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 regs[rd] = (int8_t)
5723 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5724 break;
5725 case DIF_OP_ULDSH:
5726 regs[rd] = (int16_t)
5727 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5728 break;
5729 case DIF_OP_ULDSW:
5730 regs[rd] = (int32_t)
5731 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5732 break;
5733 case DIF_OP_ULDUB:
5734 regs[rd] =
5735 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5736 break;
5737 case DIF_OP_ULDUH:
5738 regs[rd] =
5739 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5740 break;
5741 case DIF_OP_ULDUW:
5742 regs[rd] =
5743 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5744 break;
5745 case DIF_OP_ULDX:
5746 regs[rd] =
5747 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5748 break;
5749 case DIF_OP_RET:
5750 rval = regs[rd];
5751 pc = textlen;
5752 break;
5753 case DIF_OP_NOP:
5754 break;
5755 case DIF_OP_SETX:
5756 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5757 break;
5758 case DIF_OP_SETS:
5759 regs[rd] = (uint64_t)(uintptr_t)
5760 (strtab + DIF_INSTR_STRING(instr));
5761 break;
5762 case DIF_OP_SCMP: {
5763 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5764 uintptr_t s1 = regs[r1];
5765 uintptr_t s2 = regs[r2];
5766
5767 if (s1 != NULL &&
5768 !dtrace_strcanload(s1, sz, mstate, vstate))
5769 break;
5770 if (s2 != NULL &&
5771 !dtrace_strcanload(s2, sz, mstate, vstate))
5772 break;
5773
5774 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5775
5776 cc_n = cc_r < 0;
5777 cc_z = cc_r == 0;
5778 cc_v = cc_c = 0;
5779 break;
5780 }
5781 case DIF_OP_LDGA:
5782 regs[rd] = dtrace_dif_variable(mstate, state,
5783 r1, regs[r2]);
5784 break;
5785 case DIF_OP_LDGS:
5786 id = DIF_INSTR_VAR(instr);
5787
5788 if (id >= DIF_VAR_OTHER_UBASE) {
5789 uintptr_t a;
5790
5791 id -= DIF_VAR_OTHER_UBASE;
5792 svar = vstate->dtvs_globals[id];
5793 ASSERT(svar != NULL);
5794 v = &svar->dtsv_var;
5795
5796 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5797 regs[rd] = svar->dtsv_data;
5798 break;
5799 }
5800
5801 a = (uintptr_t)svar->dtsv_data;
5802
5803 if (*(uint8_t *)a == UINT8_MAX) {
5804 /*
5805 * If the 0th byte is set to UINT8_MAX
5806 * then this is to be treated as a
5807 * reference to a NULL variable.
5808 */
5809 regs[rd] = NULL;
5810 } else {
5811 regs[rd] = a + sizeof (uint64_t);
5812 }
5813
5814 break;
5815 }
5816
5817 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5818 break;
5819
5820 case DIF_OP_STGS:
5821 id = DIF_INSTR_VAR(instr);
5822
5823 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5824 id -= DIF_VAR_OTHER_UBASE;
5825
5826 svar = vstate->dtvs_globals[id];
5827 ASSERT(svar != NULL);
5828 v = &svar->dtsv_var;
5829
5830 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5831 uintptr_t a = (uintptr_t)svar->dtsv_data;
5832
5833 ASSERT(a != NULL);
5834 ASSERT(svar->dtsv_size != 0);
5835
5836 if (regs[rd] == NULL) {
5837 *(uint8_t *)a = UINT8_MAX;
5838 break;
5839 } else {
5840 *(uint8_t *)a = 0;
5841 a += sizeof (uint64_t);
5842 }
5843 if (!dtrace_vcanload(
5844 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5845 mstate, vstate))
5846 break;
5847
5848 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5849 (void *)a, &v->dtdv_type);
5850 break;
5851 }
5852
5853 svar->dtsv_data = regs[rd];
5854 break;
5855
5856 case DIF_OP_LDTA:
5857 /*
5858 * There are no DTrace built-in thread-local arrays at
5859 * present. This opcode is saved for future work.
5860 */
5861 *flags |= CPU_DTRACE_ILLOP;
5862 regs[rd] = 0;
5863 break;
5864
5865 case DIF_OP_LDLS:
5866 id = DIF_INSTR_VAR(instr);
5867
5868 if (id < DIF_VAR_OTHER_UBASE) {
5869 /*
5870 * For now, this has no meaning.
5871 */
5872 regs[rd] = 0;
5873 break;
5874 }
5875
5876 id -= DIF_VAR_OTHER_UBASE;
5877
5878 ASSERT(id < vstate->dtvs_nlocals);
5879 ASSERT(vstate->dtvs_locals != NULL);
5880
5881 svar = vstate->dtvs_locals[id];
5882 ASSERT(svar != NULL);
5883 v = &svar->dtsv_var;
5884
5885 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5886 uintptr_t a = (uintptr_t)svar->dtsv_data;
5887 size_t sz = v->dtdv_type.dtdt_size;
5888
5889 sz += sizeof (uint64_t);
5890 ASSERT(svar->dtsv_size == NCPU * sz);
5891 a += CPU->cpu_id * sz;
5892
5893 if (*(uint8_t *)a == UINT8_MAX) {
5894 /*
5895 * If the 0th byte is set to UINT8_MAX
5896 * then this is to be treated as a
5897 * reference to a NULL variable.
5898 */
5899 regs[rd] = NULL;
5900 } else {
5901 regs[rd] = a + sizeof (uint64_t);
5902 }
5903
5904 break;
5905 }
5906
5907 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5908 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5909 regs[rd] = tmp[CPU->cpu_id];
5910 break;
5911
5912 case DIF_OP_STLS:
5913 id = DIF_INSTR_VAR(instr);
5914
5915 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5916 id -= DIF_VAR_OTHER_UBASE;
5917 ASSERT(id < vstate->dtvs_nlocals);
5918
5919 ASSERT(vstate->dtvs_locals != NULL);
5920 svar = vstate->dtvs_locals[id];
5921 ASSERT(svar != NULL);
5922 v = &svar->dtsv_var;
5923
5924 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5925 uintptr_t a = (uintptr_t)svar->dtsv_data;
5926 size_t sz = v->dtdv_type.dtdt_size;
5927
5928 sz += sizeof (uint64_t);
5929 ASSERT(svar->dtsv_size == NCPU * sz);
5930 a += CPU->cpu_id * sz;
5931
5932 if (regs[rd] == NULL) {
5933 *(uint8_t *)a = UINT8_MAX;
5934 break;
5935 } else {
5936 *(uint8_t *)a = 0;
5937 a += sizeof (uint64_t);
5938 }
5939
5940 if (!dtrace_vcanload(
5941 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5942 mstate, vstate))
5943 break;
5944
5945 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5946 (void *)a, &v->dtdv_type);
5947 break;
5948 }
5949
5950 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5951 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5952 tmp[CPU->cpu_id] = regs[rd];
5953 break;
5954
5955 case DIF_OP_LDTS: {
5956 dtrace_dynvar_t *dvar;
5957 dtrace_key_t *key;
5958
5959 id = DIF_INSTR_VAR(instr);
5960 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5961 id -= DIF_VAR_OTHER_UBASE;
5962 v = &vstate->dtvs_tlocals[id];
5963
5964 key = &tupregs[DIF_DTR_NREGS];
5965 key[0].dttk_value = (uint64_t)id;
5966 key[0].dttk_size = 0;
5967 DTRACE_TLS_THRKEY(key[1].dttk_value);
5968 key[1].dttk_size = 0;
5969
5970 dvar = dtrace_dynvar(dstate, 2, key,
5971 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5972 mstate, vstate);
5973
5974 if (dvar == NULL) {
5975 regs[rd] = 0;
5976 break;
5977 }
5978
5979 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5980 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5981 } else {
5982 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5983 }
5984
5985 break;
5986 }
5987
5988 case DIF_OP_STTS: {
5989 dtrace_dynvar_t *dvar;
5990 dtrace_key_t *key;
5991
5992 id = DIF_INSTR_VAR(instr);
5993 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5994 id -= DIF_VAR_OTHER_UBASE;
5995
5996 key = &tupregs[DIF_DTR_NREGS];
5997 key[0].dttk_value = (uint64_t)id;
5998 key[0].dttk_size = 0;
5999 DTRACE_TLS_THRKEY(key[1].dttk_value);
6000 key[1].dttk_size = 0;
6001 v = &vstate->dtvs_tlocals[id];
6002
6003 dvar = dtrace_dynvar(dstate, 2, key,
6004 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6005 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6006 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6007 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6008
6009 /*
6010 * Given that we're storing to thread-local data,
6011 * we need to flush our predicate cache.
6012 */
6013 curthread->t_predcache = NULL;
6014
6015 if (dvar == NULL)
6016 break;
6017
6018 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6019 if (!dtrace_vcanload(
6020 (void *)(uintptr_t)regs[rd],
6021 &v->dtdv_type, mstate, vstate))
6022 break;
6023
6024 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6025 dvar->dtdv_data, &v->dtdv_type);
6026 } else {
6027 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6028 }
6029
6030 break;
6031 }
6032
6033 case DIF_OP_SRA:
6034 regs[rd] = (int64_t)regs[r1] >> regs[r2];
6035 break;
6036
6037 case DIF_OP_CALL:
6038 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6039 regs, tupregs, ttop, mstate, state);
6040 break;
6041
6042 case DIF_OP_PUSHTR:
6043 if (ttop == DIF_DTR_NREGS) {
6044 *flags |= CPU_DTRACE_TUPOFLOW;
6045 break;
6046 }
6047
6048 if (r1 == DIF_TYPE_STRING) {
6049 /*
6050 * If this is a string type and the size is 0,
6051 * we'll use the system-wide default string
6052 * size. Note that we are _not_ looking at
6053 * the value of the DTRACEOPT_STRSIZE option;
6054 * had this been set, we would expect to have
6055 * a non-zero size value in the "pushtr".
6056 */
6057 tupregs[ttop].dttk_size =
6058 dtrace_strlen((char *)(uintptr_t)regs[rd],
6059 regs[r2] ? regs[r2] :
6060 dtrace_strsize_default) + 1;
6061 } else {
6062 tupregs[ttop].dttk_size = regs[r2];
6063 }
6064
6065 tupregs[ttop++].dttk_value = regs[rd];
6066 break;
6067
6068 case DIF_OP_PUSHTV:
6069 if (ttop == DIF_DTR_NREGS) {
6070 *flags |= CPU_DTRACE_TUPOFLOW;
6071 break;
6072 }
6073
6074 tupregs[ttop].dttk_value = regs[rd];
6075 tupregs[ttop++].dttk_size = 0;
6076 break;
6077
6078 case DIF_OP_POPTS:
6079 if (ttop != 0)
6080 ttop--;
6081 break;
6082
6083 case DIF_OP_FLUSHTS:
6084 ttop = 0;
6085 break;
6086
6087 case DIF_OP_LDGAA:
6088 case DIF_OP_LDTAA: {
6089 dtrace_dynvar_t *dvar;
6090 dtrace_key_t *key = tupregs;
6091 uint_t nkeys = ttop;
6092
6093 id = DIF_INSTR_VAR(instr);
6094 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6095 id -= DIF_VAR_OTHER_UBASE;
6096
6097 key[nkeys].dttk_value = (uint64_t)id;
6098 key[nkeys++].dttk_size = 0;
6099
6100 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6101 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6102 key[nkeys++].dttk_size = 0;
6103 v = &vstate->dtvs_tlocals[id];
6104 } else {
6105 v = &vstate->dtvs_globals[id]->dtsv_var;
6106 }
6107
6108 dvar = dtrace_dynvar(dstate, nkeys, key,
6109 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6110 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6111 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6112
6113 if (dvar == NULL) {
6114 regs[rd] = 0;
6115 break;
6116 }
6117
6118 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6119 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6120 } else {
6121 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6122 }
6123
6124 break;
6125 }
6126
6127 case DIF_OP_STGAA:
6128 case DIF_OP_STTAA: {
6129 dtrace_dynvar_t *dvar;
6130 dtrace_key_t *key = tupregs;
6131 uint_t nkeys = ttop;
6132
6133 id = DIF_INSTR_VAR(instr);
6134 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6135 id -= DIF_VAR_OTHER_UBASE;
6136
6137 key[nkeys].dttk_value = (uint64_t)id;
6138 key[nkeys++].dttk_size = 0;
6139
6140 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6141 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6142 key[nkeys++].dttk_size = 0;
6143 v = &vstate->dtvs_tlocals[id];
6144 } else {
6145 v = &vstate->dtvs_globals[id]->dtsv_var;
6146 }
6147
6148 dvar = dtrace_dynvar(dstate, nkeys, key,
6149 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6150 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6151 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6152 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6153
6154 if (dvar == NULL)
6155 break;
6156
6157 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6158 if (!dtrace_vcanload(
6159 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6160 mstate, vstate))
6161 break;
6162
6163 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6164 dvar->dtdv_data, &v->dtdv_type);
6165 } else {
6166 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6167 }
6168
6169 break;
6170 }
6171
6172 case DIF_OP_ALLOCS: {
6173 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6174 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6175
6176 /*
6177 * Rounding up the user allocation size could have
6178 * overflowed large, bogus allocations (like -1ULL) to
6179 * 0.
6180 */
6181 if (size < regs[r1] ||
6182 !DTRACE_INSCRATCH(mstate, size)) {
6183 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6184 regs[rd] = NULL;
6185 break;
6186 }
6187
6188 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6189 mstate->dtms_scratch_ptr += size;
6190 regs[rd] = ptr;
6191 break;
6192 }
6193
6194 case DIF_OP_COPYS:
6195 if (!dtrace_canstore(regs[rd], regs[r2],
6196 mstate, vstate)) {
6197 *flags |= CPU_DTRACE_BADADDR;
6198 *illval = regs[rd];
6199 break;
6200 }
6201
6202 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6203 break;
6204
6205 dtrace_bcopy((void *)(uintptr_t)regs[r1],
6206 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6207 break;
6208
6209 case DIF_OP_STB:
6210 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6211 *flags |= CPU_DTRACE_BADADDR;
6212 *illval = regs[rd];
6213 break;
6214 }
6215 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6216 break;
6217
6218 case DIF_OP_STH:
6219 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6220 *flags |= CPU_DTRACE_BADADDR;
6221 *illval = regs[rd];
6222 break;
6223 }
6224 if (regs[rd] & 1) {
6225 *flags |= CPU_DTRACE_BADALIGN;
6226 *illval = regs[rd];
6227 break;
6228 }
6229 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6230 break;
6231
6232 case DIF_OP_STW:
6233 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6234 *flags |= CPU_DTRACE_BADADDR;
6235 *illval = regs[rd];
6236 break;
6237 }
6238 if (regs[rd] & 3) {
6239 *flags |= CPU_DTRACE_BADALIGN;
6240 *illval = regs[rd];
6241 break;
6242 }
6243 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6244 break;
6245
6246 case DIF_OP_STX:
6247 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6248 *flags |= CPU_DTRACE_BADADDR;
6249 *illval = regs[rd];
6250 break;
6251 }
6252 if (regs[rd] & 7) {
6253 *flags |= CPU_DTRACE_BADALIGN;
6254 *illval = regs[rd];
6255 break;
6256 }
6257 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6258 break;
6259 }
6260 }
6261
6262 if (!(*flags & CPU_DTRACE_FAULT))
6263 return (rval);
6264
6265 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6266 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6267
6268 return (0);
6269 }
6270
6271 static void
6272 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6273 {
6274 dtrace_probe_t *probe = ecb->dte_probe;
6275 dtrace_provider_t *prov = probe->dtpr_provider;
6276 char c[DTRACE_FULLNAMELEN + 80], *str;
6277 char *msg = "dtrace: breakpoint action at probe ";
6278 char *ecbmsg = " (ecb ";
6279 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6280 uintptr_t val = (uintptr_t)ecb;
6281 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6282
6283 if (dtrace_destructive_disallow)
6284 return;
6285
6286 /*
6287 * It's impossible to be taking action on the NULL probe.
6288 */
6289 ASSERT(probe != NULL);
6290
6291 /*
6292 * This is a poor man's (destitute man's?) sprintf(): we want to
6293 * print the provider name, module name, function name and name of
6294 * the probe, along with the hex address of the ECB with the breakpoint
6295 * action -- all of which we must place in the character buffer by
6296 * hand.
6297 */
6298 while (*msg != '\0')
6299 c[i++] = *msg++;
6300
6301 for (str = prov->dtpv_name; *str != '\0'; str++)
6302 c[i++] = *str;
6303 c[i++] = ':';
6304
6305 for (str = probe->dtpr_mod; *str != '\0'; str++)
6306 c[i++] = *str;
6307 c[i++] = ':';
6308
6309 for (str = probe->dtpr_func; *str != '\0'; str++)
6310 c[i++] = *str;
6311 c[i++] = ':';
6312
6313 for (str = probe->dtpr_name; *str != '\0'; str++)
6314 c[i++] = *str;
6315
6316 while (*ecbmsg != '\0')
6317 c[i++] = *ecbmsg++;
6318
6319 while (shift >= 0) {
6320 mask = (uintptr_t)0xf << shift;
6321
6322 if (val >= ((uintptr_t)1 << shift))
6323 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6324 shift -= 4;
6325 }
6326
6327 c[i++] = ')';
6328 c[i] = '\0';
6329
6330 debug_enter(c);
6331 }
6332
6333 static void
6334 dtrace_action_panic(dtrace_ecb_t *ecb)
6335 {
6336 dtrace_probe_t *probe = ecb->dte_probe;
6337
6338 /*
6339 * It's impossible to be taking action on the NULL probe.
6340 */
6341 ASSERT(probe != NULL);
6342
6343 if (dtrace_destructive_disallow)
6344 return;
6345
6346 if (dtrace_panicked != NULL)
6347 return;
6348
6349 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6350 return;
6351
6352 /*
6353 * We won the right to panic. (We want to be sure that only one
6354 * thread calls panic() from dtrace_probe(), and that panic() is
6355 * called exactly once.)
6356 */
6357 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6358 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6359 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6360 }
6361
6362 static void
6363 dtrace_action_raise(uint64_t sig)
6364 {
6365 if (dtrace_destructive_disallow)
6366 return;
6367
6368 if (sig >= NSIG) {
6369 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6370 return;
6371 }
6372
6373 /*
6374 * raise() has a queue depth of 1 -- we ignore all subsequent
6375 * invocations of the raise() action.
6376 */
6377 if (curthread->t_dtrace_sig == 0)
6378 curthread->t_dtrace_sig = (uint8_t)sig;
6379
6380 curthread->t_sig_check = 1;
6381 aston(curthread);
6382 }
6383
6384 static void
6385 dtrace_action_stop(void)
6386 {
6387 if (dtrace_destructive_disallow)
6388 return;
6389
6390 if (!curthread->t_dtrace_stop) {
6391 curthread->t_dtrace_stop = 1;
6392 curthread->t_sig_check = 1;
6393 aston(curthread);
6394 }
6395 }
6396
6397 static void
6398 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
6399 {
6400 hrtime_t now;
6401 volatile uint16_t *flags;
6402 cpu_t *cpu = CPU;
6403
6404 if (dtrace_destructive_disallow)
6405 return;
6406
6407 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
6408
6409 now = dtrace_gethrtime();
6410
6411 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
6412 /*
6413 * We need to advance the mark to the current time.
6414 */
6415 cpu->cpu_dtrace_chillmark = now;
6416 cpu->cpu_dtrace_chilled = 0;
6417 }
6418
6419 /*
6420 * Now check to see if the requested chill time would take us over
6421 * the maximum amount of time allowed in the chill interval. (Or
6422 * worse, if the calculation itself induces overflow.)
6423 */
6424 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
6425 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
6426 *flags |= CPU_DTRACE_ILLOP;
6427 return;
6428 }
6429
6430 while (dtrace_gethrtime() - now < val)
6431 continue;
6432
6433 /*
6434 * Normally, we assure that the value of the variable "timestamp" does
6435 * not change within an ECB. The presence of chill() represents an
6436 * exception to this rule, however.
6437 */
6438 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
6439 cpu->cpu_dtrace_chilled += val;
6440 }
6441
6442 static void
6443 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
6444 uint64_t *buf, uint64_t arg)
6445 {
6446 int nframes = DTRACE_USTACK_NFRAMES(arg);
6447 int strsize = DTRACE_USTACK_STRSIZE(arg);
6448 uint64_t *pcs = &buf[1], *fps;
6449 char *str = (char *)&pcs[nframes];
6450 int size, offs = 0, i, j;
6451 uintptr_t old = mstate->dtms_scratch_ptr, saved;
6452 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6453 char *sym;
6454
6455 /*
6456 * Should be taking a faster path if string space has not been
6457 * allocated.
6458 */
6459 ASSERT(strsize != 0);
6460
6461 /*
6462 * We will first allocate some temporary space for the frame pointers.
6463 */
6464 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6465 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
6466 (nframes * sizeof (uint64_t));
6467
6468 if (!DTRACE_INSCRATCH(mstate, size)) {
6469 /*
6470 * Not enough room for our frame pointers -- need to indicate
6471 * that we ran out of scratch space.
6472 */
6473 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6474 return;
6475 }
6476
6477 mstate->dtms_scratch_ptr += size;
6478 saved = mstate->dtms_scratch_ptr;
6479
6480 /*
6481 * Now get a stack with both program counters and frame pointers.
6482 */
6483 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6484 dtrace_getufpstack(buf, fps, nframes + 1);
6485 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6486
6487 /*
6488 * If that faulted, we're cooked.
6489 */
6490 if (*flags & CPU_DTRACE_FAULT)
6491 goto out;
6492
6493 /*
6494 * Now we want to walk up the stack, calling the USTACK helper. For
6495 * each iteration, we restore the scratch pointer.
6496 */
6497 for (i = 0; i < nframes; i++) {
6498 mstate->dtms_scratch_ptr = saved;
6499
6500 if (offs >= strsize)
6501 break;
6502
6503 sym = (char *)(uintptr_t)dtrace_helper(
6504 DTRACE_HELPER_ACTION_USTACK,
6505 mstate, state, pcs[i], fps[i]);
6506
6507 /*
6508 * If we faulted while running the helper, we're going to
6509 * clear the fault and null out the corresponding string.
6510 */
6511 if (*flags & CPU_DTRACE_FAULT) {
6512 *flags &= ~CPU_DTRACE_FAULT;
6513 str[offs++] = '\0';
6514 continue;
6515 }
6516
6517 if (sym == NULL) {
6518 str[offs++] = '\0';
6519 continue;
6520 }
6521
6522 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6523
6524 /*
6525 * Now copy in the string that the helper returned to us.
6526 */
6527 for (j = 0; offs + j < strsize; j++) {
6528 if ((str[offs + j] = sym[j]) == '\0')
6529 break;
6530 }
6531
6532 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6533
6534 offs += j + 1;
6535 }
6536
6537 if (offs >= strsize) {
6538 /*
6539 * If we didn't have room for all of the strings, we don't
6540 * abort processing -- this needn't be a fatal error -- but we
6541 * still want to increment a counter (dts_stkstroverflows) to
6542 * allow this condition to be warned about. (If this is from
6543 * a jstack() action, it is easily tuned via jstackstrsize.)
6544 */
6545 dtrace_error(&state->dts_stkstroverflows);
6546 }
6547
6548 while (offs < strsize)
6549 str[offs++] = '\0';
6550
6551 out:
6552 mstate->dtms_scratch_ptr = old;
6553 }
6554
6555 /*
6556 * If you're looking for the epicenter of DTrace, you just found it. This
6557 * is the function called by the provider to fire a probe -- from which all
6558 * subsequent probe-context DTrace activity emanates.
6559 */
6560 void
6561 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6562 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6563 {
6564 processorid_t cpuid;
6565 dtrace_icookie_t cookie;
6566 dtrace_probe_t *probe;
6567 dtrace_mstate_t mstate;
6568 dtrace_ecb_t *ecb;
6569 dtrace_action_t *act;
6570 intptr_t offs;
6571 size_t size;
6572 int vtime, onintr;
6573 volatile uint16_t *flags;
6574 hrtime_t now, end;
6575
6576 /*
6577 * Kick out immediately if this CPU is still being born (in which case
6578 * curthread will be set to -1) or the current thread can't allow
6579 * probes in its current context.
6580 */
6581 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6582 return;
6583
6584 cookie = dtrace_interrupt_disable();
6585 probe = dtrace_probes[id - 1];
6586 cpuid = CPU->cpu_id;
6587 onintr = CPU_ON_INTR(CPU);
6588
6589 CPU->cpu_dtrace_probes++;
6590
6591 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6592 probe->dtpr_predcache == curthread->t_predcache) {
6593 /*
6594 * We have hit in the predicate cache; we know that
6595 * this predicate would evaluate to be false.
6596 */
6597 dtrace_interrupt_enable(cookie);
6598 return;
6599 }
6600
6601 if (panic_quiesce) {
6602 /*
6603 * We don't trace anything if we're panicking.
6604 */
6605 dtrace_interrupt_enable(cookie);
6606 return;
6607 }
6608
6609 now = dtrace_gethrtime();
6610 vtime = dtrace_vtime_references != 0;
6611
6612 if (vtime && curthread->t_dtrace_start)
6613 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6614
6615 mstate.dtms_difo = NULL;
6616 mstate.dtms_probe = probe;
6617 mstate.dtms_strtok = NULL;
6618 mstate.dtms_arg[0] = arg0;
6619 mstate.dtms_arg[1] = arg1;
6620 mstate.dtms_arg[2] = arg2;
6621 mstate.dtms_arg[3] = arg3;
6622 mstate.dtms_arg[4] = arg4;
6623
6624 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6625
6626 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6627 dtrace_predicate_t *pred = ecb->dte_predicate;
6628 dtrace_state_t *state = ecb->dte_state;
6629 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6630 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6631 dtrace_vstate_t *vstate = &state->dts_vstate;
6632 dtrace_provider_t *prov = probe->dtpr_provider;
6633 uint64_t tracememsize = 0;
6634 int committed = 0;
6635 caddr_t tomax;
6636
6637 /*
6638 * A little subtlety with the following (seemingly innocuous)
6639 * declaration of the automatic 'val': by looking at the
6640 * code, you might think that it could be declared in the
6641 * action processing loop, below. (That is, it's only used in
6642 * the action processing loop.) However, it must be declared
6643 * out of that scope because in the case of DIF expression
6644 * arguments to aggregating actions, one iteration of the
6645 * action loop will use the last iteration's value.
6646 */
6647 #ifdef lint
6648 uint64_t val = 0;
6649 #else
6650 uint64_t val;
6651 #endif
6652
6653 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6654 mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
6655 mstate.dtms_getf = NULL;
6656
6657 *flags &= ~CPU_DTRACE_ERROR;
6658
6659 if (prov == dtrace_provider) {
6660 /*
6661 * If dtrace itself is the provider of this probe,
6662 * we're only going to continue processing the ECB if
6663 * arg0 (the dtrace_state_t) is equal to the ECB's
6664 * creating state. (This prevents disjoint consumers
6665 * from seeing one another's metaprobes.)
6666 */
6667 if (arg0 != (uint64_t)(uintptr_t)state)
6668 continue;
6669 }
6670
6671 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6672 /*
6673 * We're not currently active. If our provider isn't
6674 * the dtrace pseudo provider, we're not interested.
6675 */
6676 if (prov != dtrace_provider)
6677 continue;
6678
6679 /*
6680 * Now we must further check if we are in the BEGIN
6681 * probe. If we are, we will only continue processing
6682 * if we're still in WARMUP -- if one BEGIN enabling
6683 * has invoked the exit() action, we don't want to
6684 * evaluate subsequent BEGIN enablings.
6685 */
6686 if (probe->dtpr_id == dtrace_probeid_begin &&
6687 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6688 ASSERT(state->dts_activity ==
6689 DTRACE_ACTIVITY_DRAINING);
6690 continue;
6691 }
6692 }
6693
6694 if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
6695 continue;
6696
6697 if (now - state->dts_alive > dtrace_deadman_timeout) {
6698 /*
6699 * We seem to be dead. Unless we (a) have kernel
6700 * destructive permissions (b) have explicitly enabled
6701 * destructive actions and (c) destructive actions have
6702 * not been disabled, we're going to transition into
6703 * the KILLED state, from which no further processing
6704 * on this state will be performed.
6705 */
6706 if (!dtrace_priv_kernel_destructive(state) ||
6707 !state->dts_cred.dcr_destructive ||
6708 dtrace_destructive_disallow) {
6709 void *activity = &state->dts_activity;
6710 dtrace_activity_t current;
6711
6712 do {
6713 current = state->dts_activity;
6714 } while (dtrace_cas32(activity, current,
6715 DTRACE_ACTIVITY_KILLED) != current);
6716
6717 continue;
6718 }
6719 }
6720
6721 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6722 ecb->dte_alignment, state, &mstate)) < 0)
6723 continue;
6724
6725 tomax = buf->dtb_tomax;
6726 ASSERT(tomax != NULL);
6727
6728 if (ecb->dte_size != 0) {
6729 dtrace_rechdr_t dtrh;
6730 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6731 mstate.dtms_timestamp = dtrace_gethrtime();
6732 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6733 }
6734 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6735 dtrh.dtrh_epid = ecb->dte_epid;
6736 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6737 mstate.dtms_timestamp);
6738 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6739 }
6740
6741 mstate.dtms_epid = ecb->dte_epid;
6742 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6743
6744 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6745 mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
6746
6747 if (pred != NULL) {
6748 dtrace_difo_t *dp = pred->dtp_difo;
6749 int rval;
6750
6751 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6752
6753 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6754 dtrace_cacheid_t cid = probe->dtpr_predcache;
6755
6756 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6757 /*
6758 * Update the predicate cache...
6759 */
6760 ASSERT(cid == pred->dtp_cacheid);
6761 curthread->t_predcache = cid;
6762 }
6763
6764 continue;
6765 }
6766 }
6767
6768 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6769 act != NULL; act = act->dta_next) {
6770 size_t valoffs;
6771 dtrace_difo_t *dp;
6772 dtrace_recdesc_t *rec = &act->dta_rec;
6773
6774 size = rec->dtrd_size;
6775 valoffs = offs + rec->dtrd_offset;
6776
6777 if (DTRACEACT_ISAGG(act->dta_kind)) {
6778 uint64_t v = 0xbad;
6779 dtrace_aggregation_t *agg;
6780
6781 agg = (dtrace_aggregation_t *)act;
6782
6783 if ((dp = act->dta_difo) != NULL)
6784 v = dtrace_dif_emulate(dp,
6785 &mstate, vstate, state);
6786
6787 if (*flags & CPU_DTRACE_ERROR)
6788 continue;
6789
6790 /*
6791 * Note that we always pass the expression
6792 * value from the previous iteration of the
6793 * action loop. This value will only be used
6794 * if there is an expression argument to the
6795 * aggregating action, denoted by the
6796 * dtag_hasarg field.
6797 */
6798 dtrace_aggregate(agg, buf,
6799 offs, aggbuf, v, val);
6800 continue;
6801 }
6802
6803 switch (act->dta_kind) {
6804 case DTRACEACT_STOP:
6805 if (dtrace_priv_proc_destructive(state,
6806 &mstate))
6807 dtrace_action_stop();
6808 continue;
6809
6810 case DTRACEACT_BREAKPOINT:
6811 if (dtrace_priv_kernel_destructive(state))
6812 dtrace_action_breakpoint(ecb);
6813 continue;
6814
6815 case DTRACEACT_PANIC:
6816 if (dtrace_priv_kernel_destructive(state))
6817 dtrace_action_panic(ecb);
6818 continue;
6819
6820 case DTRACEACT_STACK:
6821 if (!dtrace_priv_kernel(state))
6822 continue;
6823
6824 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6825 size / sizeof (pc_t), probe->dtpr_aframes,
6826 DTRACE_ANCHORED(probe) ? NULL :
6827 (uint32_t *)arg0);
6828
6829 continue;
6830
6831 case DTRACEACT_JSTACK:
6832 case DTRACEACT_USTACK:
6833 if (!dtrace_priv_proc(state, &mstate))
6834 continue;
6835
6836 /*
6837 * See comment in DIF_VAR_PID.
6838 */
6839 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6840 CPU_ON_INTR(CPU)) {
6841 int depth = DTRACE_USTACK_NFRAMES(
6842 rec->dtrd_arg) + 1;
6843
6844 dtrace_bzero((void *)(tomax + valoffs),
6845 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6846 + depth * sizeof (uint64_t));
6847
6848 continue;
6849 }
6850
6851 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6852 curproc->p_dtrace_helpers != NULL) {
6853 /*
6854 * This is the slow path -- we have
6855 * allocated string space, and we're
6856 * getting the stack of a process that
6857 * has helpers. Call into a separate
6858 * routine to perform this processing.
6859 */
6860 dtrace_action_ustack(&mstate, state,
6861 (uint64_t *)(tomax + valoffs),
6862 rec->dtrd_arg);
6863 continue;
6864 }
6865
6866 /*
6867 * Clear the string space, since there's no
6868 * helper to do it for us.
6869 */
6870 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
6871 int depth = DTRACE_USTACK_NFRAMES(
6872 rec->dtrd_arg);
6873 size_t strsize = DTRACE_USTACK_STRSIZE(
6874 rec->dtrd_arg);
6875 uint64_t *buf = (uint64_t *)(tomax +
6876 valoffs);
6877 void *strspace = &buf[depth + 1];
6878
6879 dtrace_bzero(strspace,
6880 MIN(depth, strsize));
6881 }
6882
6883 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6884 dtrace_getupcstack((uint64_t *)
6885 (tomax + valoffs),
6886 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6887 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6888 continue;
6889
6890 default:
6891 break;
6892 }
6893
6894 dp = act->dta_difo;
6895 ASSERT(dp != NULL);
6896
6897 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6898
6899 if (*flags & CPU_DTRACE_ERROR)
6900 continue;
6901
6902 switch (act->dta_kind) {
6903 case DTRACEACT_SPECULATE: {
6904 dtrace_rechdr_t *dtrh;
6905
6906 ASSERT(buf == &state->dts_buffer[cpuid]);
6907 buf = dtrace_speculation_buffer(state,
6908 cpuid, val);
6909
6910 if (buf == NULL) {
6911 *flags |= CPU_DTRACE_DROP;
6912 continue;
6913 }
6914
6915 offs = dtrace_buffer_reserve(buf,
6916 ecb->dte_needed, ecb->dte_alignment,
6917 state, NULL);
6918
6919 if (offs < 0) {
6920 *flags |= CPU_DTRACE_DROP;
6921 continue;
6922 }
6923
6924 tomax = buf->dtb_tomax;
6925 ASSERT(tomax != NULL);
6926
6927 if (ecb->dte_size == 0)
6928 continue;
6929
6930 ASSERT3U(ecb->dte_size, >=,
6931 sizeof (dtrace_rechdr_t));
6932 dtrh = ((void *)(tomax + offs));
6933 dtrh->dtrh_epid = ecb->dte_epid;
6934 /*
6935 * When the speculation is committed, all of
6936 * the records in the speculative buffer will
6937 * have their timestamps set to the commit
6938 * time. Until then, it is set to a sentinel
6939 * value, for debugability.
6940 */
6941 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6942 continue;
6943 }
6944
6945 case DTRACEACT_CHILL:
6946 if (dtrace_priv_kernel_destructive(state))
6947 dtrace_action_chill(&mstate, val);
6948 continue;
6949
6950 case DTRACEACT_RAISE:
6951 if (dtrace_priv_proc_destructive(state,
6952 &mstate))
6953 dtrace_action_raise(val);
6954 continue;
6955
6956 case DTRACEACT_COMMIT:
6957 ASSERT(!committed);
6958
6959 /*
6960 * We need to commit our buffer state.
6961 */
6962 if (ecb->dte_size)
6963 buf->dtb_offset = offs + ecb->dte_size;
6964 buf = &state->dts_buffer[cpuid];
6965 dtrace_speculation_commit(state, cpuid, val);
6966 committed = 1;
6967 continue;
6968
6969 case DTRACEACT_DISCARD:
6970 dtrace_speculation_discard(state, cpuid, val);
6971 continue;
6972
6973 case DTRACEACT_DIFEXPR:
6974 case DTRACEACT_LIBACT:
6975 case DTRACEACT_PRINTF:
6976 case DTRACEACT_PRINTA:
6977 case DTRACEACT_SYSTEM:
6978 case DTRACEACT_FREOPEN:
6979 case DTRACEACT_TRACEMEM:
6980 break;
6981
6982 case DTRACEACT_TRACEMEM_DYNSIZE:
6983 tracememsize = val;
6984 break;
6985
6986 case DTRACEACT_SYM:
6987 case DTRACEACT_MOD:
6988 if (!dtrace_priv_kernel(state))
6989 continue;
6990 break;
6991
6992 case DTRACEACT_USYM:
6993 case DTRACEACT_UMOD:
6994 case DTRACEACT_UADDR: {
6995 struct pid *pid = curthread->t_procp->p_pidp;
6996
6997 if (!dtrace_priv_proc(state, &mstate))
6998 continue;
6999
7000 DTRACE_STORE(uint64_t, tomax,
7001 valoffs, (uint64_t)pid->pid_id);
7002 DTRACE_STORE(uint64_t, tomax,
7003 valoffs + sizeof (uint64_t), val);
7004
7005 continue;
7006 }
7007
7008 case DTRACEACT_EXIT: {
7009 /*
7010 * For the exit action, we are going to attempt
7011 * to atomically set our activity to be
7012 * draining. If this fails (either because
7013 * another CPU has beat us to the exit action,
7014 * or because our current activity is something
7015 * other than ACTIVE or WARMUP), we will
7016 * continue. This assures that the exit action
7017 * can be successfully recorded at most once
7018 * when we're in the ACTIVE state. If we're
7019 * encountering the exit() action while in
7020 * COOLDOWN, however, we want to honor the new
7021 * status code. (We know that we're the only
7022 * thread in COOLDOWN, so there is no race.)
7023 */
7024 void *activity = &state->dts_activity;
7025 dtrace_activity_t current = state->dts_activity;
7026
7027 if (current == DTRACE_ACTIVITY_COOLDOWN)
7028 break;
7029
7030 if (current != DTRACE_ACTIVITY_WARMUP)
7031 current = DTRACE_ACTIVITY_ACTIVE;
7032
7033 if (dtrace_cas32(activity, current,
7034 DTRACE_ACTIVITY_DRAINING) != current) {
7035 *flags |= CPU_DTRACE_DROP;
7036 continue;
7037 }
7038
7039 break;
7040 }
7041
7042 default:
7043 ASSERT(0);
7044 }
7045
7046 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
7047 uintptr_t end = valoffs + size;
7048
7049 if (tracememsize != 0 &&
7050 valoffs + tracememsize < end) {
7051 end = valoffs + tracememsize;
7052 tracememsize = 0;
7053 }
7054
7055 if (!dtrace_vcanload((void *)(uintptr_t)val,
7056 &dp->dtdo_rtype, &mstate, vstate))
7057 continue;
7058
7059 /*
7060 * If this is a string, we're going to only
7061 * load until we find the zero byte -- after
7062 * which we'll store zero bytes.
7063 */
7064 if (dp->dtdo_rtype.dtdt_kind ==
7065 DIF_TYPE_STRING) {
7066 char c = '\0' + 1;
7067 int intuple = act->dta_intuple;
7068 size_t s;
7069
7070 for (s = 0; s < size; s++) {
7071 if (c != '\0')
7072 c = dtrace_load8(val++);
7073
7074 DTRACE_STORE(uint8_t, tomax,
7075 valoffs++, c);
7076
7077 if (c == '\0' && intuple)
7078 break;
7079 }
7080
7081 continue;
7082 }
7083
7084 while (valoffs < end) {
7085 DTRACE_STORE(uint8_t, tomax, valoffs++,
7086 dtrace_load8(val++));
7087 }
7088
7089 continue;
7090 }
7091
7092 switch (size) {
7093 case 0:
7094 break;
7095
7096 case sizeof (uint8_t):
7097 DTRACE_STORE(uint8_t, tomax, valoffs, val);
7098 break;
7099 case sizeof (uint16_t):
7100 DTRACE_STORE(uint16_t, tomax, valoffs, val);
7101 break;
7102 case sizeof (uint32_t):
7103 DTRACE_STORE(uint32_t, tomax, valoffs, val);
7104 break;
7105 case sizeof (uint64_t):
7106 DTRACE_STORE(uint64_t, tomax, valoffs, val);
7107 break;
7108 default:
7109 /*
7110 * Any other size should have been returned by
7111 * reference, not by value.
