;
15285 }
15286
15287 last = new;
15288 }
15289 }
15290
15291 /*
15292 * Duplicate the helper providers and register them with the
15293 * DTrace framework.
15294 */
15295 if (help->dthps_nprovs > 0) {
15296 newhelp->dthps_nprovs = help->dthps_nprovs;
15297 newhelp->dthps_maxprovs = help->dthps_nprovs;
15298 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15299 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15300 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15301 newhelp->dthps_provs[i] = help->dthps_provs[i];
15302 newhelp->dthps_provs[i]->dthp_ref++;
15303 }
15304
15305 hasprovs = 1;
15306 }
15307
15308 mutex_exit(&dtrace_lock);
15309
15310 if (hasprovs)
15311 dtrace_helper_provider_register(to, newhelp, NULL);
15312 }
15313
15314 /*
15315 * DTrace Hook Functions
15316 */
15317 static void
15318 dtrace_module_loaded(struct modctl *ctl)
15319 {
15320 dtrace_provider_t *prv;
15321
15322 mutex_enter(&dtrace_provider_lock);
15323 mutex_enter(&mod_lock);
15324
15325 ASSERT(ctl->mod_busy);
15326
15327 /*
15328 * We're going to call each providers per-module provide operation
15329 * specifying only this module.
15330 */
15331 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15332 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15333
15334 mutex_exit(&mod_lock);
15335 mutex_exit(&dtrace_provider_lock);
15336
15337 /*
15338 * If we have any retained enablings, we need to match against them.
15339 * Enabling probes requires that cpu_lock be held, and we cannot hold
15340 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15341 * module. (In particular, this happens when loading scheduling
15342 * classes.) So if we have any retained enablings, we need to dispatch
15343 * our task queue to do the match for us.
15344 */
15345 mutex_enter(&dtrace_lock);
15346
15347 if (dtrace_retained == NULL) {
15348 mutex_exit(&dtrace_lock);
15349 return;
15350 }
15351
15352 (void) taskq_dispatch(dtrace_taskq,
15353 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15354
15355 mutex_exit(&dtrace_lock);
15356
15357 /*
15358 * And now, for a little heuristic sleaze: in general, we want to
15359 * match modules as soon as they load. However, we cannot guarantee
15360 * this, because it would lead us to the lock ordering violation
15361 * outlined above. The common case, of course, is that cpu_lock is
15362 * _not_ held -- so we delay here for a clock tick, hoping that that's
15363 * long enough for the task queue to do its work. If it's not, it's
15364 * not a serious problem -- it just means that the module that we
15365 * just loaded may not be immediately instrumentable.
15366 */
15367 delay(1);
15368 }
15369
15370 static void
15371 dtrace_module_unloaded(struct modctl *ctl)
15372 {
15373 dtrace_probe_t template, *probe, *first, *next;
15374 dtrace_provider_t *prov;
15375
15376 template.dtpr_mod = ctl->mod_modname;
15377
15378 mutex_enter(&dtrace_provider_lock);
15379 mutex_enter(&mod_lock);
15380 mutex_enter(&dtrace_lock);
15381
15382 if (dtrace_bymod == NULL) {
15383 /*
15384 * The DTrace module is loaded (obviously) but not attached;
15385 * we don't have any work to do.
15386 */
15387 mutex_exit(&dtrace_provider_lock);
15388 mutex_exit(&mod_lock);
15389 mutex_exit(&dtrace_lock);
15390 return;
15391 }
15392
15393 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15394 probe != NULL; probe = probe->dtpr_nextmod) {
15395 if (probe->dtpr_ecb != NULL) {
15396 mutex_exit(&dtrace_provider_lock);
15397 mutex_exit(&mod_lock);
15398 mutex_exit(&dtrace_lock);
15399
15400 /*
15401 * This shouldn't _actually_ be possible -- we're
15402 * unloading a module that has an enabled probe in it.
15403 * (It's normally up to the provider to make sure that
15404 * this can't happen.) However, because dtps_enable()
15405 * doesn't have a failure mode, there can be an
15406 * enable/unload race. Upshot: we don't want to
15407 * assert, but we're not going to disable the
15408 * probe, either.
15409 */
15410 if (dtrace_err_verbose) {
15411 cmn_err(CE_WARN, "unloaded module '%s' had "
15412 "enabled probes", ctl->mod_modname);
15413 }
15414
15415 return;
15416 }
15417 }
15418
15419 probe = first;
15420
15421 for (first = NULL; probe != NULL; probe = next) {
15422 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15423
15424 dtrace_probes[probe->dtpr_id - 1] = NULL;
15425
15426 next = probe->dtpr_nextmod;
15427 dtrace_hash_remove(dtrace_bymod, probe);
15428 dtrace_hash_remove(dtrace_byfunc, probe);
15429 dtrace_hash_remove(dtrace_byname, probe);
15430
15431 if (first == NULL) {
15432 first = probe;
15433 probe->dtpr_nextmod = NULL;
15434 } else {
15435 probe->dtpr_nextmod = first;
15436 first = probe;
15437 }
15438 }
15439
15440 /*
15441 * We've removed all of the module's probes from the hash chains and
15442 * from the probe array. Now issue a dtrace_sync() to be sure that
15443 * everyone has cleared out from any probe array processing.
15444 */
15445 dtrace_sync();
15446
15447 for (probe = first; probe != NULL; probe = first) {
15448 first = probe->dtpr_nextmod;
15449 prov = probe->dtpr_provider;
15450 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15451 probe->dtpr_arg);
15452 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15453 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15454 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15455 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15456 kmem_free(probe, sizeof (dtrace_probe_t));
15457 }
15458
15459 mutex_exit(&dtrace_lock);
15460 mutex_exit(&mod_lock);
15461 mutex_exit(&dtrace_provider_lock);
15462 }
15463
15464 void
15465 dtrace_suspend(void)
15466 {
15467 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15468 }
15469
15470 void
15471 dtrace_resume(void)
15472 {
15473 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15474 }
15475
15476 static int
15477 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15478 {
15479 ASSERT(MUTEX_HELD(&cpu_lock));
15480 mutex_enter(&dtrace_lock);
15481
15482 switch (what) {
15483 case CPU_CONFIG: {
15484 dtrace_state_t *state;
15485 dtrace_optval_t *opt, rs, c;
15486
15487 /*
15488 * For now, we only allocate a new buffer for anonymous state.