7112 */
7113 ASSERT(0);
7114 break;
7115 }
7116 }
7117
7118 if (*flags & CPU_DTRACE_DROP)
7119 continue;
7120
7121 if (*flags & CPU_DTRACE_FAULT) {
7122 int ndx;
7123 dtrace_action_t *err;
7124
7125 buf->dtb_errors++;
7126
7127 if (probe->dtpr_id == dtrace_probeid_error) {
7128 /*
7129 * There's nothing we can do -- we had an
7130 * error on the error probe. We bump an
7131 * error counter to at least indicate that
7132 * this condition happened.
7133 */
7134 dtrace_error(&state->dts_dblerrors);
7135 continue;
7136 }
7137
7138 if (vtime) {
7139 /*
7140 * Before recursing on dtrace_probe(), we
7141 * need to explicitly clear out our start
7142 * time to prevent it from being accumulated
7143 * into t_dtrace_vtime.
7144 */
7145 curthread->t_dtrace_start = 0;
7146 }
7147
7148 /*
7149 * Iterate over the actions to figure out which action
7150 * we were processing when we experienced the error.
7151 * Note that act points _past_ the faulting action; if
7152 * act is ecb->dte_action, the fault was in the
7153 * predicate, if it's ecb->dte_action->dta_next it's
7154 * in action #1, and so on.
7155 */
7156 for (err = ecb->dte_action, ndx = 0;
7157 err != act; err = err->dta_next, ndx++)
7158 continue;
7159
7160 dtrace_probe_error(state, ecb->dte_epid, ndx,
7161 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7162 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7163 cpu_core[cpuid].cpuc_dtrace_illval);
7164
7165 continue;
7166 }
7167
7168 if (!committed)
7169 buf->dtb_offset = offs + ecb->dte_size;
7170 }
7171
7172 end = dtrace_gethrtime();
7173 if (vtime)
7174 curthread->t_dtrace_start = end;
7175
7176 CPU->cpu_dtrace_nsec += end - now;
7177
7178 dtrace_interrupt_enable(cookie);
7179 }
7180
7181 /*
7182 * DTrace Probe Hashing Functions
7183 *
7184 * The functions in this section (and indeed, the functions in remaining
7185 * sections) are not _called_ from probe context. (Any exceptions to this are
7186 * marked with a "Note:".) Rather, they are called from elsewhere in the
7187 * DTrace framework to look-up probes in, add probes to and remove probes from
7188 * the DTrace probe hashes. (Each probe is hashed by each element of the
7189 * probe tuple -- allowing for fast lookups, regardless of what was
7190 * specified.)
7191 */
7192 static uint_t
7193 dtrace_hash_str(char *p)
7194 {
7195 unsigned int g;
7196 uint_t hval = 0;
7197
7198 while (*p) {
7199 hval = (hval << 4) + *p++;
7200 if ((g = (hval & 0xf0000000)) != 0)
7201 hval ^= g >> 24;
7202 hval &= ~g;
7203 }
7204 return (hval);
7205 }
7206
7207 static dtrace_hash_t *
7208 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
7209 {
7210 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
7211
7212 hash->dth_stroffs = stroffs;
7213 hash->dth_nextoffs = nextoffs;
7214 hash->dth_prevoffs = prevoffs;
7215
7216 hash->dth_size = 1;
7217 hash->dth_mask = hash->dth_size - 1;
7218
7219 hash->dth_tab = kmem_zalloc(hash->dth_size *
7220 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
7221
7222 return (hash);
7223 }
7224
7225 static void
7226 dtrace_hash_destroy(dtrace_hash_t *hash)
7227 {
7228 #ifdef DEBUG
7229 int i;
7230
7231 for (i = 0; i < hash->dth_size; i++)
7232 ASSERT(hash->dth_tab[i] == NULL);
7233 #endif
7234
7235 kmem_free(hash->dth_tab,
7236 hash->dth_size * sizeof (dtrace_hashbucket_t *));
7237 kmem_free(hash, sizeof (dtrace_hash_t));
7238 }
7239
7240 static void
7241 dtrace_hash_resize(dtrace_hash_t *hash)
7242 {
7243 int size = hash->dth_size, i, ndx;
7244 int new_size = hash->dth_size << 1;
7245 int new_mask = new_size - 1;
7246 dtrace_hashbucket_t **new_tab, *bucket, *next;
7247
7248 ASSERT((new_size & new_mask) == 0);
7249
7250 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
7251
7252 for (i = 0; i < size; i++) {
7253 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
7254 dtrace_probe_t *probe = bucket->dthb_chain;
7255
7256 ASSERT(probe != NULL);
7257 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
7258
7259 next = bucket->dthb_next;
7260 bucket->dthb_next = new_tab[ndx];
7261 new_tab[ndx] = bucket;
7262 }
7263 }
7264
7265 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
7266 hash->dth_tab = new_tab;
7267 hash->dth_size = new_size;
7268 hash->dth_mask = new_mask;
7269 }
7270
7271 static void
7272 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
7273 {
7274 int hashval = DTRACE_HASHSTR(hash, new);
7275 int ndx = hashval & hash->dth_mask;
7276 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7277 dtrace_probe_t **nextp, **prevp;
7278
7279 for (; bucket != NULL; bucket = bucket->dthb_next) {
7280 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
7281 goto add;
7282 }
7283
7284 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
7285 dtrace_hash_resize(hash);
7286 dtrace_hash_add(hash, new);
7287 return;
7288 }
7289
7290 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
7291 bucket->dthb_next = hash->dth_tab[ndx];
7292 hash->dth_tab[ndx] = bucket;
7293 hash->dth_nbuckets++;
7294
7295 add:
7296 nextp = DTRACE_HASHNEXT(hash, new);
7297 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
7298 *nextp = bucket->dthb_chain;
7299
7300 if (bucket->dthb_chain != NULL) {
7301 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
7302 ASSERT(*prevp == NULL);
7303 *prevp = new;
7304 }
7305
7306 bucket->dthb_chain = new;
7307 bucket->dthb_len++;
7308 }
7309
7310 static dtrace_probe_t *
7311 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
7312 {
7313 int hashval = DTRACE_HASHSTR(hash, template);
7314 int ndx = hashval & hash->dth_mask;
7315 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7316
7317 for (; bucket != NULL; bucket = bucket->dthb_next) {
7318 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7319 return (bucket->dthb_chain);
7320 }
7321
7322 return (NULL);
7323 }
7324
7325 static int
7326 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
7327 {
7328 int hashval = DTRACE_HASHSTR(hash, template);
7329 int ndx = hashval & hash->dth_mask;
7330 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7331
7332 for (; bucket != NULL; bucket = bucket->dthb_next) {
7333 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7334 return (bucket->dthb_len);
7335 }
7336
7337 return (NULL);
7338 }
7339
7340 static void
7341 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7342 {
7343 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7344 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7345
7346 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7347 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7348
7349 /*
7350 * Find the bucket that we're removing this probe from.
7351 */
7352 for (; bucket != NULL; bucket = bucket->dthb_next) {
7353 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7354 break;
7355 }
7356
7357 ASSERT(bucket != NULL);
7358
7359 if (*prevp == NULL) {
7360 if (*nextp == NULL) {
7361 /*
7362 * The removed probe was the only probe on this
7363 * bucket; we need to remove the bucket.
7364 */
7365 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7366
7367 ASSERT(bucket->dthb_chain == probe);
7368 ASSERT(b != NULL);
7369
7370 if (b == bucket) {
7371 hash->dth_tab[ndx] = bucket->dthb_next;
7372 } else {
7373 while (b->dthb_next != bucket)
7374 b = b->dthb_next;
7375 b->dthb_next = bucket->dthb_next;
7376 }
7377
7378 ASSERT(hash->dth_nbuckets > 0);
7379 hash->dth_nbuckets--;
7380 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7381 return;
7382 }
7383
7384 bucket->dthb_chain = *nextp;
7385 } else {
7386 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7387 }
7388
7389 if (*nextp != NULL)
7390 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7391 }
7392
7393 /*
7394 * DTrace Utility Functions
7395 *
7396 * These are random utility functions that are _not_ called from probe context.
7397 */
7398 static int
7399 dtrace_badattr(const dtrace_attribute_t *a)
7400 {
7401 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7402 a->dtat_data > DTRACE_STABILITY_MAX ||
7403 a->dtat_class > DTRACE_CLASS_MAX);
7404 }
7405
7406 /*
7407 * Return a duplicate copy of a string. If the specified string is NULL,
7408 * this function returns a zero-length string.
7409 */
7410 static char *
7411 dtrace_strdup(const char *str)
7412 {
7413 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7414
7415 if (str != NULL)
7416 (void) strcpy(new, str);
7417
7418 return (new);
7419 }
7420
7421 #define DTRACE_ISALPHA(c) \
7422 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7423
7424 static int
7425 dtrace_badname(const char *s)
7426 {
7427 char c;
7428
7429 if (s == NULL || (c = *s++) == '\0')
7430 return (0);
7431
7432 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7433 return (1);
7434
7435 while ((c = *s++) != '\0') {
7436 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7437 c != '-' && c != '_' && c != '.' && c != '`')
7438 return (1);
7439 }
7440
7441 return (0);
7442 }
7443
7444 static void
7445 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7446 {
7447 uint32_t priv;
7448
7449 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7450 /*
7451 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7452 */
7453 priv = DTRACE_PRIV_ALL;
7454 } else {
7455 *uidp = crgetuid(cr);
7456 *zoneidp = crgetzoneid(cr);
7457
7458 priv = 0;
7459 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7460 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7461 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7462 priv |= DTRACE_PRIV_USER;
7463 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7464 priv |= DTRACE_PRIV_PROC;
7465 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7466 priv |= DTRACE_PRIV_OWNER;
7467 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7468 priv |= DTRACE_PRIV_ZONEOWNER;
7469 }
7470
7471 *privp = priv;
7472 }
7473
7474 #ifdef DTRACE_ERRDEBUG
7475 static void
7476 dtrace_errdebug(const char *str)
7477 {
7478 int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
7479 int occupied = 0;
7480
7481 mutex_enter(&dtrace_errlock);
7482 dtrace_errlast = str;
7483 dtrace_errthread = curthread;
7484
7485 while (occupied++ < DTRACE_ERRHASHSZ) {
7486 if (dtrace_errhash[hval].dter_msg == str) {
7487 dtrace_errhash[hval].dter_count++;
7488 goto out;
7489 }
7490
7491 if (dtrace_errhash[hval].dter_msg != NULL) {
7492 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7493 continue;
7494 }
7495
7496 dtrace_errhash[hval].dter_msg = str;
7497 dtrace_errhash[hval].dter_count = 1;
7498 goto out;
7499 }
7500
7501 panic("dtrace: undersized error hash");
7502 out:
7503 mutex_exit(&dtrace_errlock);
7504 }
7505 #endif
7506
7507 /*
7508 * DTrace Matching Functions
7509 *
7510 * These functions are used to match groups of probes, given some elements of
7511 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7512 */
7513 static int
7514 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7515 zoneid_t zoneid)
7516 {
7517 if (priv != DTRACE_PRIV_ALL) {
7518 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7519 uint32_t match = priv & ppriv;
7520
7521 /*
7522 * No PRIV_DTRACE_* privileges...
7523 */
7524 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7525 DTRACE_PRIV_KERNEL)) == 0)
7526 return (0);
7527
7528 /*
7529 * No matching bits, but there were bits to match...
7530 */
7531 if (match == 0 && ppriv != 0)
7532 return (0);
7533
7534 /*
7535 * Need to have permissions to the process, but don't...
7536 */
7537 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7538 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7539 return (0);
7540 }
7541
7542 /*
7543 * Need to be in the same zone unless we possess the
7544 * privilege to examine all zones.
7545 */
7546 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7547 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7548 return (0);
7549 }
7550 }
7551
7552 return (1);
7553 }
7554
7555 /*
7556 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7557 * consists of input pattern strings and an ops-vector to evaluate them.
7558 * This function returns >0 for match, 0 for no match, and <0 for error.
7559 */
7560 static int
7561 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7562 uint32_t priv, uid_t uid, zoneid_t zoneid)
7563 {
7564 dtrace_provider_t *pvp = prp->dtpr_provider;
7565 int rv;
7566
7567 if (pvp->dtpv_defunct)
7568 return (0);
7569
7570 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7571 return (rv);
7572
7573 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7574 return (rv);
7575
7576 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7577 return (rv);
7578
7579 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7580 return (rv);
7581
7582 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7583 return (0);
7584
7585 return (rv);
7586 }
7587
7588 /*
7589 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7590 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7591 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7592 * In addition, all of the recursion cases except for '*' matching have been
7593 * unwound. For '*', we still implement recursive evaluation, but a depth
7594 * counter is maintained and matching is aborted if we recurse too deep.
7595 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7596 */
7597 static int
7598 dtrace_match_glob(const char *s, const char *p, int depth)
7599 {
7600 const char *olds;
7601 char s1, c;
7602 int gs;
7603
7604 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7605 return (-1);
7606
7607 if (s == NULL)
7608 s = ""; /* treat NULL as empty string */
7609
7610 top:
7611 olds = s;
7612 s1 = *s++;
7613
7614 if (p == NULL)
7615 return (0);
7616
7617 if ((c = *p++) == '\0')
7618 return (s1 == '\0');
7619
7620 switch (c) {
7621 case '[': {
7622 int ok = 0, notflag = 0;
7623 char lc = '\0';
7624
7625 if (s1 == '\0')
7626 return (0);
7627
7628 if (*p == '!') {
7629 notflag = 1;
7630 p++;
7631 }
7632
7633 if ((c = *p++) == '\0')
7634 return (0);
7635
7636 do {
7637 if (c == '-' && lc != '\0' && *p != ']') {
7638 if ((c = *p++) == '\0')
7639 return (0);
7640 if (c == '\\' && (c = *p++) == '\0')
7641 return (0);
7642
7643 if (notflag) {
7644 if (s1 < lc || s1 > c)
7645 ok++;
7646 else
7647 return (0);
7648 } else if (lc <= s1 && s1 <= c)
7649 ok++;
7650
7651 } else if (c == '\\' && (c = *p++) == '\0')
7652 return (0);
7653
7654 lc = c; /* save left-hand 'c' for next iteration */
7655
7656 if (notflag) {
7657 if (s1 != c)
7658 ok++;
7659 else
7660 return (0);
7661 } else if (s1 == c)
7662 ok++;
7663
7664 if ((c = *p++) == '\0')
7665 return (0);
7666
7667 } while (c != ']');
7668
7669 if (ok)
7670 goto top;
7671
7672 return (0);
7673 }
7674
7675 case '\\':
7676 if ((c = *p++) == '\0')
7677 return (0);
7678 /*FALLTHRU*/
7679
7680 default:
7681 if (c != s1)
7682 return (0);
7683 /*FALLTHRU*/
7684
7685 case '?':
7686 if (s1 != '\0')
7687 goto top;
7688 return (0);
7689
7690 case '*':
7691 while (*p == '*')
7692 p++; /* consecutive *'s are identical to a single one */
7693
7694 if (*p == '\0')
7695 return (1);
7696
7697 for (s = olds; *s != '\0'; s++) {
7698 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7699 return (gs);
7700 }
7701
7702 return (0);
7703 }
7704 }
7705
7706 /*ARGSUSED*/
7707 static int
7708 dtrace_match_string(const char *s, const char *p, int depth)
7709 {
7710 return (s != NULL && strcmp(s, p) == 0);
7711 }
7712
7713 /*ARGSUSED*/
7714 static int
7715 dtrace_match_nul(const char *s, const char *p, int depth)
7716 {
7717 return (1); /* always match the empty pattern */
7718 }
7719
7720 /*ARGSUSED*/
7721 static int
7722 dtrace_match_nonzero(const char *s, const char *p, int depth)
7723 {
7724 return (s != NULL && s[0] != '\0');
7725 }
7726
7727 static int
7728 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7729 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7730 {
7731 dtrace_probe_t template, *probe;
7732 dtrace_hash_t *hash = NULL;
7733 int len, rc, best = INT_MAX, nmatched = 0;
7734 dtrace_id_t i;
7735
7736 ASSERT(MUTEX_HELD(&dtrace_lock));
7737
7738 /*
7739 * If the probe ID is specified in the key, just lookup by ID and
7740 * invoke the match callback once if a matching probe is found.
7741 */
7742 if (pkp->dtpk_id != DTRACE_IDNONE) {
7743 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7744 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7745 if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
7746 return (DTRACE_MATCH_FAIL);
7747 nmatched++;
7748 }
7749 return (nmatched);
7750 }
7751
7752 template.dtpr_mod = (char *)pkp->dtpk_mod;
7753 template.dtpr_func = (char *)pkp->dtpk_func;
7754 template.dtpr_name = (char *)pkp->dtpk_name;
7755
7756 /*
7757 * We want to find the most distinct of the module name, function
7758 * name, and name. So for each one that is not a glob pattern or
7759 * empty string, we perform a lookup in the corresponding hash and
7760 * use the hash table with the fewest collisions to do our search.
7761 */
7762 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7763 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7764 best = len;
7765 hash = dtrace_bymod;
7766 }
7767
7768 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7769 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7770 best = len;
7771 hash = dtrace_byfunc;
7772 }
7773
7774 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7775 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7776 best = len;
7777 hash = dtrace_byname;
7778 }
7779
7780 /*
7781 * If we did not select a hash table, iterate over every probe and
7782 * invoke our callback for each one that matches our input probe key.
7783 */
7784 if (hash == NULL) {
7785 for (i = 0; i < dtrace_nprobes; i++) {
7786 if ((probe = dtrace_probes[i]) == NULL ||
7787 dtrace_match_probe(probe, pkp, priv, uid,
7788 zoneid) <= 0)
7789 continue;
7790
7791 nmatched++;
7792
7793 if ((rc = (*matched)(probe, arg)) !=
7794 DTRACE_MATCH_NEXT) {
7795 if (rc == DTRACE_MATCH_FAIL)
7796 return (DTRACE_MATCH_FAIL);
7797 break;
7798 }
7799 }
7800
7801 return (nmatched);
7802 }
7803
7804 /*
7805 * If we selected a hash table, iterate over each probe of the same key
7806 * name and invoke the callback for every probe that matches the other
7807 * attributes of our input probe key.
7808 */
7809 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7810 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7811
7812 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7813 continue;
7814
7815 nmatched++;
7816
7817 if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
7818 if (rc == DTRACE_MATCH_FAIL)
7819 return (DTRACE_MATCH_FAIL);
7820 break;
7821 }
7822 }
7823
7824 return (nmatched);
7825 }
7826
7827 /*
7828 * Return the function pointer dtrace_probecmp() should use to compare the
7829 * specified pattern with a string. For NULL or empty patterns, we select
7830 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7831 * For non-empty non-glob strings, we use dtrace_match_string().
7832 */
7833 static dtrace_probekey_f *
7834 dtrace_probekey_func(const char *p)
7835 {
7836 char c;
7837
7838 if (p == NULL || *p == '\0')
7839 return (&dtrace_match_nul);
7840
7841 while ((c = *p++) != '\0') {
7842 if (c == '[' || c == '?' || c == '*' || c == '\\')
7843 return (&dtrace_match_glob);
7844 }
7845
7846 return (&dtrace_match_string);
7847 }
7848
7849 /*
7850 * Build a probe comparison key for use with dtrace_match_probe() from the
7851 * given probe description. By convention, a null key only matches anchored
7852 * probes: if each field is the empty string, reset dtpk_fmatch to
7853 * dtrace_match_nonzero().
7854 */
7855 static void
7856 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7857 {
7858 pkp->dtpk_prov = pdp->dtpd_provider;
7859 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7860
7861 pkp->dtpk_mod = pdp->dtpd_mod;
7862 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7863
7864 pkp->dtpk_func = pdp->dtpd_func;
7865 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7866
7867 pkp->dtpk_name = pdp->dtpd_name;
7868 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7869
7870 pkp->dtpk_id = pdp->dtpd_id;
7871
7872 if (pkp->dtpk_id == DTRACE_IDNONE &&
7873 pkp->dtpk_pmatch == &dtrace_match_nul &&
7874 pkp->dtpk_mmatch == &dtrace_match_nul &&
7875 pkp->dtpk_fmatch == &dtrace_match_nul &&
7876 pkp->dtpk_nmatch == &dtrace_match_nul)
7877 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7878 }
7879
7880 /*
7881 * DTrace Provider-to-Framework API Functions
7882 *
7883 * These functions implement much of the Provider-to-Framework API, as
7884 * described in <sys/dtrace.h>. The parts of the API not in this section are
7885 * the functions in the API for probe management (found below), and
7886 * dtrace_probe() itself (found above).
7887 */
7888
7889 /*
7890 * Register the calling provider with the DTrace framework. This should
7891 * generally be called by DTrace providers in their attach(9E) entry point.
7892 */
7893 int
7894 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7895 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7896 {
7897 dtrace_provider_t *provider;
7898
7899 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7900 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7901 "arguments", name ? name : "<NULL>");
7902 return (EINVAL);
7903 }
7904
7905 if (name[0] == '\0' || dtrace_badname(name)) {
7906 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7907 "provider name", name);
7908 return (EINVAL);
7909 }
7910
7911 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7912 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7913 pops->dtps_destroy == NULL ||
7914 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7915 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7916 "provider ops", name);
7917 return (EINVAL);
7918 }
7919
7920 if (dtrace_badattr(&pap->dtpa_provider) ||
7921 dtrace_badattr(&pap->dtpa_mod) ||
7922 dtrace_badattr(&pap->dtpa_func) ||
7923 dtrace_badattr(&pap->dtpa_name) ||
7924 dtrace_badattr(&pap->dtpa_args)) {
7925 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7926 "provider attributes", name);
7927 return (EINVAL);
7928 }
7929
7930 if (priv & ~DTRACE_PRIV_ALL) {
7931 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7932 "privilege attributes", name);
7933 return (EINVAL);
7934 }
7935
7936 if ((priv & DTRACE_PRIV_KERNEL) &&
7937 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7938 pops->dtps_mode == NULL) {
7939 cmn_err(CE_WARN, "failed to register provider '%s': need "
7940 "dtps_mode() op for given privilege attributes", name);
7941 return (EINVAL);
7942 }
7943
7944 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7945 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7946 (void) strcpy(provider->dtpv_name, name);
7947
7948 provider->dtpv_attr = *pap;
7949 provider->dtpv_priv.dtpp_flags = priv;
7950 if (cr != NULL) {
7951 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7952 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7953 }
7954 provider->dtpv_pops = *pops;
7955
7956 if (pops->dtps_provide == NULL) {
7957 ASSERT(pops->dtps_provide_module != NULL);
7958 provider->dtpv_pops.dtps_provide =
7959 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
7960 }
7961
7962 if (pops->dtps_provide_module == NULL) {
7963 ASSERT(pops->dtps_provide != NULL);
7964 provider->dtpv_pops.dtps_provide_module =
7965 (void (*)(void *, struct modctl *))dtrace_nullop;
7966 }
7967
7968 if (pops->dtps_suspend == NULL) {
7969 ASSERT(pops->dtps_resume == NULL);
7970 provider->dtpv_pops.dtps_suspend =
7971 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7972 provider->dtpv_pops.dtps_resume =
7973 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7974 }
7975
7976 provider->dtpv_arg = arg;
7977 *idp = (dtrace_provider_id_t)provider;
7978
7979 if (pops == &dtrace_provider_ops) {
7980 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7981 ASSERT(MUTEX_HELD(&dtrace_lock));
7982 ASSERT(dtrace_anon.dta_enabling == NULL);
7983
7984 /*
7985 * We make sure that the DTrace provider is at the head of
7986 * the provider chain.
7987 */
7988 provider->dtpv_next = dtrace_provider;
7989 dtrace_provider = provider;
7990 return (0);
7991 }
7992
7993 mutex_enter(&dtrace_provider_lock);
7994 mutex_enter(&dtrace_lock);
7995
7996 /*
7997 * If there is at least one provider registered, we'll add this
7998 * provider after the first provider.
7999 */
8000 if (dtrace_provider != NULL) {
8001 provider->dtpv_next = dtrace_provider->dtpv_next;
8002 dtrace_provider->dtpv_next = provider;
8003 } else {
8004 dtrace_provider = provider;
8005 }
8006
8007 if (dtrace_retained != NULL) {
8008 dtrace_enabling_provide(provider);
8009
8010 /*
8011 * Now we need to call dtrace_enabling_matchall() -- which
8012 * will acquire cpu_lock and dtrace_lock. We therefore need
8013 * to drop all of our locks before calling into it...
8014 */
8015 mutex_exit(&dtrace_lock);
8016 mutex_exit(&dtrace_provider_lock);
8017 dtrace_enabling_matchall();
8018
8019 return (0);
8020 }
8021
8022 mutex_exit(&dtrace_lock);
8023 mutex_exit(&dtrace_provider_lock);
8024
8025 return (0);
8026 }
8027
8028 /*
8029 * Unregister the specified provider from the DTrace framework. This should
8030 * generally be called by DTrace providers in their detach(9E) entry point.
8031 */
8032 int
8033 dtrace_unregister(dtrace_provider_id_t id)
8034 {
8035 dtrace_provider_t *old = (dtrace_provider_t *)id;
8036 dtrace_provider_t *prev = NULL;
8037 int i, self = 0, noreap = 0;
8038 dtrace_probe_t *probe, *first = NULL;
8039
8040 if (old->dtpv_pops.dtps_enable ==
8041 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
8042 /*
8043 * If DTrace itself is the provider, we're called with locks
8044 * already held.
8045 */
8046 ASSERT(old == dtrace_provider);
8047 ASSERT(dtrace_devi != NULL);
8048 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8049 ASSERT(MUTEX_HELD(&dtrace_lock));
8050 self = 1;
8051
8052 if (dtrace_provider->dtpv_next != NULL) {
8053 /*
8054 * There's another provider here; return failure.
8055 */
8056 return (EBUSY);
8057 }
8058 } else {
8059 mutex_enter(&dtrace_provider_lock);
8060 mutex_enter(&mod_lock);
8061 mutex_enter(&dtrace_lock);
8062 }
8063
8064 /*
8065 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8066 * probes, we refuse to let providers slither away, unless this
8067 * provider has already been explicitly invalidated.
8068 */
8069 if (!old->dtpv_defunct &&
8070 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8071 dtrace_anon.dta_state->dts_necbs > 0))) {
8072 if (!self) {
8073 mutex_exit(&dtrace_lock);
8074 mutex_exit(&mod_lock);
8075 mutex_exit(&dtrace_provider_lock);
8076 }
8077 return (EBUSY);
8078 }
8079
8080 /*
8081 * Attempt to destroy the probes associated with this provider.
8082 */
8083 for (i = 0; i < dtrace_nprobes; i++) {
8084 if ((probe = dtrace_probes[i]) == NULL)
8085 continue;
8086
8087 if (probe->dtpr_provider != old)
8088 continue;
8089
8090 if (probe->dtpr_ecb == NULL)
8091 continue;
8092
8093 /*
8094 * If we are trying to unregister a defunct provider, and the
8095 * provider was made defunct within the interval dictated by
8096 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8097 * attempt to reap our enablings. To denote that the provider
8098 * should reattempt to unregister itself at some point in the
8099 * future, we will return a differentiable error code (EAGAIN
8100 * instead of EBUSY) in this case.
8101 */
8102 if (dtrace_gethrtime() - old->dtpv_defunct >
8103 dtrace_unregister_defunct_reap)
8104 noreap = 1;
8105
8106 if (!self) {
8107 mutex_exit(&dtrace_lock);
8108 mutex_exit(&mod_lock);
8109 mutex_exit(&dtrace_provider_lock);
8110 }
8111
8112 if (noreap)
8113 return (EBUSY);
8114
8115 (void) taskq_dispatch(dtrace_taskq,
8116 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8117
8118 return (EAGAIN);
8119 }
8120
8121 /*
8122 * All of the probes for this provider are disabled; we can safely
8123 * remove all of them from their hash chains and from the probe array.
8124 */
8125 for (i = 0; i < dtrace_nprobes; i++) {
8126 if ((probe = dtrace_probes[i]) == NULL)
8127 continue;
8128
8129 if (probe->dtpr_provider != old)
8130 continue;
8131
8132 dtrace_probes[i] = NULL;
8133
8134 dtrace_hash_remove(dtrace_bymod, probe);
8135 dtrace_hash_remove(dtrace_byfunc, probe);
8136 dtrace_hash_remove(dtrace_byname, probe);
8137
8138 if (first == NULL) {
8139 first = probe;
8140 probe->dtpr_nextmod = NULL;
8141 } else {
8142 probe->dtpr_nextmod = first;
8143 first = probe;
8144 }
8145 }
8146
8147 /*
8148 * The provider's probes have been removed from the hash chains and
8149 * from the probe array. Now issue a dtrace_sync() to be sure that
8150 * everyone has cleared out from any probe array processing.
8151 */
8152 dtrace_sync();
8153
8154 for (probe = first; probe != NULL; probe = first) {
8155 first = probe->dtpr_nextmod;
8156
8157 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8158 probe->dtpr_arg);
8159 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8160 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8161 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8162 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8163 kmem_free(probe, sizeof (dtrace_probe_t));
8164 }
8165
8166 if ((prev = dtrace_provider) == old) {
8167 ASSERT(self || dtrace_devi == NULL);
8168 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8169 dtrace_provider = old->dtpv_next;
8170 } else {
8171 while (prev != NULL && prev->dtpv_next != old)
8172 prev = prev->dtpv_next;
8173
8174 if (prev == NULL) {
8175 panic("attempt to unregister non-existent "
8176 "dtrace provider %p\n", (void *)id);
8177 }
8178
8179 prev->dtpv_next = old->dtpv_next;
8180 }
8181
8182 if (!self) {
8183 mutex_exit(&dtrace_lock);
8184 mutex_exit(&mod_lock);
8185 mutex_exit(&dtrace_provider_lock);
8186 }
8187
8188 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
8189 kmem_free(old, sizeof (dtrace_provider_t));
8190
8191 return (0);
8192 }
8193
8194 /*
8195 * Invalidate the specified provider. All subsequent probe lookups for the
8196 * specified provider will fail, but its probes will not be removed.
8197 */
8198 void
8199 dtrace_invalidate(dtrace_provider_id_t id)
8200 {
8201 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
8202
8203 ASSERT(pvp->dtpv_pops.dtps_enable !=
8204 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8205
8206 mutex_enter(&dtrace_provider_lock);
8207 mutex_enter(&dtrace_lock);
8208
8209 pvp->dtpv_defunct = dtrace_gethrtime();
8210
8211 mutex_exit(&dtrace_lock);
8212 mutex_exit(&dtrace_provider_lock);
8213 }
8214
8215 /*
8216 * Indicate whether or not DTrace has attached.
8217 */
8218 int
8219 dtrace_attached(void)
8220 {
8221 /*
8222 * dtrace_provider will be non-NULL iff the DTrace driver has
8223 * attached. (It's non-NULL because DTrace is always itself a
8224 * provider.)
8225 */
8226 return (dtrace_provider != NULL);
8227 }
8228
8229 /*
8230 * Remove all the unenabled probes for the given provider. This function is
8231 * not unlike dtrace_unregister(), except that it doesn't remove the provider
8232 * -- just as many of its associated probes as it can.
8233 */
8234 int
8235 dtrace_condense(dtrace_provider_id_t id)
8236 {
8237 dtrace_provider_t *prov = (dtrace_provider_t *)id;
8238 int i;
8239 dtrace_probe_t *probe;
8240
8241 /*
8242 * Make sure this isn't the dtrace provider itself.
8243 */
8244 ASSERT(prov->dtpv_pops.dtps_enable !=
8245 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8246
8247 mutex_enter(&dtrace_provider_lock);
8248 mutex_enter(&dtrace_lock);
8249
8250 /*
8251 * Attempt to destroy the probes associated with this provider.
8252 */
8253 for (i = 0; i < dtrace_nprobes; i++) {
8254 if ((probe = dtrace_probes[i]) == NULL)
8255 continue;
8256
8257 if (probe->dtpr_provider != prov)
8258 continue;
8259
8260 if (probe->dtpr_ecb != NULL)
8261 continue;
8262
8263 dtrace_probes[i] = NULL;
8264
8265 dtrace_hash_remove(dtrace_bymod, probe);
8266 dtrace_hash_remove(dtrace_byfunc, probe);
8267 dtrace_hash_remove(dtrace_byname, probe);
8268
8269 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
8270 probe->dtpr_arg);
8271 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8272 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8273 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8274 kmem_free(probe, sizeof (dtrace_probe_t));
8275 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
8276 }
8277
8278 mutex_exit(&dtrace_lock);
8279 mutex_exit(&dtrace_provider_lock);
8280
8281 return (0);
8282 }
8283
8284 /*
8285 * DTrace Probe Management Functions
8286 *
8287 * The functions in this section perform the DTrace probe management,
8288 * including functions to create probes, look-up probes, and call into the
8289 * providers to request that probes be provided. Some of these functions are
8290 * in the Provider-to-Framework API; these functions can be identified by the
8291 * fact that they are not declared "static".
8292 */
8293
8294 /*
8295 * Create a probe with the specified module name, function name, and name.
8296 */
8297 dtrace_id_t
8298 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
8299 const char *func, const char *name, int aframes, void *arg)
8300 {
8301 dtrace_probe_t *probe, **probes;
8302 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
8303 dtrace_id_t id;
8304
8305 if (provider == dtrace_provider) {
8306 ASSERT(MUTEX_HELD(&dtrace_lock));
8307 } else {
8308 mutex_enter(&dtrace_lock);
8309 }
8310
8311 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
8312 VM_BESTFIT | VM_SLEEP);
8313 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8314
8315 probe->dtpr_id = id;
8316 probe->dtpr_gen = dtrace_probegen++;
8317 probe->dtpr_mod = dtrace_strdup(mod);
8318 probe->dtpr_func = dtrace_strdup(func);
8319 probe->dtpr_name = dtrace_strdup(name);
8320 probe->dtpr_arg = arg;
8321 probe->dtpr_aframes = aframes;
8322 probe->dtpr_provider = provider;
8323
8324 dtrace_hash_add(dtrace_bymod, probe);
8325 dtrace_hash_add(dtrace_byfunc, probe);
8326 dtrace_hash_add(dtrace_byname, probe);
8327
8328 if (id - 1 >= dtrace_nprobes) {
8329 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8330 size_t nsize = osize << 1;
8331
8332 if (nsize == 0) {
8333 ASSERT(osize == 0);
8334 ASSERT(dtrace_probes == NULL);
8335 nsize = sizeof (dtrace_probe_t *);
8336 }
8337
8338 probes = kmem_zalloc(nsize, KM_SLEEP);
8339
8340 if (dtrace_probes == NULL) {
8341 ASSERT(osize == 0);
8342 dtrace_probes = probes;
8343 dtrace_nprobes = 1;
8344 } else {
8345 dtrace_probe_t **oprobes = dtrace_probes;
8346
8347 bcopy(oprobes, probes, osize);
8348 dtrace_membar_producer();
8349 dtrace_probes = probes;
8350
8351 dtrace_sync();
8352
8353 /*
8354 * All CPUs are now seeing the new probes array; we can
8355 * safely free the old array.
8356 */
8357 kmem_free(oprobes, osize);
8358 dtrace_nprobes <<= 1;
8359 }
8360
8361 ASSERT(id - 1 < dtrace_nprobes);
8362 }
8363
8364 ASSERT(dtrace_probes[id - 1] == NULL);
8365 dtrace_probes[id - 1] = probe;
8366
8367 if (provider != dtrace_provider)
8368 mutex_exit(&dtrace_lock);
8369
8370 return (id);
8371 }
8372
8373 static dtrace_probe_t *
8374 dtrace_probe_lookup_id(dtrace_id_t id)
8375 {
8376 ASSERT(MUTEX_HELD(&dtrace_lock));
8377
8378 if (id == 0 || id > dtrace_nprobes)
8379 return (NULL);
8380
8381 return (dtrace_probes[id - 1]);
8382 }
8383
8384 static int
8385 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8386 {
8387 *((dtrace_id_t *)arg) = probe->dtpr_id;
8388
8389 return (DTRACE_MATCH_DONE);
8390 }
8391
8392 /*
8393 * Look up a probe based on provider and one or more of module name, function
8394 * name and probe name.
8395 */
8396 dtrace_id_t
8397 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
8398 const char *func, const char *name)
8399 {
8400 dtrace_probekey_t pkey;
8401 dtrace_id_t id;
8402 int match;
8403
8404 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8405 pkey.dtpk_pmatch = &dtrace_match_string;
8406 pkey.dtpk_mod = mod;
8407 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8408 pkey.dtpk_func = func;
8409 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8410 pkey.dtpk_name = name;
8411 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8412 pkey.dtpk_id = DTRACE_IDNONE;
8413
8414 mutex_enter(&dtrace_lock);
8415 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8416 dtrace_probe_lookup_match, &id);
8417 mutex_exit(&dtrace_lock);
8418
8419 ASSERT(match == 1 || match == 0);
8420 return (match ? id : 0);
8421 }
8422
8423 /*
8424 * Returns the probe argument associated with the specified probe.
8425 */
8426 void *
8427 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8428 {
8429 dtrace_probe_t *probe;
8430 void *rval = NULL;
8431
8432 mutex_enter(&dtrace_lock);
8433
8434 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8435 probe->dtpr_provider == (dtrace_provider_t *)id)
8436 rval = probe->dtpr_arg;
8437
8438 mutex_exit(&dtrace_lock);
8439
8440 return (rval);
8441 }
8442
8443 /*
8444 * Copy a probe into a probe description.
8445 */
8446 static void
8447 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8448 {
8449 bzero(pdp, sizeof (dtrace_probedesc_t));
8450 pdp->dtpd_id = prp->dtpr_id;
8451
8452 (void) strncpy(pdp->dtpd_provider,
8453 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8454
8455 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8456 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8457 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8458 }
8459
8460 /*
8461 * Called to indicate that a probe -- or probes -- should be provided by a
8462 * specfied provider. If the specified description is NULL, the provider will
8463 * be told to provide all of its probes. (This is done whenever a new
8464 * consumer comes along, or whenever a retained enabling is to be matched.) If
8465 * the specified description is non-NULL, the provider is given the
8466 * opportunity to dynamically provide the specified probe, allowing providers
8467 * to support the creation of probes on-the-fly. (So-called _autocreated_
8468 * probes.) If the provider is NULL, the operations will be applied to all
8469 * providers; if the provider is non-NULL the operations will only be applied
8470 * to the specified provider. The dtrace_provider_lock must be held, and the
8471 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8472 * will need to grab the dtrace_lock when it reenters the framework through
8473 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8474 */
8475 static void
8476 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8477 {
8478 struct modctl *ctl;
8479 int all = 0;
8480
8481 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8482
8483 if (prv == NULL) {
8484 all = 1;
8485 prv = dtrace_provider;
8486 }
8487
8488 do {
8489 /*
8490 * First, call the blanket provide operation.
8491 */
8492 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8493
8494 /*
8495 * Now call the per-module provide operation. We will grab
8496 * mod_lock to prevent the list from being modified. Note
8497 * that this also prevents the mod_busy bits from changing.
8498 * (mod_busy can only be changed with mod_lock held.)
8499 */
8500 mutex_enter(&mod_lock);
8501
8502 ctl = &modules;
8503 do {
8504 if (ctl->mod_busy || ctl->mod_mp == NULL)
8505 continue;
8506
8507 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8508
8509 } while ((ctl = ctl->mod_next) != &modules);
8510
8511 mutex_exit(&mod_lock);
8512 } while (all && (prv = prv->dtpv_next) != NULL);
8513 }
8514
8515 /*
8516 * Iterate over each probe, and call the Framework-to-Provider API function
8517 * denoted by offs.
8518 */
8519 static void
8520 dtrace_probe_foreach(uintptr_t offs)
8521 {
8522 dtrace_provider_t *prov;
8523 void (*func)(void *, dtrace_id_t, void *);
8524 dtrace_probe_t *probe;
8525 dtrace_icookie_t cookie;
8526 int i;
8527
8528 /*
8529 * We disable interrupts to walk through the probe array. This is
8530 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8531 * won't see stale data.