15489 */
15490 if ((state = dtrace_anon.dta_state) == NULL)
15491 break;
15492
15493 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15494 break;
15495
15496 opt = state->dts_options;
15497 c = opt[DTRACEOPT_CPU];
15498
15499 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15500 break;
15501
15502 /*
15503 * Regardless of what the actual policy is, we're going to
15504 * temporarily set our resize policy to be manual. We're
15505 * also going to temporarily set our CPU option to denote
15506 * the newly configured CPU.
15507 */
15508 rs = opt[DTRACEOPT_BUFRESIZE];
15509 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15510 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15511
15512 (void) dtrace_state_buffers(state);
15513
15514 opt[DTRACEOPT_BUFRESIZE] = rs;
15515 opt[DTRACEOPT_CPU] = c;
15516
15517 break;
15518 }
15519
15520 case CPU_UNCONFIG:
15521 /*
15522 * We don't free the buffer in the CPU_UNCONFIG case. (The
15523 * buffer will be freed when the consumer exits.)
15524 */
15525 break;
15526
15527 default:
15528 break;
15529 }
15530
15531 mutex_exit(&dtrace_lock);
15532 return (0);
15533 }
15534
15535 static void
15536 dtrace_cpu_setup_initial(processorid_t cpu)
15537 {
15538 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15539 }
15540
15541 static void
15542 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15543 {
15544 if (dtrace_toxranges >= dtrace_toxranges_max) {
15545 int osize, nsize;
15546 dtrace_toxrange_t *range;
15547
15548 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15549
15550 if (osize == 0) {
15551 ASSERT(dtrace_toxrange == NULL);
15552 ASSERT(dtrace_toxranges_max == 0);
15553 dtrace_toxranges_max = 1;
15554 } else {
15555 dtrace_toxranges_max <<= 1;
15556 }
15557
15558 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15559 range = kmem_zalloc(nsize, KM_SLEEP);
15560
15561 if (dtrace_toxrange != NULL) {
15562 ASSERT(osize != 0);
15563 bcopy(dtrace_toxrange, range, osize);
15564 kmem_free(dtrace_toxrange, osize);
15565 }
15566
15567 dtrace_toxrange = range;
15568 }
15569
15570 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
15571 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
15572
15573 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15574 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15575 dtrace_toxranges++;
15576 }
15577
15578 static void
15579 dtrace_getf_barrier()
15580 {
15581 /*
15582 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
15583 * that contain calls to getf(), this routine will be called on every
15584 * closef() before either the underlying vnode is released or the
15585 * file_t itself is freed. By the time we are here, it is essential
15586 * that the file_t can no longer be accessed from a call to getf()
15587 * in probe context -- that assures that a dtrace_sync() can be used
15588 * to clear out any enablings referring to the old structures.
15589 */
15590 if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
15591 kcred->cr_zone->zone_dtrace_getf != 0)
15592 dtrace_sync();
15593 }
15594
15595 /*
15596 * DTrace Driver Cookbook Functions
15597 */
15598 /*ARGSUSED*/
15599 static int
15600 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15601 {
15602 dtrace_provider_id_t id;
15603 dtrace_state_t *state = NULL;
15604 dtrace_enabling_t *enab;
15605
15606 mutex_enter(&cpu_lock);
15607 mutex_enter(&dtrace_provider_lock);
15608 mutex_enter(&dtrace_lock);
15609
15610 if (ddi_soft_state_init(&dtrace_softstate,
15611 sizeof (dtrace_state_t), 0) != 0) {
15612 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15613 mutex_exit(&cpu_lock);
15614 mutex_exit(&dtrace_provider_lock);
15615 mutex_exit(&dtrace_lock);
15616 return (DDI_FAILURE);
15617 }
15618
15619 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15620 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15621 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15622 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15623 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15624 ddi_remove_minor_node(devi, NULL);
15625 ddi_soft_state_fini(&dtrace_softstate);
15626 mutex_exit(&cpu_lock);
15627 mutex_exit(&dtrace_provider_lock);
15628 mutex_exit(&dtrace_lock);
15629 return (DDI_FAILURE);
15630 }
15631
15632 ddi_report_dev(devi);
15633 dtrace_devi = devi;
15634
15635 dtrace_modload = dtrace_module_loaded;
15636 dtrace_modunload = dtrace_module_unloaded;
15637 dtrace_cpu_init = dtrace_cpu_setup_initial;
15638 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15639 dtrace_helpers_fork = dtrace_helpers_duplicate;
15640 dtrace_cpustart_init = dtrace_suspend;
15641 dtrace_cpustart_fini = dtrace_resume;
15642 dtrace_debugger_init = dtrace_suspend;
15643 dtrace_debugger_fini = dtrace_resume;
15644
15645 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15646
15647 ASSERT(MUTEX_HELD(&cpu_lock));
15648
15649 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15650 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15651 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15652 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15653 VM_SLEEP | VMC_IDENTIFIER);
15654 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15655 1, INT_MAX, 0);
15656
15657 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15658 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15659 NULL, NULL, NULL, NULL, NULL, 0);
15660
15661 ASSERT(MUTEX_HELD(&cpu_lock));
15662 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15663 offsetof(dtrace_probe_t, dtpr_nextmod),
15664 offsetof(dtrace_probe_t, dtpr_prevmod));
15665
15666 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15667 offsetof(dtrace_probe_t, dtpr_nextfunc),
15668 offsetof(dtrace_probe_t, dtpr_prevfunc));
15669
15670 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15671 offsetof(dtrace_probe_t, dtpr_nextname),
15672 offsetof(dtrace_probe_t, dtpr_prevname));
15673
15674 if (dtrace_retain_max < 1) {
15675 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15676 "setting to 1", dtrace_retain_max);
15677 dtrace_retain_max = 1;
15678 }
15679
15680 /*
15681 * Now discover our toxic ranges.
15682 */
15683 dtrace_toxic_ranges(dtrace_toxrange_add);
15684
15685 /*
15686 * Before we register ourselves as a provider to our own framework,
15687 * we would like to assert that dtrace_provider is NULL -- but that's
15688 * not true if we were loaded as a dependency of a DTrace provider.
15689 * Once we've registered, we can assert that dtrace_provider is our
15690 * pseudo provider.
15691 */
15692 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15693 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15694
15695 ASSERT(dtrace_provider != NULL);
15696 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15697
15698 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15699 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15700 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15701 dtrace_provider, NULL, NULL, "END", 0, NULL);
15702 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15703 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15704
15705 dtrace_anon_property();
15706 mutex_exit(&cpu_lock);
15707
15708 /*
15709 * If DTrace helper tracing is enabled, we need to allocate the
15710 * trace buffer and initialize the values.
15711 */
15712 if (dtrace_helptrace_enabled) {
15713 ASSERT(dtrace_helptrace_buffer == NULL);
15714 dtrace_helptrace_buffer =
15715 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15716 dtrace_helptrace_next = 0;
15717 }
15718
15719 /*
15720 * If there are already providers, we must ask them to provide their
15721 * probes, and then match any anonymous enabling against them. Note
15722 * that there should be no other retained enablings at this time:
15723 * the only retained enablings at this time should be the anonymous
15724 * enabling.
15725 */
15726 if (dtrace_anon.dta_enabling != NULL) {
15727 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15728
15729 dtrace_enabling_provide(NULL);
15730 state = dtrace_anon.dta_state;
15731
15732 /*
15733 * We couldn't hold cpu_lock across the above call to
15734 * dtrace_enabling_provide(), but we must hold it to actually
15735 * enable the probes. We have to drop all of our locks, pick
15736 * up cpu_lock, and regain our locks before matching the
15737 * retained anonymous enabling.
15738 */
15739 mutex_exit(&dtrace_lock);
15740 mutex_exit(&dtrace_provider_lock);
15741
15742 mutex_enter(&cpu_lock);
15743 mutex_enter(&dtrace_provider_lock);
15744 mutex_enter(&dtrace_lock);
15745
15746 if ((enab = dtrace_anon.dta_enabling) != NULL)
15747 (void) dtrace_enabling_match(enab, NULL);
15748
15749 mutex_exit(&cpu_lock);
15750 }
15751
15752 mutex_exit(&dtrace_lock);
15753 mutex_exit(&dtrace_provider_lock);
15754
15755 if (state != NULL) {
15756 /*
15757 * If we created any anonymous state, set it going now.
15758 */
15759 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15760 }
15761
15762 return (DDI_SUCCESS);
15763 }
15764
15765 /*ARGSUSED*/
15766 static int
15767 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15768 {
15769 dtrace_state_t *state;
15770 uint32_t priv;
15771 uid_t uid;
15772 zoneid_t zoneid;
15773
15774 if (getminor(*devp) == DTRACEMNRN_HELPER)
15775 return (0);
15776
15777 /*
15778 * If this wasn't an open with the "helper" minor, then it must be
15779 * the "dtrace" minor.
15780 */
15781 if (getminor(*devp) != DTRACEMNRN_DTRACE)
15782 return (ENXIO);
15783
15784 /*
15785 * If no DTRACE_PRIV_* bits are set in the credential, then the
15786 * caller lacks sufficient permission to do anything with DTrace.
15787 */
15788 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15789 if (priv == DTRACE_PRIV_NONE)
15790 return (EACCES);
15791
15792 /*
15793 * Ask all providers to provide all their probes.
15794 */
15795 mutex_enter(&dtrace_provider_lock);
15796 dtrace_probe_provide(NULL, NULL);
15797 mutex_exit(&dtrace_provider_lock);
15798
15799 mutex_enter(&cpu_lock);
15800 mutex_enter(&dtrace_lock);
15801 dtrace_opens++;
15802 dtrace_membar_producer();
15803
15804 /*
15805 * If the kernel debugger is active (that is, if the kernel debugger
15806 * modified text in some way), we won't allow the open.
15807 */
15808 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15809 dtrace_opens--;
15810 mutex_exit(&cpu_lock);
15811 mutex_exit(&dtrace_lock);
15812 return (EBUSY);
15813 }
15814
15815 state = dtrace_state_create(devp, cred_p);
15816 mutex_exit(&cpu_lock);
15817
15818 if (state == NULL) {
15819 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15820 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15821 mutex_exit(&dtrace_lock);
15822 return (EAGAIN);
15823 }
15824
15825 mutex_exit(&dtrace_lock);
15826
15827 return (0);
15828 }
15829
15830 /*ARGSUSED*/
15831 static int
15832 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15833 {
15834 minor_t minor = getminor(dev);
15835 dtrace_state_t *state;
15836
15837 if (minor == DTRACEMNRN_HELPER)
15838 return (0);
15839
15840 state = ddi_get_soft_state(dtrace_softstate, minor);
15841
15842 mutex_enter(&cpu_lock);
15843 mutex_enter(&dtrace_lock);
15844
15845 if (state->dts_anon) {
15846 /*
15847 * There is anonymous state. Destroy that first.
15848 */
15849 ASSERT(dtrace_anon.dta_state == NULL);
15850 dtrace_state_destroy(state->dts_anon);
15851 }
15852
15853 dtrace_state_destroy(state);
15854 ASSERT(dtrace_opens > 0);
15855
15856 /*
15857 * Only relinquish control of the kernel debugger interface when there
15858 * are no consumers and no anonymous enablings.