8532 */
8533 cookie = dtrace_interrupt_disable();
8534
8535 for (i = 0; i < dtrace_nprobes; i++) {
8536 if ((probe = dtrace_probes[i]) == NULL)
8537 continue;
8538
8539 if (probe->dtpr_ecb == NULL) {
8540 /*
8541 * This probe isn't enabled -- don't call the function.
8542 */
8543 continue;
8544 }
8545
8546 prov = probe->dtpr_provider;
8547 func = *((void(**)(void *, dtrace_id_t, void *))
8548 ((uintptr_t)&prov->dtpv_pops + offs));
8549
8550 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8551 }
8552
8553 dtrace_interrupt_enable(cookie);
8554 }
8555
8556 static int
8557 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8558 {
8559 dtrace_probekey_t pkey;
8560 uint32_t priv;
8561 uid_t uid;
8562 zoneid_t zoneid;
8563
8564 ASSERT(MUTEX_HELD(&dtrace_lock));
8565 dtrace_ecb_create_cache = NULL;
8566
8567 if (desc == NULL) {
8568 /*
8569 * If we're passed a NULL description, we're being asked to
8570 * create an ECB with a NULL probe.
8571 */
8572 (void) dtrace_ecb_create_enable(NULL, enab);
8573 return (0);
8574 }
8575
8576 dtrace_probekey(desc, &pkey);
8577 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8578 &priv, &uid, &zoneid);
8579
8580 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8581 enab));
8582 }
8583
8584 /*
8585 * DTrace Helper Provider Functions
8586 */
8587 static void
8588 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8589 {
8590 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8591 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8592 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8593 }
8594
8595 static void
8596 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8597 const dof_provider_t *dofprov, char *strtab)
8598 {
8599 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8600 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8601 dofprov->dofpv_provattr);
8602 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8603 dofprov->dofpv_modattr);
8604 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8605 dofprov->dofpv_funcattr);
8606 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8607 dofprov->dofpv_nameattr);
8608 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8609 dofprov->dofpv_argsattr);
8610 }
8611
8612 static void
8613 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8614 {
8615 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8616 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8617 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8618 dof_provider_t *provider;
8619 dof_probe_t *probe;
8620 uint32_t *off, *enoff;
8621 uint8_t *arg;
8622 char *strtab;
8623 uint_t i, nprobes;
8624 dtrace_helper_provdesc_t dhpv;
8625 dtrace_helper_probedesc_t dhpb;
8626 dtrace_meta_t *meta = dtrace_meta_pid;
8627 dtrace_mops_t *mops = &meta->dtm_mops;
8628 void *parg;
8629
8630 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8631 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8632 provider->dofpv_strtab * dof->dofh_secsize);
8633 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8634 provider->dofpv_probes * dof->dofh_secsize);
8635 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8636 provider->dofpv_prargs * dof->dofh_secsize);
8637 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8638 provider->dofpv_proffs * dof->dofh_secsize);
8639
8640 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8641 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8642 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8643 enoff = NULL;
8644
8645 /*
8646 * See dtrace_helper_provider_validate().
8647 */
8648 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8649 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8650 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8651 provider->dofpv_prenoffs * dof->dofh_secsize);
8652 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8653 }
8654
8655 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8656
8657 /*
8658 * Create the provider.
8659 */
8660 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8661
8662 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8663 return;
8664
8665 meta->dtm_count++;
8666
8667 /*
8668 * Create the probes.
8669 */
8670 for (i = 0; i < nprobes; i++) {
8671 probe = (dof_probe_t *)(uintptr_t)(daddr +
8672 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8673
8674 dhpb.dthpb_mod = dhp->dofhp_mod;
8675 dhpb.dthpb_func = strtab + probe->dofpr_func;
8676 dhpb.dthpb_name = strtab + probe->dofpr_name;
8677 dhpb.dthpb_base = probe->dofpr_addr;
8678 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8679 dhpb.dthpb_noffs = probe->dofpr_noffs;
8680 if (enoff != NULL) {
8681 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8682 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8683 } else {
8684 dhpb.dthpb_enoffs = NULL;
8685 dhpb.dthpb_nenoffs = 0;
8686 }
8687 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8688 dhpb.dthpb_nargc = probe->dofpr_nargc;
8689 dhpb.dthpb_xargc = probe->dofpr_xargc;
8690 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8691 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8692
8693 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8694 }
8695 }
8696
8697 static void
8698 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8699 {
8700 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8701 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8702 int i;
8703
8704 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8705
8706 for (i = 0; i < dof->dofh_secnum; i++) {
8707 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8708 dof->dofh_secoff + i * dof->dofh_secsize);
8709
8710 if (sec->dofs_type != DOF_SECT_PROVIDER)
8711 continue;
8712
8713 dtrace_helper_provide_one(dhp, sec, pid);
8714 }
8715
8716 /*
8717 * We may have just created probes, so we must now rematch against
8718 * any retained enablings. Note that this call will acquire both
8719 * cpu_lock and dtrace_lock; the fact that we are holding
8720 * dtrace_meta_lock now is what defines the ordering with respect to
8721 * these three locks.
8722 */
8723 dtrace_enabling_matchall();
8724 }
8725
8726 static void
8727 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8728 {
8729 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8730 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8731 dof_sec_t *str_sec;
8732 dof_provider_t *provider;
8733 char *strtab;
8734 dtrace_helper_provdesc_t dhpv;
8735 dtrace_meta_t *meta = dtrace_meta_pid;
8736 dtrace_mops_t *mops = &meta->dtm_mops;
8737
8738 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8739 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8740 provider->dofpv_strtab * dof->dofh_secsize);
8741
8742 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8743
8744 /*
8745 * Create the provider.
8746 */
8747 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8748
8749 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8750
8751 meta->dtm_count--;
8752 }
8753
8754 static void
8755 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8756 {
8757 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8758 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8759 int i;
8760
8761 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8762
8763 for (i = 0; i < dof->dofh_secnum; i++) {
8764 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8765 dof->dofh_secoff + i * dof->dofh_secsize);
8766
8767 if (sec->dofs_type != DOF_SECT_PROVIDER)
8768 continue;
8769
8770 dtrace_helper_provider_remove_one(dhp, sec, pid);
8771 }
8772 }
8773
8774 /*
8775 * DTrace Meta Provider-to-Framework API Functions
8776 *
8777 * These functions implement the Meta Provider-to-Framework API, as described
8778 * in <sys/dtrace.h>.
8779 */
8780 int
8781 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8782 dtrace_meta_provider_id_t *idp)
8783 {
8784 dtrace_meta_t *meta;
8785 dtrace_helpers_t *help, *next;
8786 int i;
8787
8788 *idp = DTRACE_METAPROVNONE;
8789
8790 /*
8791 * We strictly don't need the name, but we hold onto it for
8792 * debuggability. All hail error queues!
8793 */
8794 if (name == NULL) {
8795 cmn_err(CE_WARN, "failed to register meta-provider: "
8796 "invalid name");
8797 return (EINVAL);
8798 }
8799
8800 if (mops == NULL ||
8801 mops->dtms_create_probe == NULL ||
8802 mops->dtms_provide_pid == NULL ||
8803 mops->dtms_remove_pid == NULL) {
8804 cmn_err(CE_WARN, "failed to register meta-register %s: "
8805 "invalid ops", name);
8806 return (EINVAL);
8807 }
8808
8809 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8810 meta->dtm_mops = *mops;
8811 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8812 (void) strcpy(meta->dtm_name, name);
8813 meta->dtm_arg = arg;
8814
8815 mutex_enter(&dtrace_meta_lock);
8816 mutex_enter(&dtrace_lock);
8817
8818 if (dtrace_meta_pid != NULL) {
8819 mutex_exit(&dtrace_lock);
8820 mutex_exit(&dtrace_meta_lock);
8821 cmn_err(CE_WARN, "failed to register meta-register %s: "
8822 "user-land meta-provider exists", name);
8823 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8824 kmem_free(meta, sizeof (dtrace_meta_t));
8825 return (EINVAL);
8826 }
8827
8828 dtrace_meta_pid = meta;
8829 *idp = (dtrace_meta_provider_id_t)meta;
8830
8831 /*
8832 * If there are providers and probes ready to go, pass them
8833 * off to the new meta provider now.
8834 */
8835
8836 help = dtrace_deferred_pid;
8837 dtrace_deferred_pid = NULL;
8838
8839 mutex_exit(&dtrace_lock);
8840
8841 while (help != NULL) {
8842 for (i = 0; i < help->dthps_nprovs; i++) {
8843 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8844 help->dthps_pid);
8845 }
8846
8847 next = help->dthps_next;
8848 help->dthps_next = NULL;
8849 help->dthps_prev = NULL;
8850 help->dthps_deferred = 0;
8851 help = next;
8852 }
8853
8854 mutex_exit(&dtrace_meta_lock);
8855
8856 return (0);
8857 }
8858
8859 int
8860 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8861 {
8862 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8863
8864 mutex_enter(&dtrace_meta_lock);
8865 mutex_enter(&dtrace_lock);
8866
8867 if (old == dtrace_meta_pid) {
8868 pp = &dtrace_meta_pid;
8869 } else {
8870 panic("attempt to unregister non-existent "
8871 "dtrace meta-provider %p\n", (void *)old);
8872 }
8873
8874 if (old->dtm_count != 0) {
8875 mutex_exit(&dtrace_lock);
8876 mutex_exit(&dtrace_meta_lock);
8877 return (EBUSY);
8878 }
8879
8880 *pp = NULL;
8881
8882 mutex_exit(&dtrace_lock);
8883 mutex_exit(&dtrace_meta_lock);
8884
8885 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8886 kmem_free(old, sizeof (dtrace_meta_t));
8887
8888 return (0);
8889 }
8890
8891
8892 /*
8893 * DTrace DIF Object Functions
8894 */
8895 static int
8896 dtrace_difo_err(uint_t pc, const char *format, ...)
8897 {
8898 if (dtrace_err_verbose) {
8899 va_list alist;
8900
8901 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8902 va_start(alist, format);
8903 (void) vuprintf(format, alist);
8904 va_end(alist);
8905 }
8906
8907 #ifdef DTRACE_ERRDEBUG
8908 dtrace_errdebug(format);
8909 #endif
8910 return (1);
8911 }
8912
8913 /*
8914 * Validate a DTrace DIF object by checking the IR instructions. The following
8915 * rules are currently enforced by dtrace_difo_validate():
8916 *
8917 * 1. Each instruction must have a valid opcode
8918 * 2. Each register, string, variable, or subroutine reference must be valid
8919 * 3. No instruction can modify register %r0 (must be zero)
8920 * 4. All instruction reserved bits must be set to zero
8921 * 5. The last instruction must be a "ret" instruction
8922 * 6. All branch targets must reference a valid instruction _after_ the branch
8923 */
8924 static int
8925 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8926 cred_t *cr)
8927 {
8928 int err = 0, i;
8929 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8930 int kcheckload;
8931 uint_t pc;
8932
8933 kcheckload = cr == NULL ||
8934 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8935
8936 dp->dtdo_destructive = 0;
8937
8938 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8939 dif_instr_t instr = dp->dtdo_buf[pc];
8940
8941 uint_t r1 = DIF_INSTR_R1(instr);
8942 uint_t r2 = DIF_INSTR_R2(instr);
8943 uint_t rd = DIF_INSTR_RD(instr);
8944 uint_t rs = DIF_INSTR_RS(instr);
8945 uint_t label = DIF_INSTR_LABEL(instr);
8946 uint_t v = DIF_INSTR_VAR(instr);
8947 uint_t subr = DIF_INSTR_SUBR(instr);
8948 uint_t type = DIF_INSTR_TYPE(instr);
8949 uint_t op = DIF_INSTR_OP(instr);
8950
8951 switch (op) {
8952 case DIF_OP_OR:
8953 case DIF_OP_XOR:
8954 case DIF_OP_AND:
8955 case DIF_OP_SLL:
8956 case DIF_OP_SRL:
8957 case DIF_OP_SRA:
8958 case DIF_OP_SUB:
8959 case DIF_OP_ADD:
8960 case DIF_OP_MUL:
8961 case DIF_OP_SDIV:
8962 case DIF_OP_UDIV:
8963 case DIF_OP_SREM:
8964 case DIF_OP_UREM:
8965 case DIF_OP_COPYS:
8966 if (r1 >= nregs)
8967 err += efunc(pc, "invalid register %u\n", r1);
8968 if (r2 >= nregs)
8969 err += efunc(pc, "invalid register %u\n", r2);
8970 if (rd >= nregs)
8971 err += efunc(pc, "invalid register %u\n", rd);
8972 if (rd == 0)
8973 err += efunc(pc, "cannot write to %r0\n");
8974 break;
8975 case DIF_OP_NOT:
8976 case DIF_OP_MOV:
8977 case DIF_OP_ALLOCS:
8978 if (r1 >= nregs)
8979 err += efunc(pc, "invalid register %u\n", r1);
8980 if (r2 != 0)
8981 err += efunc(pc, "non-zero reserved bits\n");
8982 if (rd >= nregs)
8983 err += efunc(pc, "invalid register %u\n", rd);
8984 if (rd == 0)
8985 err += efunc(pc, "cannot write to %r0\n");
8986 break;
8987 case DIF_OP_LDSB:
8988 case DIF_OP_LDSH:
8989 case DIF_OP_LDSW:
8990 case DIF_OP_LDUB:
8991 case DIF_OP_LDUH:
8992 case DIF_OP_LDUW:
8993 case DIF_OP_LDX:
8994 if (r1 >= nregs)
8995 err += efunc(pc, "invalid register %u\n", r1);
8996 if (r2 != 0)
8997 err += efunc(pc, "non-zero reserved bits\n");
8998 if (rd >= nregs)
8999 err += efunc(pc, "invalid register %u\n", rd);
9000 if (rd == 0)
9001 err += efunc(pc, "cannot write to %r0\n");
9002 if (kcheckload)
9003 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9004 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9005 break;
9006 case DIF_OP_RLDSB:
9007 case DIF_OP_RLDSH:
9008 case DIF_OP_RLDSW:
9009 case DIF_OP_RLDUB:
9010 case DIF_OP_RLDUH:
9011 case DIF_OP_RLDUW:
9012 case DIF_OP_RLDX:
9013 if (r1 >= nregs)
9014 err += efunc(pc, "invalid register %u\n", r1);
9015 if (r2 != 0)
9016 err += efunc(pc, "non-zero reserved bits\n");
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_ULDSB:
9023 case DIF_OP_ULDSH:
9024 case DIF_OP_ULDSW:
9025 case DIF_OP_ULDUB:
9026 case DIF_OP_ULDUH:
9027 case DIF_OP_ULDUW:
9028 case DIF_OP_ULDX:
9029 if (r1 >= nregs)
9030 err += efunc(pc, "invalid register %u\n", r1);
9031 if (r2 != 0)
9032 err += efunc(pc, "non-zero reserved bits\n");
9033 if (rd >= nregs)
9034 err += efunc(pc, "invalid register %u\n", rd);
9035 if (rd == 0)
9036 err += efunc(pc, "cannot write to %r0\n");
9037 break;
9038 case DIF_OP_STB:
9039 case DIF_OP_STH:
9040 case DIF_OP_STW:
9041 case DIF_OP_STX:
9042 if (r1 >= nregs)
9043 err += efunc(pc, "invalid register %u\n", r1);
9044 if (r2 != 0)
9045 err += efunc(pc, "non-zero reserved bits\n");
9046 if (rd >= nregs)
9047 err += efunc(pc, "invalid register %u\n", rd);
9048 if (rd == 0)
9049 err += efunc(pc, "cannot write to 0 address\n");
9050 break;
9051 case DIF_OP_CMP:
9052 case DIF_OP_SCMP:
9053 if (r1 >= nregs)
9054 err += efunc(pc, "invalid register %u\n", r1);
9055 if (r2 >= nregs)
9056 err += efunc(pc, "invalid register %u\n", r2);
9057 if (rd != 0)
9058 err += efunc(pc, "non-zero reserved bits\n");
9059 break;
9060 case DIF_OP_TST:
9061 if (r1 >= nregs)
9062 err += efunc(pc, "invalid register %u\n", r1);
9063 if (r2 != 0 || rd != 0)
9064 err += efunc(pc, "non-zero reserved bits\n");
9065 break;
9066 case DIF_OP_BA:
9067 case DIF_OP_BE:
9068 case DIF_OP_BNE:
9069 case DIF_OP_BG:
9070 case DIF_OP_BGU:
9071 case DIF_OP_BGE:
9072 case DIF_OP_BGEU:
9073 case DIF_OP_BL:
9074 case DIF_OP_BLU:
9075 case DIF_OP_BLE:
9076 case DIF_OP_BLEU:
9077 if (label >= dp->dtdo_len) {
9078 err += efunc(pc, "invalid branch target %u\n",
9079 label);
9080 }
9081 if (label <= pc) {
9082 err += efunc(pc, "backward branch to %u\n",
9083 label);
9084 }
9085 break;
9086 case DIF_OP_RET:
9087 if (r1 != 0 || r2 != 0)
9088 err += efunc(pc, "non-zero reserved bits\n");
9089 if (rd >= nregs)
9090 err += efunc(pc, "invalid register %u\n", rd);
9091 break;
9092 case DIF_OP_NOP:
9093 case DIF_OP_POPTS:
9094 case DIF_OP_FLUSHTS:
9095 if (r1 != 0 || r2 != 0 || rd != 0)
9096 err += efunc(pc, "non-zero reserved bits\n");
9097 break;
9098 case DIF_OP_SETX:
9099 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9100 err += efunc(pc, "invalid integer ref %u\n",
9101 DIF_INSTR_INTEGER(instr));
9102 }
9103 if (rd >= nregs)
9104 err += efunc(pc, "invalid register %u\n", rd);
9105 if (rd == 0)
9106 err += efunc(pc, "cannot write to %r0\n");
9107 break;
9108 case DIF_OP_SETS:
9109 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9110 err += efunc(pc, "invalid string ref %u\n",
9111 DIF_INSTR_STRING(instr));
9112 }
9113 if (rd >= nregs)
9114 err += efunc(pc, "invalid register %u\n", rd);
9115 if (rd == 0)
9116 err += efunc(pc, "cannot write to %r0\n");
9117 break;
9118 case DIF_OP_LDGA:
9119 case DIF_OP_LDTA:
9120 if (r1 > DIF_VAR_ARRAY_MAX)
9121 err += efunc(pc, "invalid array %u\n", r1);
9122 if (r2 >= nregs)
9123 err += efunc(pc, "invalid register %u\n", r2);
9124 if (rd >= nregs)
9125 err += efunc(pc, "invalid register %u\n", rd);
9126 if (rd == 0)
9127 err += efunc(pc, "cannot write to %r0\n");
9128 break;
9129 case DIF_OP_LDGS:
9130 case DIF_OP_LDTS:
9131 case DIF_OP_LDLS:
9132 case DIF_OP_LDGAA:
9133 case DIF_OP_LDTAA:
9134 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9135 err += efunc(pc, "invalid variable %u\n", v);
9136 if (rd >= nregs)
9137 err += efunc(pc, "invalid register %u\n", rd);
9138 if (rd == 0)
9139 err += efunc(pc, "cannot write to %r0\n");
9140 break;
9141 case DIF_OP_STGS:
9142 case DIF_OP_STTS:
9143 case DIF_OP_STLS:
9144 case DIF_OP_STGAA:
9145 case DIF_OP_STTAA:
9146 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9147 err += efunc(pc, "invalid variable %u\n", v);
9148 if (rs >= nregs)
9149 err += efunc(pc, "invalid register %u\n", rd);
9150 break;
9151 case DIF_OP_CALL:
9152 if (subr > DIF_SUBR_MAX)
9153 err += efunc(pc, "invalid subr %u\n", subr);
9154 if (rd >= nregs)
9155 err += efunc(pc, "invalid register %u\n", rd);
9156 if (rd == 0)
9157 err += efunc(pc, "cannot write to %r0\n");
9158
9159 if (subr == DIF_SUBR_COPYOUT ||
9160 subr == DIF_SUBR_COPYOUTSTR) {
9161 dp->dtdo_destructive = 1;
9162 }
9163
9164 if (subr == DIF_SUBR_GETF) {
9165 /*
9166 * If we have a getf() we need to record that
9167 * in our state. Note that our state can be
9168 * NULL if this is a helper -- but in that
9169 * case, the call to getf() is itself illegal,
9170 * and will be caught (slightly later) when
9171 * the helper is validated.
9172 */
9173 if (vstate->dtvs_state != NULL)
9174 vstate->dtvs_state->dts_getf++;
9175 }
9176
9177 break;
9178 case DIF_OP_PUSHTR:
9179 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
9180 err += efunc(pc, "invalid ref type %u\n", type);
9181 if (r2 >= nregs)
9182 err += efunc(pc, "invalid register %u\n", r2);
9183 if (rs >= nregs)
9184 err += efunc(pc, "invalid register %u\n", rs);
9185 break;
9186 case DIF_OP_PUSHTV:
9187 if (type != DIF_TYPE_CTF)
9188 err += efunc(pc, "invalid val type %u\n", type);
9189 if (r2 >= nregs)
9190 err += efunc(pc, "invalid register %u\n", r2);
9191 if (rs >= nregs)
9192 err += efunc(pc, "invalid register %u\n", rs);
9193 break;
9194 default:
9195 err += efunc(pc, "invalid opcode %u\n",
9196 DIF_INSTR_OP(instr));
9197 }
9198 }
9199
9200 if (dp->dtdo_len != 0 &&
9201 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
9202 err += efunc(dp->dtdo_len - 1,
9203 "expected 'ret' as last DIF instruction\n");
9204 }
9205
9206 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
9207 /*
9208 * If we're not returning by reference, the size must be either
9209 * 0 or the size of one of the base types.
9210 */
9211 switch (dp->dtdo_rtype.dtdt_size) {
9212 case 0:
9213 case sizeof (uint8_t):
9214 case sizeof (uint16_t):
9215 case sizeof (uint32_t):
9216 case sizeof (uint64_t):
9217 break;
9218
9219 default:
9220 err += efunc(dp->dtdo_len - 1, "bad return size\n");
9221 }
9222 }
9223
9224 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
9225 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
9226 dtrace_diftype_t *vt, *et;
9227 uint_t id, ndx;
9228
9229 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
9230 v->dtdv_scope != DIFV_SCOPE_THREAD &&
9231 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
9232 err += efunc(i, "unrecognized variable scope %d\n",
9233 v->dtdv_scope);
9234 break;
9235 }
9236
9237 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
9238 v->dtdv_kind != DIFV_KIND_SCALAR) {
9239 err += efunc(i, "unrecognized variable type %d\n",
9240 v->dtdv_kind);
9241 break;
9242 }
9243
9244 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
9245 err += efunc(i, "%d exceeds variable id limit\n", id);
9246 break;
9247 }
9248
9249 if (id < DIF_VAR_OTHER_UBASE)
9250 continue;
9251
9252 /*
9253 * For user-defined variables, we need to check that this
9254 * definition is identical to any previous definition that we
9255 * encountered.
9256 */
9257 ndx = id - DIF_VAR_OTHER_UBASE;
9258
9259 switch (v->dtdv_scope) {
9260 case DIFV_SCOPE_GLOBAL:
9261 if (ndx < vstate->dtvs_nglobals) {
9262 dtrace_statvar_t *svar;
9263
9264 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
9265 existing = &svar->dtsv_var;
9266 }
9267
9268 break;
9269
9270 case DIFV_SCOPE_THREAD:
9271 if (ndx < vstate->dtvs_ntlocals)
9272 existing = &vstate->dtvs_tlocals[ndx];
9273 break;
9274
9275 case DIFV_SCOPE_LOCAL:
9276 if (ndx < vstate->dtvs_nlocals) {
9277 dtrace_statvar_t *svar;
9278
9279 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
9280 existing = &svar->dtsv_var;
9281 }
9282
9283 break;
9284 }
9285
9286 vt = &v->dtdv_type;
9287
9288 if (vt->dtdt_flags & DIF_TF_BYREF) {
9289 if (vt->dtdt_size == 0) {
9290 err += efunc(i, "zero-sized variable\n");
9291 break;
9292 }
9293
9294 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
9295 vt->dtdt_size > dtrace_global_maxsize) {
9296 err += efunc(i, "oversized by-ref global\n");
9297 break;
9298 }
9299 }
9300
9301 if (existing == NULL || existing->dtdv_id == 0)
9302 continue;
9303
9304 ASSERT(existing->dtdv_id == v->dtdv_id);
9305 ASSERT(existing->dtdv_scope == v->dtdv_scope);
9306
9307 if (existing->dtdv_kind != v->dtdv_kind)
9308 err += efunc(i, "%d changed variable kind\n", id);
9309
9310 et = &existing->dtdv_type;
9311
9312 if (vt->dtdt_flags != et->dtdt_flags) {
9313 err += efunc(i, "%d changed variable type flags\n", id);
9314 break;
9315 }
9316
9317 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
9318 err += efunc(i, "%d changed variable type size\n", id);
9319 break;
9320 }
9321 }
9322
9323 return (err);
9324 }
9325
9326 /*
9327 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
9328 * are much more constrained than normal DIFOs. Specifically, they may
9329 * not:
9330 *
9331 * 1. Make calls to subroutines other than copyin(), copyinstr() or
9332 * miscellaneous string routines
9333 * 2. Access DTrace variables other than the args[] array, and the
9334 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9335 * 3. Have thread-local variables.
9336 * 4. Have dynamic variables.
9337 */
9338 static int
9339 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9340 {
9341 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9342 int err = 0;
9343 uint_t pc;
9344
9345 for (pc = 0; pc < dp->dtdo_len; pc++) {
9346 dif_instr_t instr = dp->dtdo_buf[pc];
9347
9348 uint_t v = DIF_INSTR_VAR(instr);
9349 uint_t subr = DIF_INSTR_SUBR(instr);
9350 uint_t op = DIF_INSTR_OP(instr);
9351
9352 switch (op) {
9353 case DIF_OP_OR:
9354 case DIF_OP_XOR:
9355 case DIF_OP_AND:
9356 case DIF_OP_SLL:
9357 case DIF_OP_SRL:
9358 case DIF_OP_SRA:
9359 case DIF_OP_SUB:
9360 case DIF_OP_ADD:
9361 case DIF_OP_MUL:
9362 case DIF_OP_SDIV:
9363 case DIF_OP_UDIV:
9364 case DIF_OP_SREM:
9365 case DIF_OP_UREM:
9366 case DIF_OP_COPYS:
9367 case DIF_OP_NOT:
9368 case DIF_OP_MOV:
9369 case DIF_OP_RLDSB:
9370 case DIF_OP_RLDSH:
9371 case DIF_OP_RLDSW:
9372 case DIF_OP_RLDUB:
9373 case DIF_OP_RLDUH:
9374 case DIF_OP_RLDUW:
9375 case DIF_OP_RLDX:
9376 case DIF_OP_ULDSB:
9377 case DIF_OP_ULDSH:
9378 case DIF_OP_ULDSW:
9379 case DIF_OP_ULDUB:
9380 case DIF_OP_ULDUH:
9381 case DIF_OP_ULDUW:
9382 case DIF_OP_ULDX:
9383 case DIF_OP_STB:
9384 case DIF_OP_STH:
9385 case DIF_OP_STW:
9386 case DIF_OP_STX:
9387 case DIF_OP_ALLOCS:
9388 case DIF_OP_CMP:
9389 case DIF_OP_SCMP:
9390 case DIF_OP_TST:
9391 case DIF_OP_BA:
9392 case DIF_OP_BE:
9393 case DIF_OP_BNE:
9394 case DIF_OP_BG:
9395 case DIF_OP_BGU:
9396 case DIF_OP_BGE:
9397 case DIF_OP_BGEU:
9398 case DIF_OP_BL:
9399 case DIF_OP_BLU:
9400 case DIF_OP_BLE:
9401 case DIF_OP_BLEU:
9402 case DIF_OP_RET:
9403 case DIF_OP_NOP:
9404 case DIF_OP_POPTS:
9405 case DIF_OP_FLUSHTS:
9406 case DIF_OP_SETX:
9407 case DIF_OP_SETS:
9408 case DIF_OP_LDGA:
9409 case DIF_OP_LDLS:
9410 case DIF_OP_STGS:
9411 case DIF_OP_STLS:
9412 case DIF_OP_PUSHTR:
9413 case DIF_OP_PUSHTV:
9414 break;
9415
9416 case DIF_OP_LDGS:
9417 if (v >= DIF_VAR_OTHER_UBASE)
9418 break;
9419
9420 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9421 break;
9422
9423 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9424 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9425 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9426 v == DIF_VAR_UID || v == DIF_VAR_GID)
9427 break;
9428
9429 err += efunc(pc, "illegal variable %u\n", v);
9430 break;
9431
9432 case DIF_OP_LDTA:
9433 case DIF_OP_LDTS:
9434 case DIF_OP_LDGAA:
9435 case DIF_OP_LDTAA:
9436 err += efunc(pc, "illegal dynamic variable load\n");
9437 break;
9438
9439 case DIF_OP_STTS:
9440 case DIF_OP_STGAA:
9441 case DIF_OP_STTAA:
9442 err += efunc(pc, "illegal dynamic variable store\n");
9443 break;
9444
9445 case DIF_OP_CALL:
9446 if (subr == DIF_SUBR_ALLOCA ||
9447 subr == DIF_SUBR_BCOPY ||
9448 subr == DIF_SUBR_COPYIN ||
9449 subr == DIF_SUBR_COPYINTO ||
9450 subr == DIF_SUBR_COPYINSTR ||
9451 subr == DIF_SUBR_INDEX ||
9452 subr == DIF_SUBR_INET_NTOA ||
9453 subr == DIF_SUBR_INET_NTOA6 ||
9454 subr == DIF_SUBR_INET_NTOP ||
9455 subr == DIF_SUBR_JSON ||
9456 subr == DIF_SUBR_LLTOSTR ||
9457 subr == DIF_SUBR_STRTOLL ||
9458 subr == DIF_SUBR_RINDEX ||
9459 subr == DIF_SUBR_STRCHR ||
9460 subr == DIF_SUBR_STRJOIN ||
9461 subr == DIF_SUBR_STRRCHR ||
9462 subr == DIF_SUBR_STRSTR ||
9463 subr == DIF_SUBR_HTONS ||
9464 subr == DIF_SUBR_HTONL ||
9465 subr == DIF_SUBR_HTONLL ||
9466 subr == DIF_SUBR_NTOHS ||
9467 subr == DIF_SUBR_NTOHL ||
9468 subr == DIF_SUBR_NTOHLL)
9469 break;
9470
9471 err += efunc(pc, "invalid subr %u\n", subr);
9472 break;
9473
9474 default:
9475 err += efunc(pc, "invalid opcode %u\n",
9476 DIF_INSTR_OP(instr));
9477 }
9478 }
9479
9480 return (err);
9481 }
9482
9483 /*
9484 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9485 * basis; 0 if not.
9486 */
9487 static int
9488 dtrace_difo_cacheable(dtrace_difo_t *dp)
9489 {
9490 int i;
9491
9492 if (dp == NULL)
9493 return (0);
9494
9495 for (i = 0; i < dp->dtdo_varlen; i++) {
9496 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9497
9498 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9499 continue;
9500
9501 switch (v->dtdv_id) {
9502 case DIF_VAR_CURTHREAD:
9503 case DIF_VAR_PID:
9504 case DIF_VAR_TID:
9505 case DIF_VAR_EXECNAME:
9506 case DIF_VAR_ZONENAME:
9507 break;
9508
9509 default:
9510 return (0);
9511 }
9512 }
9513
9514 /*
9515 * This DIF object may be cacheable. Now we need to look for any
9516 * array loading instructions, any memory loading instructions, or
9517 * any stores to thread-local variables.
9518 */
9519 for (i = 0; i < dp->dtdo_len; i++) {
9520 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9521
9522 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9523 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9524 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9525 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9526 return (0);
9527 }
9528
9529 return (1);
9530 }
9531
9532 static void
9533 dtrace_difo_hold(dtrace_difo_t *dp)
9534 {
9535 int i;
9536
9537 ASSERT(MUTEX_HELD(&dtrace_lock));
9538
9539 dp->dtdo_refcnt++;
9540 ASSERT(dp->dtdo_refcnt != 0);
9541
9542 /*
9543 * We need to check this DIF object for references to the variable
9544 * DIF_VAR_VTIMESTAMP.
9545 */
9546 for (i = 0; i < dp->dtdo_varlen; i++) {
9547 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9548
9549 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9550 continue;
9551
9552 if (dtrace_vtime_references++ == 0)
9553 dtrace_vtime_enable();
9554 }
9555 }
9556
9557 /*
9558 * This routine calculates the dynamic variable chunksize for a given DIF
9559 * object. The calculation is not fool-proof, and can probably be tricked by
9560 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9561 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9562 * if a dynamic variable size exceeds the chunksize.
9563 */
9564 static void
9565 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9566 {
9567 uint64_t sval;
9568 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9569 const dif_instr_t *text = dp->dtdo_buf;
9570 uint_t pc, srd = 0;
9571 uint_t ttop = 0;
9572 size_t size, ksize;
9573 uint_t id, i;
9574
9575 for (pc = 0; pc < dp->dtdo_len; pc++) {
9576 dif_instr_t instr = text[pc];
9577 uint_t op = DIF_INSTR_OP(instr);
9578 uint_t rd = DIF_INSTR_RD(instr);
9579 uint_t r1 = DIF_INSTR_R1(instr);
9580 uint_t nkeys = 0;
9581 uchar_t scope;
9582
9583 dtrace_key_t *key = tupregs;
9584
9585 switch (op) {
9586 case DIF_OP_SETX:
9587 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9588 srd = rd;
9589 continue;
9590
9591 case DIF_OP_STTS:
9592 key = &tupregs[DIF_DTR_NREGS];
9593 key[0].dttk_size = 0;
9594 key[1].dttk_size = 0;
9595 nkeys = 2;
9596 scope = DIFV_SCOPE_THREAD;
9597 break;
9598
9599 case DIF_OP_STGAA:
9600 case DIF_OP_STTAA:
9601 nkeys = ttop;
9602
9603 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9604 key[nkeys++].dttk_size = 0;
9605
9606 key[nkeys++].dttk_size = 0;
9607
9608 if (op == DIF_OP_STTAA) {
9609 scope = DIFV_SCOPE_THREAD;
9610 } else {
9611 scope = DIFV_SCOPE_GLOBAL;
9612 }
9613
9614 break;
9615
9616 case DIF_OP_PUSHTR:
9617 if (ttop == DIF_DTR_NREGS)
9618 return;
9619
9620 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9621 /*
9622 * If the register for the size of the "pushtr"
9623 * is %r0 (or the value is 0) and the type is
9624 * a string, we'll use the system-wide default
9625 * string size.
9626 */
9627 tupregs[ttop++].dttk_size =
9628 dtrace_strsize_default;
9629 } else {
9630 if (srd == 0)
9631 return;
9632
9633 tupregs[ttop++].dttk_size = sval;
9634 }
9635
9636 break;
9637
9638 case DIF_OP_PUSHTV:
9639 if (ttop == DIF_DTR_NREGS)
9640 return;
9641
9642 tupregs[ttop++].dttk_size = 0;
9643 break;
9644
9645 case DIF_OP_FLUSHTS:
9646 ttop = 0;
9647 break;
9648
9649 case DIF_OP_POPTS:
9650 if (ttop != 0)
9651 ttop--;
9652 break;
9653 }
9654
9655 sval = 0;
9656 srd = 0;
9657
9658 if (nkeys == 0)
9659 continue;
9660
9661 /*
9662 * We have a dynamic variable allocation; calculate its size.
9663 */
9664 for (ksize = 0, i = 0; i < nkeys; i++)
9665 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9666
9667 size = sizeof (dtrace_dynvar_t);
9668 size += sizeof (dtrace_key_t) * (nkeys - 1);
9669 size += ksize;
9670
9671 /*
9672 * Now we need to determine the size of the stored data.
9673 */
9674 id = DIF_INSTR_VAR(instr);
9675
9676 for (i = 0; i < dp->dtdo_varlen; i++) {
9677 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9678
9679 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9680 size += v->dtdv_type.dtdt_size;
9681 break;
9682 }
9683 }
9684
9685 if (i == dp->dtdo_varlen)
9686 return;
9687
9688 /*
9689 * We have the size. If this is larger than the chunk size
9690 * for our dynamic variable state, reset the chunk size.
9691 */
9692 size = P2ROUNDUP(size, sizeof (uint64_t));
9693
9694 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9695 vstate->dtvs_dynvars.dtds_chunksize = size;
9696 }
9697 }
9698
9699 static void
9700 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9701 {
9702 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9703 uint_t id;
9704
9705 ASSERT(MUTEX_HELD(&dtrace_lock));
9706 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9707
9708 for (i = 0; i < dp->dtdo_varlen; i++) {
9709 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9710 dtrace_statvar_t *svar, ***svarp;
9711 size_t dsize = 0;
9712 uint8_t scope = v->dtdv_scope;
9713 int *np;
9714
9715 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9716 continue;
9717
9718 id -= DIF_VAR_OTHER_UBASE;
9719
9720 switch (scope) {
9721 case DIFV_SCOPE_THREAD:
9722 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9723 dtrace_difv_t *tlocals;
9724
9725 if ((ntlocals = (otlocals << 1)) == 0)
9726 ntlocals = 1;
9727
9728 osz = otlocals * sizeof (dtrace_difv_t);
9729 nsz = ntlocals * sizeof (dtrace_difv_t);
9730
9731 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9732
9733 if (osz != 0) {
9734 bcopy(vstate->dtvs_tlocals,
9735 tlocals, osz);
9736 kmem_free(vstate->dtvs_tlocals, osz);
9737 }
9738
9739 vstate->dtvs_tlocals = tlocals;
9740 vstate->dtvs_ntlocals = ntlocals;
9741 }
9742
9743 vstate->dtvs_tlocals[id] = *v;
9744 continue;
9745
9746 case DIFV_SCOPE_LOCAL:
9747 np = &vstate->dtvs_nlocals;
9748 svarp = &vstate->dtvs_locals;
9749
9750 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9751 dsize = NCPU * (v->dtdv_type.dtdt_size +
9752 sizeof (uint64_t));
9753 else
9754 dsize = NCPU * sizeof (uint64_t);
9755
9756 break;
9757
9758 case DIFV_SCOPE_GLOBAL:
9759 np = &vstate->dtvs_nglobals;
9760 svarp = &vstate->dtvs_globals;
9761
9762 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9763 dsize = v->dtdv_type.dtdt_size +
9764 sizeof (uint64_t);
9765
9766 break;
9767
9768 default:
9769 ASSERT(0);
9770 }
9771
9772 while (id >= (oldsvars = *np)) {
9773 dtrace_statvar_t **statics;
9774 int newsvars, oldsize, newsize;
9775
9776 if ((newsvars = (oldsvars << 1)) == 0)
9777 newsvars = 1;
9778
9779 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9780 newsize = newsvars * sizeof (dtrace_statvar_t *);
9781
9782 statics = kmem_zalloc(newsize, KM_SLEEP);
9783
9784 if (oldsize != 0) {
9785 bcopy(*svarp, statics, oldsize);
9786 kmem_free(*svarp, oldsize);
9787 }
9788
9789 *svarp = statics;
9790 *np = newsvars;
9791 }
9792
9793 if ((svar = (*svarp)[id]) == NULL) {
9794 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9795 svar->dtsv_var = *v;
9796
9797 if ((svar->dtsv_size = dsize) != 0) {
9798 svar->dtsv_data = (uint64_t)(uintptr_t)
9799 kmem_zalloc(dsize, KM_SLEEP);
9800 }
9801
9802 (*svarp)[id] = svar;
9803 }
9804
9805 svar->dtsv_refcnt++;
9806 }
9807
9808 dtrace_difo_chunksize(dp, vstate);
9809 dtrace_difo_hold(dp);
9810 }
9811
9812 static dtrace_difo_t *
9813 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9814 {
9815 dtrace_difo_t *new;
9816 size_t sz;
9817
9818 ASSERT(dp->dtdo_buf != NULL);
9819 ASSERT(dp->dtdo_refcnt != 0);
9820
9821 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9822
9823 ASSERT(dp->dtdo_buf != NULL);
9824 sz = dp->dtdo_len * sizeof (dif_instr_t);
9825 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9826 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9827 new->dtdo_len = dp->dtdo_len;
9828
9829 if (dp->dtdo_strtab != NULL) {
9830 ASSERT(dp->dtdo_strlen != 0);
9831 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9832 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9833 new->dtdo_strlen = dp->dtdo_strlen;
9834 }
9835
9836 if (dp->dtdo_inttab != NULL) {
9837 ASSERT(dp->dtdo_intlen != 0);
9838 sz = dp->dtdo_intlen * sizeof (uint64_t);
9839 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9840 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9841 new->dtdo_intlen = dp->dtdo_intlen;
9842 }
9843
9844 if (dp->dtdo_vartab != NULL) {
9845 ASSERT(dp->dtdo_varlen != 0);
9846 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9847 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9848 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9849 new->dtdo_varlen = dp->dtdo_varlen;
9850 }
9851
9852 dtrace_difo_init(new, vstate);
9853 return (new);
9854 }
9855
9856 static void
9857 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9858 {
9859 int i;
9860
9861 ASSERT(dp->dtdo_refcnt == 0);
9862
9863 for (i = 0; i < dp->dtdo_varlen; i++) {
9864 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9865 dtrace_statvar_t *svar, **svarp;
9866 uint_t id;
9867 uint8_t scope = v->dtdv_scope;
9868 int *np;
9869
9870 switch (scope) {
9871 case DIFV_SCOPE_THREAD:
9872 continue;
9873
9874 case DIFV_SCOPE_LOCAL:
9875 np = &vstate->dtvs_nlocals;
9876 svarp = vstate->dtvs_locals;
9877 break;
9878
9879 case DIFV_SCOPE_GLOBAL:
9880 np = &vstate->dtvs_nglobals;
9881 svarp = vstate->dtvs_globals;
9882 break;
9883
9884 default:
9885 ASSERT(0);
9886 }
9887
9888 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9889 continue;
9890
9891 id -= DIF_VAR_OTHER_UBASE;
9892 ASSERT(id < *np);
9893
9894 svar = svarp[id];
9895 ASSERT(svar != NULL);
9896 ASSERT(svar->dtsv_refcnt > 0);
9897
9898 if (--svar->dtsv_refcnt > 0)
9899 continue;
9900
9901 if (svar->dtsv_size != 0) {
9902 ASSERT(svar->dtsv_data != NULL);
9903 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9904 svar->dtsv_size);
9905 }
9906
9907 kmem_free(svar, sizeof (dtrace_statvar_t));
9908 svarp[id] = NULL;
9909 }
9910
9911 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9912 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9913 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9914 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9915
9916 kmem_free(dp, sizeof (dtrace_difo_t));
9917 }
9918
9919 static void
9920 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9921 {
9922 int i;
9923
9924 ASSERT(MUTEX_HELD(&dtrace_lock));
9925 ASSERT(dp->dtdo_refcnt != 0);
9926
9927 for (i = 0; i < dp->dtdo_varlen; i++) {
9928 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9929
9930 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9931 continue;
9932
9933 ASSERT(dtrace_vtime_references > 0);
9934 if (--dtrace_vtime_references == 0)
9935 dtrace_vtime_disable();
9936 }
9937
9938 if (--dp->dtdo_refcnt == 0)
9939 dtrace_difo_destroy(dp, vstate);
9940 }
9941
9942 /*
9943 * DTrace Format Functions
9944 */
9945 static uint16_t
9946 dtrace_format_add(dtrace_state_t *state, char *str)
9947 {
9948 char *fmt, **new;
9949 uint16_t ndx, len = strlen(str) + 1;
9950
9951 fmt = kmem_zalloc(len, KM_SLEEP);
9952 bcopy(str, fmt, len);
9953
9954 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9955 if (state->dts_formats[ndx] == NULL) {
9956 state->dts_formats[ndx] = fmt;
9957 return (ndx + 1);
9958 }
9959 }
9960
9961 if (state->dts_nformats == USHRT_MAX) {
9962 /*
9963 * This is only likely if a denial-of-service attack is being
9964 * attempted. As such, it's okay to fail silently here.
9965 */
9966 kmem_free(fmt, len);
9967 return (0);
9968 }
9969
9970 /*
9971 * For simplicity, we always resize the formats array to be exactly the
9972 * number of formats.
9973 */
9974 ndx = state->dts_nformats++;
9975 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9976
9977 if (state->dts_formats != NULL) {
9978 ASSERT(ndx != 0);
9979 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9980 kmem_free(state->dts_formats, ndx * sizeof (char *));
9981 }
9982
9983 state->dts_formats = new;
9984 state->dts_formats[ndx] = fmt;
9985
9986 return (ndx + 1);
9987 }
9988
9989 static void
9990 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9991 {
9992 char *fmt;
9993
9994 ASSERT(state->dts_formats != NULL);
9995 ASSERT(format <= state->dts_nformats);
9996 ASSERT(state->dts_formats[format - 1] != NULL);
9997
9998 fmt = state->dts_formats[format - 1];
9999 kmem_free(fmt, strlen(fmt) + 1);
10000 state->dts_formats[format - 1] = NULL;
10001 }
10002
10003 static void
10004 dtrace_format_destroy(dtrace_state_t *state)
10005 {
10006 int i;
10007
10008 if (state->dts_nformats == 0) {
10009 ASSERT(state->dts_formats == NULL);
10010 return;
10011 }
10012
10013 ASSERT(state->dts_formats != NULL);
10014
10015 for (i = 0; i < state->dts_nformats; i++) {
10016 char *fmt = state->dts_formats[i];
10017
10018 if (fmt == NULL)
10019 continue;
10020
10021 kmem_free(fmt, strlen(fmt) + 1);
10022 }
10023
10024 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10025 state->dts_nformats = 0;
10026 state->dts_formats = NULL;
10027 }
10028
10029 /*
10030 * DTrace Predicate Functions
10031 */
10032 static dtrace_predicate_t *
10033 dtrace_predicate_create(dtrace_difo_t *dp)
10034 {
10035 dtrace_predicate_t *pred;
10036
10037 ASSERT(MUTEX_HELD(&dtrace_lock));
10038 ASSERT(dp->dtdo_refcnt != 0);
10039
10040 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10041 pred->dtp_difo = dp;
10042 pred->dtp_refcnt = 1;
10043
10044 if (!dtrace_difo_cacheable(dp))
10045 return (pred);
10046
10047 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10048 /*
10049 * This is only theoretically possible -- we have had 2^32
10050 * cacheable predicates on this machine. We cannot allow any
10051 * more predicates to become cacheable: as unlikely as it is,
10052 * there may be a thread caching a (now stale) predicate cache
10053 * ID. (N.B.: the temptation is being successfully resisted to
10054 * have this cmn_err() "Holy shit -- we executed this code!")
10055 */
10056 return (pred);
10057 }
10058
10059 pred->dtp_cacheid = dtrace_predcache_id++;
10060
10061 return (pred);
10062 }
10063
10064 static void
10065 dtrace_predicate_hold(dtrace_predicate_t *pred)
10066 {
10067 ASSERT(MUTEX_HELD(&dtrace_lock));
10068 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10069 ASSERT(pred->dtp_refcnt > 0);
10070
10071 pred->dtp_refcnt++;
10072 }
10073
10074 static void
10075 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10076 {
10077 dtrace_difo_t *dp = pred->dtp_difo;
10078
10079 ASSERT(MUTEX_HELD(&dtrace_lock));
10080 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
10081 ASSERT(pred->dtp_refcnt > 0);
10082
10083 if (--pred->dtp_refcnt == 0) {
10084 dtrace_difo_release(pred->dtp_difo, vstate);
10085 kmem_free(pred, sizeof (dtrace_predicate_t));
10086 }
10087 }
10088
10089 /*
10090 * DTrace Action Description Functions
10091 */
10092 static dtrace_actdesc_t *
10093 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
10094 uint64_t uarg, uint64_t arg)
10095 {
10096 dtrace_actdesc_t *act;
10097
10098 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
10099 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
10100
10101 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
10102 act->dtad_kind = kind;
10103 act->dtad_ntuple = ntuple;
10104 act->dtad_uarg = uarg;
10105 act->dtad_arg = arg;
10106 act->dtad_refcnt = 1;
10107
10108 return (act);
10109 }
10110
10111 static void
10112 dtrace_actdesc_hold(dtrace_actdesc_t *act)
10113 {
10114 ASSERT(act->dtad_refcnt >= 1);
10115 act->dtad_refcnt++;
10116 }
10117
10118 static void
10119 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
10120 {
10121 dtrace_actkind_t kind = act->dtad_kind;
10122 dtrace_difo_t *dp;
10123
10124 ASSERT(act->dtad_refcnt >= 1);
10125
10126 if (--act->dtad_refcnt != 0)
10127 return;
10128
10129 if ((dp = act->dtad_difo) != NULL)
10130 dtrace_difo_release(dp, vstate);
10131
10132 if (DTRACEACT_ISPRINTFLIKE(kind)) {
10133 char *str = (char *)(uintptr_t)act->dtad_arg;
10134
10135 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
10136 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
10137
10138 if (str != NULL)
10139 kmem_free(str, strlen(str) + 1);
10140 }
10141
10142 kmem_free(act, sizeof (dtrace_actdesc_t));
10143 }
10144
10145 /*
10146 * DTrace ECB Functions
10147 */
10148 static dtrace_ecb_t *
10149 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
10150 {
10151 dtrace_ecb_t *ecb;
10152 dtrace_epid_t epid;
10153
10154 ASSERT(MUTEX_HELD(&dtrace_lock));
10155
10156 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
10157 ecb->dte_predicate = NULL;
10158 ecb->dte_probe = probe;
10159
10160 /*
10161 * The default size is the size of the default action: recording
10162 * the header.
10163 */
10164 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
10165 ecb->dte_alignment = sizeof (dtrace_epid_t);
10166
10167 epid = state->dts_epid++;
10168
10169 if (epid - 1 >= state->dts_necbs) {
10170 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
10171 int necbs = state->dts_necbs << 1;
10172
10173 ASSERT(epid == state->dts_necbs + 1);
10174
10175 if (necbs == 0) {
10176 ASSERT(oecbs == NULL);
10177 necbs = 1;
10178 }
10179
10180 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
10181
10182 if (oecbs != NULL)
10183 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
10184
10185 dtrace_membar_producer();
10186 state->dts_ecbs = ecbs;
10187
10188 if (oecbs != NULL) {
10189 /*
10190 * If this state is active, we must dtrace_sync()
10191 * before we can free the old dts_ecbs array: we're
10192 * coming in hot, and there may be active ring
10193 * buffer processing (which indexes into the dts_ecbs
10194 * array) on another CPU.
10195 */
10196 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
10197 dtrace_sync();
10198
10199 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
10200 }
10201
10202 dtrace_membar_producer();
10203 state->dts_necbs = necbs;
10204 }
10205
10206 ecb->dte_state = state;
10207
10208 ASSERT(state->dts_ecbs[epid - 1] == NULL);
10209 dtrace_membar_producer();
10210 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
10211
10212 return (ecb);
10213 }
10214
10215 static int
10216 dtrace_ecb_enable(dtrace_ecb_t *ecb)
10217 {
10218 dtrace_probe_t *probe = ecb->dte_probe;
10219
10220 ASSERT(MUTEX_HELD(&cpu_lock));
10221 ASSERT(MUTEX_HELD(&dtrace_lock));
10222 ASSERT(ecb->dte_next == NULL);
10223
10224 if (probe == NULL) {
10225 /*
10226 * This is the NULL probe -- there's nothing to do.
10227 */
10228 return (0);
10229 }
10230
10231 if (probe->dtpr_ecb == NULL) {
10232 dtrace_provider_t *prov = probe->dtpr_provider;
10233
10234 /*
10235 * We're the first ECB on this probe.
10236 */
10237 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
10238
10239 if (ecb->dte_predicate != NULL)
10240 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
10241
10242 return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
10243 probe->dtpr_id, probe->dtpr_arg));
10244 } else {
10245 /*
10246 * This probe is already active. Swing the last pointer to
10247 * point to the new ECB, and issue a dtrace_sync() to assure
10248 * that all CPUs have seen the change.
10249 */
10250 ASSERT(probe->dtpr_ecb_last != NULL);
10251 probe->dtpr_ecb_last->dte_next = ecb;
10252 probe->dtpr_ecb_last = ecb;
10253 probe->dtpr_predcache = 0;
10254
10255 dtrace_sync();
10256 return (0);
10257 }
10258 }
10259
10260 static void
10261 dtrace_ecb_resize(dtrace_ecb_t *ecb)
10262 {
10263 dtrace_action_t *act;
10264 uint32_t curneeded = UINT32_MAX;
10265 uint32_t aggbase = UINT32_MAX;
10266
10267 /*
10268 * If we record anything, we always record the dtrace_rechdr_t. (And
10269 * we always record it first.)
10270 */
10271 ecb->dte_size = sizeof (dtrace_rechdr_t);
10272 ecb->dte_alignment = sizeof (dtrace_epid_t);
10273
10274 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10275 dtrace_recdesc_t *rec = &act->dta_rec;
10276 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
10277
10278 ecb->dte_alignment = MAX(ecb->dte_alignment,
10279 rec->dtrd_alignment);
10280
10281 if (DTRACEACT_ISAGG(act->dta_kind)) {
10282 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10283
10284 ASSERT(rec->dtrd_size != 0);
10285 ASSERT(agg->dtag_first != NULL);
10286 ASSERT(act->dta_prev->dta_intuple);
10287 ASSERT(aggbase != UINT32_MAX);
10288 ASSERT(curneeded != UINT32_MAX);
10289
10290 agg->dtag_base = aggbase;
10291
10292 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10293 rec->dtrd_offset = curneeded;
10294 curneeded += rec->dtrd_size;
10295 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
10296
10297 aggbase = UINT32_MAX;
10298 curneeded = UINT32_MAX;
10299 } else if (act->dta_intuple) {
10300 if (curneeded == UINT32_MAX) {
10301 /*
10302 * This is the first record in a tuple. Align
10303 * curneeded to be at offset 4 in an 8-byte
10304 * aligned block.
10305 */
10306 ASSERT(act->dta_prev == NULL ||
10307 !act->dta_prev->dta_intuple);
10308 ASSERT3U(aggbase, ==, UINT32_MAX);
10309 curneeded = P2PHASEUP(ecb->dte_size,
10310 sizeof (uint64_t), sizeof (dtrace_aggid_t));
10311
10312 aggbase = curneeded - sizeof (dtrace_aggid_t);
10313 ASSERT(IS_P2ALIGNED(aggbase,
10314 sizeof (uint64_t)));
10315 }
10316 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10317 rec->dtrd_offset = curneeded;
10318 curneeded += rec->dtrd_size;
10319 } else {
10320 /* tuples must be followed by an aggregation */
10321 ASSERT(act->dta_prev == NULL ||
10322 !act->dta_prev->dta_intuple);
10323
10324 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10325 rec->dtrd_alignment);
10326 rec->dtrd_offset = ecb->dte_size;
10327 ecb->dte_size += rec->dtrd_size;
10328 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10329 }
10330 }
10331
10332 if ((act = ecb->dte_action) != NULL &&
10333 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10334 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10335 /*
10336 * If the size is still sizeof (dtrace_rechdr_t), then all
10337 * actions store no data; set the size to 0.
10338 */
10339 ecb->dte_size = 0;
10340 }
10341
10342 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10343 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10344 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10345 ecb->dte_needed);
10346 }
10347
10348 static dtrace_action_t *
10349 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10350 {
10351 dtrace_aggregation_t *agg;
10352 size_t size = sizeof (uint64_t);
10353 int ntuple = desc->dtad_ntuple;
10354 dtrace_action_t *act;
10355 dtrace_recdesc_t *frec;
10356 dtrace_aggid_t aggid;
10357 dtrace_state_t *state = ecb->dte_state;
10358
10359 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10360 agg->dtag_ecb = ecb;
10361
10362 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10363
10364 switch (desc->dtad_kind) {
10365 case DTRACEAGG_MIN:
10366 agg->dtag_initial = INT64_MAX;
10367 agg->dtag_aggregate = dtrace_aggregate_min;
10368 break;
10369
10370 case DTRACEAGG_MAX:
10371 agg->dtag_initial = INT64_MIN;
10372 agg->dtag_aggregate = dtrace_aggregate_max;
10373 break;
10374
10375 case DTRACEAGG_COUNT:
10376 agg->dtag_aggregate = dtrace_aggregate_count;
10377 break;
10378
10379 case DTRACEAGG_QUANTIZE:
10380 agg->dtag_aggregate = dtrace_aggregate_quantize;
10381 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10382 sizeof (uint64_t);
10383 break;
10384
10385 case DTRACEAGG_LQUANTIZE: {
10386 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10387 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10388
10389 agg->dtag_initial = desc->dtad_arg;
10390 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10391
10392 if (step == 0 || levels == 0)
10393 goto err;
10394
10395 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10396 break;
10397 }
10398
10399 case DTRACEAGG_LLQUANTIZE: {
10400 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10401 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10402 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10403 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10404 int64_t v;
10405
10406 agg->dtag_initial = desc->dtad_arg;
10407 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10408
10409 if (factor < 2 || low >= high || nsteps < factor)
10410 goto err;
10411
10412 /*
10413 * Now check that the number of steps evenly divides a power
10414 * of the factor. (This assures both integer bucket size and
10415 * linearity within each magnitude.)
10416 */
10417 for (v = factor; v < nsteps; v *= factor)
10418 continue;
10419
10420 if ((v % nsteps) || (nsteps % factor))
10421 goto err;
10422
10423 size = (dtrace_aggregate_llquantize_bucket(factor,
10424 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10425 break;
10426 }
10427
10428 case DTRACEAGG_AVG:
10429 agg->dtag_aggregate = dtrace_aggregate_avg;
10430 size = sizeof (uint64_t) * 2;
10431 break;
10432
10433 case DTRACEAGG_STDDEV:
10434 agg->dtag_aggregate = dtrace_aggregate_stddev;
10435 size = sizeof (uint64_t) * 4;
10436 break;
10437
10438 case DTRACEAGG_SUM:
10439 agg->dtag_aggregate = dtrace_aggregate_sum;
10440 break;
10441
10442 default:
10443 goto err;
10444 }
10445
10446 agg->dtag_action.dta_rec.dtrd_size = size;
10447
10448 if (ntuple == 0)
10449 goto err;
10450
10451 /*
10452 * We must make sure that we have enough actions for the n-tuple.
10453 */
10454 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10455 if (DTRACEACT_ISAGG(act->dta_kind))
10456 break;
10457
10458 if (--ntuple == 0) {
10459 /*
10460 * This is the action with which our n-tuple begins.
10461 */
10462 agg->dtag_first = act;
10463 goto success;
10464 }
10465 }
10466
10467 /*
10468 * This n-tuple is short by ntuple elements. Return failure.
10469 */
10470 ASSERT(ntuple != 0);
10471 err:
10472 kmem_free(agg, sizeof (dtrace_aggregation_t));
10473 return (NULL);
10474
10475 success:
10476 /*
10477 * If the last action in the tuple has a size of zero, it's actually
10478 * an expression argument for the aggregating action.
10479 */
10480 ASSERT(ecb->dte_action_last != NULL);
10481 act = ecb->dte_action_last;
10482
10483 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10484 ASSERT(act->dta_difo != NULL);
10485
10486 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10487 agg->dtag_hasarg = 1;
10488 }
10489
10490 /*
10491 * We need to allocate an id for this aggregation.
10492 */
10493 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10494 VM_BESTFIT | VM_SLEEP);
10495
10496 if (aggid - 1 >= state->dts_naggregations) {
10497 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10498 dtrace_aggregation_t **aggs;
10499 int naggs = state->dts_naggregations << 1;
10500 int onaggs = state->dts_naggregations;
10501
10502 ASSERT(aggid == state->dts_naggregations + 1);
10503
10504 if (naggs == 0) {
10505 ASSERT(oaggs == NULL);
10506 naggs = 1;
10507 }
10508
10509 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10510
10511 if (oaggs != NULL) {
10512 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10513 kmem_free(oaggs, onaggs * sizeof (*aggs));
10514 }
10515
10516 state->dts_aggregations = aggs;
10517 state->dts_naggregations = naggs;
10518 }
10519
10520 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10521 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10522
10523 frec = &agg->dtag_first->dta_rec;
10524 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10525 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10526
10527 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10528 ASSERT(!act->dta_intuple);
10529 act->dta_intuple = 1;
10530 }
10531
10532 return (&agg->dtag_action);
10533 }
10534
10535 static void
10536 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10537 {
10538 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10539 dtrace_state_t *state = ecb->dte_state;
10540 dtrace_aggid_t aggid = agg->dtag_id;
10541
10542 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10543 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10544
10545 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10546 state->dts_aggregations[aggid - 1] = NULL;
10547
10548 kmem_free(agg, sizeof (dtrace_aggregation_t));
10549 }
10550
10551 static int
10552 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10553 {
10554 dtrace_action_t *action, *last;
10555 dtrace_difo_t *dp = desc->dtad_difo;
10556 uint32_t size = 0, align = sizeof (uint8_t), mask;
10557 uint16_t format = 0;
10558 dtrace_recdesc_t *rec;
10559 dtrace_state_t *state = ecb->dte_state;
10560 dtrace_optval_t *opt = state->dts_options, nframes, strsize;
10561 uint64_t arg = desc->dtad_arg;
10562
10563 ASSERT(MUTEX_HELD(&dtrace_lock));
10564 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10565
10566 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10567 /*
10568 * If this is an aggregating action, there must be neither
10569 * a speculate nor a commit on the action chain.
10570 */
10571 dtrace_action_t *act;
10572
10573 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10574 if (act->dta_kind == DTRACEACT_COMMIT)
10575 return (EINVAL);
10576
10577 if (act->dta_kind == DTRACEACT_SPECULATE)
10578 return (EINVAL);
10579 }
10580
10581 action = dtrace_ecb_aggregation_create(ecb, desc);
10582
10583 if (action == NULL)
10584 return (EINVAL);
10585 } else {
10586 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10587 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10588 dp != NULL && dp->dtdo_destructive)) {
10589 state->dts_destructive = 1;
10590 }
10591
10592 switch (desc->dtad_kind) {
10593 case DTRACEACT_PRINTF:
10594 case DTRACEACT_PRINTA:
10595 case DTRACEACT_SYSTEM:
10596 case DTRACEACT_FREOPEN:
10597 case DTRACEACT_DIFEXPR:
10598 /*
10599 * We know that our arg is a string -- turn it into a
10600 * format.
10601 */
10602 if (arg == NULL) {
10603 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10604 desc->dtad_kind == DTRACEACT_DIFEXPR);
10605 format = 0;
10606 } else {
10607 ASSERT(arg != NULL);
10608 ASSERT(arg > KERNELBASE);
10609 format = dtrace_format_add(state,
10610 (char *)(uintptr_t)arg);
10611 }
10612
10613 /*FALLTHROUGH*/
10614 case DTRACEACT_LIBACT:
10615 case DTRACEACT_TRACEMEM:
10616 case DTRACEACT_TRACEMEM_DYNSIZE:
10617 if (dp == NULL)
10618 return (EINVAL);
10619
10620 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10621 break;
10622
10623 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10624 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10625 return (EINVAL);
10626
10627 size = opt[DTRACEOPT_STRSIZE];
10628 }
10629
10630 break;
10631
10632 case DTRACEACT_STACK:
10633 if ((nframes = arg) == 0) {
10634 nframes = opt[DTRACEOPT_STACKFRAMES];
10635 ASSERT(nframes > 0);
10636 arg = nframes;
10637 }
10638
10639 size = nframes * sizeof (pc_t);
10640 break;
10641
10642 case DTRACEACT_JSTACK:
10643 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10644 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10645
10646 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10647 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10648
10649 arg = DTRACE_USTACK_ARG(nframes, strsize);
10650
10651 /*FALLTHROUGH*/
10652 case DTRACEACT_USTACK:
10653 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10654 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10655 strsize = DTRACE_USTACK_STRSIZE(arg);
10656 nframes = opt[DTRACEOPT_USTACKFRAMES];
10657 ASSERT(nframes > 0);
10658 arg = DTRACE_USTACK_ARG(nframes, strsize);
10659 }
10660
10661 /*
10662 * Save a slot for the pid.
10663 */
10664 size = (nframes + 1) * sizeof (uint64_t);
10665 size += DTRACE_USTACK_STRSIZE(arg);
10666 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10667
10668 break;
10669
10670 case DTRACEACT_SYM:
10671 case DTRACEACT_MOD:
10672 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10673 sizeof (uint64_t)) ||
10674 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10675 return (EINVAL);
10676 break;
10677
10678 case DTRACEACT_USYM:
10679 case DTRACEACT_UMOD:
10680 case DTRACEACT_UADDR:
10681 if (dp == NULL ||
10682 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10683 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10684 return (EINVAL);
10685
10686 /*
10687 * We have a slot for the pid, plus a slot for the
10688 * argument. To keep things simple (aligned with
10689 * bitness-neutral sizing), we store each as a 64-bit
10690 * quantity.
10691 */
10692 size = 2 * sizeof (uint64_t);
10693 break;
10694
10695 case DTRACEACT_STOP:
10696 case DTRACEACT_BREAKPOINT:
10697 case DTRACEACT_PANIC:
10698 break;
10699
10700 case DTRACEACT_CHILL:
10701 case DTRACEACT_DISCARD:
10702 case DTRACEACT_RAISE:
10703 if (dp == NULL)
10704 return (EINVAL);
10705 break;
10706
10707 case DTRACEACT_EXIT:
10708 if (dp == NULL ||
10709 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10710 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10711 return (EINVAL);
10712 break;
10713
10714 case DTRACEACT_SPECULATE:
10715 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10716 return (EINVAL);
10717
10718 if (dp == NULL)
10719 return (EINVAL);
10720
10721 state->dts_speculates = 1;
10722 break;
10723
10724 case DTRACEACT_COMMIT: {
10725 dtrace_action_t *act = ecb->dte_action;
10726
10727 for (; act != NULL; act = act->dta_next) {
10728 if (act->dta_kind == DTRACEACT_COMMIT)
10729 return (EINVAL);
10730 }
10731
10732 if (dp == NULL)
10733 return (EINVAL);
10734 break;
10735 }
10736
10737 default:
10738 return (EINVAL);
10739 }
10740
10741 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10742 /*
10743 * If this is a data-storing action or a speculate,
10744 * we must be sure that there isn't a commit on the
10745 * action chain.
10746 */
10747 dtrace_action_t *act = ecb->dte_action;
10748
10749 for (; act != NULL; act = act->dta_next) {
10750 if (act->dta_kind == DTRACEACT_COMMIT)
10751 return (EINVAL);
10752 }
10753 }
10754
10755 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10756 action->dta_rec.dtrd_size = size;
10757 }
10758
10759 action->dta_refcnt = 1;
10760 rec = &action->dta_rec;
10761 size = rec->dtrd_size;
10762
10763 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10764 if (!(size & mask)) {
10765 align = mask + 1;
10766 break;
10767 }
10768 }
10769
10770 action->dta_kind = desc->dtad_kind;
10771
10772 if ((action->dta_difo = dp) != NULL)
10773 dtrace_difo_hold(dp);
10774
10775 rec->dtrd_action = action->dta_kind;
10776 rec->dtrd_arg = arg;
10777 rec->dtrd_uarg = desc->dtad_uarg;
10778 rec->dtrd_alignment = (uint16_t)align;
10779 rec->dtrd_format = format;
10780
10781 if ((last = ecb->dte_action_last) != NULL) {
10782 ASSERT(ecb->dte_action != NULL);
10783 action->dta_prev = last;
10784 last->dta_next = action;
10785 } else {
10786 ASSERT(ecb->dte_action == NULL);
10787 ecb->dte_action = action;
10788 }
10789
10790 ecb->dte_action_last = action;
10791
10792 return (0);
10793 }
10794
10795 static void
10796 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10797 {
10798 dtrace_action_t *act = ecb->dte_action, *next;
10799 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10800 dtrace_difo_t *dp;
10801 uint16_t format;
10802
10803 if (act != NULL && act->dta_refcnt > 1) {
10804 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10805 act->dta_refcnt--;
10806 } else {
10807 for (; act != NULL; act = next) {
10808 next = act->dta_next;
10809 ASSERT(next != NULL || act == ecb->dte_action_last);
10810 ASSERT(act->dta_refcnt == 1);
10811
10812 if ((format = act->dta_rec.dtrd_format) != 0)
10813 dtrace_format_remove(ecb->dte_state, format);
10814
10815 if ((dp = act->dta_difo) != NULL)
10816 dtrace_difo_release(dp, vstate);
10817
10818 if (DTRACEACT_ISAGG(act->dta_kind)) {
10819 dtrace_ecb_aggregation_destroy(ecb, act);
10820 } else {
10821 kmem_free(act, sizeof (dtrace_action_t));
10822 }
10823 }
10824 }
10825
10826 ecb->dte_action = NULL;
10827 ecb->dte_action_last = NULL;
10828 ecb->dte_size = 0;
10829 }
10830
10831 static void
10832 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10833 {
10834 /*
10835 * We disable the ECB by removing it from its probe.
10836 */
10837 dtrace_ecb_t *pecb, *prev = NULL;
10838 dtrace_probe_t *probe = ecb->dte_probe;
10839
10840 ASSERT(MUTEX_HELD(&dtrace_lock));
10841
10842 if (probe == NULL) {
10843 /*
10844 * This is the NULL probe; there is nothing to disable.
10845 */
10846 return;
10847 }
10848
10849 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10850 if (pecb == ecb)
10851 break;
10852 prev = pecb;
10853 }
10854
10855 ASSERT(pecb != NULL);
10856
10857 if (prev == NULL) {
10858 probe->dtpr_ecb = ecb->dte_next;
10859 } else {
10860 prev->dte_next = ecb->dte_next;
10861 }
10862
10863 if (ecb == probe->dtpr_ecb_last) {
10864 ASSERT(ecb->dte_next == NULL);
10865 probe->dtpr_ecb_last = prev;
10866 }
10867
10868 /*
10869 * The ECB has been disconnected from the probe; now sync to assure
10870 * that all CPUs have seen the change before returning.
10871 */
10872 dtrace_sync();
10873
10874 if (probe->dtpr_ecb == NULL) {
10875 /*
10876 * That was the last ECB on the probe; clear the predicate
10877 * cache ID for the probe, disable it and sync one more time
10878 * to assure that we'll never hit it again.
10879 */
10880 dtrace_provider_t *prov = probe->dtpr_provider;
10881
10882 ASSERT(ecb->dte_next == NULL);
10883 ASSERT(probe->dtpr_ecb_last == NULL);
10884 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10885 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10886 probe->dtpr_id, probe->dtpr_arg);
10887 dtrace_sync();
10888 } else {
10889 /*
10890 * There is at least one ECB remaining on the probe. If there
10891 * is _exactly_ one, set the probe's predicate cache ID to be
10892 * the predicate cache ID of the remaining ECB.
10893 */
10894 ASSERT(probe->dtpr_ecb_last != NULL);
10895 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10896
10897 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10898 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10899
10900 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10901
10902 if (p != NULL)
10903 probe->dtpr_predcache = p->dtp_cacheid;
10904 }
10905
10906 ecb->dte_next = NULL;
10907 }
10908 }
10909
10910 static void
10911 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10912 {
10913 dtrace_state_t *state = ecb->dte_state;
10914 dtrace_vstate_t *vstate = &state->dts_vstate;
10915 dtrace_predicate_t *pred;
10916 dtrace_epid_t epid = ecb->dte_epid;
10917
10918 ASSERT(MUTEX_HELD(&dtrace_lock));
10919 ASSERT(ecb->dte_next == NULL);
10920 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10921
10922 if ((pred = ecb->dte_predicate) != NULL)
10923 dtrace_predicate_release(pred, vstate);
10924
10925 dtrace_ecb_action_remove(ecb);
10926
10927 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10928 state->dts_ecbs[epid - 1] = NULL;
10929
10930 kmem_free(ecb, sizeof (dtrace_ecb_t));
10931 }
10932
10933 static dtrace_ecb_t *
10934 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10935 dtrace_enabling_t *enab)
10936 {
10937 dtrace_ecb_t *ecb;
10938 dtrace_predicate_t *pred;
10939 dtrace_actdesc_t *act;
10940 dtrace_provider_t *prov;
10941 dtrace_ecbdesc_t *desc = enab->dten_current;
10942
10943 ASSERT(MUTEX_HELD(&dtrace_lock));
10944 ASSERT(state != NULL);
10945
10946 ecb = dtrace_ecb_add(state, probe);
10947 ecb->dte_uarg = desc->dted_uarg;
10948
10949 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10950 dtrace_predicate_hold(pred);
10951 ecb->dte_predicate = pred;
10952 }
10953
10954 if (probe != NULL) {
10955 /*
10956 * If the provider shows more leg than the consumer is old
10957 * enough to see, we need to enable the appropriate implicit
10958 * predicate bits to prevent the ecb from activating at
10959 * revealing times.
10960 *
10961 * Providers specifying DTRACE_PRIV_USER at register time
10962 * are stating that they need the /proc-style privilege
10963 * model to be enforced, and this is what DTRACE_COND_OWNER
10964 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10965 */
10966 prov = probe->dtpr_provider;
10967 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10968 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10969 ecb->dte_cond |= DTRACE_COND_OWNER;
10970
10971 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10972 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10973 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10974
10975 /*
10976 * If the provider shows us kernel innards and the user
10977 * is lacking sufficient privilege, enable the
10978 * DTRACE_COND_USERMODE implicit predicate.
10979 */
10980 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10981 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10982 ecb->dte_cond |= DTRACE_COND_USERMODE;
10983 }
10984
10985 if (dtrace_ecb_create_cache != NULL) {
10986 /*
10987 * If we have a cached ecb, we'll use its action list instead
10988 * of creating our own (saving both time and space).
10989 */
10990 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10991 dtrace_action_t *act = cached->dte_action;
10992
10993 if (act != NULL) {
10994 ASSERT(act->dta_refcnt > 0);
10995 act->dta_refcnt++;
10996 ecb->dte_action = act;
10997 ecb->dte_action_last = cached->dte_action_last;
10998 ecb->dte_needed = cached->dte_needed;
10999 ecb->dte_size = cached->dte_size;
11000 ecb->dte_alignment = cached->dte_alignment;
11001 }
11002
11003 return (ecb);
11004 }
11005
11006 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11007 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11008 dtrace_ecb_destroy(ecb);
11009 return (NULL);
11010 }
11011 }
11012
11013 dtrace_ecb_resize(ecb);
11014
11015 return (dtrace_ecb_create_cache = ecb);
11016 }
11017
11018 static int
11019 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11020 {
11021 dtrace_ecb_t *ecb;
11022 dtrace_enabling_t *enab = arg;
11023 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11024
11025 ASSERT(state != NULL);
11026
11027 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11028 /*
11029 * This probe was created in a generation for which this
11030 * enabling has previously created ECBs; we don't want to
11031 * enable it again, so just kick out.
11032 */
11033 return (DTRACE_MATCH_NEXT);
11034 }
11035
11036 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11037 return (DTRACE_MATCH_DONE);
11038
11039 if (dtrace_ecb_enable(ecb) < 0)
11040 return (DTRACE_MATCH_FAIL);
11041
11042 return (DTRACE_MATCH_NEXT);
11043 }
11044
11045 static dtrace_ecb_t *
11046 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11047 {
11048 dtrace_ecb_t *ecb;
11049
11050 ASSERT(MUTEX_HELD(&dtrace_lock));
11051
11052 if (id == 0 || id > state->dts_necbs)
11053 return (NULL);
11054
11055 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
11056 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
11057
11058 return (state->dts_ecbs[id - 1]);
11059 }
11060
11061 static dtrace_aggregation_t *
11062 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
11063 {
11064 dtrace_aggregation_t *agg;
11065
11066 ASSERT(MUTEX_HELD(&dtrace_lock));
11067
11068 if (id == 0 || id > state->dts_naggregations)
11069 return (NULL);
11070
11071 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
11072 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
11073 agg->dtag_id == id);
11074
11075 return (state->dts_aggregations[id - 1]);
11076 }
11077
11078 /*
11079 * DTrace Buffer Functions
11080 *
11081 * The following functions manipulate DTrace buffers. Most of these functions
11082 * are called in the context of establishing or processing consumer state;
11083 * exceptions are explicitly noted.
11084 */
11085
11086 /*
11087 * Note: called from cross call context. This function switches the two
11088 * buffers on a given CPU. The atomicity of this operation is assured by
11089 * disabling interrupts while the actual switch takes place; the disabling of
11090 * interrupts serializes the execution with any execution of dtrace_probe() on
11091 * the same CPU.
11092 */
11093 static void
11094 dtrace_buffer_switch(dtrace_buffer_t *buf)
11095 {
11096 caddr_t tomax = buf->dtb_tomax;
11097 caddr_t xamot = buf->dtb_xamot;
11098 dtrace_icookie_t cookie;
11099 hrtime_t now;
11100
11101 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11102 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
11103
11104 cookie = dtrace_interrupt_disable();
11105 now = dtrace_gethrtime();
11106 buf->dtb_tomax = xamot;
11107 buf->dtb_xamot = tomax;
11108 buf->dtb_xamot_drops = buf->dtb_drops;
11109 buf->dtb_xamot_offset = buf->dtb_offset;
11110 buf->dtb_xamot_errors = buf->dtb_errors;
11111 buf->dtb_xamot_flags = buf->dtb_flags;
11112 buf->dtb_offset = 0;
11113 buf->dtb_drops = 0;
11114 buf->dtb_errors = 0;
11115 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
11116 buf->dtb_interval = now - buf->dtb_switched;
11117 buf->dtb_switched = now;
11118 dtrace_interrupt_enable(cookie);
11119 }
11120
11121 /*
11122 * Note: called from cross call context. This function activates a buffer
11123 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
11124 * is guaranteed by the disabling of interrupts.
11125 */
11126 static void
11127 dtrace_buffer_activate(dtrace_state_t *state)
11128 {
11129 dtrace_buffer_t *buf;
11130 dtrace_icookie_t cookie = dtrace_interrupt_disable();
11131
11132 buf = &state->dts_buffer[CPU->cpu_id];
11133
11134 if (buf->dtb_tomax != NULL) {
11135 /*
11136 * We might like to assert that the buffer is marked inactive,
11137 * but this isn't necessarily true: the buffer for the CPU
11138 * that processes the BEGIN probe has its buffer activated
11139 * manually. In this case, we take the (harmless) action
11140 * re-clearing the bit INACTIVE bit.
11141 */
11142 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
11143 }
11144
11145 dtrace_interrupt_enable(cookie);
11146 }
11147
11148 static int
11149 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
11150 processorid_t cpu, int *factor)
11151 {
11152 cpu_t *cp;
11153 dtrace_buffer_t *buf;
11154 int allocated = 0, desired = 0;
11155
11156 ASSERT(MUTEX_HELD(&cpu_lock));
11157 ASSERT(MUTEX_HELD(&dtrace_lock));
11158
11159 *factor = 1;
11160
11161 if (size > dtrace_nonroot_maxsize &&
11162 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
11163 return (EFBIG);
11164
11165 cp = cpu_list;
11166
11167 do {
11168 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11169 continue;
11170
11171 buf = &bufs[cp->cpu_id];
11172
11173 /*
11174 * If there is already a buffer allocated for this CPU, it
11175 * is only possible that this is a DR event. In this case,
11176 * the buffer size must match our specified size.