15859 */
15860 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15861 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15862
15863 mutex_exit(&dtrace_lock);
15864 mutex_exit(&cpu_lock);
15865
15866 return (0);
15867 }
15868
15869 /*ARGSUSED*/
15870 static int
15871 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15872 {
15873 int rval;
15874 dof_helper_t help, *dhp = NULL;
15875
15876 switch (cmd) {
15877 case DTRACEHIOC_ADDDOF:
15878 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15879 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15880 return (EFAULT);
15881 }
15882
15883 dhp = &help;
15884 arg = (intptr_t)help.dofhp_dof;
15885 /*FALLTHROUGH*/
15886
15887 case DTRACEHIOC_ADD: {
15888 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15889
15890 if (dof == NULL)
15891 return (rval);
15892
15893 mutex_enter(&dtrace_lock);
15894
15895 /*
15896 * dtrace_helper_slurp() takes responsibility for the dof --
15897 * it may free it now or it may save it and free it later.
15898 */
15899 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15900 *rv = rval;
15901 rval = 0;
15902 } else {
15903 rval = EINVAL;
15904 }
15905
15906 mutex_exit(&dtrace_lock);
15907 return (rval);
15908 }
15909
15910 case DTRACEHIOC_REMOVE: {
15911 mutex_enter(&dtrace_lock);
15912 rval = dtrace_helper_destroygen(arg);
15913 mutex_exit(&dtrace_lock);
15914
15915 return (rval);
15916 }
15917
15918 default:
15919 break;
15920 }
15921
15922 return (ENOTTY);
15923 }
15924
15925 /*ARGSUSED*/
15926 static int
15927 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15928 {
15929 minor_t minor = getminor(dev);
15930 dtrace_state_t *state;
15931 int rval;
15932
15933 if (minor == DTRACEMNRN_HELPER)
15934 return (dtrace_ioctl_helper(cmd, arg, rv));
15935
15936 state = ddi_get_soft_state(dtrace_softstate, minor);
15937
15938 if (state->dts_anon) {
15939 ASSERT(dtrace_anon.dta_state == NULL);
15940 state = state->dts_anon;
15941 }
15942
15943 switch (cmd) {
15944 case DTRACEIOC_PROVIDER: {
15945 dtrace_providerdesc_t pvd;
15946 dtrace_provider_t *pvp;
15947
15948 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15949 return (EFAULT);
15950
15951 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15952 mutex_enter(&dtrace_provider_lock);
15953
15954 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15955 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15956 break;
15957 }
15958
15959 mutex_exit(&dtrace_provider_lock);
15960
15961 if (pvp == NULL)
15962 return (ESRCH);
15963
15964 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15965 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15966 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15967 return (EFAULT);
15968
15969 return (0);
15970 }
15971
15972 case DTRACEIOC_EPROBE: {
15973 dtrace_eprobedesc_t epdesc;
15974 dtrace_ecb_t *ecb;
15975 dtrace_action_t *act;
15976 void *buf;
15977 size_t size;
15978 uintptr_t dest;
15979 int nrecs;
15980
15981 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15982 return (EFAULT);
15983
15984 mutex_enter(&dtrace_lock);
15985
15986 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15987 mutex_exit(&dtrace_lock);
15988 return (EINVAL);
15989 }
15990
15991 if (ecb->dte_probe == NULL) {
15992 mutex_exit(&dtrace_lock);
15993 return (EINVAL);
15994 }
15995
15996 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15997 epdesc.dtepd_uarg = ecb->dte_uarg;
15998 epdesc.dtepd_size = ecb->dte_size;
15999
16000 nrecs = epdesc.dtepd_nrecs;
16001 epdesc.dtepd_nrecs = 0;
16002 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16003 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16004 continue;
16005
16006 epdesc.dtepd_nrecs++;
16007 }
16008
16009 /*
16010 * Now that we have the size, we need to allocate a temporary
16011 * buffer in which to store the complete description. We need
16012 * the temporary buffer to be able to drop dtrace_lock()
16013 * across the copyout(), below.
16014 */
16015 size = sizeof (dtrace_eprobedesc_t) +
16016 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
16017
16018 buf = kmem_alloc(size, KM_SLEEP);
16019 dest = (uintptr_t)buf;
16020
16021 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
16022 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
16023
16024 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16025 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16026 continue;
16027
16028 if (nrecs-- == 0)
16029 break;
16030
16031 bcopy(&act->dta_rec, (void *)dest,
16032 sizeof (dtrace_recdesc_t));
16033 dest += sizeof (dtrace_recdesc_t);
16034 }
16035
16036 mutex_exit(&dtrace_lock);
16037
16038 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16039 kmem_free(buf, size);
16040 return (EFAULT);
16041 }
16042
16043 kmem_free(buf, size);
16044 return (0);
16045 }
16046
16047 case DTRACEIOC_AGGDESC: {
16048 dtrace_aggdesc_t aggdesc;
16049 dtrace_action_t *act;
16050 dtrace_aggregation_t *agg;
16051 int nrecs;
16052 uint32_t offs;
16053 dtrace_recdesc_t *lrec;
16054 void *buf;
16055 size_t size;
16056 uintptr_t dest;
16057
16058 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16059 return (EFAULT);
16060
16061 mutex_enter(&dtrace_lock);
16062
16063 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16064 mutex_exit(&dtrace_lock);
16065 return (EINVAL);
16066 }
16067
16068 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16069
16070 nrecs = aggdesc.dtagd_nrecs;
16071 aggdesc.dtagd_nrecs = 0;
16072
16073 offs = agg->dtag_base;
16074 lrec = &agg->dtag_action.dta_rec;
16075 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16076
16077 for (act = agg->dtag_first; ; act = act->dta_next) {
16078 ASSERT(act->dta_intuple ||
16079 DTRACEACT_ISAGG(act->dta_kind));
16080
16081 /*
16082 * If this action has a record size of zero, it
16083 * denotes an argument to the aggregating action.
16084 * Because the presence of this record doesn't (or
16085 * shouldn't) affect the way the data is interpreted,
16086 * we don't copy it out to save user-level the
16087 * confusion of dealing with a zero-length record.