11177 */
11178 if (buf->dtb_tomax != NULL) {
11179 ASSERT(buf->dtb_size == size);
11180 continue;
11181 }
11182
11183 ASSERT(buf->dtb_xamot == NULL);
11184
11185 if ((buf->dtb_tomax = kmem_zalloc(size,
11186 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11187 goto err;
11188
11189 buf->dtb_size = size;
11190 buf->dtb_flags = flags;
11191 buf->dtb_offset = 0;
11192 buf->dtb_drops = 0;
11193
11194 if (flags & DTRACEBUF_NOSWITCH)
11195 continue;
11196
11197 if ((buf->dtb_xamot = kmem_zalloc(size,
11198 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11199 goto err;
11200 } while ((cp = cp->cpu_next) != cpu_list);
11201
11202 return (0);
11203
11204 err:
11205 cp = cpu_list;
11206
11207 do {
11208 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11209 continue;
11210
11211 buf = &bufs[cp->cpu_id];
11212 desired += 2;
11213
11214 if (buf->dtb_xamot != NULL) {
11215 ASSERT(buf->dtb_tomax != NULL);
11216 ASSERT(buf->dtb_size == size);
11217 kmem_free(buf->dtb_xamot, size);
11218 allocated++;
11219 }
11220
11221 if (buf->dtb_tomax != NULL) {
11222 ASSERT(buf->dtb_size == size);
11223 kmem_free(buf->dtb_tomax, size);
11224 allocated++;
11225 }
11226
11227 buf->dtb_tomax = NULL;
11228 buf->dtb_xamot = NULL;
11229 buf->dtb_size = 0;
11230 } while ((cp = cp->cpu_next) != cpu_list);
11231
11232 *factor = desired / (allocated > 0 ? allocated : 1);
11233
11234 return (ENOMEM);
11235 }
11236
11237 /*
11238 * Note: called from probe context. This function just increments the drop
11239 * count on a buffer. It has been made a function to allow for the
11240 * possibility of understanding the source of mysterious drop counts. (A
11241 * problem for which one may be particularly disappointed that DTrace cannot
11242 * be used to understand DTrace.)
11243 */
11244 static void
11245 dtrace_buffer_drop(dtrace_buffer_t *buf)
11246 {
11247 buf->dtb_drops++;
11248 }
11249
11250 /*
11251 * Note: called from probe context. This function is called to reserve space
11252 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
11253 * mstate. Returns the new offset in the buffer, or a negative value if an
11254 * error has occurred.
11255 */
11256 static intptr_t
11257 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11258 dtrace_state_t *state, dtrace_mstate_t *mstate)
11259 {
11260 intptr_t offs = buf->dtb_offset, soffs;
11261 intptr_t woffs;
11262 caddr_t tomax;
11263 size_t total;
11264
11265 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11266 return (-1);
11267
11268 if ((tomax = buf->dtb_tomax) == NULL) {
11269 dtrace_buffer_drop(buf);
11270 return (-1);
11271 }
11272
11273 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11274 while (offs & (align - 1)) {
11275 /*
11276 * Assert that our alignment is off by a number which
11277 * is itself sizeof (uint32_t) aligned.
11278 */
11279 ASSERT(!((align - (offs & (align - 1))) &
11280 (sizeof (uint32_t) - 1)));
11281 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11282 offs += sizeof (uint32_t);
11283 }
11284
11285 if ((soffs = offs + needed) > buf->dtb_size) {
11286 dtrace_buffer_drop(buf);
11287 return (-1);
11288 }
11289
11290 if (mstate == NULL)
11291 return (offs);
11292
11293 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11294 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11295 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11296
11297 return (offs);
11298 }
11299
11300 if (buf->dtb_flags & DTRACEBUF_FILL) {
11301 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11302 (buf->dtb_flags & DTRACEBUF_FULL))
11303 return (-1);
11304 goto out;
11305 }
11306
11307 total = needed + (offs & (align - 1));
11308
11309 /*
11310 * For a ring buffer, life is quite a bit more complicated. Before
11311 * we can store any padding, we need to adjust our wrapping offset.
11312 * (If we've never before wrapped or we're not about to, no adjustment
11313 * is required.)
11314 */
11315 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11316 offs + total > buf->dtb_size) {
11317 woffs = buf->dtb_xamot_offset;
11318
11319 if (offs + total > buf->dtb_size) {
11320 /*
11321 * We can't fit in the end of the buffer. First, a
11322 * sanity check that we can fit in the buffer at all.
11323 */
11324 if (total > buf->dtb_size) {
11325 dtrace_buffer_drop(buf);
11326 return (-1);
11327 }
11328
11329 /*
11330 * We're going to be storing at the top of the buffer,
11331 * so now we need to deal with the wrapped offset. We
11332 * only reset our wrapped offset to 0 if it is
11333 * currently greater than the current offset. If it
11334 * is less than the current offset, it is because a
11335 * previous allocation induced a wrap -- but the
11336 * allocation didn't subsequently take the space due
11337 * to an error or false predicate evaluation. In this
11338 * case, we'll just leave the wrapped offset alone: if
11339 * the wrapped offset hasn't been advanced far enough
11340 * for this allocation, it will be adjusted in the
11341 * lower loop.
11342 */
11343 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11344 if (woffs >= offs)
11345 woffs = 0;
11346 } else {
11347 woffs = 0;
11348 }
11349
11350 /*
11351 * Now we know that we're going to be storing to the
11352 * top of the buffer and that there is room for us
11353 * there. We need to clear the buffer from the current
11354 * offset to the end (there may be old gunk there).
11355 */
11356 while (offs < buf->dtb_size)
11357 tomax[offs++] = 0;
11358
11359 /*
11360 * We need to set our offset to zero. And because we
11361 * are wrapping, we need to set the bit indicating as
11362 * much. We can also adjust our needed space back
11363 * down to the space required by the ECB -- we know
11364 * that the top of the buffer is aligned.
11365 */
11366 offs = 0;
11367 total = needed;
11368 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11369 } else {
11370 /*
11371 * There is room for us in the buffer, so we simply
11372 * need to check the wrapped offset.
11373 */
11374 if (woffs < offs) {
11375 /*
11376 * The wrapped offset is less than the offset.
11377 * This can happen if we allocated buffer space
11378 * that induced a wrap, but then we didn't
11379 * subsequently take the space due to an error
11380 * or false predicate evaluation. This is
11381 * okay; we know that _this_ allocation isn't
11382 * going to induce a wrap. We still can't
11383 * reset the wrapped offset to be zero,
11384 * however: the space may have been trashed in
11385 * the previous failed probe attempt. But at
11386 * least the wrapped offset doesn't need to
11387 * be adjusted at all...
11388 */
11389 goto out;
11390 }
11391 }
11392
11393 while (offs + total > woffs) {
11394 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11395 size_t size;
11396
11397 if (epid == DTRACE_EPIDNONE) {
11398 size = sizeof (uint32_t);
11399 } else {
11400 ASSERT3U(epid, <=, state->dts_necbs);
11401 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11402
11403 size = state->dts_ecbs[epid - 1]->dte_size;
11404 }
11405
11406 ASSERT(woffs + size <= buf->dtb_size);
11407 ASSERT(size != 0);
11408
11409 if (woffs + size == buf->dtb_size) {
11410 /*
11411 * We've reached the end of the buffer; we want
11412 * to set the wrapped offset to 0 and break
11413 * out. However, if the offs is 0, then we're
11414 * in a strange edge-condition: the amount of
11415 * space that we want to reserve plus the size
11416 * of the record that we're overwriting is
11417 * greater than the size of the buffer. This
11418 * is problematic because if we reserve the
11419 * space but subsequently don't consume it (due
11420 * to a failed predicate or error) the wrapped
11421 * offset will be 0 -- yet the EPID at offset 0
11422 * will not be committed. This situation is
11423 * relatively easy to deal with: if we're in
11424 * this case, the buffer is indistinguishable
11425 * from one that hasn't wrapped; we need only
11426 * finish the job by clearing the wrapped bit,
11427 * explicitly setting the offset to be 0, and
11428 * zero'ing out the old data in the buffer.
11429 */
11430 if (offs == 0) {
11431 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11432 buf->dtb_offset = 0;
11433 woffs = total;
11434
11435 while (woffs < buf->dtb_size)
11436 tomax[woffs++] = 0;
11437 }
11438
11439 woffs = 0;
11440 break;
11441 }
11442
11443 woffs += size;
11444 }
11445
11446 /*
11447 * We have a wrapped offset. It may be that the wrapped offset
11448 * has become zero -- that's okay.
11449 */
11450 buf->dtb_xamot_offset = woffs;
11451 }
11452
11453 out:
11454 /*
11455 * Now we can plow the buffer with any necessary padding.
11456 */
11457 while (offs & (align - 1)) {
11458 /*
11459 * Assert that our alignment is off by a number which
11460 * is itself sizeof (uint32_t) aligned.
11461 */
11462 ASSERT(!((align - (offs & (align - 1))) &
11463 (sizeof (uint32_t) - 1)));
11464 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11465 offs += sizeof (uint32_t);
11466 }
11467
11468 if (buf->dtb_flags & DTRACEBUF_FILL) {
11469 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11470 buf->dtb_flags |= DTRACEBUF_FULL;
11471 return (-1);
11472 }
11473 }
11474
11475 if (mstate == NULL)
11476 return (offs);
11477
11478 /*
11479 * For ring buffers and fill buffers, the scratch space is always
11480 * the inactive buffer.
11481 */
11482 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11483 mstate->dtms_scratch_size = buf->dtb_size;
11484 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11485
11486 return (offs);
11487 }
11488
11489 static void
11490 dtrace_buffer_polish(dtrace_buffer_t *buf)
11491 {
11492 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11493 ASSERT(MUTEX_HELD(&dtrace_lock));
11494
11495 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11496 return;
11497
11498 /*
11499 * We need to polish the ring buffer. There are three cases:
11500 *
11501 * - The first (and presumably most common) is that there is no gap
11502 * between the buffer offset and the wrapped offset. In this case,
11503 * there is nothing in the buffer that isn't valid data; we can
11504 * mark the buffer as polished and return.
11505 *
11506 * - The second (less common than the first but still more common
11507 * than the third) is that there is a gap between the buffer offset
11508 * and the wrapped offset, and the wrapped offset is larger than the
11509 * buffer offset. This can happen because of an alignment issue, or
11510 * can happen because of a call to dtrace_buffer_reserve() that
11511 * didn't subsequently consume the buffer space. In this case,
11512 * we need to zero the data from the buffer offset to the wrapped
11513 * offset.
11514 *
11515 * - The third (and least common) is that there is a gap between the
11516 * buffer offset and the wrapped offset, but the wrapped offset is
11517 * _less_ than the buffer offset. This can only happen because a
11518 * call to dtrace_buffer_reserve() induced a wrap, but the space
11519 * was not subsequently consumed. In this case, we need to zero the
11520 * space from the offset to the end of the buffer _and_ from the
11521 * top of the buffer to the wrapped offset.
11522 */
11523 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11524 bzero(buf->dtb_tomax + buf->dtb_offset,
11525 buf->dtb_xamot_offset - buf->dtb_offset);
11526 }
11527
11528 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11529 bzero(buf->dtb_tomax + buf->dtb_offset,
11530 buf->dtb_size - buf->dtb_offset);
11531 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11532 }
11533 }
11534
11535 /*
11536 * This routine determines if data generated at the specified time has likely
11537 * been entirely consumed at user-level. This routine is called to determine
11538 * if an ECB on a defunct probe (but for an active enabling) can be safely
11539 * disabled and destroyed.
11540 */
11541 static int
11542 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11543 {
11544 int i;
11545
11546 for (i = 0; i < NCPU; i++) {
11547 dtrace_buffer_t *buf = &bufs[i];
11548
11549 if (buf->dtb_size == 0)
11550 continue;
11551
11552 if (buf->dtb_flags & DTRACEBUF_RING)
11553 return (0);
11554
11555 if (!buf->dtb_switched && buf->dtb_offset != 0)
11556 return (0);
11557
11558 if (buf->dtb_switched - buf->dtb_interval < when)
11559 return (0);
11560 }
11561
11562 return (1);
11563 }
11564
11565 static void
11566 dtrace_buffer_free(dtrace_buffer_t *bufs)
11567 {
11568 int i;
11569
11570 for (i = 0; i < NCPU; i++) {
11571 dtrace_buffer_t *buf = &bufs[i];
11572
11573 if (buf->dtb_tomax == NULL) {
11574 ASSERT(buf->dtb_xamot == NULL);
11575 ASSERT(buf->dtb_size == 0);
11576 continue;
11577 }
11578
11579 if (buf->dtb_xamot != NULL) {
11580 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11581 kmem_free(buf->dtb_xamot, buf->dtb_size);
11582 }
11583
11584 kmem_free(buf->dtb_tomax, buf->dtb_size);
11585 buf->dtb_size = 0;
11586 buf->dtb_tomax = NULL;
11587 buf->dtb_xamot = NULL;
11588 }
11589 }
11590
11591 /*
11592 * DTrace Enabling Functions
11593 */
11594 static dtrace_enabling_t *
11595 dtrace_enabling_create(dtrace_vstate_t *vstate)
11596 {
11597 dtrace_enabling_t *enab;
11598
11599 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11600 enab->dten_vstate = vstate;
11601
11602 return (enab);
11603 }
11604
11605 static void
11606 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11607 {
11608 dtrace_ecbdesc_t **ndesc;
11609 size_t osize, nsize;
11610
11611 /*
11612 * We can't add to enablings after we've enabled them, or after we've
11613 * retained them.
11614 */
11615 ASSERT(enab->dten_probegen == 0);
11616 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11617
11618 if (enab->dten_ndesc < enab->dten_maxdesc) {
11619 enab->dten_desc[enab->dten_ndesc++] = ecb;
11620 return;
11621 }
11622
11623 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11624
11625 if (enab->dten_maxdesc == 0) {
11626 enab->dten_maxdesc = 1;
11627 } else {
11628 enab->dten_maxdesc <<= 1;
11629 }
11630
11631 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11632
11633 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11634 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11635 bcopy(enab->dten_desc, ndesc, osize);
11636 kmem_free(enab->dten_desc, osize);
11637
11638 enab->dten_desc = ndesc;
11639 enab->dten_desc[enab->dten_ndesc++] = ecb;
11640 }
11641
11642 static void
11643 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11644 dtrace_probedesc_t *pd)
11645 {
11646 dtrace_ecbdesc_t *new;
11647 dtrace_predicate_t *pred;
11648 dtrace_actdesc_t *act;
11649
11650 /*
11651 * We're going to create a new ECB description that matches the
11652 * specified ECB in every way, but has the specified probe description.
11653 */
11654 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11655
11656 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11657 dtrace_predicate_hold(pred);
11658
11659 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11660 dtrace_actdesc_hold(act);
11661
11662 new->dted_action = ecb->dted_action;
11663 new->dted_pred = ecb->dted_pred;
11664 new->dted_probe = *pd;
11665 new->dted_uarg = ecb->dted_uarg;
11666
11667 dtrace_enabling_add(enab, new);
11668 }
11669
11670 static void
11671 dtrace_enabling_dump(dtrace_enabling_t *enab)
11672 {
11673 int i;
11674
11675 for (i = 0; i < enab->dten_ndesc; i++) {
11676 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11677
11678 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11679 desc->dtpd_provider, desc->dtpd_mod,
11680 desc->dtpd_func, desc->dtpd_name);
11681 }
11682 }
11683
11684 static void
11685 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11686 {
11687 int i;
11688 dtrace_ecbdesc_t *ep;
11689 dtrace_vstate_t *vstate = enab->dten_vstate;
11690
11691 ASSERT(MUTEX_HELD(&dtrace_lock));
11692
11693 for (i = 0; i < enab->dten_ndesc; i++) {
11694 dtrace_actdesc_t *act, *next;
11695 dtrace_predicate_t *pred;
11696
11697 ep = enab->dten_desc[i];
11698
11699 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11700 dtrace_predicate_release(pred, vstate);
11701
11702 for (act = ep->dted_action; act != NULL; act = next) {
11703 next = act->dtad_next;
11704 dtrace_actdesc_release(act, vstate);
11705 }
11706
11707 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11708 }
11709
11710 kmem_free(enab->dten_desc,
11711 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11712
11713 /*
11714 * If this was a retained enabling, decrement the dts_nretained count
11715 * and take it off of the dtrace_retained list.
11716 */
11717 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11718 dtrace_retained == enab) {
11719 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11720 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11721 enab->dten_vstate->dtvs_state->dts_nretained--;
11722 dtrace_retained_gen++;
11723 }
11724
11725 if (enab->dten_prev == NULL) {
11726 if (dtrace_retained == enab) {
11727 dtrace_retained = enab->dten_next;
11728
11729 if (dtrace_retained != NULL)
11730 dtrace_retained->dten_prev = NULL;
11731 }
11732 } else {
11733 ASSERT(enab != dtrace_retained);
11734 ASSERT(dtrace_retained != NULL);
11735 enab->dten_prev->dten_next = enab->dten_next;
11736 }
11737
11738 if (enab->dten_next != NULL) {
11739 ASSERT(dtrace_retained != NULL);
11740 enab->dten_next->dten_prev = enab->dten_prev;
11741 }
11742
11743 kmem_free(enab, sizeof (dtrace_enabling_t));
11744 }
11745
11746 static int
11747 dtrace_enabling_retain(dtrace_enabling_t *enab)
11748 {
11749 dtrace_state_t *state;
11750
11751 ASSERT(MUTEX_HELD(&dtrace_lock));
11752 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11753 ASSERT(enab->dten_vstate != NULL);
11754
11755 state = enab->dten_vstate->dtvs_state;
11756 ASSERT(state != NULL);
11757
11758 /*
11759 * We only allow each state to retain dtrace_retain_max enablings.
11760 */
11761 if (state->dts_nretained >= dtrace_retain_max)
11762 return (ENOSPC);
11763
11764 state->dts_nretained++;
11765 dtrace_retained_gen++;
11766
11767 if (dtrace_retained == NULL) {
11768 dtrace_retained = enab;
11769 return (0);
11770 }
11771
11772 enab->dten_next = dtrace_retained;
11773 dtrace_retained->dten_prev = enab;
11774 dtrace_retained = enab;
11775
11776 return (0);
11777 }
11778
11779 static int
11780 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11781 dtrace_probedesc_t *create)
11782 {
11783 dtrace_enabling_t *new, *enab;
11784 int found = 0, err = ENOENT;
11785
11786 ASSERT(MUTEX_HELD(&dtrace_lock));
11787 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11788 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11789 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11790 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11791
11792 new = dtrace_enabling_create(&state->dts_vstate);
11793
11794 /*
11795 * Iterate over all retained enablings, looking for enablings that
11796 * match the specified state.
11797 */
11798 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11799 int i;
11800
11801 /*
11802 * dtvs_state can only be NULL for helper enablings -- and
11803 * helper enablings can't be retained.
11804 */
11805 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11806
11807 if (enab->dten_vstate->dtvs_state != state)
11808 continue;
11809
11810 /*
11811 * Now iterate over each probe description; we're looking for
11812 * an exact match to the specified probe description.
11813 */
11814 for (i = 0; i < enab->dten_ndesc; i++) {
11815 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11816 dtrace_probedesc_t *pd = &ep->dted_probe;
11817
11818 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11819 continue;
11820
11821 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11822 continue;
11823
11824 if (strcmp(pd->dtpd_func, match->dtpd_func))
11825 continue;
11826
11827 if (strcmp(pd->dtpd_name, match->dtpd_name))
11828 continue;
11829
11830 /*
11831 * We have a winning probe! Add it to our growing
11832 * enabling.
11833 */
11834 found = 1;
11835 dtrace_enabling_addlike(new, ep, create);
11836 }
11837 }
11838
11839 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11840 dtrace_enabling_destroy(new);
11841 return (err);
11842 }
11843
11844 return (0);
11845 }
11846
11847 static void
11848 dtrace_enabling_retract(dtrace_state_t *state)
11849 {
11850 dtrace_enabling_t *enab, *next;
11851
11852 ASSERT(MUTEX_HELD(&dtrace_lock));
11853
11854 /*
11855 * Iterate over all retained enablings, destroy the enablings retained
11856 * for the specified state.
11857 */
11858 for (enab = dtrace_retained; enab != NULL; enab = next) {
11859 next = enab->dten_next;
11860
11861 /*
11862 * dtvs_state can only be NULL for helper enablings -- and
11863 * helper enablings can't be retained.
11864 */
11865 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11866
11867 if (enab->dten_vstate->dtvs_state == state) {
11868 ASSERT(state->dts_nretained > 0);
11869 dtrace_enabling_destroy(enab);
11870 }
11871 }
11872
11873 ASSERT(state->dts_nretained == 0);
11874 }
11875
11876 static int
11877 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11878 {
11879 int i = 0;
11880 int total_matched = 0, matched = 0;
11881
11882 ASSERT(MUTEX_HELD(&cpu_lock));
11883 ASSERT(MUTEX_HELD(&dtrace_lock));
11884
11885 for (i = 0; i < enab->dten_ndesc; i++) {
11886 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11887
11888 enab->dten_current = ep;
11889 enab->dten_error = 0;
11890
11891 /*
11892 * If a provider failed to enable a probe then get out and
11893 * let the consumer know we failed.
11894 */
11895 if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
11896 return (EBUSY);
11897
11898 total_matched += matched;
11899
11900 if (enab->dten_error != 0) {
11901 /*
11902 * If we get an error half-way through enabling the
11903 * probes, we kick out -- perhaps with some number of
11904 * them enabled. Leaving enabled probes enabled may
11905 * be slightly confusing for user-level, but we expect
11906 * that no one will attempt to actually drive on in
11907 * the face of such errors. If this is an anonymous
11908 * enabling (indicated with a NULL nmatched pointer),
11909 * we cmn_err() a message. We aren't expecting to
11910 * get such an error -- such as it can exist at all,
11911 * it would be a result of corrupted DOF in the driver
11912 * properties.
11913 */
11914 if (nmatched == NULL) {
11915 cmn_err(CE_WARN, "dtrace_enabling_match() "
11916 "error on %p: %d", (void *)ep,
11917 enab->dten_error);
11918 }
11919
11920 return (enab->dten_error);
11921 }
11922 }
11923
11924 enab->dten_probegen = dtrace_probegen;
11925 if (nmatched != NULL)
11926 *nmatched = total_matched;
11927
11928 return (0);
11929 }
11930
11931 static void
11932 dtrace_enabling_matchall(void)
11933 {
11934 dtrace_enabling_t *enab;
11935
11936 mutex_enter(&cpu_lock);
11937 mutex_enter(&dtrace_lock);
11938
11939 /*
11940 * Iterate over all retained enablings to see if any probes match
11941 * against them. We only perform this operation on enablings for which
11942 * we have sufficient permissions by virtue of being in the global zone
11943 * or in the same zone as the DTrace client. Because we can be called
11944 * after dtrace_detach() has been called, we cannot assert that there
11945 * are retained enablings. We can safely load from dtrace_retained,
11946 * however: the taskq_destroy() at the end of dtrace_detach() will
11947 * block pending our completion.
11948 */
11949 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11950 dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
11951 cred_t *cr = dcr->dcr_cred;
11952 zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
11953
11954 if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
11955 (zone == GLOBAL_ZONEID || getzoneid() == zone)))
11956 (void) dtrace_enabling_match(enab, NULL);
11957 }
11958
11959 mutex_exit(&dtrace_lock);
11960 mutex_exit(&cpu_lock);
11961 }
11962
11963 /*
11964 * If an enabling is to be enabled without having matched probes (that is, if
11965 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11966 * enabling must be _primed_ by creating an ECB for every ECB description.
11967 * This must be done to assure that we know the number of speculations, the
11968 * number of aggregations, the minimum buffer size needed, etc. before we
11969 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11970 * enabling any probes, we create ECBs for every ECB decription, but with a
11971 * NULL probe -- which is exactly what this function does.
11972 */
11973 static void
11974 dtrace_enabling_prime(dtrace_state_t *state)
11975 {
11976 dtrace_enabling_t *enab;
11977 int i;
11978
11979 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11980 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11981
11982 if (enab->dten_vstate->dtvs_state != state)
11983 continue;
11984
11985 /*
11986 * We don't want to prime an enabling more than once, lest
11987 * we allow a malicious user to induce resource exhaustion.
11988 * (The ECBs that result from priming an enabling aren't
11989 * leaked -- but they also aren't deallocated until the
11990 * consumer state is destroyed.)
11991 */
11992 if (enab->dten_primed)
11993 continue;
11994
11995 for (i = 0; i < enab->dten_ndesc; i++) {
11996 enab->dten_current = enab->dten_desc[i];
11997 (void) dtrace_probe_enable(NULL, enab);
11998 }
11999
12000 enab->dten_primed = 1;
12001 }
12002 }
12003
12004 /*
12005 * Called to indicate that probes should be provided due to retained
12006 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
12007 * must take an initial lap through the enabling calling the dtps_provide()
12008 * entry point explicitly to allow for autocreated probes.
12009 */
12010 static void
12011 dtrace_enabling_provide(dtrace_provider_t *prv)
12012 {
12013 int i, all = 0;
12014 dtrace_probedesc_t desc;
12015 dtrace_genid_t gen;
12016
12017 ASSERT(MUTEX_HELD(&dtrace_lock));
12018 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
12019
12020 if (prv == NULL) {
12021 all = 1;
12022 prv = dtrace_provider;
12023 }
12024
12025 do {
12026 dtrace_enabling_t *enab;
12027 void *parg = prv->dtpv_arg;
12028
12029 retry:
12030 gen = dtrace_retained_gen;
12031 for (enab = dtrace_retained; enab != NULL;
12032 enab = enab->dten_next) {
12033 for (i = 0; i < enab->dten_ndesc; i++) {
12034 desc = enab->dten_desc[i]->dted_probe;
12035 mutex_exit(&dtrace_lock);
12036 prv->dtpv_pops.dtps_provide(parg, &desc);
12037 mutex_enter(&dtrace_lock);
12038 /*
12039 * Process the retained enablings again if
12040 * they have changed while we weren't holding
12041 * dtrace_lock.
12042 */
12043 if (gen != dtrace_retained_gen)
12044 goto retry;
12045 }
12046 }
12047 } while (all && (prv = prv->dtpv_next) != NULL);
12048
12049 mutex_exit(&dtrace_lock);
12050 dtrace_probe_provide(NULL, all ? NULL : prv);
12051 mutex_enter(&dtrace_lock);
12052 }
12053
12054 /*
12055 * Called to reap ECBs that are attached to probes from defunct providers.
12056 */
12057 static void
12058 dtrace_enabling_reap(void)
12059 {
12060 dtrace_provider_t *prov;
12061 dtrace_probe_t *probe;
12062 dtrace_ecb_t *ecb;
12063 hrtime_t when;
12064 int i;
12065
12066 mutex_enter(&cpu_lock);
12067 mutex_enter(&dtrace_lock);
12068
12069 for (i = 0; i < dtrace_nprobes; i++) {
12070 if ((probe = dtrace_probes[i]) == NULL)
12071 continue;
12072
12073 if (probe->dtpr_ecb == NULL)
12074 continue;
12075
12076 prov = probe->dtpr_provider;
12077
12078 if ((when = prov->dtpv_defunct) == 0)
12079 continue;
12080
12081 /*
12082 * We have ECBs on a defunct provider: we want to reap these
12083 * ECBs to allow the provider to unregister. The destruction
12084 * of these ECBs must be done carefully: if we destroy the ECB
12085 * and the consumer later wishes to consume an EPID that
12086 * corresponds to the destroyed ECB (and if the EPID metadata
12087 * has not been previously consumed), the consumer will abort
12088 * processing on the unknown EPID. To reduce (but not, sadly,
12089 * eliminate) the possibility of this, we will only destroy an
12090 * ECB for a defunct provider if, for the state that
12091 * corresponds to the ECB:
12092 *
12093 * (a) There is no speculative tracing (which can effectively
12094 * cache an EPID for an arbitrary amount of time).
12095 *
12096 * (b) The principal buffers have been switched twice since the
12097 * provider became defunct.
12098 *
12099 * (c) The aggregation buffers are of zero size or have been
12100 * switched twice since the provider became defunct.
12101 *
12102 * We use dts_speculates to determine (a) and call a function
12103 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
12104 * that as soon as we've been unable to destroy one of the ECBs
12105 * associated with the probe, we quit trying -- reaping is only
12106 * fruitful in as much as we can destroy all ECBs associated
12107 * with the defunct provider's probes.
12108 */
12109 while ((ecb = probe->dtpr_ecb) != NULL) {
12110 dtrace_state_t *state = ecb->dte_state;
12111 dtrace_buffer_t *buf = state->dts_buffer;
12112 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
12113
12114 if (state->dts_speculates)
12115 break;
12116
12117 if (!dtrace_buffer_consumed(buf, when))
12118 break;
12119
12120 if (!dtrace_buffer_consumed(aggbuf, when))
12121 break;
12122
12123 dtrace_ecb_disable(ecb);
12124 ASSERT(probe->dtpr_ecb != ecb);
12125 dtrace_ecb_destroy(ecb);
12126 }
12127 }
12128
12129 mutex_exit(&dtrace_lock);
12130 mutex_exit(&cpu_lock);
12131 }
12132
12133 /*
12134 * DTrace DOF Functions
12135 */
12136 /*ARGSUSED*/
12137 static void
12138 dtrace_dof_error(dof_hdr_t *dof, const char *str)
12139 {
12140 if (dtrace_err_verbose)
12141 cmn_err(CE_WARN, "failed to process DOF: %s", str);
12142
12143 #ifdef DTRACE_ERRDEBUG
12144 dtrace_errdebug(str);
12145 #endif
12146 }
12147
12148 /*
12149 * Create DOF out of a currently enabled state. Right now, we only create
12150 * DOF containing the run-time options -- but this could be expanded to create
12151 * complete DOF representing the enabled state.
12152 */
12153 static dof_hdr_t *
12154 dtrace_dof_create(dtrace_state_t *state)
12155 {
12156 dof_hdr_t *dof;
12157 dof_sec_t *sec;
12158 dof_optdesc_t *opt;
12159 int i, len = sizeof (dof_hdr_t) +
12160 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
12161 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12162
12163 ASSERT(MUTEX_HELD(&dtrace_lock));
12164
12165 dof = kmem_zalloc(len, KM_SLEEP);
12166 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
12167 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
12168 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
12169 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
12170
12171 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
12172 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
12173 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
12174 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
12175 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
12176 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
12177
12178 dof->dofh_flags = 0;
12179 dof->dofh_hdrsize = sizeof (dof_hdr_t);
12180 dof->dofh_secsize = sizeof (dof_sec_t);
12181 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
12182 dof->dofh_secoff = sizeof (dof_hdr_t);
12183 dof->dofh_loadsz = len;
12184 dof->dofh_filesz = len;
12185 dof->dofh_pad = 0;
12186
12187 /*
12188 * Fill in the option section header...
12189 */
12190 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
12191 sec->dofs_type = DOF_SECT_OPTDESC;
12192 sec->dofs_align = sizeof (uint64_t);
12193 sec->dofs_flags = DOF_SECF_LOAD;
12194 sec->dofs_entsize = sizeof (dof_optdesc_t);
12195
12196 opt = (dof_optdesc_t *)((uintptr_t)sec +
12197 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
12198
12199 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
12200 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12201
12202 for (i = 0; i < DTRACEOPT_MAX; i++) {
12203 opt[i].dofo_option = i;
12204 opt[i].dofo_strtab = DOF_SECIDX_NONE;
12205 opt[i].dofo_value = state->dts_options[i];
12206 }
12207
12208 return (dof);
12209 }
12210
12211 static dof_hdr_t *
12212 dtrace_dof_copyin(uintptr_t uarg, int *errp)
12213 {
12214 dof_hdr_t hdr, *dof;
12215
12216 ASSERT(!MUTEX_HELD(&dtrace_lock));
12217
12218 /*
12219 * First, we're going to copyin() the sizeof (dof_hdr_t).
12220 */
12221 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
12222 dtrace_dof_error(NULL, "failed to copyin DOF header");
12223 *errp = EFAULT;
12224 return (NULL);
12225 }
12226
12227 /*
12228 * Now we'll allocate the entire DOF and copy it in -- provided
12229 * that the length isn't outrageous.
12230 */
12231 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
12232 dtrace_dof_error(&hdr, "load size exceeds maximum");
12233 *errp = E2BIG;
12234 return (NULL);
12235 }
12236
12237 if (hdr.dofh_loadsz < sizeof (hdr)) {
12238 dtrace_dof_error(&hdr, "invalid load size");
12239 *errp = EINVAL;
12240 return (NULL);
12241 }
12242
12243 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12244
12245 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
12246 dof->dofh_loadsz != hdr.dofh_loadsz) {
12247 kmem_free(dof, hdr.dofh_loadsz);
12248 *errp = EFAULT;
12249 return (NULL);
12250 }
12251
12252 return (dof);
12253 }
12254
12255 static dof_hdr_t *
12256 dtrace_dof_property(const char *name)
12257 {
12258 uchar_t *buf;
12259 uint64_t loadsz;
12260 unsigned int len, i;
12261 dof_hdr_t *dof;
12262
12263 /*
12264 * Unfortunately, array of values in .conf files are always (and
12265 * only) interpreted to be integer arrays. We must read our DOF
12266 * as an integer array, and then squeeze it into a byte array.
12267 */
12268 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12269 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12270 return (NULL);
12271
12272 for (i = 0; i < len; i++)
12273 buf[i] = (uchar_t)(((int *)buf)[i]);
12274
12275 if (len < sizeof (dof_hdr_t)) {
12276 ddi_prop_free(buf);
12277 dtrace_dof_error(NULL, "truncated header");
12278 return (NULL);
12279 }
12280
12281 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12282 ddi_prop_free(buf);
12283 dtrace_dof_error(NULL, "truncated DOF");
12284 return (NULL);
12285 }
12286
12287 if (loadsz >= dtrace_dof_maxsize) {
12288 ddi_prop_free(buf);
12289 dtrace_dof_error(NULL, "oversized DOF");
12290 return (NULL);
12291 }
12292
12293 dof = kmem_alloc(loadsz, KM_SLEEP);
12294 bcopy(buf, dof, loadsz);
12295 ddi_prop_free(buf);
12296
12297 return (dof);
12298 }
12299
12300 static void
12301 dtrace_dof_destroy(dof_hdr_t *dof)
12302 {
12303 kmem_free(dof, dof->dofh_loadsz);
12304 }
12305
12306 /*
12307 * Return the dof_sec_t pointer corresponding to a given section index. If the
12308 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12309 * a type other than DOF_SECT_NONE is specified, the header is checked against
12310 * this type and NULL is returned if the types do not match.
12311 */
12312 static dof_sec_t *
12313 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12314 {
12315 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12316 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12317
12318 if (i >= dof->dofh_secnum) {
12319 dtrace_dof_error(dof, "referenced section index is invalid");
12320 return (NULL);
12321 }
12322
12323 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12324 dtrace_dof_error(dof, "referenced section is not loadable");
12325 return (NULL);
12326 }
12327
12328 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12329 dtrace_dof_error(dof, "referenced section is the wrong type");
12330 return (NULL);
12331 }
12332
12333 return (sec);
12334 }
12335
12336 static dtrace_probedesc_t *
12337 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12338 {
12339 dof_probedesc_t *probe;
12340 dof_sec_t *strtab;
12341 uintptr_t daddr = (uintptr_t)dof;
12342 uintptr_t str;
12343 size_t size;
12344
12345 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12346 dtrace_dof_error(dof, "invalid probe section");
12347 return (NULL);
12348 }
12349
12350 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12351 dtrace_dof_error(dof, "bad alignment in probe description");
12352 return (NULL);
12353 }
12354
12355 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12356 dtrace_dof_error(dof, "truncated probe description");
12357 return (NULL);
12358 }
12359
12360 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12361 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12362
12363 if (strtab == NULL)
12364 return (NULL);
12365
12366 str = daddr + strtab->dofs_offset;
12367 size = strtab->dofs_size;
12368
12369 if (probe->dofp_provider >= strtab->dofs_size) {
12370 dtrace_dof_error(dof, "corrupt probe provider");
12371 return (NULL);
12372 }
12373
12374 (void) strncpy(desc->dtpd_provider,
12375 (char *)(str + probe->dofp_provider),
12376 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12377
12378 if (probe->dofp_mod >= strtab->dofs_size) {
12379 dtrace_dof_error(dof, "corrupt probe module");
12380 return (NULL);
12381 }
12382
12383 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12384 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12385
12386 if (probe->dofp_func >= strtab->dofs_size) {
12387 dtrace_dof_error(dof, "corrupt probe function");
12388 return (NULL);
12389 }
12390
12391 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12392 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12393
12394 if (probe->dofp_name >= strtab->dofs_size) {
12395 dtrace_dof_error(dof, "corrupt probe name");
12396 return (NULL);
12397 }
12398
12399 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12400 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12401
12402 return (desc);
12403 }
12404
12405 static dtrace_difo_t *
12406 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12407 cred_t *cr)
12408 {
12409 dtrace_difo_t *dp;
12410 size_t ttl = 0;
12411 dof_difohdr_t *dofd;
12412 uintptr_t daddr = (uintptr_t)dof;
12413 size_t max = dtrace_difo_maxsize;
12414 int i, l, n;
12415
12416 static const struct {
12417 int section;
12418 int bufoffs;
12419 int lenoffs;
12420 int entsize;
12421 int align;
12422 const char *msg;
12423 } difo[] = {
12424 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12425 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12426 sizeof (dif_instr_t), "multiple DIF sections" },
12427
12428 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12429 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12430 sizeof (uint64_t), "multiple integer tables" },
12431
12432 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12433 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12434 sizeof (char), "multiple string tables" },
12435
12436 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12437 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12438 sizeof (uint_t), "multiple variable tables" },
12439
12440 { DOF_SECT_NONE, 0, 0, 0, NULL }
12441 };
12442
12443 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12444 dtrace_dof_error(dof, "invalid DIFO header section");
12445 return (NULL);
12446 }
12447
12448 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12449 dtrace_dof_error(dof, "bad alignment in DIFO header");
12450 return (NULL);
12451 }
12452
12453 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12454 sec->dofs_size % sizeof (dof_secidx_t)) {
12455 dtrace_dof_error(dof, "bad size in DIFO header");
12456 return (NULL);
12457 }
12458
12459 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12460 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12461
12462 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12463 dp->dtdo_rtype = dofd->dofd_rtype;
12464
12465 for (l = 0; l < n; l++) {
12466 dof_sec_t *subsec;
12467 void **bufp;
12468 uint32_t *lenp;
12469
12470 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12471 dofd->dofd_links[l])) == NULL)
12472 goto err; /* invalid section link */
12473
12474 if (ttl + subsec->dofs_size > max) {
12475 dtrace_dof_error(dof, "exceeds maximum size");
12476 goto err;
12477 }
12478
12479 ttl += subsec->dofs_size;
12480
12481 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12482 if (subsec->dofs_type != difo[i].section)
12483 continue;
12484
12485 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12486 dtrace_dof_error(dof, "section not loaded");
12487 goto err;
12488 }
12489
12490 if (subsec->dofs_align != difo[i].align) {
12491 dtrace_dof_error(dof, "bad alignment");
12492 goto err;
12493 }
12494
12495 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12496 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12497
12498 if (*bufp != NULL) {
12499 dtrace_dof_error(dof, difo[i].msg);
12500 goto err;
12501 }
12502
12503 if (difo[i].entsize != subsec->dofs_entsize) {
12504 dtrace_dof_error(dof, "entry size mismatch");
12505 goto err;
12506 }
12507
12508 if (subsec->dofs_entsize != 0 &&
12509 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12510 dtrace_dof_error(dof, "corrupt entry size");
12511 goto err;
12512 }
12513
12514 *lenp = subsec->dofs_size;
12515 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12516 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12517 *bufp, subsec->dofs_size);
12518
12519 if (subsec->dofs_entsize != 0)
12520 *lenp /= subsec->dofs_entsize;
12521
12522 break;
12523 }
12524
12525 /*
12526 * If we encounter a loadable DIFO sub-section that is not
12527 * known to us, assume this is a broken program and fail.
12528 */
12529 if (difo[i].section == DOF_SECT_NONE &&
12530 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12531 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12532 goto err;
12533 }
12534 }
12535
12536 if (dp->dtdo_buf == NULL) {
12537 /*
12538 * We can't have a DIF object without DIF text.
12539 */
12540 dtrace_dof_error(dof, "missing DIF text");
12541 goto err;
12542 }
12543
12544 /*
12545 * Before we validate the DIF object, run through the variable table
12546 * looking for the strings -- if any of their size are under, we'll set
12547 * their size to be the system-wide default string size. Note that
12548 * this should _not_ happen if the "strsize" option has been set --
12549 * in this case, the compiler should have set the size to reflect the
12550 * setting of the option.
12551 */
12552 for (i = 0; i < dp->dtdo_varlen; i++) {
12553 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12554 dtrace_diftype_t *t = &v->dtdv_type;
12555
12556 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12557 continue;
12558
12559 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12560 t->dtdt_size = dtrace_strsize_default;
12561 }
12562
12563 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12564 goto err;
12565
12566 dtrace_difo_init(dp, vstate);
12567 return (dp);
12568
12569 err:
12570 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12571 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12572 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12573 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12574
12575 kmem_free(dp, sizeof (dtrace_difo_t));
12576 return (NULL);
12577 }
12578
12579 static dtrace_predicate_t *
12580 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12581 cred_t *cr)
12582 {
12583 dtrace_difo_t *dp;
12584
12585 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12586 return (NULL);
12587
12588 return (dtrace_predicate_create(dp));
12589 }
12590
12591 static dtrace_actdesc_t *
12592 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12593 cred_t *cr)
12594 {
12595 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12596 dof_actdesc_t *desc;
12597 dof_sec_t *difosec;
12598 size_t offs;
12599 uintptr_t daddr = (uintptr_t)dof;
12600 uint64_t arg;
12601 dtrace_actkind_t kind;
12602
12603 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12604 dtrace_dof_error(dof, "invalid action section");
12605 return (NULL);
12606 }
12607
12608 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12609 dtrace_dof_error(dof, "truncated action description");
12610 return (NULL);
12611 }
12612
12613 if (sec->dofs_align != sizeof (uint64_t)) {
12614 dtrace_dof_error(dof, "bad alignment in action description");
12615 return (NULL);
12616 }
12617
12618 if (sec->dofs_size < sec->dofs_entsize) {
12619 dtrace_dof_error(dof, "section entry size exceeds total size");
12620 return (NULL);
12621 }
12622
12623 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12624 dtrace_dof_error(dof, "bad entry size in action description");
12625 return (NULL);
12626 }
12627
12628 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12629 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12630 return (NULL);
12631 }
12632
12633 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12634 desc = (dof_actdesc_t *)(daddr +
12635 (uintptr_t)sec->dofs_offset + offs);
12636 kind = (dtrace_actkind_t)desc->dofa_kind;
12637
12638 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12639 (kind != DTRACEACT_PRINTA ||
12640 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12641 (kind == DTRACEACT_DIFEXPR &&
12642 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12643 dof_sec_t *strtab;
12644 char *str, *fmt;
12645 uint64_t i;
12646
12647 /*
12648 * The argument to these actions is an index into the
12649 * DOF string table. For printf()-like actions, this
12650 * is the format string. For print(), this is the
12651 * CTF type of the expression result.
12652 */
12653 if ((strtab = dtrace_dof_sect(dof,
12654 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12655 goto err;
12656
12657 str = (char *)((uintptr_t)dof +
12658 (uintptr_t)strtab->dofs_offset);
12659
12660 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12661 if (str[i] == '\0')
12662 break;
12663 }
12664
12665 if (i >= strtab->dofs_size) {
12666 dtrace_dof_error(dof, "bogus format string");
12667 goto err;
12668 }
12669
12670 if (i == desc->dofa_arg) {
12671 dtrace_dof_error(dof, "empty format string");
12672 goto err;
12673 }
12674
12675 i -= desc->dofa_arg;
12676 fmt = kmem_alloc(i + 1, KM_SLEEP);
12677 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12678 arg = (uint64_t)(uintptr_t)fmt;
12679 } else {
12680 if (kind == DTRACEACT_PRINTA) {
12681 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12682 arg = 0;
12683 } else {
12684 arg = desc->dofa_arg;
12685 }
12686 }
12687
12688 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12689 desc->dofa_uarg, arg);
12690
12691 if (last != NULL) {
12692 last->dtad_next = act;
12693 } else {
12694 first = act;
12695 }
12696
12697 last = act;
12698
12699 if (desc->dofa_difo == DOF_SECIDX_NONE)
12700 continue;
12701
12702 if ((difosec = dtrace_dof_sect(dof,
12703 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12704 goto err;
12705
12706 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12707
12708 if (act->dtad_difo == NULL)
12709 goto err;
12710 }
12711
12712 ASSERT(first != NULL);
12713 return (first);
12714
12715 err:
12716 for (act = first; act != NULL; act = next) {
12717 next = act->dtad_next;
12718 dtrace_actdesc_release(act, vstate);
12719 }
12720
12721 return (NULL);
12722 }
12723
12724 static dtrace_ecbdesc_t *
12725 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12726 cred_t *cr)
12727 {
12728 dtrace_ecbdesc_t *ep;
12729 dof_ecbdesc_t *ecb;
12730 dtrace_probedesc_t *desc;
12731 dtrace_predicate_t *pred = NULL;
12732
12733 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12734 dtrace_dof_error(dof, "truncated ECB description");
12735 return (NULL);
12736 }
12737
12738 if (sec->dofs_align != sizeof (uint64_t)) {
12739 dtrace_dof_error(dof, "bad alignment in ECB description");
12740 return (NULL);
12741 }
12742
12743 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12744 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12745
12746 if (sec == NULL)
12747 return (NULL);
12748
12749 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12750 ep->dted_uarg = ecb->dofe_uarg;
12751 desc = &ep->dted_probe;
12752
12753 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12754 goto err;
12755
12756 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12757 if ((sec = dtrace_dof_sect(dof,
12758 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12759 goto err;
12760
12761 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12762 goto err;
12763
12764 ep->dted_pred.dtpdd_predicate = pred;
12765 }
12766
12767 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12768 if ((sec = dtrace_dof_sect(dof,
12769 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12770 goto err;
12771
12772 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12773
12774 if (ep->dted_action == NULL)
12775 goto err;
12776 }
12777
12778 return (ep);
12779
12780 err:
12781 if (pred != NULL)
12782 dtrace_predicate_release(pred, vstate);
12783 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12784 return (NULL);
12785 }
12786
12787 /*
12788 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12789 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12790 * site of any user SETX relocations to account for load object base address.
12791 * In the future, if we need other relocations, this function can be extended.
12792 */
12793 static int
12794 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12795 {
12796 uintptr_t daddr = (uintptr_t)dof;
12797 dof_relohdr_t *dofr =
12798 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12799 dof_sec_t *ss, *rs, *ts;
12800 dof_relodesc_t *r;
12801 uint_t i, n;
12802
12803 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12804 sec->dofs_align != sizeof (dof_secidx_t)) {
12805 dtrace_dof_error(dof, "invalid relocation header");
12806 return (-1);
12807 }
12808
12809 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12810 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12811 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12812
12813 if (ss == NULL || rs == NULL || ts == NULL)
12814 return (-1); /* dtrace_dof_error() has been called already */
12815
12816 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12817 rs->dofs_align != sizeof (uint64_t)) {
12818 dtrace_dof_error(dof, "invalid relocation section");
12819 return (-1);
12820 }
12821
12822 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12823 n = rs->dofs_size / rs->dofs_entsize;
12824
12825 for (i = 0; i < n; i++) {
12826 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12827
12828 switch (r->dofr_type) {
12829 case DOF_RELO_NONE:
12830 break;
12831 case DOF_RELO_SETX:
12832 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12833 sizeof (uint64_t) > ts->dofs_size) {
12834 dtrace_dof_error(dof, "bad relocation offset");
12835 return (-1);
12836 }
12837
12838 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12839 dtrace_dof_error(dof, "misaligned setx relo");
12840 return (-1);
12841 }
12842
12843 *(uint64_t *)taddr += ubase;
12844 break;
12845 default:
12846 dtrace_dof_error(dof, "invalid relocation type");
12847 return (-1);
12848 }
12849
12850 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12851 }
12852
12853 return (0);
12854 }
12855
12856 /*
12857 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12858 * header: it should be at the front of a memory region that is at least
12859 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12860 * size. It need not be validated in any other way.
12861 */
12862 static int
12863 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12864 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12865 {
12866 uint64_t len = dof->dofh_loadsz, seclen;
12867 uintptr_t daddr = (uintptr_t)dof;
12868 dtrace_ecbdesc_t *ep;
12869 dtrace_enabling_t *enab;
12870 uint_t i;
12871
12872 ASSERT(MUTEX_HELD(&dtrace_lock));
12873 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12874
12875 /*
12876 * Check the DOF header identification bytes. In addition to checking
12877 * valid settings, we also verify that unused bits/bytes are zeroed so
12878 * we can use them later without fear of regressing existing binaries.
12879 */
12880 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12881 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12882 dtrace_dof_error(dof, "DOF magic string mismatch");
12883 return (-1);
12884 }
12885
12886 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12887 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12888 dtrace_dof_error(dof, "DOF has invalid data model");
12889 return (-1);
12890 }
12891
12892 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12893 dtrace_dof_error(dof, "DOF encoding mismatch");
12894 return (-1);
12895 }
12896
12897 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12898 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12899 dtrace_dof_error(dof, "DOF version mismatch");
12900 return (-1);
12901 }
12902
12903 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12904 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12905 return (-1);
12906 }
12907
12908 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12909 dtrace_dof_error(dof, "DOF uses too many integer registers");
12910 return (-1);
12911 }
12912
12913 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12914 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12915 return (-1);
12916 }
12917
12918 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12919 if (dof->dofh_ident[i] != 0) {
12920 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12921 return (-1);
12922 }
12923 }
12924
12925 if (dof->dofh_flags & ~DOF_FL_VALID) {
12926 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12927 return (-1);
12928 }
12929
12930 if (dof->dofh_secsize == 0) {
12931 dtrace_dof_error(dof, "zero section header size");
12932 return (-1);
12933 }
12934
12935 /*
12936 * Check that the section headers don't exceed the amount of DOF
12937 * data. Note that we cast the section size and number of sections
12938 * to uint64_t's to prevent possible overflow in the multiplication.
12939 */
12940 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12941
12942 if (dof->dofh_secoff > len || seclen > len ||
12943 dof->dofh_secoff + seclen > len) {
12944 dtrace_dof_error(dof, "truncated section headers");
12945 return (-1);
12946 }
12947
12948 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12949 dtrace_dof_error(dof, "misaligned section headers");
12950 return (-1);
12951 }
12952
12953 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12954 dtrace_dof_error(dof, "misaligned section size");
12955 return (-1);
12956 }
12957
12958 /*
12959 * Take an initial pass through the section headers to be sure that
12960 * the headers don't have stray offsets. If the 'noprobes' flag is
12961 * set, do not permit sections relating to providers, probes, or args.
12962 */
12963 for (i = 0; i < dof->dofh_secnum; i++) {
12964 dof_sec_t *sec = (dof_sec_t *)(daddr +
12965 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12966
12967 if (noprobes) {
12968 switch (sec->dofs_type) {
12969 case DOF_SECT_PROVIDER:
12970 case DOF_SECT_PROBES:
12971 case DOF_SECT_PRARGS:
12972 case DOF_SECT_PROFFS:
12973 dtrace_dof_error(dof, "illegal sections "
12974 "for enabling");
12975 return (-1);
12976 }
12977 }
12978
12979 if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
12980 !(sec->dofs_flags & DOF_SECF_LOAD)) {
12981 dtrace_dof_error(dof, "loadable section with load "
12982 "flag unset");
12983 return (-1);
12984 }
12985
12986 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12987 continue; /* just ignore non-loadable sections */
12988
12989 if (sec->dofs_align & (sec->dofs_align - 1)) {
12990 dtrace_dof_error(dof, "bad section alignment");
12991 return (-1);
12992 }
12993
12994 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12995 dtrace_dof_error(dof, "misaligned section");
12996 return (-1);
12997 }
12998
12999 if (sec->dofs_offset > len || sec->dofs_size > len ||
13000 sec->dofs_offset + sec->dofs_size > len) {
13001 dtrace_dof_error(dof, "corrupt section header");
13002 return (-1);
13003 }
13004
13005 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
13006 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
13007 dtrace_dof_error(dof, "non-terminating string table");
13008 return (-1);
13009 }
13010 }
13011
13012 /*
13013 * Take a second pass through the sections and locate and perform any
13014 * relocations that are present. We do this after the first pass to
13015 * be sure that all sections have had their headers validated.
13016 */
13017 for (i = 0; i < dof->dofh_secnum; i++) {
13018 dof_sec_t *sec = (dof_sec_t *)(daddr +
13019 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13020
13021 if (!(sec->dofs_flags & DOF_SECF_LOAD))
13022 continue; /* skip sections that are not loadable */
13023
13024 switch (sec->dofs_type) {
13025 case DOF_SECT_URELHDR:
13026 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
13027 return (-1);
13028 break;
13029 }
13030 }
13031
13032 if ((enab = *enabp) == NULL)
13033 enab = *enabp = dtrace_enabling_create(vstate);
13034
13035 for (i = 0; i < dof->dofh_secnum; i++) {
13036 dof_sec_t *sec = (dof_sec_t *)(daddr +
13037 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13038
13039 if (sec->dofs_type != DOF_SECT_ECBDESC)
13040 continue;
13041
13042 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
13043 dtrace_enabling_destroy(enab);
13044 *enabp = NULL;
13045 return (-1);
13046 }
13047
13048 dtrace_enabling_add(enab, ep);
13049 }
13050
13051 return (0);
13052 }
13053
13054 /*
13055 * Process DOF for any options. This routine assumes that the DOF has been
13056 * at least processed by dtrace_dof_slurp().
13057 */
13058 static int
13059 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
13060 {
13061 int i, rval;
13062 uint32_t entsize;
13063 size_t offs;
13064 dof_optdesc_t *desc;
13065
13066 for (i = 0; i < dof->dofh_secnum; i++) {
13067 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
13068 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13069
13070 if (sec->dofs_type != DOF_SECT_OPTDESC)
13071 continue;
13072
13073 if (sec->dofs_align != sizeof (uint64_t)) {
13074 dtrace_dof_error(dof, "bad alignment in "
13075 "option description");
13076 return (EINVAL);
13077 }
13078
13079 if ((entsize = sec->dofs_entsize) == 0) {
13080 dtrace_dof_error(dof, "zeroed option entry size");
13081 return (EINVAL);
13082 }
13083
13084 if (entsize < sizeof (dof_optdesc_t)) {
13085 dtrace_dof_error(dof, "bad option entry size");
13086 return (EINVAL);
13087 }
13088
13089 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
13090 desc = (dof_optdesc_t *)((uintptr_t)dof +
13091 (uintptr_t)sec->dofs_offset + offs);
13092
13093 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
13094 dtrace_dof_error(dof, "non-zero option string");
13095 return (EINVAL);
13096 }
13097
13098 if (desc->dofo_value == DTRACEOPT_UNSET) {
13099 dtrace_dof_error(dof, "unset option");
13100 return (EINVAL);
13101 }
13102
13103 if ((rval = dtrace_state_option(state,
13104 desc->dofo_option, desc->dofo_value)) != 0) {
13105 dtrace_dof_error(dof, "rejected option");
13106 return (rval);
13107 }
13108 }
13109 }
13110
13111 return (0);
13112 }
13113
13114 /*
13115 * DTrace Consumer State Functions
13116 */
13117 int
13118 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
13119 {
13120 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
13121 void *base;
13122 uintptr_t limit;
13123 dtrace_dynvar_t *dvar, *next, *start;
13124 int i;
13125
13126 ASSERT(MUTEX_HELD(&dtrace_lock));
13127 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
13128
13129 bzero(dstate, sizeof (dtrace_dstate_t));
13130
13131 if ((dstate->dtds_chunksize = chunksize) == 0)
13132 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
13133
13134 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
13135 size = min;
13136
13137 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
13138 return (ENOMEM);
13139
13140 dstate->dtds_size = size;
13141 dstate->dtds_base = base;
13142 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
13143 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
13144
13145 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
13146
13147 if (hashsize != 1 && (hashsize & 1))
13148 hashsize--;
13149
13150 dstate->dtds_hashsize = hashsize;
13151 dstate->dtds_hash = dstate->dtds_base;
13152
13153 /*
13154 * Set all of our hash buckets to point to the single sink, and (if
13155 * it hasn't already been set), set the sink's hash value to be the
13156 * sink sentinel value. The sink is needed for dynamic variable
13157 * lookups to know that they have iterated over an entire, valid hash
13158 * chain.
13159 */
13160 for (i = 0; i < hashsize; i++)
13161 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
13162
13163 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
13164 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
13165
13166 /*
13167 * Determine number of active CPUs. Divide free list evenly among
13168 * active CPUs.
13169 */
13170 start = (dtrace_dynvar_t *)
13171 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
13172 limit = (uintptr_t)base + size;
13173
13174 maxper = (limit - (uintptr_t)start) / NCPU;
13175 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13176
13177 for (i = 0; i < NCPU; i++) {
13178 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13179
13180 /*
13181 * If we don't even have enough chunks to make it once through
13182 * NCPUs, we're just going to allocate everything to the first
13183 * CPU. And if we're on the last CPU, we're going to allocate
13184 * whatever is left over. In either case, we set the limit to
13185 * be the limit of the dynamic variable space.
13186 */
13187 if (maxper == 0 || i == NCPU - 1) {
13188 limit = (uintptr_t)base + size;
13189 start = NULL;
13190 } else {
13191 limit = (uintptr_t)start + maxper;
13192 start = (dtrace_dynvar_t *)limit;
13193 }
13194
13195 ASSERT(limit <= (uintptr_t)base + size);
13196
13197 for (;;) {
13198 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13199 dstate->dtds_chunksize);
13200
13201 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13202 break;
13203
13204 dvar->dtdv_next = next;
13205 dvar = next;
13206 }
13207
13208 if (maxper == 0)
13209 break;
13210 }
13211
13212 return (0);
13213 }
13214
13215 void
13216 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13217 {
13218 ASSERT(MUTEX_HELD(&cpu_lock));
13219
13220 if (dstate->dtds_base == NULL)
13221 return;
13222
13223 kmem_free(dstate->dtds_base, dstate->dtds_size);
13224 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13225 }
13226
13227 static void
13228 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13229 {
13230 /*
13231 * Logical XOR, where are you?
13232 */
13233 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13234
13235 if (vstate->dtvs_nglobals > 0) {
13236 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13237 sizeof (dtrace_statvar_t *));
13238 }
13239
13240 if (vstate->dtvs_ntlocals > 0) {
13241 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13242 sizeof (dtrace_difv_t));
13243 }
13244
13245 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13246
13247 if (vstate->dtvs_nlocals > 0) {
13248 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13249 sizeof (dtrace_statvar_t *));
13250 }
13251 }
13252
13253 static void
13254 dtrace_state_clean(dtrace_state_t *state)
13255 {
13256 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13257 return;
13258
13259 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13260 dtrace_speculation_clean(state);
13261 }
13262
13263 static void
13264 dtrace_state_deadman(dtrace_state_t *state)
13265 {
13266 hrtime_t now;
13267
13268 dtrace_sync();
13269
13270 now = dtrace_gethrtime();
13271
13272 if (state != dtrace_anon.dta_state &&
13273 now - state->dts_laststatus >= dtrace_deadman_user)
13274 return;
13275
13276 /*
13277 * We must be sure that dts_alive never appears to be less than the
13278 * value upon entry to dtrace_state_deadman(), and because we lack a
13279 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13280 * store INT64_MAX to it, followed by a memory barrier, followed by
13281 * the new value. This assures that dts_alive never appears to be
13282 * less than its true value, regardless of the order in which the
13283 * stores to the underlying storage are issued.
13284 */
13285 state->dts_alive = INT64_MAX;
13286 dtrace_membar_producer();
13287 state->dts_alive = now;
13288 }
13289
13290 dtrace_state_t *
13291 dtrace_state_create(dev_t *devp, cred_t *cr)
13292 {
13293 minor_t minor;
13294 major_t major;
13295 char c[30];
13296 dtrace_state_t *state;
13297 dtrace_optval_t *opt;
13298 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13299
13300 ASSERT(MUTEX_HELD(&dtrace_lock));
13301 ASSERT(MUTEX_HELD(&cpu_lock));
13302
13303 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13304 VM_BESTFIT | VM_SLEEP);
13305
13306 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13307 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13308 return (NULL);
13309 }
13310
13311 state = ddi_get_soft_state(dtrace_softstate, minor);
13312 state->dts_epid = DTRACE_EPIDNONE + 1;
13313
13314 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
13315 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13316 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13317
13318 if (devp != NULL) {
13319 major = getemajor(*devp);
13320 } else {
13321 major = ddi_driver_major(dtrace_devi);
13322 }
13323
13324 state->dts_dev = makedevice(major, minor);
13325
13326 if (devp != NULL)
13327 *devp = state->dts_dev;
13328
13329 /*
13330 * We allocate NCPU buffers. On the one hand, this can be quite
13331 * a bit of memory per instance (nearly 36K on a Starcat). On the
13332 * other hand, it saves an additional memory reference in the probe
13333 * path.
13334 */
13335 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13336 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13337 state->dts_cleaner = CYCLIC_NONE;
13338 state->dts_deadman = CYCLIC_NONE;
13339 state->dts_vstate.dtvs_state = state;
13340
13341 for (i = 0; i < DTRACEOPT_MAX; i++)
13342 state->dts_options[i] = DTRACEOPT_UNSET;
13343
13344 /*
13345 * Set the default options.
13346 */
13347 opt = state->dts_options;
13348 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13349 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13350 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13351 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13352 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13353 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13354 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13355 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13356 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13357 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13358 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13359 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13360 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13361 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13362
13363 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13364
13365 /*
13366 * Depending on the user credentials, we set flag bits which alter probe
13367 * visibility or the amount of destructiveness allowed. In the case of
13368 * actual anonymous tracing, or the possession of all privileges, all of
13369 * the normal checks are bypassed.
13370 */
13371 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13372 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13373 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13374 } else {
13375 /*
13376 * Set up the credentials for this instantiation. We take a
13377 * hold on the credential to prevent it from disappearing on
13378 * us; this in turn prevents the zone_t referenced by this
13379 * credential from disappearing. This means that we can
13380 * examine the credential and the zone from probe context.
13381 */
13382 crhold(cr);
13383 state->dts_cred.dcr_cred = cr;
13384
13385 /*
13386 * CRA_PROC means "we have *some* privilege for dtrace" and
13387 * unlocks the use of variables like pid, zonename, etc.
13388 */
13389 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13390 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13391 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13392 }
13393
13394 /*
13395 * dtrace_user allows use of syscall and profile providers.
13396 * If the user also has proc_owner and/or proc_zone, we
13397 * extend the scope to include additional visibility and
13398 * destructive power.
13399 */
13400 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13401 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13402 state->dts_cred.dcr_visible |=
13403 DTRACE_CRV_ALLPROC;
13404
13405 state->dts_cred.dcr_action |=
13406 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13407 }
13408
13409 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13410 state->dts_cred.dcr_visible |=
13411 DTRACE_CRV_ALLZONE;
13412
13413 state->dts_cred.dcr_action |=
13414 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13415 }
13416
13417 /*
13418 * If we have all privs in whatever zone this is,
13419 * we can do destructive things to processes which
13420 * have altered credentials.
13421 */
13422 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13423 cr->cr_zone->zone_privset)) {
13424 state->dts_cred.dcr_action |=
13425 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13426 }
13427 }
13428
13429 /*
13430 * Holding the dtrace_kernel privilege also implies that
13431 * the user has the dtrace_user privilege from a visibility
13432 * perspective. But without further privileges, some
13433 * destructive actions are not available.
13434 */
13435 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13436 /*
13437 * Make all probes in all zones visible. However,
13438 * this doesn't mean that all actions become available
13439 * to all zones.
13440 */
13441 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13442 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13443
13444 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13445 DTRACE_CRA_PROC;
13446 /*
13447 * Holding proc_owner means that destructive actions
13448 * for *this* zone are allowed.
13449 */
13450 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13451 state->dts_cred.dcr_action |=
13452 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13453
13454 /*
13455 * Holding proc_zone means that destructive actions
13456 * for this user/group ID in all zones is allowed.
13457 */
13458 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13459 state->dts_cred.dcr_action |=
13460 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13461
13462 /*
13463 * If we have all privs in whatever zone this is,
13464 * we can do destructive things to processes which
13465 * have altered credentials.
13466 */
13467 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13468 cr->cr_zone->zone_privset)) {
13469 state->dts_cred.dcr_action |=
13470 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13471 }
13472 }
13473
13474 /*
13475 * Holding the dtrace_proc privilege gives control over fasttrap
13476 * and pid providers. We need to grant wider destructive
13477 * privileges in the event that the user has proc_owner and/or
13478 * proc_zone.
13479 */
13480 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13481 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13482 state->dts_cred.dcr_action |=
13483 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13484
13485 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13486 state->dts_cred.dcr_action |=
13487 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13488 }
13489 }
13490
13491 return (state);
13492 }
13493
13494 static int
13495 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13496 {
13497 dtrace_optval_t *opt = state->dts_options, size;
13498 processorid_t cpu;
13499 int flags = 0, rval, factor, divisor = 1;
13500
13501 ASSERT(MUTEX_HELD(&dtrace_lock));
13502 ASSERT(MUTEX_HELD(&cpu_lock));
13503 ASSERT(which < DTRACEOPT_MAX);
13504 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13505 (state == dtrace_anon.dta_state &&
13506 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13507
13508 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13509 return (0);
13510
13511 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13512 cpu = opt[DTRACEOPT_CPU];
13513
13514 if (which == DTRACEOPT_SPECSIZE)
13515 flags |= DTRACEBUF_NOSWITCH;
13516
13517 if (which == DTRACEOPT_BUFSIZE) {
13518 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13519 flags |= DTRACEBUF_RING;
13520
13521 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13522 flags |= DTRACEBUF_FILL;
13523
13524 if (state != dtrace_anon.dta_state ||
13525 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13526 flags |= DTRACEBUF_INACTIVE;
13527 }
13528
13529 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
13530 /*
13531 * The size must be 8-byte aligned. If the size is not 8-byte
13532 * aligned, drop it down by the difference.
13533 */
13534 if (size & (sizeof (uint64_t) - 1))
13535 size -= size & (sizeof (uint64_t) - 1);
13536
13537 if (size < state->dts_reserve) {
13538 /*
13539 * Buffers always must be large enough to accommodate
13540 * their prereserved space. We return E2BIG instead
13541 * of ENOMEM in this case to allow for user-level
13542 * software to differentiate the cases.
13543 */
13544 return (E2BIG);
13545 }
13546
13547 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
13548
13549 if (rval != ENOMEM) {
13550 opt[which] = size;
13551 return (rval);
13552 }
13553
13554 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13555 return (rval);
13556
13557 for (divisor = 2; divisor < factor; divisor <<= 1)
13558 continue;
13559 }
13560
13561 return (ENOMEM);
13562 }
13563
13564 static int
13565 dtrace_state_buffers(dtrace_state_t *state)
13566 {
13567 dtrace_speculation_t *spec = state->dts_speculations;
13568 int rval, i;
13569
13570 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13571 DTRACEOPT_BUFSIZE)) != 0)
13572 return (rval);
13573
13574 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13575 DTRACEOPT_AGGSIZE)) != 0)
13576 return (rval);
13577
13578 for (i = 0; i < state->dts_nspeculations; i++) {
13579 if ((rval = dtrace_state_buffer(state,
13580 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13581 return (rval);
13582 }
13583
13584 return (0);
13585 }
13586
13587 static void
13588 dtrace_state_prereserve(dtrace_state_t *state)
13589 {
13590 dtrace_ecb_t *ecb;
13591 dtrace_probe_t *probe;
13592
13593 state->dts_reserve = 0;
13594
13595 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13596 return;
13597
13598 /*
13599 * If our buffer policy is a "fill" buffer policy, we need to set the
13600 * prereserved space to be the space required by the END probes.
13601 */
13602 probe = dtrace_probes[dtrace_probeid_end - 1];
13603 ASSERT(probe != NULL);
13604
13605 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13606 if (ecb->dte_state != state)
13607 continue;
13608
13609 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13610 }
13611 }
13612
13613 static int
13614 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13615 {
13616 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13617 dtrace_speculation_t *spec;
13618 dtrace_buffer_t *buf;
13619 cyc_handler_t hdlr;
13620 cyc_time_t when;
13621 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13622 dtrace_icookie_t cookie;
13623
13624 mutex_enter(&cpu_lock);
13625 mutex_enter(&dtrace_lock);
13626
13627 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13628 rval = EBUSY;
13629 goto out;
13630 }
13631
13632 /*
13633 * Before we can perform any checks, we must prime all of the
13634 * retained enablings that correspond to this state.
13635 */
13636 dtrace_enabling_prime(state);
13637
13638 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13639 rval = EACCES;
13640 goto out;
13641 }
13642
13643 dtrace_state_prereserve(state);
13644
13645 /*
13646 * Now we want to do is try to allocate our speculations.
13647 * We do not automatically resize the number of speculations; if
13648 * this fails, we will fail the operation.
13649 */
13650 nspec = opt[DTRACEOPT_NSPEC];
13651 ASSERT(nspec != DTRACEOPT_UNSET);
13652
13653 if (nspec > INT_MAX) {
13654 rval = ENOMEM;
13655 goto out;
13656 }
13657
13658 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
13659 KM_NOSLEEP | KM_NORMALPRI);
13660
13661 if (spec == NULL) {
13662 rval = ENOMEM;
13663 goto out;
13664 }
13665
13666 state->dts_speculations = spec;
13667 state->dts_nspeculations = (int)nspec;
13668
13669 for (i = 0; i < nspec; i++) {
13670 if ((buf = kmem_zalloc(bufsize,
13671 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
13672 rval = ENOMEM;
13673 goto err;
13674 }
13675
13676 spec[i].dtsp_buffer = buf;
13677 }
13678
13679 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13680 if (dtrace_anon.dta_state == NULL) {
13681 rval = ENOENT;
13682 goto out;
13683 }
13684
13685 if (state->dts_necbs != 0) {
13686 rval = EALREADY;
13687 goto out;
13688 }
13689
13690 state->dts_anon = dtrace_anon_grab();
13691 ASSERT(state->dts_anon != NULL);
13692 state = state->dts_anon;
13693
13694 /*
13695 * We want "grabanon" to be set in the grabbed state, so we'll
13696 * copy that option value from the grabbing state into the
13697 * grabbed state.
13698 */
13699 state->dts_options[DTRACEOPT_GRABANON] =
13700 opt[DTRACEOPT_GRABANON];
13701
13702 *cpu = dtrace_anon.dta_beganon;
13703
13704 /*
13705 * If the anonymous state is active (as it almost certainly
13706 * is if the anonymous enabling ultimately matched anything),
13707 * we don't allow any further option processing -- but we
13708 * don't return failure.
13709 */
13710 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13711 goto out;
13712 }
13713
13714 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13715 opt[DTRACEOPT_AGGSIZE] != 0) {
13716 if (state->dts_aggregations == NULL) {
13717 /*
13718 * We're not going to create an aggregation buffer
13719 * because we don't have any ECBs that contain
13720 * aggregations -- set this option to 0.
13721 */
13722 opt[DTRACEOPT_AGGSIZE] = 0;
13723 } else {
13724 /*
13725 * If we have an aggregation buffer, we must also have
13726 * a buffer to use as scratch.
13727 */
13728 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13729 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13730 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13731 }
13732 }
13733 }
13734
13735 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13736 opt[DTRACEOPT_SPECSIZE] != 0) {
13737 if (!state->dts_speculates) {
13738 /*
13739 * We're not going to create speculation buffers
13740 * because we don't have any ECBs that actually
13741 * speculate -- set the speculation size to 0.
13742 */
13743 opt[DTRACEOPT_SPECSIZE] = 0;
13744 }
13745 }
13746
13747 /*
13748 * The bare minimum size for any buffer that we're actually going to
13749 * do anything to is sizeof (uint64_t).
13750 */
13751 sz = sizeof (uint64_t);
13752
13753 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13754 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13755 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13756 /*
13757 * A buffer size has been explicitly set to 0 (or to a size
13758 * that will be adjusted to 0) and we need the space -- we
13759 * need to return failure. We return ENOSPC to differentiate
13760 * it from failing to allocate a buffer due to failure to meet
13761 * the reserve (for which we return E2BIG).
13762 */
13763 rval = ENOSPC;
13764 goto out;
13765 }
13766
13767 if ((rval = dtrace_state_buffers(state)) != 0)
13768 goto err;
13769
13770 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13771 sz = dtrace_dstate_defsize;
13772
13773 do {
13774 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13775
13776 if (rval == 0)
13777 break;
13778
13779 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13780 goto err;
13781 } while (sz >>= 1);
13782
13783 opt[DTRACEOPT_DYNVARSIZE] = sz;
13784
13785 if (rval != 0)
13786 goto err;
13787
13788 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13789 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13790
13791 if (opt[DTRACEOPT_CLEANRATE] == 0)
13792 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13793
13794 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13795 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13796
13797 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13798 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13799
13800 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13801 hdlr.cyh_arg = state;
13802 hdlr.cyh_level = CY_LOW_LEVEL;
13803
13804 when.cyt_when = 0;
13805 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13806
13807 state->dts_cleaner = cyclic_add(&hdlr, &when);
13808
13809 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13810 hdlr.cyh_arg = state;
13811 hdlr.cyh_level = CY_LOW_LEVEL;
13812
13813 when.cyt_when = 0;
13814 when.cyt_interval = dtrace_deadman_interval;
13815
13816 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13817 state->dts_deadman = cyclic_add(&hdlr, &when);
13818
13819 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13820
13821 if (state->dts_getf != 0 &&
13822 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13823 /*
13824 * We don't have kernel privs but we have at least one call
13825 * to getf(); we need to bump our zone's count, and (if
13826 * this is the first enabling to have an unprivileged call
13827 * to getf()) we need to hook into closef().
13828 */
13829 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
13830
13831 if (dtrace_getf++ == 0) {
13832 ASSERT(dtrace_closef == NULL);
13833 dtrace_closef = dtrace_getf_barrier;
13834 }
13835 }
13836
13837 /*
13838 * Now it's time to actually fire the BEGIN probe. We need to disable
13839 * interrupts here both to record the CPU on which we fired the BEGIN
13840 * probe (the data from this CPU will be processed first at user
13841 * level) and to manually activate the buffer for this CPU.
13842 */
13843 cookie = dtrace_interrupt_disable();
13844 *cpu = CPU->cpu_id;
13845 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13846 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13847
13848 dtrace_probe(dtrace_probeid_begin,
13849 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13850 dtrace_interrupt_enable(cookie);
13851 /*
13852 * We may have had an exit action from a BEGIN probe; only change our
13853 * state to ACTIVE if we're still in WARMUP.
13854 */
13855 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13856 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13857
13858 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13859 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13860
13861 /*
13862 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13863 * want each CPU to transition its principal buffer out of the
13864 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13865 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13866 * atomically transition from processing none of a state's ECBs to
13867 * processing all of them.
13868 */
13869 dtrace_xcall(DTRACE_CPUALL,
13870 (dtrace_xcall_t)dtrace_buffer_activate, state);
13871 goto out;
13872
13873 err:
13874 dtrace_buffer_free(state->dts_buffer);
13875 dtrace_buffer_free(state->dts_aggbuffer);
13876
13877 if ((nspec = state->dts_nspeculations) == 0) {
13878 ASSERT(state->dts_speculations == NULL);
13879 goto out;
13880 }
13881
13882 spec = state->dts_speculations;
13883 ASSERT(spec != NULL);
13884
13885 for (i = 0; i < state->dts_nspeculations; i++) {
13886 if ((buf = spec[i].dtsp_buffer) == NULL)
13887 break;
13888
13889 dtrace_buffer_free(buf);
13890 kmem_free(buf, bufsize);
13891 }
13892
13893 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13894 state->dts_nspeculations = 0;
13895 state->dts_speculations = NULL;
13896
13897 out:
13898 mutex_exit(&dtrace_lock);
13899 mutex_exit(&cpu_lock);
13900
13901 return (rval);
13902 }
13903
13904 static int
13905 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13906 {
13907 dtrace_icookie_t cookie;
13908
13909 ASSERT(MUTEX_HELD(&dtrace_lock));
13910
13911 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13912 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13913 return (EINVAL);
13914
13915 /*
13916 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13917 * to be sure that every CPU has seen it. See below for the details
13918 * on why this is done.
13919 */
13920 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13921 dtrace_sync();
13922
13923 /*
13924 * By this point, it is impossible for any CPU to be still processing
13925 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13926 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13927 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13928 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13929 * iff we're in the END probe.
13930 */
13931 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13932 dtrace_sync();
13933 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13934
13935 /*
13936 * Finally, we can release the reserve and call the END probe. We
13937 * disable interrupts across calling the END probe to allow us to
13938 * return the CPU on which we actually called the END probe. This
13939 * allows user-land to be sure that this CPU's principal buffer is
13940 * processed last.
13941 */
13942 state->dts_reserve = 0;
13943
13944 cookie = dtrace_interrupt_disable();
13945 *cpu = CPU->cpu_id;
13946 dtrace_probe(dtrace_probeid_end,
13947 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13948 dtrace_interrupt_enable(cookie);
13949
13950 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13951 dtrace_sync();
13952
13953 if (state->dts_getf != 0 &&
13954 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13955 /*
13956 * We don't have kernel privs but we have at least one call
13957 * to getf(); we need to lower our zone's count, and (if
13958 * this is the last enabling to have an unprivileged call
13959 * to getf()) we need to clear the closef() hook.
13960 */
13961 ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
13962 ASSERT(dtrace_closef == dtrace_getf_barrier);
13963 ASSERT(dtrace_getf > 0);
13964
13965 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
13966
13967 if (--dtrace_getf == 0)
13968 dtrace_closef = NULL;
13969 }
13970
13971 return (0);
13972 }
13973
13974 static int
13975 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13976 dtrace_optval_t val)
13977 {
13978 ASSERT(MUTEX_HELD(&dtrace_lock));
13979
13980 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13981 return (EBUSY);
13982
13983 if (option >= DTRACEOPT_MAX)
13984 return (EINVAL);
13985
13986 if (option != DTRACEOPT_CPU && val < 0)
13987 return (EINVAL);
13988
13989 switch (option) {
13990 case DTRACEOPT_DESTRUCTIVE:
13991 if (dtrace_destructive_disallow)
13992 return (EACCES);
13993
13994 state->dts_cred.dcr_destructive = 1;
13995 break;
13996
13997 case DTRACEOPT_BUFSIZE:
13998 case DTRACEOPT_DYNVARSIZE:
13999 case DTRACEOPT_AGGSIZE:
14000 case DTRACEOPT_SPECSIZE:
14001 case DTRACEOPT_STRSIZE:
14002 if (val < 0)
14003 return (EINVAL);
14004
14005 if (val >= LONG_MAX) {
14006 /*
14007 * If this is an otherwise negative value, set it to
14008 * the highest multiple of 128m less than LONG_MAX.
14009 * Technically, we're adjusting the size without
14010 * regard to the buffer resizing policy, but in fact,
14011 * this has no effect -- if we set the buffer size to
14012 * ~LONG_MAX and the buffer policy is ultimately set to
14013 * be "manual", the buffer allocation is guaranteed to
14014 * fail, if only because the allocation requires two
14015 * buffers. (We set the the size to the highest
14016 * multiple of 128m because it ensures that the size
14017 * will remain a multiple of a megabyte when
14018 * repeatedly halved -- all the way down to 15m.)
14019 */
14020 val = LONG_MAX - (1 << 27) + 1;
14021 }
14022 }
14023
14024 state->dts_options[option] = val;
14025
14026 return (0);
14027 }
14028
14029 static void
14030 dtrace_state_destroy(dtrace_state_t *state)
14031 {
14032 dtrace_ecb_t *ecb;
14033 dtrace_vstate_t *vstate = &state->dts_vstate;
14034 minor_t minor = getminor(state->dts_dev);
14035 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14036 dtrace_speculation_t *spec = state->dts_speculations;
14037 int nspec = state->dts_nspeculations;
14038 uint32_t match;
14039
14040 ASSERT(MUTEX_HELD(&dtrace_lock));
14041 ASSERT(MUTEX_HELD(&cpu_lock));
14042
14043 /*
14044 * First, retract any retained enablings for this state.
14045 */
14046 dtrace_enabling_retract(state);
14047 ASSERT(state->dts_nretained == 0);
14048
14049 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
14050 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
14051 /*
14052 * We have managed to come into dtrace_state_destroy() on a
14053 * hot enabling -- almost certainly because of a disorderly
14054 * shutdown of a consumer. (That is, a consumer that is
14055 * exiting without having called dtrace_stop().) In this case,
14056 * we're going to set our activity to be KILLED, and then
14057 * issue a sync to be sure that everyone is out of probe
14058 * context before we start blowing away ECBs.
14059 */
14060 state->dts_activity = DTRACE_ACTIVITY_KILLED;
14061 dtrace_sync();
14062 }
14063
14064 /*
14065 * Release the credential hold we took in dtrace_state_create().