16088 */
16089 if (act->dta_rec.dtrd_size == 0) {
16090 ASSERT(agg->dtag_hasarg);
16091 continue;
16092 }
16093
16094 aggdesc.dtagd_nrecs++;
16095
16096 if (act == &agg->dtag_action)
16097 break;
16098 }
16099
16100 /*
16101 * Now that we have the size, we need to allocate a temporary
16102 * buffer in which to store the complete description. We need
16103 * the temporary buffer to be able to drop dtrace_lock()
16104 * across the copyout(), below.
16105 */
16106 size = sizeof (dtrace_aggdesc_t) +
16107 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16108
16109 buf = kmem_alloc(size, KM_SLEEP);
16110 dest = (uintptr_t)buf;
16111
16112 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16113 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16114
16115 for (act = agg->dtag_first; ; act = act->dta_next) {
16116 dtrace_recdesc_t rec = act->dta_rec;
16117
16118 /*
16119 * See the comment in the above loop for why we pass
16120 * over zero-length records.
16121 */
16122 if (rec.dtrd_size == 0) {
16123 ASSERT(agg->dtag_hasarg);
16124 continue;
16125 }
16126
16127 if (nrecs-- == 0)
16128 break;
16129
16130 rec.dtrd_offset -= offs;
16131 bcopy(&rec, (void *)dest, sizeof (rec));
16132 dest += sizeof (dtrace_recdesc_t);
16133
16134 if (act == &agg->dtag_action)
16135 break;
16136 }
16137
16138 mutex_exit(&dtrace_lock);
16139
16140 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16141 kmem_free(buf, size);
16142 return (EFAULT);
16143 }
16144
16145 kmem_free(buf, size);
16146 return (0);
16147 }
16148
16149 case DTRACEIOC_ENABLE: {
16150 dof_hdr_t *dof;
16151 dtrace_enabling_t *enab = NULL;
16152 dtrace_vstate_t *vstate;
16153 int err = 0;
16154
16155 *rv = 0;
16156
16157 /*
16158 * If a NULL argument has been passed, we take this as our
16159 * cue to reevaluate our enablings.
16160 */
16161 if (arg == NULL) {
16162 dtrace_enabling_matchall();
16163
16164 return (0);
16165 }
16166
16167 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16168 return (rval);
16169
16170 mutex_enter(&cpu_lock);
16171 mutex_enter(&dtrace_lock);
16172 vstate = &state->dts_vstate;
16173
16174 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16175 mutex_exit(&dtrace_lock);
16176 mutex_exit(&cpu_lock);
16177 dtrace_dof_destroy(dof);
16178 return (EBUSY);
16179 }
16180
16181 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16182 mutex_exit(&dtrace_lock);
16183 mutex_exit(&cpu_lock);
16184 dtrace_dof_destroy(dof);
16185 return (EINVAL);
16186 }
16187
16188 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16189 dtrace_enabling_destroy(enab);
16190 mutex_exit(&dtrace_lock);
16191 mutex_exit(&cpu_lock);
16192 dtrace_dof_destroy(dof);
16193 return (rval);
16194 }
16195
16196 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16197 err = dtrace_enabling_retain(enab);
16198 } else {
16199 dtrace_enabling_destroy(enab);
16200 }
16201
16202 mutex_exit(&cpu_lock);
16203 mutex_exit(&dtrace_lock);
16204 dtrace_dof_destroy(dof);
16205
16206 return (err);
16207 }
16208
16209 case DTRACEIOC_REPLICATE: {
16210 dtrace_repldesc_t desc;
16211 dtrace_probedesc_t *match = &desc.dtrpd_match;
16212 dtrace_probedesc_t *create = &desc.dtrpd_create;
16213 int err;
16214
16215 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16216 return (EFAULT);
16217
16218 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16219 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16220 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16221 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16222
16223 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16224 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16225 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16226 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16227
16228 mutex_enter(&dtrace_lock);
16229 err = dtrace_enabling_replicate(state, match, create);
16230 mutex_exit(&dtrace_lock);
16231
16232 return (err);
16233 }
16234
16235 case DTRACEIOC_PROBEMATCH:
16236 case DTRACEIOC_PROBES: {
16237 dtrace_probe_t *probe = NULL;
16238 dtrace_probedesc_t desc;
16239 dtrace_probekey_t pkey;
16240 dtrace_id_t i;
16241 int m = 0;
16242 uint32_t priv;
16243 uid_t uid;
16244 zoneid_t zoneid;
16245
16246 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16247 return (EFAULT);
16248
16249 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16250 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16251 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16252 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16253
16254 /*
16255 * Before we attempt to match this probe, we want to give
16256 * all providers the opportunity to provide it.