14066 */
14067 if (state->dts_cred.dcr_cred != NULL)
14068 crfree(state->dts_cred.dcr_cred);
14069
14070 /*
14071 * Now we can safely disable and destroy any enabled probes. Because
14072 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
14073 * (especially if they're all enabled), we take two passes through the
14074 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
14075 * in the second we disable whatever is left over.
14076 */
14077 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
14078 for (i = 0; i < state->dts_necbs; i++) {
14079 if ((ecb = state->dts_ecbs[i]) == NULL)
14080 continue;
14081
14082 if (match && ecb->dte_probe != NULL) {
14083 dtrace_probe_t *probe = ecb->dte_probe;
14084 dtrace_provider_t *prov = probe->dtpr_provider;
14085
14086 if (!(prov->dtpv_priv.dtpp_flags & match))
14087 continue;
14088 }
14089
14090 dtrace_ecb_disable(ecb);
14091 dtrace_ecb_destroy(ecb);
14092 }
14093
14094 if (!match)
14095 break;
14096 }
14097
14098 /*
14099 * Before we free the buffers, perform one more sync to assure that
14100 * every CPU is out of probe context.
14101 */
14102 dtrace_sync();
14103
14104 dtrace_buffer_free(state->dts_buffer);
14105 dtrace_buffer_free(state->dts_aggbuffer);
14106
14107 for (i = 0; i < nspec; i++)
14108 dtrace_buffer_free(spec[i].dtsp_buffer);
14109
14110 if (state->dts_cleaner != CYCLIC_NONE)
14111 cyclic_remove(state->dts_cleaner);
14112
14113 if (state->dts_deadman != CYCLIC_NONE)
14114 cyclic_remove(state->dts_deadman);
14115
14116 dtrace_dstate_fini(&vstate->dtvs_dynvars);
14117 dtrace_vstate_fini(vstate);
14118 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14119
14120 if (state->dts_aggregations != NULL) {
14121 #ifdef DEBUG
14122 for (i = 0; i < state->dts_naggregations; i++)
14123 ASSERT(state->dts_aggregations[i] == NULL);
14124 #endif
14125 ASSERT(state->dts_naggregations > 0);
14126 kmem_free(state->dts_aggregations,
14127 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14128 }
14129
14130 kmem_free(state->dts_buffer, bufsize);
14131 kmem_free(state->dts_aggbuffer, bufsize);
14132
14133 for (i = 0; i < nspec; i++)
14134 kmem_free(spec[i].dtsp_buffer, bufsize);
14135
14136 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14137
14138 dtrace_format_destroy(state);
14139
14140 vmem_destroy(state->dts_aggid_arena);
14141 ddi_soft_state_free(dtrace_softstate, minor);
14142 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14143 }
14144
14145 /*
14146 * DTrace Anonymous Enabling Functions
14147 */
14148 static dtrace_state_t *
14149 dtrace_anon_grab(void)
14150 {
14151 dtrace_state_t *state;
14152
14153 ASSERT(MUTEX_HELD(&dtrace_lock));
14154
14155 if ((state = dtrace_anon.dta_state) == NULL) {
14156 ASSERT(dtrace_anon.dta_enabling == NULL);
14157 return (NULL);
14158 }
14159
14160 ASSERT(dtrace_anon.dta_enabling != NULL);
14161 ASSERT(dtrace_retained != NULL);
14162
14163 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14164 dtrace_anon.dta_enabling = NULL;
14165 dtrace_anon.dta_state = NULL;
14166
14167 return (state);
14168 }
14169
14170 static void
14171 dtrace_anon_property(void)
14172 {
14173 int i, rv;
14174 dtrace_state_t *state;
14175 dof_hdr_t *dof;
14176 char c[32]; /* enough for "dof-data-" + digits */
14177
14178 ASSERT(MUTEX_HELD(&dtrace_lock));
14179 ASSERT(MUTEX_HELD(&cpu_lock));
14180
14181 for (i = 0; ; i++) {
14182 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14183
14184 dtrace_err_verbose = 1;
14185
14186 if ((dof = dtrace_dof_property(c)) == NULL) {
14187 dtrace_err_verbose = 0;
14188 break;
14189 }
14190
14191 /*
14192 * We want to create anonymous state, so we need to transition
14193 * the kernel debugger to indicate that DTrace is active. If
14194 * this fails (e.g. because the debugger has modified text in
14195 * some way), we won't continue with the processing.
14196 */
14197 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14198 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14199 "enabling ignored.");
14200 dtrace_dof_destroy(dof);
14201 break;
14202 }
14203
14204 /*
14205 * If we haven't allocated an anonymous state, we'll do so now.
14206 */
14207 if ((state = dtrace_anon.dta_state) == NULL) {
14208 state = dtrace_state_create(NULL, NULL);
14209 dtrace_anon.dta_state = state;
14210
14211 if (state == NULL) {
14212 /*
14213 * This basically shouldn't happen: the only
14214 * failure mode from dtrace_state_create() is a
14215 * failure of ddi_soft_state_zalloc() that
14216 * itself should never happen. Still, the
14217 * interface allows for a failure mode, and
14218 * we want to fail as gracefully as possible:
14219 * we'll emit an error message and cease
14220 * processing anonymous state in this case.
14221 */
14222 cmn_err(CE_WARN, "failed to create "
14223 "anonymous state");
14224 dtrace_dof_destroy(dof);
14225 break;
14226 }
14227 }
14228
14229 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14230 &dtrace_anon.dta_enabling, 0, B_TRUE);
14231
14232 if (rv == 0)
14233 rv = dtrace_dof_options(dof, state);
14234
14235 dtrace_err_verbose = 0;
14236 dtrace_dof_destroy(dof);
14237
14238 if (rv != 0) {
14239 /*
14240 * This is malformed DOF; chuck any anonymous state
14241 * that we created.
14242 */
14243 ASSERT(dtrace_anon.dta_enabling == NULL);
14244 dtrace_state_destroy(state);
14245 dtrace_anon.dta_state = NULL;
14246 break;
14247 }
14248
14249 ASSERT(dtrace_anon.dta_enabling != NULL);
14250 }
14251
14252 if (dtrace_anon.dta_enabling != NULL) {
14253 int rval;
14254
14255 /*
14256 * dtrace_enabling_retain() can only fail because we are
14257 * trying to retain more enablings than are allowed -- but
14258 * we only have one anonymous enabling, and we are guaranteed
14259 * to be allowed at least one retained enabling; we assert
14260 * that dtrace_enabling_retain() returns success.
14261 */
14262 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14263 ASSERT(rval == 0);
14264
14265 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14266 }
14267 }
14268
14269 /*
14270 * DTrace Helper Functions
14271 */
14272 static void
14273 dtrace_helper_trace(dtrace_helper_action_t *helper,
14274 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14275 {
14276 uint32_t size, next, nnext, i;
14277 dtrace_helptrace_t *ent;
14278 uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14279
14280 if (!dtrace_helptrace_enabled)
14281 return;
14282
14283 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14284
14285 /*
14286 * What would a tracing framework be without its own tracing
14287 * framework? (Well, a hell of a lot simpler, for starters...)
14288 */
14289 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14290 sizeof (uint64_t) - sizeof (uint64_t);
14291
14292 /*
14293 * Iterate until we can allocate a slot in the trace buffer.
14294 */
14295 do {
14296 next = dtrace_helptrace_next;
14297
14298 if (next + size < dtrace_helptrace_bufsize) {
14299 nnext = next + size;
14300 } else {
14301 nnext = size;
14302 }
14303 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14304
14305 /*
14306 * We have our slot; fill it in.
14307 */
14308 if (nnext == size)
14309 next = 0;
14310
14311 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
14312 ent->dtht_helper = helper;
14313 ent->dtht_where = where;
14314 ent->dtht_nlocals = vstate->dtvs_nlocals;
14315
14316 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14317 mstate->dtms_fltoffs : -1;
14318 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14319 ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
14320
14321 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14322 dtrace_statvar_t *svar;
14323
14324 if ((svar = vstate->dtvs_locals[i]) == NULL)
14325 continue;
14326
14327 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14328 ent->dtht_locals[i] =
14329 ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
14330 }
14331 }
14332
14333 static uint64_t
14334 dtrace_helper(int which, dtrace_mstate_t *mstate,
14335 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14336 {
14337 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14338 uint64_t sarg0 = mstate->dtms_arg[0];
14339 uint64_t sarg1 = mstate->dtms_arg[1];
14340 uint64_t rval;
14341 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14342 dtrace_helper_action_t *helper;
14343 dtrace_vstate_t *vstate;
14344 dtrace_difo_t *pred;
14345 int i, trace = dtrace_helptrace_enabled;
14346
14347 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14348
14349 if (helpers == NULL)
14350 return (0);
14351
14352 if ((helper = helpers->dthps_actions[which]) == NULL)
14353 return (0);
14354
14355 vstate = &helpers->dthps_vstate;
14356 mstate->dtms_arg[0] = arg0;
14357 mstate->dtms_arg[1] = arg1;
14358
14359 /*
14360 * Now iterate over each helper. If its predicate evaluates to 'true',
14361 * we'll call the corresponding actions. Note that the below calls
14362 * to dtrace_dif_emulate() may set faults in machine state. This is
14363 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14364 * the stored DIF offset with its own (which is the desired behavior).
14365 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14366 * from machine state; this is okay, too.
14367 */
14368 for (; helper != NULL; helper = helper->dtha_next) {
14369 if ((pred = helper->dtha_predicate) != NULL) {
14370 if (trace)
14371 dtrace_helper_trace(helper, mstate, vstate, 0);
14372
14373 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14374 goto next;
14375
14376 if (*flags & CPU_DTRACE_FAULT)
14377 goto err;
14378 }
14379
14380 for (i = 0; i < helper->dtha_nactions; i++) {
14381 if (trace)
14382 dtrace_helper_trace(helper,
14383 mstate, vstate, i + 1);
14384
14385 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14386 mstate, vstate, state);
14387
14388 if (*flags & CPU_DTRACE_FAULT)
14389 goto err;
14390 }
14391
14392 next:
14393 if (trace)
14394 dtrace_helper_trace(helper, mstate, vstate,
14395 DTRACE_HELPTRACE_NEXT);
14396 }
14397
14398 if (trace)
14399 dtrace_helper_trace(helper, mstate, vstate,
14400 DTRACE_HELPTRACE_DONE);
14401
14402 /*
14403 * Restore the arg0 that we saved upon entry.
14404 */
14405 mstate->dtms_arg[0] = sarg0;
14406 mstate->dtms_arg[1] = sarg1;
14407
14408 return (rval);
14409
14410 err:
14411 if (trace)
14412 dtrace_helper_trace(helper, mstate, vstate,
14413 DTRACE_HELPTRACE_ERR);
14414
14415 /*
14416 * Restore the arg0 that we saved upon entry.
14417 */
14418 mstate->dtms_arg[0] = sarg0;
14419 mstate->dtms_arg[1] = sarg1;
14420
14421 return (NULL);
14422 }
14423
14424 static void
14425 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14426 dtrace_vstate_t *vstate)
14427 {
14428 int i;
14429
14430 if (helper->dtha_predicate != NULL)
14431 dtrace_difo_release(helper->dtha_predicate, vstate);
14432
14433 for (i = 0; i < helper->dtha_nactions; i++) {
14434 ASSERT(helper->dtha_actions[i] != NULL);
14435 dtrace_difo_release(helper->dtha_actions[i], vstate);
14436 }
14437
14438 kmem_free(helper->dtha_actions,
14439 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14440 kmem_free(helper, sizeof (dtrace_helper_action_t));
14441 }
14442
14443 static int
14444 dtrace_helper_destroygen(int gen)
14445 {
14446 proc_t *p = curproc;
14447 dtrace_helpers_t *help = p->p_dtrace_helpers;
14448 dtrace_vstate_t *vstate;
14449 int i;
14450
14451 ASSERT(MUTEX_HELD(&dtrace_lock));
14452
14453 if (help == NULL || gen > help->dthps_generation)
14454 return (EINVAL);
14455
14456 vstate = &help->dthps_vstate;
14457
14458 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14459 dtrace_helper_action_t *last = NULL, *h, *next;
14460
14461 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14462 next = h->dtha_next;
14463
14464 if (h->dtha_generation == gen) {
14465 if (last != NULL) {
14466 last->dtha_next = next;
14467 } else {
14468 help->dthps_actions[i] = next;
14469 }
14470
14471 dtrace_helper_action_destroy(h, vstate);
14472 } else {
14473 last = h;
14474 }
14475 }
14476 }
14477
14478 /*
14479 * Interate until we've cleared out all helper providers with the
14480 * given generation number.
14481 */
14482 for (;;) {
14483 dtrace_helper_provider_t *prov;
14484
14485 /*
14486 * Look for a helper provider with the right generation. We
14487 * have to start back at the beginning of the list each time
14488 * because we drop dtrace_lock. It's unlikely that we'll make
14489 * more than two passes.
14490 */
14491 for (i = 0; i < help->dthps_nprovs; i++) {
14492 prov = help->dthps_provs[i];
14493
14494 if (prov->dthp_generation == gen)
14495 break;
14496 }
14497
14498 /*
14499 * If there were no matches, we're done.
14500 */
14501 if (i == help->dthps_nprovs)
14502 break;
14503
14504 /*
14505 * Move the last helper provider into this slot.
14506 */
14507 help->dthps_nprovs--;
14508 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14509 help->dthps_provs[help->dthps_nprovs] = NULL;
14510
14511 mutex_exit(&dtrace_lock);
14512
14513 /*
14514 * If we have a meta provider, remove this helper provider.
14515 */
14516 mutex_enter(&dtrace_meta_lock);
14517 if (dtrace_meta_pid != NULL) {
14518 ASSERT(dtrace_deferred_pid == NULL);
14519 dtrace_helper_provider_remove(&prov->dthp_prov,
14520 p->p_pid);
14521 }
14522 mutex_exit(&dtrace_meta_lock);
14523
14524 dtrace_helper_provider_destroy(prov);
14525
14526 mutex_enter(&dtrace_lock);
14527 }
14528
14529 return (0);
14530 }
14531
14532 static int
14533 dtrace_helper_validate(dtrace_helper_action_t *helper)
14534 {
14535 int err = 0, i;
14536 dtrace_difo_t *dp;
14537
14538 if ((dp = helper->dtha_predicate) != NULL)
14539 err += dtrace_difo_validate_helper(dp);
14540
14541 for (i = 0; i < helper->dtha_nactions; i++)
14542 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14543
14544 return (err == 0);
14545 }
14546
14547 static int
14548 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14549 {
14550 dtrace_helpers_t *help;
14551 dtrace_helper_action_t *helper, *last;
14552 dtrace_actdesc_t *act;
14553 dtrace_vstate_t *vstate;
14554 dtrace_predicate_t *pred;
14555 int count = 0, nactions = 0, i;
14556
14557 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14558 return (EINVAL);
14559
14560 help = curproc->p_dtrace_helpers;
14561 last = help->dthps_actions[which];
14562 vstate = &help->dthps_vstate;
14563
14564 for (count = 0; last != NULL; last = last->dtha_next) {
14565 count++;
14566 if (last->dtha_next == NULL)
14567 break;
14568 }
14569
14570 /*
14571 * If we already have dtrace_helper_actions_max helper actions for this
14572 * helper action type, we'll refuse to add a new one.
14573 */
14574 if (count >= dtrace_helper_actions_max)
14575 return (ENOSPC);
14576
14577 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14578 helper->dtha_generation = help->dthps_generation;
14579
14580 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14581 ASSERT(pred->dtp_difo != NULL);
14582 dtrace_difo_hold(pred->dtp_difo);
14583 helper->dtha_predicate = pred->dtp_difo;
14584 }
14585
14586 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14587 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14588 goto err;
14589
14590 if (act->dtad_difo == NULL)
14591 goto err;
14592
14593 nactions++;
14594 }
14595
14596 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14597 (helper->dtha_nactions = nactions), KM_SLEEP);
14598
14599 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14600 dtrace_difo_hold(act->dtad_difo);
14601 helper->dtha_actions[i++] = act->dtad_difo;
14602 }
14603
14604 if (!dtrace_helper_validate(helper))
14605 goto err;
14606
14607 if (last == NULL) {
14608 help->dthps_actions[which] = helper;
14609 } else {
14610 last->dtha_next = helper;
14611 }
14612
14613 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14614 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14615 dtrace_helptrace_next = 0;
14616 }
14617
14618 return (0);
14619 err:
14620 dtrace_helper_action_destroy(helper, vstate);
14621 return (EINVAL);
14622 }
14623
14624 static void
14625 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14626 dof_helper_t *dofhp)
14627 {
14628 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14629
14630 mutex_enter(&dtrace_meta_lock);
14631 mutex_enter(&dtrace_lock);
14632
14633 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14634 /*
14635 * If the dtrace module is loaded but not attached, or if
14636 * there aren't isn't a meta provider registered to deal with
14637 * these provider descriptions, we need to postpone creating
14638 * the actual providers until later.
14639 */
14640
14641 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14642 dtrace_deferred_pid != help) {
14643 help->dthps_deferred = 1;
14644 help->dthps_pid = p->p_pid;
14645 help->dthps_next = dtrace_deferred_pid;
14646 help->dthps_prev = NULL;
14647 if (dtrace_deferred_pid != NULL)
14648 dtrace_deferred_pid->dthps_prev = help;
14649 dtrace_deferred_pid = help;
14650 }
14651
14652 mutex_exit(&dtrace_lock);
14653
14654 } else if (dofhp != NULL) {
14655 /*
14656 * If the dtrace module is loaded and we have a particular
14657 * helper provider description, pass that off to the
14658 * meta provider.
14659 */
14660
14661 mutex_exit(&dtrace_lock);
14662
14663 dtrace_helper_provide(dofhp, p->p_pid);
14664
14665 } else {
14666 /*
14667 * Otherwise, just pass all the helper provider descriptions
14668 * off to the meta provider.
14669 */
14670
14671 int i;
14672 mutex_exit(&dtrace_lock);
14673
14674 for (i = 0; i < help->dthps_nprovs; i++) {
14675 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14676 p->p_pid);
14677 }
14678 }
14679
14680 mutex_exit(&dtrace_meta_lock);
14681 }
14682
14683 static int
14684 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14685 {
14686 dtrace_helpers_t *help;
14687 dtrace_helper_provider_t *hprov, **tmp_provs;
14688 uint_t tmp_maxprovs, i;
14689
14690 ASSERT(MUTEX_HELD(&dtrace_lock));
14691
14692 help = curproc->p_dtrace_helpers;
14693 ASSERT(help != NULL);
14694
14695 /*
14696 * If we already have dtrace_helper_providers_max helper providers,
14697 * we're refuse to add a new one.
14698 */
14699 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14700 return (ENOSPC);
14701
14702 /*
14703 * Check to make sure this isn't a duplicate.
14704 */
14705 for (i = 0; i < help->dthps_nprovs; i++) {
14706 if (dofhp->dofhp_dof ==
14707 help->dthps_provs[i]->dthp_prov.dofhp_dof)
14708 return (EALREADY);
14709 }
14710
14711 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14712 hprov->dthp_prov = *dofhp;
14713 hprov->dthp_ref = 1;
14714 hprov->dthp_generation = gen;
14715
14716 /*
14717 * Allocate a bigger table for helper providers if it's already full.
14718 */
14719 if (help->dthps_maxprovs == help->dthps_nprovs) {
14720 tmp_maxprovs = help->dthps_maxprovs;
14721 tmp_provs = help->dthps_provs;
14722
14723 if (help->dthps_maxprovs == 0)
14724 help->dthps_maxprovs = 2;
14725 else
14726 help->dthps_maxprovs *= 2;
14727 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14728 help->dthps_maxprovs = dtrace_helper_providers_max;
14729
14730 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14731
14732 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14733 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14734
14735 if (tmp_provs != NULL) {
14736 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14737 sizeof (dtrace_helper_provider_t *));
14738 kmem_free(tmp_provs, tmp_maxprovs *
14739 sizeof (dtrace_helper_provider_t *));
14740 }
14741 }
14742
14743 help->dthps_provs[help->dthps_nprovs] = hprov;
14744 help->dthps_nprovs++;
14745
14746 return (0);
14747 }
14748
14749 static void
14750 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14751 {
14752 mutex_enter(&dtrace_lock);
14753
14754 if (--hprov->dthp_ref == 0) {
14755 dof_hdr_t *dof;
14756 mutex_exit(&dtrace_lock);
14757 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14758 dtrace_dof_destroy(dof);
14759 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14760 } else {
14761 mutex_exit(&dtrace_lock);
14762 }
14763 }
14764
14765 static int
14766 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14767 {
14768 uintptr_t daddr = (uintptr_t)dof;
14769 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14770 dof_provider_t *provider;
14771 dof_probe_t *probe;
14772 uint8_t *arg;
14773 char *strtab, *typestr;
14774 dof_stridx_t typeidx;
14775 size_t typesz;
14776 uint_t nprobes, j, k;
14777
14778 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14779
14780 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14781 dtrace_dof_error(dof, "misaligned section offset");
14782 return (-1);
14783 }
14784
14785 /*
14786 * The section needs to be large enough to contain the DOF provider
14787 * structure appropriate for the given version.
14788 */
14789 if (sec->dofs_size <
14790 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14791 offsetof(dof_provider_t, dofpv_prenoffs) :
14792 sizeof (dof_provider_t))) {
14793 dtrace_dof_error(dof, "provider section too small");
14794 return (-1);
14795 }
14796
14797 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14798 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14799 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14800 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14801 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14802
14803 if (str_sec == NULL || prb_sec == NULL ||
14804 arg_sec == NULL || off_sec == NULL)
14805 return (-1);
14806
14807 enoff_sec = NULL;
14808
14809 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14810 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14811 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14812 provider->dofpv_prenoffs)) == NULL)
14813 return (-1);
14814
14815 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14816
14817 if (provider->dofpv_name >= str_sec->dofs_size ||
14818 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14819 dtrace_dof_error(dof, "invalid provider name");
14820 return (-1);
14821 }
14822
14823 if (prb_sec->dofs_entsize == 0 ||
14824 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14825 dtrace_dof_error(dof, "invalid entry size");
14826 return (-1);
14827 }
14828
14829 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14830 dtrace_dof_error(dof, "misaligned entry size");
14831 return (-1);
14832 }
14833
14834 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14835 dtrace_dof_error(dof, "invalid entry size");
14836 return (-1);
14837 }
14838
14839 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14840 dtrace_dof_error(dof, "misaligned section offset");
14841 return (-1);
14842 }
14843
14844 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14845 dtrace_dof_error(dof, "invalid entry size");
14846 return (-1);
14847 }
14848
14849 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14850
14851 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14852
14853 /*
14854 * Take a pass through the probes to check for errors.
14855 */
14856 for (j = 0; j < nprobes; j++) {
14857 probe = (dof_probe_t *)(uintptr_t)(daddr +
14858 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14859
14860 if (probe->dofpr_func >= str_sec->dofs_size) {
14861 dtrace_dof_error(dof, "invalid function name");
14862 return (-1);
14863 }
14864
14865 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14866 dtrace_dof_error(dof, "function name too long");
14867 return (-1);
14868 }
14869
14870 if (probe->dofpr_name >= str_sec->dofs_size ||
14871 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14872 dtrace_dof_error(dof, "invalid probe name");
14873 return (-1);
14874 }
14875
14876 /*
14877 * The offset count must not wrap the index, and the offsets
14878 * must also not overflow the section's data.
14879 */
14880 if (probe->dofpr_offidx + probe->dofpr_noffs <
14881 probe->dofpr_offidx ||
14882 (probe->dofpr_offidx + probe->dofpr_noffs) *
14883 off_sec->dofs_entsize > off_sec->dofs_size) {
14884 dtrace_dof_error(dof, "invalid probe offset");
14885 return (-1);
14886 }
14887
14888 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14889 /*
14890 * If there's no is-enabled offset section, make sure
14891 * there aren't any is-enabled offsets. Otherwise
14892 * perform the same checks as for probe offsets
14893 * (immediately above).
14894 */
14895 if (enoff_sec == NULL) {
14896 if (probe->dofpr_enoffidx != 0 ||
14897 probe->dofpr_nenoffs != 0) {
14898 dtrace_dof_error(dof, "is-enabled "
14899 "offsets with null section");
14900 return (-1);
14901 }
14902 } else if (probe->dofpr_enoffidx +
14903 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14904 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14905 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14906 dtrace_dof_error(dof, "invalid is-enabled "
14907 "offset");
14908 return (-1);
14909 }
14910
14911 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14912 dtrace_dof_error(dof, "zero probe and "
14913 "is-enabled offsets");
14914 return (-1);
14915 }
14916 } else if (probe->dofpr_noffs == 0) {
14917 dtrace_dof_error(dof, "zero probe offsets");
14918 return (-1);
14919 }
14920
14921 if (probe->dofpr_argidx + probe->dofpr_xargc <
14922 probe->dofpr_argidx ||
14923 (probe->dofpr_argidx + probe->dofpr_xargc) *
14924 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14925 dtrace_dof_error(dof, "invalid args");
14926 return (-1);
14927 }
14928
14929 typeidx = probe->dofpr_nargv;
14930 typestr = strtab + probe->dofpr_nargv;
14931 for (k = 0; k < probe->dofpr_nargc; k++) {
14932 if (typeidx >= str_sec->dofs_size) {
14933 dtrace_dof_error(dof, "bad "
14934 "native argument type");
14935 return (-1);
14936 }
14937
14938 typesz = strlen(typestr) + 1;
14939 if (typesz > DTRACE_ARGTYPELEN) {
14940 dtrace_dof_error(dof, "native "
14941 "argument type too long");
14942 return (-1);
14943 }
14944 typeidx += typesz;
14945 typestr += typesz;
14946 }
14947
14948 typeidx = probe->dofpr_xargv;
14949 typestr = strtab + probe->dofpr_xargv;
14950 for (k = 0; k < probe->dofpr_xargc; k++) {
14951 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14952 dtrace_dof_error(dof, "bad "
14953 "native argument index");
14954 return (-1);
14955 }
14956
14957 if (typeidx >= str_sec->dofs_size) {
14958 dtrace_dof_error(dof, "bad "
14959 "translated argument type");
14960 return (-1);
14961 }
14962
14963 typesz = strlen(typestr) + 1;
14964 if (typesz > DTRACE_ARGTYPELEN) {
14965 dtrace_dof_error(dof, "translated argument "
14966 "type too long");
14967 return (-1);
14968 }
14969
14970 typeidx += typesz;
14971 typestr += typesz;
14972 }
14973 }
14974
14975 return (0);
14976 }
14977
14978 static int
14979 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14980 {
14981 dtrace_helpers_t *help;
14982 dtrace_vstate_t *vstate;
14983 dtrace_enabling_t *enab = NULL;
14984 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14985 uintptr_t daddr = (uintptr_t)dof;
14986
14987 ASSERT(MUTEX_HELD(&dtrace_lock));
14988
14989 if ((help = curproc->p_dtrace_helpers) == NULL)
14990 help = dtrace_helpers_create(curproc);
14991
14992 vstate = &help->dthps_vstate;
14993
14994 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14995 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14996 dtrace_dof_destroy(dof);
14997 return (rv);
14998 }
14999
15000 /*
15001 * Look for helper providers and validate their descriptions.
15002 */
15003 if (dhp != NULL) {
15004 for (i = 0; i < dof->dofh_secnum; i++) {
15005 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
15006 dof->dofh_secoff + i * dof->dofh_secsize);
15007
15008 if (sec->dofs_type != DOF_SECT_PROVIDER)
15009 continue;
15010
15011 if (dtrace_helper_provider_validate(dof, sec) != 0) {
15012 dtrace_enabling_destroy(enab);
15013 dtrace_dof_destroy(dof);
15014 return (-1);
15015 }
15016
15017 nprovs++;
15018 }
15019 }
15020
15021 /*
15022 * Now we need to walk through the ECB descriptions in the enabling.
15023 */
15024 for (i = 0; i < enab->dten_ndesc; i++) {
15025 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
15026 dtrace_probedesc_t *desc = &ep->dted_probe;
15027
15028 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
15029 continue;
15030
15031 if (strcmp(desc->dtpd_mod, "helper") != 0)
15032 continue;
15033
15034 if (strcmp(desc->dtpd_func, "ustack") != 0)
15035 continue;
15036
15037 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
15038 ep)) != 0) {
15039 /*
15040 * Adding this helper action failed -- we are now going
15041 * to rip out the entire generation and return failure.
15042 */
15043 (void) dtrace_helper_destroygen(help->dthps_generation);
15044 dtrace_enabling_destroy(enab);
15045 dtrace_dof_destroy(dof);
15046 return (-1);
15047 }
15048
15049 nhelpers++;
15050 }
15051
15052 if (nhelpers < enab->dten_ndesc)
15053 dtrace_dof_error(dof, "unmatched helpers");
15054
15055 gen = help->dthps_generation++;
15056 dtrace_enabling_destroy(enab);
15057
15058 if (dhp != NULL && nprovs > 0) {
15059 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
15060 if (dtrace_helper_provider_add(dhp, gen) == 0) {
15061 mutex_exit(&dtrace_lock);
15062 dtrace_helper_provider_register(curproc, help, dhp);
15063 mutex_enter(&dtrace_lock);
15064
15065 destroy = 0;
15066 }
15067 }
15068
15069 if (destroy)
15070 dtrace_dof_destroy(dof);
15071
15072 return (gen);
15073 }
15074
15075 static dtrace_helpers_t *
15076 dtrace_helpers_create(proc_t *p)
15077 {
15078 dtrace_helpers_t *help;
15079
15080 ASSERT(MUTEX_HELD(&dtrace_lock));
15081 ASSERT(p->p_dtrace_helpers == NULL);
15082
15083 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
15084 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
15085 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
15086
15087 p->p_dtrace_helpers = help;
15088 dtrace_helpers++;
15089
15090 return (help);
15091 }
15092
15093 static void
15094 dtrace_helpers_destroy(void)
15095 {
15096 dtrace_helpers_t *help;
15097 dtrace_vstate_t *vstate;
15098 proc_t *p = curproc;
15099 int i;
15100
15101 mutex_enter(&dtrace_lock);
15102
15103 ASSERT(p->p_dtrace_helpers != NULL);
15104 ASSERT(dtrace_helpers > 0);
15105
15106 help = p->p_dtrace_helpers;
15107 vstate = &help->dthps_vstate;
15108
15109 /*
15110 * We're now going to lose the help from this process.
15111 */
15112 p->p_dtrace_helpers = NULL;
15113 dtrace_sync();
15114
15115 /*
15116 * Destory the helper actions.
15117 */
15118 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15119 dtrace_helper_action_t *h, *next;
15120
15121 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15122 next = h->dtha_next;
15123 dtrace_helper_action_destroy(h, vstate);
15124 h = next;
15125 }
15126 }
15127
15128 mutex_exit(&dtrace_lock);
15129
15130 /*
15131 * Destroy the helper providers.
15132 */
15133 if (help->dthps_maxprovs > 0) {
15134 mutex_enter(&dtrace_meta_lock);
15135 if (dtrace_meta_pid != NULL) {
15136 ASSERT(dtrace_deferred_pid == NULL);
15137
15138 for (i = 0; i < help->dthps_nprovs; i++) {
15139 dtrace_helper_provider_remove(
15140 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15141 }
15142 } else {
15143 mutex_enter(&dtrace_lock);
15144 ASSERT(help->dthps_deferred == 0 ||
15145 help->dthps_next != NULL ||
15146 help->dthps_prev != NULL ||
15147 help == dtrace_deferred_pid);
15148
15149 /*
15150 * Remove the helper from the deferred list.
15151 */
15152 if (help->dthps_next != NULL)
15153 help->dthps_next->dthps_prev = help->dthps_prev;
15154 if (help->dthps_prev != NULL)
15155 help->dthps_prev->dthps_next = help->dthps_next;
15156 if (dtrace_deferred_pid == help) {
15157 dtrace_deferred_pid = help->dthps_next;
15158 ASSERT(help->dthps_prev == NULL);
15159 }
15160
15161 mutex_exit(&dtrace_lock);
15162 }
15163
15164 mutex_exit(&dtrace_meta_lock);
15165
15166 for (i = 0; i < help->dthps_nprovs; i++) {
15167 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15168 }
15169
15170 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15171 sizeof (dtrace_helper_provider_t *));
15172 }
15173
15174 mutex_enter(&dtrace_lock);
15175
15176 dtrace_vstate_fini(&help->dthps_vstate);
15177 kmem_free(help->dthps_actions,
15178 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15179 kmem_free(help, sizeof (dtrace_helpers_t));
15180
15181 --dtrace_helpers;
15182 mutex_exit(&dtrace_lock);
15183 }
15184
15185 static void
15186 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15187 {
15188 dtrace_helpers_t *help, *newhelp;
15189 dtrace_helper_action_t *helper, *new, *last;
15190 dtrace_difo_t *dp;
15191 dtrace_vstate_t *vstate;
15192 int i, j, sz, hasprovs = 0;
15193
15194 mutex_enter(&dtrace_lock);
15195 ASSERT(from->p_dtrace_helpers != NULL);
15196 ASSERT(dtrace_helpers > 0);
15197
15198 help = from->p_dtrace_helpers;
15199 newhelp = dtrace_helpers_create(to);
15200 ASSERT(to->p_dtrace_helpers != NULL);
15201
15202 newhelp->dthps_generation = help->dthps_generation;
15203 vstate = &newhelp->dthps_vstate;
15204
15205 /*
15206 * Duplicate the helper actions.
15207 */
15208 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15209 if ((helper = help->dthps_actions[i]) == NULL)
15210 continue;
15211
15212 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15213 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15214 KM_SLEEP);
15215 new->dtha_generation = helper->dtha_generation;
15216
15217 if ((dp = helper->dtha_predicate) != NULL) {
15218 dp = dtrace_difo_duplicate(dp, vstate);
15219 new->dtha_predicate = dp;
15220 }
15221
15222 new->dtha_nactions = helper->dtha_nactions;
15223 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15224 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15225
15226 for (j = 0; j < new->dtha_nactions; j++) {
15227 dtrace_difo_t *dp = helper->dtha_actions[j];
15228
15229 ASSERT(dp != NULL);
15230 dp = dtrace_difo_duplicate(dp, vstate);
15231 new->dtha_actions[j] = dp;
15232 }
15233
15234 if (last != NULL) {
15235 last->dtha_next = new;
15236 } else {
15237 newhelp->dthps_actions[i] = new;
15238 }
15239
15240 last = new;
15241 }
15242 }
15243
15244 /*
15245 * Duplicate the helper providers and register them with the
15246 * DTrace framework.
15247 */
15248 if (help->dthps_nprovs > 0) {
15249 newhelp->dthps_nprovs = help->dthps_nprovs;
15250 newhelp->dthps_maxprovs = help->dthps_nprovs;
15251 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15252 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15253 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15254 newhelp->dthps_provs[i] = help->dthps_provs[i];
15255 newhelp->dthps_provs[i]->dthp_ref++;
15256 }
15257
15258 hasprovs = 1;
15259 }
15260
15261 mutex_exit(&dtrace_lock);
15262
15263 if (hasprovs)
15264 dtrace_helper_provider_register(to, newhelp, NULL);
15265 }
15266
15267 /*
15268 * DTrace Hook Functions
15269 */
15270 static void
15271 dtrace_module_loaded(struct modctl *ctl)
15272 {
15273 dtrace_provider_t *prv;
15274
15275 mutex_enter(&dtrace_provider_lock);
15276 mutex_enter(&mod_lock);
15277
15278 ASSERT(ctl->mod_busy);
15279
15280 /*
15281 * We're going to call each providers per-module provide operation
15282 * specifying only this module.
15283 */
15284 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15285 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15286
15287 mutex_exit(&mod_lock);
15288 mutex_exit(&dtrace_provider_lock);
15289
15290 /*
15291 * If we have any retained enablings, we need to match against them.
15292 * Enabling probes requires that cpu_lock be held, and we cannot hold
15293 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15294 * module. (In particular, this happens when loading scheduling
15295 * classes.) So if we have any retained enablings, we need to dispatch
15296 * our task queue to do the match for us.
15297 */
15298 mutex_enter(&dtrace_lock);
15299
15300 if (dtrace_retained == NULL) {
15301 mutex_exit(&dtrace_lock);
15302 return;
15303 }
15304
15305 (void) taskq_dispatch(dtrace_taskq,
15306 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15307
15308 mutex_exit(&dtrace_lock);
15309
15310 /*
15311 * And now, for a little heuristic sleaze: in general, we want to
15312 * match modules as soon as they load. However, we cannot guarantee
15313 * this, because it would lead us to the lock ordering violation
15314 * outlined above. The common case, of course, is that cpu_lock is
15315 * _not_ held -- so we delay here for a clock tick, hoping that that's
15316 * long enough for the task queue to do its work. If it's not, it's
15317 * not a serious problem -- it just means that the module that we
15318 * just loaded may not be immediately instrumentable.
15319 */
15320 delay(1);
15321 }
15322
15323 static void
15324 dtrace_module_unloaded(struct modctl *ctl)
15325 {
15326 dtrace_probe_t template, *probe, *first, *next;
15327 dtrace_provider_t *prov;
15328
15329 template.dtpr_mod = ctl->mod_modname;
15330
15331 mutex_enter(&dtrace_provider_lock);
15332 mutex_enter(&mod_lock);
15333 mutex_enter(&dtrace_lock);
15334
15335 if (dtrace_bymod == NULL) {
15336 /*
15337 * The DTrace module is loaded (obviously) but not attached;
15338 * we don't have any work to do.
15339 */
15340 mutex_exit(&dtrace_provider_lock);
15341 mutex_exit(&mod_lock);
15342 mutex_exit(&dtrace_lock);
15343 return;
15344 }
15345
15346 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15347 probe != NULL; probe = probe->dtpr_nextmod) {
15348 if (probe->dtpr_ecb != NULL) {
15349 mutex_exit(&dtrace_provider_lock);
15350 mutex_exit(&mod_lock);
15351 mutex_exit(&dtrace_lock);
15352
15353 /*
15354 * This shouldn't _actually_ be possible -- we're
15355 * unloading a module that has an enabled probe in it.
15356 * (It's normally up to the provider to make sure that
15357 * this can't happen.) However, because dtps_enable()
15358 * doesn't have a failure mode, there can be an
15359 * enable/unload race. Upshot: we don't want to
15360 * assert, but we're not going to disable the
15361 * probe, either.
15362 */
15363 if (dtrace_err_verbose) {
15364 cmn_err(CE_WARN, "unloaded module '%s' had "
15365 "enabled probes", ctl->mod_modname);
15366 }
15367
15368 return;
15369 }
15370 }
15371
15372 probe = first;
15373
15374 for (first = NULL; probe != NULL; probe = next) {
15375 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15376
15377 dtrace_probes[probe->dtpr_id - 1] = NULL;
15378
15379 next = probe->dtpr_nextmod;
15380 dtrace_hash_remove(dtrace_bymod, probe);
15381 dtrace_hash_remove(dtrace_byfunc, probe);
15382 dtrace_hash_remove(dtrace_byname, probe);
15383
15384 if (first == NULL) {
15385 first = probe;
15386 probe->dtpr_nextmod = NULL;
15387 } else {
15388 probe->dtpr_nextmod = first;
15389 first = probe;
15390 }
15391 }
15392
15393 /*
15394 * We've removed all of the module's probes from the hash chains and
15395 * from the probe array. Now issue a dtrace_sync() to be sure that
15396 * everyone has cleared out from any probe array processing.
15397 */
15398 dtrace_sync();
15399
15400 for (probe = first; probe != NULL; probe = first) {
15401 first = probe->dtpr_nextmod;
15402 prov = probe->dtpr_provider;
15403 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15404 probe->dtpr_arg);
15405 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15406 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15407 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15408 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15409 kmem_free(probe, sizeof (dtrace_probe_t));
15410 }
15411
15412 mutex_exit(&dtrace_lock);
15413 mutex_exit(&mod_lock);
15414 mutex_exit(&dtrace_provider_lock);
15415 }
15416
15417 void
15418 dtrace_suspend(void)
15419 {
15420 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15421 }
15422
15423 void
15424 dtrace_resume(void)
15425 {
15426 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15427 }
15428
15429 static int
15430 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15431 {
15432 ASSERT(MUTEX_HELD(&cpu_lock));
15433 mutex_enter(&dtrace_lock);
15434
15435 switch (what) {
15436 case CPU_CONFIG: {
15437 dtrace_state_t *state;
15438 dtrace_optval_t *opt, rs, c;
15439
15440 /*
15441 * For now, we only allocate a new buffer for anonymous state.