16257 */
16258 if (desc.dtpd_id == DTRACE_IDNONE) {
16259 mutex_enter(&dtrace_provider_lock);
16260 dtrace_probe_provide(&desc, NULL);
16261 mutex_exit(&dtrace_provider_lock);
16262 desc.dtpd_id++;
16263 }
16264
16265 if (cmd == DTRACEIOC_PROBEMATCH) {
16266 dtrace_probekey(&desc, &pkey);
16267 pkey.dtpk_id = DTRACE_IDNONE;
16268 }
16269
16270 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16271
16272 mutex_enter(&dtrace_lock);
16273
16274 if (cmd == DTRACEIOC_PROBEMATCH) {
16275 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16276 if ((probe = dtrace_probes[i - 1]) != NULL &&
16277 (m = dtrace_match_probe(probe, &pkey,
16278 priv, uid, zoneid)) != 0)
16279 break;
16280 }
16281
16282 if (m < 0) {
16283 mutex_exit(&dtrace_lock);
16284 return (EINVAL);
16285 }
16286
16287 } else {
16288 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16289 if ((probe = dtrace_probes[i - 1]) != NULL &&
16290 dtrace_match_priv(probe, priv, uid, zoneid))
16291 break;
16292 }
16293 }
16294
16295 if (probe == NULL) {
16296 mutex_exit(&dtrace_lock);
16297 return (ESRCH);
16298 }
16299
16300 dtrace_probe_description(probe, &desc);
16301 mutex_exit(&dtrace_lock);
16302
16303 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16304 return (EFAULT);
16305
16306 return (0);
16307 }
16308
16309 case DTRACEIOC_PROBEARG: {
16310 dtrace_argdesc_t desc;
16311 dtrace_probe_t *probe;
16312 dtrace_provider_t *prov;
16313
16314 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16315 return (EFAULT);
16316
16317 if (desc.dtargd_id == DTRACE_IDNONE)
16318 return (EINVAL);
16319
16320 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16321 return (EINVAL);
16322
16323 mutex_enter(&dtrace_provider_lock);
16324 mutex_enter(&mod_lock);
16325 mutex_enter(&dtrace_lock);
16326
16327 if (desc.dtargd_id > dtrace_nprobes) {
16328 mutex_exit(&dtrace_lock);
16329 mutex_exit(&mod_lock);
16330 mutex_exit(&dtrace_provider_lock);
16331 return (EINVAL);
16332 }
16333
16334 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16335 mutex_exit(&dtrace_lock);
16336 mutex_exit(&mod_lock);
16337 mutex_exit(&dtrace_provider_lock);
16338 return (EINVAL);
16339 }
16340
16341 mutex_exit(&dtrace_lock);
16342
16343 prov = probe->dtpr_provider;
16344
16345 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16346 /*
16347 * There isn't any typed information for this probe.
16348 * Set the argument number to DTRACE_ARGNONE.
16349 */
16350 desc.dtargd_ndx = DTRACE_ARGNONE;
16351 } else {
16352 desc.dtargd_native[0] = '\0';
16353 desc.dtargd_xlate[0] = '\0';
16354 desc.dtargd_mapping = desc.dtargd_ndx;
16355
16356 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16357 probe->dtpr_id, probe->dtpr_arg, &desc);
16358 }
16359
16360 mutex_exit(&mod_lock);
16361 mutex_exit(&dtrace_provider_lock);
16362
16363 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16364 return (EFAULT);
16365
16366 return (0);
16367 }
16368
16369 case DTRACEIOC_GO: {
16370 processorid_t cpuid;
16371 rval = dtrace_state_go(state, &cpuid);
16372
16373 if (rval != 0)
16374 return (rval);
16375
16376 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16377 return (EFAULT);
16378
16379 return (0);
16380 }
16381
16382 case DTRACEIOC_STOP: {
16383 processorid_t cpuid;
16384
16385 mutex_enter(&dtrace_lock);
16386 rval = dtrace_state_stop(state, &cpuid);
16387 mutex_exit(&dtrace_lock);
16388
16389 if (rval != 0)
16390 return (rval);
16391
16392 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16393 return (EFAULT);
16394
16395 return (0);
16396 }
16397
16398 case DTRACEIOC_DOFGET: {
16399 dof_hdr_t hdr, *dof;
16400 uint64_t len;
16401
16402 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16403 return (EFAULT);
16404
16405 mutex_enter(&dtrace_lock);
16406 dof = dtrace_dof_create(state);
16407 mutex_exit(&dtrace_lock);
16408
16409 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16410 rval = copyout(dof, (void *)arg, len);
16411 dtrace_dof_destroy(dof);
16412
16413 return (rval == 0 ? 0 : EFAULT);
16414 }
16415
16416 case DTRACEIOC_AGGSNAP:
16417 case DTRACEIOC_BUFSNAP: {
16418 dtrace_bufdesc_t desc;
16419 caddr_t cached;
16420 dtrace_buffer_t *buf;
16421
16422 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16423 return (EFAULT);
16424
16425 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16426 return (EINVAL);
16427
16428 mutex_enter(&dtrace_lock);
16429
16430 if (cmd == DTRACEIOC_BUFSNAP) {
16431 buf = &state->dts_buffer[desc.dtbd_cpu];
16432 } else {
16433 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16434 }
16435
16436 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16437 size_t sz = buf->dtb_offset;
16438
16439 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16440 mutex_exit(&dtrace_lock);
16441 return (EBUSY);
16442 }
16443
16444 /*
16445 * If this buffer has already been consumed, we're
16446 * going to indicate that there's nothing left here
16447 * to consume.
16448 */
16449 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16450 mutex_exit(&dtrace_lock);
16451
16452 desc.dtbd_size = 0;
16453 desc.dtbd_drops = 0;
16454 desc.dtbd_errors = 0;
16455 desc.dtbd_oldest = 0;
16456 sz = sizeof (desc);
16457
16458 if (copyout(&desc, (void *)arg, sz) != 0)
16459 return (EFAULT);
16460
16461 return (0);
16462 }
16463
16464 /*
16465 * If this is a ring buffer that has wrapped, we want
16466 * to copy the whole thing out.
16467 */
16468 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16469 dtrace_buffer_polish(buf);
16470 sz = buf->dtb_size;
16471 }
16472
16473 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16474 mutex_exit(&dtrace_lock);
16475 return (EFAULT);
16476 }
16477
16478 desc.dtbd_size = sz;
16479 desc.dtbd_drops = buf->dtb_drops;
16480 desc.dtbd_errors = buf->dtb_errors;
16481 desc.dtbd_oldest = buf->dtb_xamot_offset;
16482 desc.dtbd_timestamp = dtrace_gethrtime();
16483
16484 mutex_exit(&dtrace_lock);
16485
16486 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16487 return (EFAULT);
16488
16489 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16490
16491 return (0);
16492 }
16493
16494 if (buf->dtb_tomax == NULL) {
16495 ASSERT(buf->dtb_xamot == NULL);
16496 mutex_exit(&dtrace_lock);
16497 return (ENOENT);
16498 }
16499
16500 cached = buf->dtb_tomax;
16501 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16502
16503 dtrace_xcall(desc.dtbd_cpu,
16504 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16505
16506 state->dts_errors += buf->dtb_xamot_errors;
16507
16508 /*
16509 * If the buffers did not actually switch, then the cross call
16510 * did not take place -- presumably because the given CPU is
16511 * not in the ready set. If this is the case, we'll return
16512 * ENOENT.