15442 */
15443 if ((state = dtrace_anon.dta_state) == NULL)
15444 break;
15445
15446 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15447 break;
15448
15449 opt = state->dts_options;
15450 c = opt[DTRACEOPT_CPU];
15451
15452 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15453 break;
15454
15455 /*
15456 * Regardless of what the actual policy is, we're going to
15457 * temporarily set our resize policy to be manual. We're
15458 * also going to temporarily set our CPU option to denote
15459 * the newly configured CPU.
15460 */
15461 rs = opt[DTRACEOPT_BUFRESIZE];
15462 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15463 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15464
15465 (void) dtrace_state_buffers(state);
15466
15467 opt[DTRACEOPT_BUFRESIZE] = rs;
15468 opt[DTRACEOPT_CPU] = c;
15469
15470 break;
15471 }
15472
15473 case CPU_UNCONFIG:
15474 /*
15475 * We don't free the buffer in the CPU_UNCONFIG case. (The
15476 * buffer will be freed when the consumer exits.)
15477 */
15478 break;
15479
15480 default:
15481 break;
15482 }
15483
15484 mutex_exit(&dtrace_lock);
15485 return (0);
15486 }
15487
15488 static void
15489 dtrace_cpu_setup_initial(processorid_t cpu)
15490 {
15491 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15492 }
15493
15494 static void
15495 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15496 {
15497 if (dtrace_toxranges >= dtrace_toxranges_max) {
15498 int osize, nsize;
15499 dtrace_toxrange_t *range;
15500
15501 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15502
15503 if (osize == 0) {
15504 ASSERT(dtrace_toxrange == NULL);
15505 ASSERT(dtrace_toxranges_max == 0);
15506 dtrace_toxranges_max = 1;
15507 } else {
15508 dtrace_toxranges_max <<= 1;
15509 }
15510
15511 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15512 range = kmem_zalloc(nsize, KM_SLEEP);
15513
15514 if (dtrace_toxrange != NULL) {
15515 ASSERT(osize != 0);
15516 bcopy(dtrace_toxrange, range, osize);
15517 kmem_free(dtrace_toxrange, osize);
15518 }
15519
15520 dtrace_toxrange = range;
15521 }
15522
15523 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
15524 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
15525
15526 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15527 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15528 dtrace_toxranges++;
15529 }
15530
15531 static void
15532 dtrace_getf_barrier()
15533 {
15534 /*
15535 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
15536 * that contain calls to getf(), this routine will be called on every
15537 * closef() before either the underlying vnode is released or the
15538 * file_t itself is freed. By the time we are here, it is essential
15539 * that the file_t can no longer be accessed from a call to getf()
15540 * in probe context -- that assures that a dtrace_sync() can be used
15541 * to clear out any enablings referring to the old structures.
15542 */
15543 if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
15544 kcred->cr_zone->zone_dtrace_getf != 0)
15545 dtrace_sync();
15546 }
15547
15548 /*
15549 * DTrace Driver Cookbook Functions
15550 */
15551 /*ARGSUSED*/
15552 static int
15553 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15554 {
15555 dtrace_provider_id_t id;
15556 dtrace_state_t *state = NULL;
15557 dtrace_enabling_t *enab;
15558
15559 mutex_enter(&cpu_lock);
15560 mutex_enter(&dtrace_provider_lock);
15561 mutex_enter(&dtrace_lock);
15562
15563 if (ddi_soft_state_init(&dtrace_softstate,
15564 sizeof (dtrace_state_t), 0) != 0) {
15565 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15566 mutex_exit(&cpu_lock);
15567 mutex_exit(&dtrace_provider_lock);
15568 mutex_exit(&dtrace_lock);
15569 return (DDI_FAILURE);
15570 }
15571
15572 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15573 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15574 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15575 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15576 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15577 ddi_remove_minor_node(devi, NULL);
15578 ddi_soft_state_fini(&dtrace_softstate);
15579 mutex_exit(&cpu_lock);
15580 mutex_exit(&dtrace_provider_lock);
15581 mutex_exit(&dtrace_lock);
15582 return (DDI_FAILURE);
15583 }
15584
15585 ddi_report_dev(devi);
15586 dtrace_devi = devi;
15587
15588 dtrace_modload = dtrace_module_loaded;
15589 dtrace_modunload = dtrace_module_unloaded;
15590 dtrace_cpu_init = dtrace_cpu_setup_initial;
15591 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15592 dtrace_helpers_fork = dtrace_helpers_duplicate;
15593 dtrace_cpustart_init = dtrace_suspend;
15594 dtrace_cpustart_fini = dtrace_resume;
15595 dtrace_debugger_init = dtrace_suspend;
15596 dtrace_debugger_fini = dtrace_resume;
15597
15598 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15599
15600 ASSERT(MUTEX_HELD(&cpu_lock));
15601
15602 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15603 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15604 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15605 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15606 VM_SLEEP | VMC_IDENTIFIER);
15607 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15608 1, INT_MAX, 0);
15609
15610 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15611 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15612 NULL, NULL, NULL, NULL, NULL, 0);
15613
15614 ASSERT(MUTEX_HELD(&cpu_lock));
15615 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15616 offsetof(dtrace_probe_t, dtpr_nextmod),
15617 offsetof(dtrace_probe_t, dtpr_prevmod));
15618
15619 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15620 offsetof(dtrace_probe_t, dtpr_nextfunc),
15621 offsetof(dtrace_probe_t, dtpr_prevfunc));
15622
15623 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15624 offsetof(dtrace_probe_t, dtpr_nextname),
15625 offsetof(dtrace_probe_t, dtpr_prevname));
15626
15627 if (dtrace_retain_max < 1) {
15628 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15629 "setting to 1", dtrace_retain_max);
15630 dtrace_retain_max = 1;
15631 }
15632
15633 /*
15634 * Now discover our toxic ranges.
15635 */
15636 dtrace_toxic_ranges(dtrace_toxrange_add);
15637
15638 /*
15639 * Before we register ourselves as a provider to our own framework,
15640 * we would like to assert that dtrace_provider is NULL -- but that's
15641 * not true if we were loaded as a dependency of a DTrace provider.
15642 * Once we've registered, we can assert that dtrace_provider is our
15643 * pseudo provider.
15644 */
15645 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15646 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15647
15648 ASSERT(dtrace_provider != NULL);
15649 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15650
15651 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15652 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15653 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15654 dtrace_provider, NULL, NULL, "END", 0, NULL);
15655 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15656 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15657
15658 dtrace_anon_property();
15659 mutex_exit(&cpu_lock);
15660
15661 /*
15662 * If DTrace helper tracing is enabled, we need to allocate the
15663 * trace buffer and initialize the values.
15664 */
15665 if (dtrace_helptrace_enabled) {
15666 ASSERT(dtrace_helptrace_buffer == NULL);
15667 dtrace_helptrace_buffer =
15668 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15669 dtrace_helptrace_next = 0;
15670 }
15671
15672 /*
15673 * If there are already providers, we must ask them to provide their
15674 * probes, and then match any anonymous enabling against them. Note
15675 * that there should be no other retained enablings at this time:
15676 * the only retained enablings at this time should be the anonymous
15677 * enabling.
15678 */
15679 if (dtrace_anon.dta_enabling != NULL) {
15680 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15681
15682 dtrace_enabling_provide(NULL);
15683 state = dtrace_anon.dta_state;
15684
15685 /*
15686 * We couldn't hold cpu_lock across the above call to
15687 * dtrace_enabling_provide(), but we must hold it to actually
15688 * enable the probes. We have to drop all of our locks, pick
15689 * up cpu_lock, and regain our locks before matching the
15690 * retained anonymous enabling.
15691 */
15692 mutex_exit(&dtrace_lock);
15693 mutex_exit(&dtrace_provider_lock);
15694
15695 mutex_enter(&cpu_lock);
15696 mutex_enter(&dtrace_provider_lock);
15697 mutex_enter(&dtrace_lock);
15698
15699 if ((enab = dtrace_anon.dta_enabling) != NULL)
15700 (void) dtrace_enabling_match(enab, NULL);
15701
15702 mutex_exit(&cpu_lock);
15703 }
15704
15705 mutex_exit(&dtrace_lock);
15706 mutex_exit(&dtrace_provider_lock);
15707
15708 if (state != NULL) {
15709 /*
15710 * If we created any anonymous state, set it going now.
15711 */
15712 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15713 }
15714
15715 return (DDI_SUCCESS);
15716 }
15717
15718 /*ARGSUSED*/
15719 static int
15720 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15721 {
15722 dtrace_state_t *state;
15723 uint32_t priv;
15724 uid_t uid;
15725 zoneid_t zoneid;
15726
15727 if (getminor(*devp) == DTRACEMNRN_HELPER)
15728 return (0);
15729
15730 /*
15731 * If this wasn't an open with the "helper" minor, then it must be
15732 * the "dtrace" minor.
15733 */
15734 if (getminor(*devp) != DTRACEMNRN_DTRACE)
15735 return (ENXIO);
15736
15737 /*
15738 * If no DTRACE_PRIV_* bits are set in the credential, then the
15739 * caller lacks sufficient permission to do anything with DTrace.
15740 */
15741 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15742 if (priv == DTRACE_PRIV_NONE)
15743 return (EACCES);
15744
15745 /*
15746 * Ask all providers to provide all their probes.
15747 */
15748 mutex_enter(&dtrace_provider_lock);
15749 dtrace_probe_provide(NULL, NULL);
15750 mutex_exit(&dtrace_provider_lock);
15751
15752 mutex_enter(&cpu_lock);
15753 mutex_enter(&dtrace_lock);
15754 dtrace_opens++;
15755 dtrace_membar_producer();
15756
15757 /*
15758 * If the kernel debugger is active (that is, if the kernel debugger
15759 * modified text in some way), we won't allow the open.
15760 */
15761 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15762 dtrace_opens--;
15763 mutex_exit(&cpu_lock);
15764 mutex_exit(&dtrace_lock);
15765 return (EBUSY);
15766 }
15767
15768 state = dtrace_state_create(devp, cred_p);
15769 mutex_exit(&cpu_lock);
15770
15771 if (state == NULL) {
15772 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15773 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15774 mutex_exit(&dtrace_lock);
15775 return (EAGAIN);
15776 }
15777
15778 mutex_exit(&dtrace_lock);
15779
15780 return (0);
15781 }
15782
15783 /*ARGSUSED*/
15784 static int
15785 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15786 {
15787 minor_t minor = getminor(dev);
15788 dtrace_state_t *state;
15789
15790 if (minor == DTRACEMNRN_HELPER)
15791 return (0);
15792
15793 state = ddi_get_soft_state(dtrace_softstate, minor);
15794
15795 mutex_enter(&cpu_lock);
15796 mutex_enter(&dtrace_lock);
15797
15798 if (state->dts_anon) {
15799 /*
15800 * There is anonymous state. Destroy that first.
15801 */
15802 ASSERT(dtrace_anon.dta_state == NULL);
15803 dtrace_state_destroy(state->dts_anon);
15804 }
15805
15806 dtrace_state_destroy(state);
15807 ASSERT(dtrace_opens > 0);
15808
15809 /*
15810 * Only relinquish control of the kernel debugger interface when there
15811 * are no consumers and no anonymous enablings.
15812 */
15813 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15814 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15815
15816 mutex_exit(&dtrace_lock);
15817 mutex_exit(&cpu_lock);
15818
15819 return (0);
15820 }
15821
15822 /*ARGSUSED*/
15823 static int
15824 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15825 {
15826 int rval;
15827 dof_helper_t help, *dhp = NULL;
15828
15829 switch (cmd) {
15830 case DTRACEHIOC_ADDDOF:
15831 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15832 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15833 return (EFAULT);
15834 }
15835
15836 dhp = &help;
15837 arg = (intptr_t)help.dofhp_dof;
15838 /*FALLTHROUGH*/
15839
15840 case DTRACEHIOC_ADD: {
15841 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15842
15843 if (dof == NULL)
15844 return (rval);
15845
15846 mutex_enter(&dtrace_lock);
15847
15848 /*
15849 * dtrace_helper_slurp() takes responsibility for the dof --
15850 * it may free it now or it may save it and free it later.
15851 */
15852 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15853 *rv = rval;
15854 rval = 0;
15855 } else {
15856 rval = EINVAL;
15857 }
15858
15859 mutex_exit(&dtrace_lock);
15860 return (rval);
15861 }
15862
15863 case DTRACEHIOC_REMOVE: {
15864 mutex_enter(&dtrace_lock);
15865 rval = dtrace_helper_destroygen(arg);
15866 mutex_exit(&dtrace_lock);
15867
15868 return (rval);
15869 }
15870
15871 default:
15872 break;
15873 }
15874
15875 return (ENOTTY);
15876 }
15877
15878 /*ARGSUSED*/
15879 static int
15880 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15881 {
15882 minor_t minor = getminor(dev);
15883 dtrace_state_t *state;
15884 int rval;
15885
15886 if (minor == DTRACEMNRN_HELPER)
15887 return (dtrace_ioctl_helper(cmd, arg, rv));
15888
15889 state = ddi_get_soft_state(dtrace_softstate, minor);
15890
15891 if (state->dts_anon) {
15892 ASSERT(dtrace_anon.dta_state == NULL);
15893 state = state->dts_anon;
15894 }
15895
15896 switch (cmd) {
15897 case DTRACEIOC_PROVIDER: {
15898 dtrace_providerdesc_t pvd;
15899 dtrace_provider_t *pvp;
15900
15901 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15902 return (EFAULT);
15903
15904 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15905 mutex_enter(&dtrace_provider_lock);
15906
15907 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15908 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15909 break;
15910 }
15911
15912 mutex_exit(&dtrace_provider_lock);
15913
15914 if (pvp == NULL)
15915 return (ESRCH);
15916
15917 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15918 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15919 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15920 return (EFAULT);
15921
15922 return (0);
15923 }
15924
15925 case DTRACEIOC_EPROBE: {
15926 dtrace_eprobedesc_t epdesc;
15927 dtrace_ecb_t *ecb;
15928 dtrace_action_t *act;
15929 void *buf;
15930 size_t size;
15931 uintptr_t dest;
15932 int nrecs;
15933
15934 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15935 return (EFAULT);
15936
15937 mutex_enter(&dtrace_lock);
15938
15939 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15940 mutex_exit(&dtrace_lock);
15941 return (EINVAL);
15942 }
15943
15944 if (ecb->dte_probe == NULL) {
15945 mutex_exit(&dtrace_lock);
15946 return (EINVAL);
15947 }
15948
15949 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15950 epdesc.dtepd_uarg = ecb->dte_uarg;
15951 epdesc.dtepd_size = ecb->dte_size;
15952
15953 nrecs = epdesc.dtepd_nrecs;
15954 epdesc.dtepd_nrecs = 0;
15955 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15956 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15957 continue;
15958
15959 epdesc.dtepd_nrecs++;
15960 }
15961
15962 /*
15963 * Now that we have the size, we need to allocate a temporary
15964 * buffer in which to store the complete description. We need
15965 * the temporary buffer to be able to drop dtrace_lock()
15966 * across the copyout(), below.
15967 */
15968 size = sizeof (dtrace_eprobedesc_t) +
15969 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15970
15971 buf = kmem_alloc(size, KM_SLEEP);
15972 dest = (uintptr_t)buf;
15973
15974 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15975 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15976
15977 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15978 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15979 continue;
15980
15981 if (nrecs-- == 0)
15982 break;
15983
15984 bcopy(&act->dta_rec, (void *)dest,
15985 sizeof (dtrace_recdesc_t));
15986 dest += sizeof (dtrace_recdesc_t);
15987 }
15988
15989 mutex_exit(&dtrace_lock);
15990
15991 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15992 kmem_free(buf, size);
15993 return (EFAULT);
15994 }
15995
15996 kmem_free(buf, size);
15997 return (0);
15998 }
15999
16000 case DTRACEIOC_AGGDESC: {
16001 dtrace_aggdesc_t aggdesc;
16002 dtrace_action_t *act;
16003 dtrace_aggregation_t *agg;
16004 int nrecs;
16005 uint32_t offs;
16006 dtrace_recdesc_t *lrec;
16007 void *buf;
16008 size_t size;
16009 uintptr_t dest;
16010
16011 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16012 return (EFAULT);
16013
16014 mutex_enter(&dtrace_lock);
16015
16016 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16017 mutex_exit(&dtrace_lock);
16018 return (EINVAL);
16019 }
16020
16021 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16022
16023 nrecs = aggdesc.dtagd_nrecs;
16024 aggdesc.dtagd_nrecs = 0;
16025
16026 offs = agg->dtag_base;
16027 lrec = &agg->dtag_action.dta_rec;
16028 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16029
16030 for (act = agg->dtag_first; ; act = act->dta_next) {
16031 ASSERT(act->dta_intuple ||
16032 DTRACEACT_ISAGG(act->dta_kind));
16033
16034 /*
16035 * If this action has a record size of zero, it
16036 * denotes an argument to the aggregating action.
16037 * Because the presence of this record doesn't (or
16038 * shouldn't) affect the way the data is interpreted,
16039 * we don't copy it out to save user-level the
16040 * confusion of dealing with a zero-length record.
16041 */
16042 if (act->dta_rec.dtrd_size == 0) {
16043 ASSERT(agg->dtag_hasarg);
16044 continue;
16045 }
16046
16047 aggdesc.dtagd_nrecs++;
16048
16049 if (act == &agg->dtag_action)
16050 break;
16051 }
16052
16053 /*
16054 * Now that we have the size, we need to allocate a temporary
16055 * buffer in which to store the complete description. We need
16056 * the temporary buffer to be able to drop dtrace_lock()
16057 * across the copyout(), below.
16058 */
16059 size = sizeof (dtrace_aggdesc_t) +
16060 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16061
16062 buf = kmem_alloc(size, KM_SLEEP);
16063 dest = (uintptr_t)buf;
16064
16065 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16066 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16067
16068 for (act = agg->dtag_first; ; act = act->dta_next) {
16069 dtrace_recdesc_t rec = act->dta_rec;
16070
16071 /*
16072 * See the comment in the above loop for why we pass
16073 * over zero-length records.
16074 */
16075 if (rec.dtrd_size == 0) {
16076 ASSERT(agg->dtag_hasarg);
16077 continue;
16078 }
16079
16080 if (nrecs-- == 0)
16081 break;
16082
16083 rec.dtrd_offset -= offs;
16084 bcopy(&rec, (void *)dest, sizeof (rec));
16085 dest += sizeof (dtrace_recdesc_t);
16086
16087 if (act == &agg->dtag_action)
16088 break;
16089 }
16090
16091 mutex_exit(&dtrace_lock);
16092
16093 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16094 kmem_free(buf, size);
16095 return (EFAULT);
16096 }
16097
16098 kmem_free(buf, size);
16099 return (0);
16100 }
16101
16102 case DTRACEIOC_ENABLE: {
16103 dof_hdr_t *dof;
16104 dtrace_enabling_t *enab = NULL;
16105 dtrace_vstate_t *vstate;
16106 int err = 0;
16107
16108 *rv = 0;
16109
16110 /*
16111 * If a NULL argument has been passed, we take this as our
16112 * cue to reevaluate our enablings.
16113 */
16114 if (arg == NULL) {
16115 dtrace_enabling_matchall();
16116
16117 return (0);
16118 }
16119
16120 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16121 return (rval);
16122
16123 mutex_enter(&cpu_lock);
16124 mutex_enter(&dtrace_lock);
16125 vstate = &state->dts_vstate;
16126
16127 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16128 mutex_exit(&dtrace_lock);
16129 mutex_exit(&cpu_lock);
16130 dtrace_dof_destroy(dof);
16131 return (EBUSY);
16132 }
16133
16134 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16135 mutex_exit(&dtrace_lock);
16136 mutex_exit(&cpu_lock);
16137 dtrace_dof_destroy(dof);
16138 return (EINVAL);
16139 }
16140
16141 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16142 dtrace_enabling_destroy(enab);
16143 mutex_exit(&dtrace_lock);
16144 mutex_exit(&cpu_lock);
16145 dtrace_dof_destroy(dof);
16146 return (rval);
16147 }
16148
16149 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16150 err = dtrace_enabling_retain(enab);
16151 } else {
16152 dtrace_enabling_destroy(enab);
16153 }
16154
16155 mutex_exit(&cpu_lock);
16156 mutex_exit(&dtrace_lock);
16157 dtrace_dof_destroy(dof);
16158
16159 return (err);
16160 }
16161
16162 case DTRACEIOC_REPLICATE: {
16163 dtrace_repldesc_t desc;
16164 dtrace_probedesc_t *match = &desc.dtrpd_match;
16165 dtrace_probedesc_t *create = &desc.dtrpd_create;
16166 int err;
16167
16168 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16169 return (EFAULT);
16170
16171 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16172 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16173 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16174 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16175
16176 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16177 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16178 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16179 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16180
16181 mutex_enter(&dtrace_lock);
16182 err = dtrace_enabling_replicate(state, match, create);
16183 mutex_exit(&dtrace_lock);
16184
16185 return (err);
16186 }
16187
16188 case DTRACEIOC_PROBEMATCH:
16189 case DTRACEIOC_PROBES: {
16190 dtrace_probe_t *probe = NULL;
16191 dtrace_probedesc_t desc;
16192 dtrace_probekey_t pkey;
16193 dtrace_id_t i;
16194 int m = 0;
16195 uint32_t priv;
16196 uid_t uid;
16197 zoneid_t zoneid;
16198
16199 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16200 return (EFAULT);
16201
16202 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16203 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16204 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16205 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16206
16207 /*
16208 * Before we attempt to match this probe, we want to give
16209 * all providers the opportunity to provide it.
16210 */
16211 if (desc.dtpd_id == DTRACE_IDNONE) {
16212 mutex_enter(&dtrace_provider_lock);
16213 dtrace_probe_provide(&desc, NULL);
16214 mutex_exit(&dtrace_provider_lock);
16215 desc.dtpd_id++;
16216 }
16217
16218 if (cmd == DTRACEIOC_PROBEMATCH) {
16219 dtrace_probekey(&desc, &pkey);
16220 pkey.dtpk_id = DTRACE_IDNONE;
16221 }
16222
16223 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16224
16225 mutex_enter(&dtrace_lock);
16226
16227 if (cmd == DTRACEIOC_PROBEMATCH) {
16228 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16229 if ((probe = dtrace_probes[i - 1]) != NULL &&
16230 (m = dtrace_match_probe(probe, &pkey,
16231 priv, uid, zoneid)) != 0)
16232 break;
16233 }
16234
16235 if (m < 0) {
16236 mutex_exit(&dtrace_lock);
16237 return (EINVAL);
16238 }
16239
16240 } else {
16241 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16242 if ((probe = dtrace_probes[i - 1]) != NULL &&
16243 dtrace_match_priv(probe, priv, uid, zoneid))
16244 break;
16245 }
16246 }
16247
16248 if (probe == NULL) {
16249 mutex_exit(&dtrace_lock);
16250 return (ESRCH);
16251 }
16252
16253 dtrace_probe_description(probe, &desc);
16254 mutex_exit(&dtrace_lock);
16255
16256 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16257 return (EFAULT);
16258
16259 return (0);
16260 }
16261
16262 case DTRACEIOC_PROBEARG: {
16263 dtrace_argdesc_t desc;
16264 dtrace_probe_t *probe;
16265 dtrace_provider_t *prov;
16266
16267 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16268 return (EFAULT);
16269
16270 if (desc.dtargd_id == DTRACE_IDNONE)
16271 return (EINVAL);
16272
16273 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16274 return (EINVAL);
16275
16276 mutex_enter(&dtrace_provider_lock);
16277 mutex_enter(&mod_lock);
16278 mutex_enter(&dtrace_lock);
16279
16280 if (desc.dtargd_id > dtrace_nprobes) {
16281 mutex_exit(&dtrace_lock);
16282 mutex_exit(&mod_lock);
16283 mutex_exit(&dtrace_provider_lock);
16284 return (EINVAL);
16285 }
16286
16287 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16288 mutex_exit(&dtrace_lock);
16289 mutex_exit(&mod_lock);
16290 mutex_exit(&dtrace_provider_lock);
16291 return (EINVAL);
16292 }
16293
16294 mutex_exit(&dtrace_lock);
16295
16296 prov = probe->dtpr_provider;
16297
16298 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16299 /*
16300 * There isn't any typed information for this probe.
16301 * Set the argument number to DTRACE_ARGNONE.
16302 */
16303 desc.dtargd_ndx = DTRACE_ARGNONE;
16304 } else {
16305 desc.dtargd_native[0] = '\0';
16306 desc.dtargd_xlate[0] = '\0';
16307 desc.dtargd_mapping = desc.dtargd_ndx;
16308
16309 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16310 probe->dtpr_id, probe->dtpr_arg, &desc);
16311 }
16312
16313 mutex_exit(&mod_lock);
16314 mutex_exit(&dtrace_provider_lock);
16315
16316 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16317 return (EFAULT);
16318
16319 return (0);
16320 }
16321
16322 case DTRACEIOC_GO: {
16323 processorid_t cpuid;
16324 rval = dtrace_state_go(state, &cpuid);
16325
16326 if (rval != 0)
16327 return (rval);
16328
16329 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16330 return (EFAULT);
16331
16332 return (0);
16333 }
16334
16335 case DTRACEIOC_STOP: {
16336 processorid_t cpuid;
16337
16338 mutex_enter(&dtrace_lock);
16339 rval = dtrace_state_stop(state, &cpuid);
16340 mutex_exit(&dtrace_lock);
16341
16342 if (rval != 0)
16343 return (rval);
16344
16345 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16346 return (EFAULT);
16347
16348 return (0);
16349 }
16350
16351 case DTRACEIOC_DOFGET: {
16352 dof_hdr_t hdr, *dof;
16353 uint64_t len;
16354
16355 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16356 return (EFAULT);
16357
16358 mutex_enter(&dtrace_lock);
16359 dof = dtrace_dof_create(state);
16360 mutex_exit(&dtrace_lock);
16361
16362 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16363 rval = copyout(dof, (void *)arg, len);
16364 dtrace_dof_destroy(dof);
16365
16366 return (rval == 0 ? 0 : EFAULT);
16367 }
16368
16369 case DTRACEIOC_AGGSNAP:
16370 case DTRACEIOC_BUFSNAP: {
16371 dtrace_bufdesc_t desc;
16372 caddr_t cached;
16373 dtrace_buffer_t *buf;
16374
16375 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16376 return (EFAULT);
16377
16378 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16379 return (EINVAL);
16380
16381 mutex_enter(&dtrace_lock);
16382
16383 if (cmd == DTRACEIOC_BUFSNAP) {
16384 buf = &state->dts_buffer[desc.dtbd_cpu];
16385 } else {
16386 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16387 }
16388
16389 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16390 size_t sz = buf->dtb_offset;
16391
16392 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16393 mutex_exit(&dtrace_lock);
16394 return (EBUSY);
16395 }
16396
16397 /*
16398 * If this buffer has already been consumed, we're
16399 * going to indicate that there's nothing left here
16400 * to consume.
16401 */
16402 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16403 mutex_exit(&dtrace_lock);
16404
16405 desc.dtbd_size = 0;
16406 desc.dtbd_drops = 0;
16407 desc.dtbd_errors = 0;
16408 desc.dtbd_oldest = 0;
16409 sz = sizeof (desc);
16410
16411 if (copyout(&desc, (void *)arg, sz) != 0)
16412 return (EFAULT);
16413
16414 return (0);
16415 }
16416
16417 /*
16418 * If this is a ring buffer that has wrapped, we want
16419 * to copy the whole thing out.
16420 */
16421 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16422 dtrace_buffer_polish(buf);
16423 sz = buf->dtb_size;
16424 }
16425
16426 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16427 mutex_exit(&dtrace_lock);
16428 return (EFAULT);
16429 }
16430
16431 desc.dtbd_size = sz;
16432 desc.dtbd_drops = buf->dtb_drops;
16433 desc.dtbd_errors = buf->dtb_errors;
16434 desc.dtbd_oldest = buf->dtb_xamot_offset;
16435 desc.dtbd_timestamp = dtrace_gethrtime();
16436
16437 mutex_exit(&dtrace_lock);
16438
16439 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16440 return (EFAULT);
16441
16442 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16443
16444 return (0);
16445 }
16446
16447 if (buf->dtb_tomax == NULL) {
16448 ASSERT(buf->dtb_xamot == NULL);
16449 mutex_exit(&dtrace_lock);
16450 return (ENOENT);
16451 }
16452
16453 cached = buf->dtb_tomax;
16454 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16455
16456 dtrace_xcall(desc.dtbd_cpu,
16457 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16458
16459 state->dts_errors += buf->dtb_xamot_errors;
16460
16461 /*
16462 * If the buffers did not actually switch, then the cross call
16463 * did not take place -- presumably because the given CPU is
16464 * not in the ready set. If this is the case, we'll return
16465 * ENOENT.
16466 */
16467 if (buf->dtb_tomax == cached) {
16468 ASSERT(buf->dtb_xamot != cached);
16469 mutex_exit(&dtrace_lock);
16470 return (ENOENT);
16471 }
16472
16473 ASSERT(cached == buf->dtb_xamot);
16474
16475 /*
16476 * We have our snapshot; now copy it out.
16477 */
16478 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16479 buf->dtb_xamot_offset) != 0) {
16480 mutex_exit(&dtrace_lock);
16481 return (EFAULT);
16482 }
16483
16484 desc.dtbd_size = buf->dtb_xamot_offset;
16485 desc.dtbd_drops = buf->dtb_xamot_drops;
16486 desc.dtbd_errors = buf->dtb_xamot_errors;
16487 desc.dtbd_oldest = 0;
16488 desc.dtbd_timestamp = buf->dtb_switched;
16489
16490 mutex_exit(&dtrace_lock);
16491
16492 /*
16493 * Finally, copy out the buffer description.
16494 */
16495 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16496 return (EFAULT);
16497
16498 return (0);
16499 }
16500
16501 case DTRACEIOC_CONF: {
16502 dtrace_conf_t conf;
16503
16504 bzero(&conf, sizeof (conf));
16505 conf.dtc_difversion = DIF_VERSION;
16506 conf.dtc_difintregs = DIF_DIR_NREGS;
16507 conf.dtc_diftupregs = DIF_DTR_NREGS;
16508 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16509
16510 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16511 return (EFAULT);
16512
16513 return (0);
16514 }
16515
16516 case DTRACEIOC_STATUS: {
16517 dtrace_status_t stat;
16518 dtrace_dstate_t *dstate;
16519 int i, j;
16520 uint64_t nerrs;
16521
16522 /*
16523 * See the comment in dtrace_state_deadman() for the reason
16524 * for setting dts_laststatus to INT64_MAX before setting
16525 * it to the correct value.
16526 */
16527 state->dts_laststatus = INT64_MAX;
16528 dtrace_membar_producer();
16529 state->dts_laststatus = dtrace_gethrtime();
16530
16531 bzero(&stat, sizeof (stat));
16532
16533 mutex_enter(&dtrace_lock);
16534
16535 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16536 mutex_exit(&dtrace_lock);
16537 return (ENOENT);
16538 }
16539
16540 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16541 stat.dtst_exiting = 1;
16542
16543 nerrs = state->dts_errors;
16544 dstate = &state->dts_vstate.dtvs_dynvars;
16545
16546 for (i = 0; i < NCPU; i++) {
16547 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16548
16549 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16550 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16551 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16552
16553 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16554 stat.dtst_filled++;
16555
16556 nerrs += state->dts_buffer[i].dtb_errors;
16557
16558 for (j = 0; j < state->dts_nspeculations; j++) {
16559 dtrace_speculation_t *spec;
16560 dtrace_buffer_t *buf;
16561
16562 spec = &state->dts_speculations[j];
16563 buf = &spec->dtsp_buffer[i];
16564 stat.dtst_specdrops += buf->dtb_xamot_drops;
16565 }
16566 }
16567
16568 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16569 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16570 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16571 stat.dtst_dblerrors = state->dts_dblerrors;
16572 stat.dtst_killed =
16573 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16574 stat.dtst_errors = nerrs;
16575
16576 mutex_exit(&dtrace_lock);
16577
16578 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16579 return (EFAULT);
16580
16581 return (0);
16582 }
16583
16584 case DTRACEIOC_FORMAT: {
16585 dtrace_fmtdesc_t fmt;
16586 char *str;
16587 int len;
16588
16589 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16590 return (EFAULT);
16591
16592 mutex_enter(&dtrace_lock);
16593
16594 if (fmt.dtfd_format == 0 ||
16595 fmt.dtfd_format > state->dts_nformats) {
16596 mutex_exit(&dtrace_lock);
16597 return (EINVAL);
16598 }
16599
16600 /*
16601 * Format strings are allocated contiguously and they are
16602 * never freed; if a format index is less than the number
16603 * of formats, we can assert that the format map is non-NULL
16604 * and that the format for the specified index is non-NULL.
16605 */
16606 ASSERT(state->dts_formats != NULL);
16607 str = state->dts_formats[fmt.dtfd_format - 1];
16608 ASSERT(str != NULL);
16609
16610 len = strlen(str) + 1;
16611
16612 if (len > fmt.dtfd_length) {
16613 fmt.dtfd_length = len;
16614
16615 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16616 mutex_exit(&dtrace_lock);
16617 return (EINVAL);
16618 }
16619 } else {
16620 if (copyout(str, fmt.dtfd_string, len) != 0) {
16621 mutex_exit(&dtrace_lock);
16622 return (EINVAL);
16623 }
16624 }
16625
16626 mutex_exit(&dtrace_lock);
16627 return (0);
16628 }
16629
16630 default:
16631 break;
16632 }
16633
16634 return (ENOTTY);
16635 }
16636
16637 /*ARGSUSED*/
16638 static int
16639 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16640 {
16641 dtrace_state_t *state;
16642
16643 switch (cmd) {
16644 case DDI_DETACH:
16645 break;
16646
16647 case DDI_SUSPEND:
16648 return (DDI_SUCCESS);
16649
16650 default:
16651 return (DDI_FAILURE);
16652 }
16653
16654 mutex_enter(&cpu_lock);
16655 mutex_enter(&dtrace_provider_lock);
16656 mutex_enter(&dtrace_lock);
16657
16658 ASSERT(dtrace_opens == 0);
16659
16660 if (dtrace_helpers > 0) {
16661 mutex_exit(&dtrace_provider_lock);
16662 mutex_exit(&dtrace_lock);
16663 mutex_exit(&cpu_lock);
16664 return (DDI_FAILURE);
16665 }
16666
16667 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16668 mutex_exit(&dtrace_provider_lock);
16669 mutex_exit(&dtrace_lock);
16670 mutex_exit(&cpu_lock);
16671 return (DDI_FAILURE);
16672 }
16673
16674 dtrace_provider = NULL;
16675
16676 if ((state = dtrace_anon_grab()) != NULL) {
16677 /*
16678 * If there were ECBs on this state, the provider should
16679 * have not been allowed to detach; assert that there is
16680 * none.
16681 */
16682 ASSERT(state->dts_necbs == 0);
16683 dtrace_state_destroy(state);
16684
16685 /*
16686 * If we're being detached with anonymous state, we need to
16687 * indicate to the kernel debugger that DTrace is now inactive.
16688 */
16689 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16690 }
16691
16692 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16693 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16694 dtrace_cpu_init = NULL;
16695 dtrace_helpers_cleanup = NULL;
16696 dtrace_helpers_fork = NULL;
16697 dtrace_cpustart_init = NULL;
16698 dtrace_cpustart_fini = NULL;
16699 dtrace_debugger_init = NULL;
16700 dtrace_debugger_fini = NULL;
16701 dtrace_modload = NULL;
16702 dtrace_modunload = NULL;
16703
16704 ASSERT(dtrace_getf == 0);
16705 ASSERT(dtrace_closef == NULL);
16706
16707 mutex_exit(&cpu_lock);
16708
16709 if (dtrace_helptrace_enabled) {
16710 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16711 dtrace_helptrace_buffer = NULL;
16712 }
16713
16714 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16715 dtrace_probes = NULL;
16716 dtrace_nprobes = 0;
16717
16718 dtrace_hash_destroy(dtrace_bymod);
16719 dtrace_hash_destroy(dtrace_byfunc);
16720 dtrace_hash_destroy(dtrace_byname);
16721 dtrace_bymod = NULL;
16722 dtrace_byfunc = NULL;
16723 dtrace_byname = NULL;
16724
16725 kmem_cache_destroy(dtrace_state_cache);
16726 vmem_destroy(dtrace_minor);
16727 vmem_destroy(dtrace_arena);
16728
16729 if (dtrace_toxrange != NULL) {
16730 kmem_free(dtrace_toxrange,
16731 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16732 dtrace_toxrange = NULL;
16733 dtrace_toxranges = 0;
16734 dtrace_toxranges_max = 0;
16735 }
16736
16737 ddi_remove_minor_node(dtrace_devi, NULL);
16738 dtrace_devi = NULL;
16739
16740 ddi_soft_state_fini(&dtrace_softstate);
16741
16742 ASSERT(dtrace_vtime_references == 0);
16743 ASSERT(dtrace_opens == 0);
16744 ASSERT(dtrace_retained == NULL);
16745
16746 mutex_exit(&dtrace_lock);
16747 mutex_exit(&dtrace_provider_lock);
16748
16749 /*
16750 * We don't destroy the task queue until after we have dropped our
16751 * locks (taskq_destroy() may block on running tasks). To prevent
16752 * attempting to do work after we have effectively detached but before
16753 * the task queue has been destroyed, all tasks dispatched via the
16754 * task queue must check that DTrace is still attached before
16755 * performing any operation.
16756 */
16757 taskq_destroy(dtrace_taskq);
16758 dtrace_taskq = NULL;
16759
16760 return (DDI_SUCCESS);
16761 }
16762
16763 /*ARGSUSED*/
16764 static int
16765 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16766 {
16767 int error;
16768
16769 switch (infocmd) {
16770 case DDI_INFO_DEVT2DEVINFO:
16771 *result = (void *)dtrace_devi;
16772 error = DDI_SUCCESS;
16773 break;
16774 case DDI_INFO_DEVT2INSTANCE:
16775 *result = (void *)0;
16776 error = DDI_SUCCESS;
16777 break;
16778 default:
16779 error = DDI_FAILURE;
16780 }
16781 return (error);
16782 }
16783
16784 static struct cb_ops dtrace_cb_ops = {
16785 dtrace_open, /* open */
16786 dtrace_close, /* close */
16787 nulldev, /* strategy */
16788 nulldev, /* print */
16789 nodev, /* dump */
16790 nodev, /* read */
16791 nodev, /* write */
16792 dtrace_ioctl, /* ioctl */
16793 nodev, /* devmap */
16794 nodev, /* mmap */
16795 nodev, /* segmap */
16796 nochpoll, /* poll */
16797 ddi_prop_op, /* cb_prop_op */
16798 0, /* streamtab */
16799 D_NEW | D_MP /* Driver compatibility flag */
16800 };
16801
16802 static struct dev_ops dtrace_ops = {
16803 DEVO_REV, /* devo_rev */
16804 0, /* refcnt */
16805 dtrace_info, /* get_dev_info */
16806 nulldev, /* identify */
16807 nulldev, /* probe */
16808 dtrace_attach, /* attach */
16809 dtrace_detach, /* detach */
16810 nodev, /* reset */
16811 &dtrace_cb_ops, /* driver operations */
16812 NULL, /* bus operations */
16813 nodev, /* dev power */
16814 ddi_quiesce_not_needed, /* quiesce */
16815 };
16816
16817 static struct modldrv modldrv = {
16818 &mod_driverops, /* module type (this is a pseudo driver) */
16819 "Dynamic Tracing", /* name of module */
16820 &dtrace_ops, /* driver ops */
16821 };
16822
16823 static struct modlinkage modlinkage = {
16824 MODREV_1,
16825 (void *)&modldrv,
16826 NULL
16827 };
16828
16829 int
16830 _init(void)
16831 {
16832 return (mod_install(&modlinkage));
16833 }
16834
16835 int
16836 _info(struct modinfo *modinfop)
16837 {
16838 return (mod_info(&modlinkage, modinfop));
16839 }
16840
16841 int
16842 _fini(void)
16843 {
16844 return (mod_remove(&modlinkage));
16845 }