16513 */
16514 if (buf->dtb_tomax == cached) {
16515 ASSERT(buf->dtb_xamot != cached);
16516 mutex_exit(&dtrace_lock);
16517 return (ENOENT);
16518 }
16519
16520 ASSERT(cached == buf->dtb_xamot);
16521
16522 /*
16523 * We have our snapshot; now copy it out.
16524 */
16525 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16526 buf->dtb_xamot_offset) != 0) {
16527 mutex_exit(&dtrace_lock);
16528 return (EFAULT);
16529 }
16530
16531 desc.dtbd_size = buf->dtb_xamot_offset;
16532 desc.dtbd_drops = buf->dtb_xamot_drops;
16533 desc.dtbd_errors = buf->dtb_xamot_errors;
16534 desc.dtbd_oldest = 0;
16535 desc.dtbd_timestamp = buf->dtb_switched;
16536
16537 mutex_exit(&dtrace_lock);
16538
16539 /*
16540 * Finally, copy out the buffer description.
16541 */
16542 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16543 return (EFAULT);
16544
16545 return (0);
16546 }
16547
16548 case DTRACEIOC_CONF: {
16549 dtrace_conf_t conf;
16550
16551 bzero(&conf, sizeof (conf));
16552 conf.dtc_difversion = DIF_VERSION;
16553 conf.dtc_difintregs = DIF_DIR_NREGS;
16554 conf.dtc_diftupregs = DIF_DTR_NREGS;
16555 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16556
16557 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16558 return (EFAULT);
16559
16560 return (0);
16561 }
16562
16563 case DTRACEIOC_STATUS: {
16564 dtrace_status_t stat;
16565 dtrace_dstate_t *dstate;
16566 int i, j;
16567 uint64_t nerrs;
16568
16569 /*
16570 * See the comment in dtrace_state_deadman() for the reason
16571 * for setting dts_laststatus to INT64_MAX before setting
16572 * it to the correct value.
16573 */
16574 state->dts_laststatus = INT64_MAX;
16575 dtrace_membar_producer();
16576 state->dts_laststatus = dtrace_gethrtime();
16577
16578 bzero(&stat, sizeof (stat));
16579
16580 mutex_enter(&dtrace_lock);
16581
16582 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16583 mutex_exit(&dtrace_lock);
16584 return (ENOENT);
16585 }
16586
16587 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16588 stat.dtst_exiting = 1;
16589
16590 nerrs = state->dts_errors;
16591 dstate = &state->dts_vstate.dtvs_dynvars;
16592
16593 for (i = 0; i < NCPU; i++) {
16594 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16595
16596 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16597 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16598 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16599
16600 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16601 stat.dtst_filled++;
16602
16603 nerrs += state->dts_buffer[i].dtb_errors;
16604
16605 for (j = 0; j < state->dts_nspeculations; j++) {
16606 dtrace_speculation_t *spec;
16607 dtrace_buffer_t *buf;
16608
16609 spec = &state->dts_speculations[j];
16610 buf = &spec->dtsp_buffer[i];
16611 stat.dtst_specdrops += buf->dtb_xamot_drops;
16612 }
16613 }
16614
16615 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16616 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16617 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16618 stat.dtst_dblerrors = state->dts_dblerrors;
16619 stat.dtst_killed =
16620 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16621 stat.dtst_errors = nerrs;
16622
16623 mutex_exit(&dtrace_lock);
16624
16625 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16626 return (EFAULT);
16627
16628 return (0);
16629 }
16630
16631 case DTRACEIOC_FORMAT: {
16632 dtrace_fmtdesc_t fmt;
16633 char *str;
16634 int len;
16635
16636 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16637 return (EFAULT);
16638
16639 mutex_enter(&dtrace_lock);
16640
16641 if (fmt.dtfd_format == 0 ||
16642 fmt.dtfd_format > state->dts_nformats) {
16643 mutex_exit(&dtrace_lock);
16644 return (EINVAL);
16645 }
16646
16647 /*
16648 * Format strings are allocated contiguously and they are
16649 * never freed; if a format index is less than the number
16650 * of formats, we can assert that the format map is non-NULL
16651 * and that the format for the specified index is non-NULL.
16652 */
16653 ASSERT(state->dts_formats != NULL);
16654 str = state->dts_formats[fmt.dtfd_format - 1];
16655 ASSERT(str != NULL);
16656
16657 len = strlen(str) + 1;
16658
16659 if (len > fmt.dtfd_length) {
16660 fmt.dtfd_length = len;
16661
16662 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16663 mutex_exit(&dtrace_lock);
16664 return (EINVAL);
16665 }
16666 } else {
16667 if (copyout(str, fmt.dtfd_string, len) != 0) {
16668 mutex_exit(&dtrace_lock);
16669 return (EINVAL);
16670 }
16671 }
16672
16673 mutex_exit(&dtrace_lock);
16674 return (0);
16675 }
16676
16677 default:
16678 break;
16679 }
16680
16681 return (ENOTTY);
16682 }
16683
16684 /*ARGSUSED*/
16685 static int
16686 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16687 {
16688 dtrace_state_t *state;
16689
16690 switch (cmd) {
16691 case DDI_DETACH:
16692 break;
16693
16694 case DDI_SUSPEND:
16695 return (DDI_SUCCESS);
16696
16697 default:
16698 return (DDI_FAILURE);
16699 }
16700
16701 mutex_enter(&cpu_lock);
16702 mutex_enter(&dtrace_provider_lock);
16703 mutex_enter(&dtrace_lock);
16704
16705 ASSERT(dtrace_opens == 0);
16706
16707 if (dtrace_helpers > 0) {
16708 mutex_exit(&dtrace_provider_lock);
16709 mutex_exit(&dtrace_lock);
16710 mutex_exit(&cpu_lock);
16711 return (DDI_FAILURE);
16712 }
16713
16714 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16715 mutex_exit(&dtrace_provider_lock);
16716 mutex_exit(&dtrace_lock);
16717 mutex_exit(&cpu_lock);
16718 return (DDI_FAILURE);
16719 }
16720
16721 dtrace_provider = NULL;
16722
16723 if ((state = dtrace_anon_grab()) != NULL) {
16724 /*
16725 * If there were ECBs on this state, the provider should
16726 * have not been allowed to detach; assert that there is
16727 * none.
16728 */
16729 ASSERT(state->dts_necbs == 0);
16730 dtrace_state_destroy(state);
16731
16732 /*
16733 * If we're being detached with anonymous state, we need to
16734 * indicate to the kernel debugger that DTrace is now inactive.
16735 */
16736 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16737 }
16738
16739 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16740 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16741 dtrace_cpu_init = NULL;
16742 dtrace_helpers_cleanup = NULL;
16743 dtrace_helpers_fork = NULL;
16744 dtrace_cpustart_init = NULL;
16745 dtrace_cpustart_fini = NULL;
16746 dtrace_debugger_init = NULL;
16747 dtrace_debugger_fini = NULL;
16748 dtrace_modload = NULL;
16749 dtrace_modunload = NULL;
16750
16751 ASSERT(dtrace_getf == 0);
16752 ASSERT(dtrace_closef == NULL);
16753
16754 mutex_exit(&cpu_lock);
16755
16756 if (dtrace_helptrace_enabled) {
16757 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16758 dtrace_helptrace_buffer = NULL;
16759 }
16760
16761 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16762 dtrace_probes = NULL;
16763 dtrace_nprobes = 0;
16764
16765 dtrace_hash_destroy(dtrace_bymod);
16766 dtrace_hash_destroy(dtrace_byfunc);
16767 dtrace_hash_destroy(dtrace_byname);
16768 dtrace_bymod = NULL;
16769 dtrace_byfunc = NULL;
16770 dtrace_byname = NULL;
16771
16772 kmem_cache_destroy(dtrace_state_cache);
16773 vmem_destroy(dtrace_minor);
16774 vmem_destroy(dtrace_arena);
16775
16776 if (dtrace_toxrange != NULL) {
16777 kmem_free(dtrace_toxrange,
16778 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16779 dtrace_toxrange = NULL;
16780 dtrace_toxranges = 0;
16781 dtrace_toxranges_max = 0;
16782 }
16783
16784 ddi_remove_minor_node(dtrace_devi, NULL);
16785 dtrace_devi = NULL;
16786
16787 ddi_soft_state_fini(&dtrace_softstate);
16788
16789 ASSERT(dtrace_vtime_references == 0);
16790 ASSERT(dtrace_opens == 0);
16791 ASSERT(dtrace_retained == NULL);
16792
16793 mutex_exit(&dtrace_lock);
16794 mutex_exit(&dtrace_provider_lock);
16795
16796 /*
16797 * We don't destroy the task queue until after we have dropped our
16798 * locks (taskq_destroy() may block on running tasks). To prevent
16799 * attempting to do work after we have effectively detached but before
16800 * the task queue has been destroyed, all tasks dispatched via the
16801 * task queue must check that DTrace is still attached before
16802 * performing any operation.
16803 */
16804 taskq_destroy(dtrace_taskq);
16805 dtrace_taskq = NULL;
16806
16807 return (DDI_SUCCESS);
16808 }
16809
16810 /*ARGSUSED*/
16811 static int
16812 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16813 {
16814 int error;
16815
16816 switch (infocmd) {
16817 case DDI_INFO_DEVT2DEVINFO:
16818 *result = (void *)dtrace_devi;
16819 error = DDI_SUCCESS;
16820 break;
16821 case DDI_INFO_DEVT2INSTANCE:
16822 *result = (void *)0;
16823 error = DDI_SUCCESS;
16824 break;
16825 default:
16826 error = DDI_FAILURE;
16827 }
16828 return (error);
16829 }
16830
16831 static struct cb_ops dtrace_cb_ops = {
16832 dtrace_open, /* open */
16833 dtrace_close, /* close */
16834 nulldev, /* strategy */
16835 nulldev, /* print */
16836 nodev, /* dump */
16837 nodev, /* read */
16838 nodev, /* write */
16839 dtrace_ioctl, /* ioctl */
16840 nodev, /* devmap */
16841 nodev, /* mmap */
16842 nodev, /* segmap */
16843 nochpoll, /* poll */
16844 ddi_prop_op, /* cb_prop_op */
16845 0, /* streamtab */
16846 D_NEW | D_MP /* Driver compatibility flag */
16847 };
16848
16849 static struct dev_ops dtrace_ops = {
16850 DEVO_REV, /* devo_rev */
16851 0, /* refcnt */
16852 dtrace_info, /* get_dev_info */
16853 nulldev, /* identify */
16854 nulldev, /* probe */
16855 dtrace_attach, /* attach */
16856 dtrace_detach, /* detach */
16857 nodev, /* reset */
16858 &dtrace_cb_ops, /* driver operations */
16859 NULL, /* bus operations */
16860 nodev, /* dev power */
16861 ddi_quiesce_not_needed, /* quiesce */
16862 };
16863
16864 static struct modldrv modldrv = {
16865 &mod_driverops, /* module type (this is a pseudo driver) */
16866 "Dynamic Tracing", /* name of module */
16867 &dtrace_ops, /* driver ops */
16868 };
16869
16870 static struct modlinkage modlinkage = {
16871 MODREV_1,
16872 (void *)&modldrv,
16873 NULL
16874 };
16875
16876 int
16877 _init(void)
16878 {
16879 return (mod_install(&modlinkage));
16880 }
16881
16882 int
16883 _info(struct modinfo *modinfop)
16884 {
16885 return (mod_info(&modlinkage, modinfop));
16886 }
16887
16888 int
16889 _fini(void)
16890 {
16891 return (mod_remove(&modlinkage));
16892 }