1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
25 * Copyright (c) 2012 by Delphix. All rights reserved.
26 */
27
28 /*
29 * DTrace - Dynamic Tracing for Solaris
30 *
31 * This is the implementation of the Solaris Dynamic Tracing framework
32 * (DTrace). The user-visible interface to DTrace is described at length in
33 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
34 * library, the in-kernel DTrace framework, and the DTrace providers are
35 * described in the block comments in the <sys/dtrace.h> header file. The
36 * internal architecture of DTrace is described in the block comments in the
37 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
38 * implementation very much assume mastery of all of these sources; if one has
39 * an unanswered question about the implementation, one should consult them
40 * first.
41 *
42 * The functions here are ordered roughly as follows:
43 *
44 * - Probe context functions
45 * - Probe hashing functions
46 * - Non-probe context utility functions
47 * - Matching functions
48 * - Provider-to-Framework API functions
49 * - Probe management functions
50 * - DIF object functions
51 * - Format functions
52 * - Predicate functions
53 * - ECB functions
54 * - Buffer functions
55 * - Enabling functions
56 * - DOF functions
57 * - Anonymous enabling functions
58 * - Consumer state functions
59 * - Helper functions
60 * - Hook functions
61 * - Driver cookbook functions
62 *
63 * Each group of functions begins with a block comment labelled the "DTrace
64 * [Group] Functions", allowing one to find each block by searching forward
65 * on capital-f functions.
66 */
67 #include <sys/errno.h>
68 #include <sys/stat.h>
69 #include <sys/modctl.h>
70 #include <sys/conf.h>
71 #include <sys/systm.h>
72 #include <sys/ddi.h>
73 #include <sys/sunddi.h>
74 #include <sys/cpuvar.h>
75 #include <sys/kmem.h>
76 #include <sys/strsubr.h>
77 #include <sys/sysmacros.h>
78 #include <sys/dtrace_impl.h>
79 #include <sys/atomic.h>
80 #include <sys/cmn_err.h>
81 #include <sys/mutex_impl.h>
82 #include <sys/rwlock_impl.h>
83 #include <sys/ctf_api.h>
84 #include <sys/panic.h>
85 #include <sys/priv_impl.h>
86 #include <sys/policy.h>
87 #include <sys/cred_impl.h>
88 #include <sys/procfs_isa.h>
89 #include <sys/taskq.h>
90 #include <sys/mkdev.h>
91 #include <sys/kdi.h>
92 #include <sys/zone.h>
93 #include <sys/socket.h>
94 #include <netinet/in.h>
95
96 /*
97 * DTrace Tunable Variables
98 *
99 * The following variables may be tuned by adding a line to /etc/system that
100 * includes both the name of the DTrace module ("dtrace") and the name of the
101 * variable. For example:
102 *
103 * set dtrace:dtrace_destructive_disallow = 1
104 *
105 * In general, the only variables that one should be tuning this way are those
106 * that affect system-wide DTrace behavior, and for which the default behavior
107 * is undesirable. Most of these variables are tunable on a per-consumer
108 * basis using DTrace options, and need not be tuned on a system-wide basis.
109 * When tuning these variables, avoid pathological values; while some attempt
110 * is made to verify the integrity of these variables, they are not considered
111 * part of the supported interface to DTrace, and they are therefore not
112 * checked comprehensively. Further, these variables should not be tuned
113 * dynamically via "mdb -kw" or other means; they should only be tuned via
114 * /etc/system.
115 */
116 int dtrace_destructive_disallow = 0;
117 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
118 size_t dtrace_difo_maxsize = (256 * 1024);
119 dtrace_optval_t dtrace_dof_maxsize = (8 * 1024 * 1024);
120 size_t dtrace_global_maxsize = (16 * 1024);
121 size_t dtrace_actions_max = (16 * 1024);
122 size_t dtrace_retain_max = 1024;
123 dtrace_optval_t dtrace_helper_actions_max = 1024;
124 dtrace_optval_t dtrace_helper_providers_max = 32;
125 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
126 size_t dtrace_strsize_default = 256;
127 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
128 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
129 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
130 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
131 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
132 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
133 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
134 dtrace_optval_t dtrace_nspec_default = 1;
135 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
136 dtrace_optval_t dtrace_stackframes_default = 20;
137 dtrace_optval_t dtrace_ustackframes_default = 20;
138 dtrace_optval_t dtrace_jstackframes_default = 50;
139 dtrace_optval_t dtrace_jstackstrsize_default = 512;
140 int dtrace_msgdsize_max = 128;
141 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
142 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
143 int dtrace_devdepth_max = 32;
144 int dtrace_err_verbose;
145 hrtime_t dtrace_deadman_interval = NANOSEC;
146 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
147 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
148 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
149
150 /*
151 * DTrace External Variables
152 *
153 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
154 * available to DTrace consumers via the backtick (`) syntax. One of these,
155 * dtrace_zero, is made deliberately so: it is provided as a source of
156 * well-known, zero-filled memory. While this variable is not documented,
157 * it is used by some translators as an implementation detail.
158 */
159 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
160
161 /*
162 * DTrace Internal Variables
163 */
164 static dev_info_t *dtrace_devi; /* device info */
165 static vmem_t *dtrace_arena; /* probe ID arena */
166 static vmem_t *dtrace_minor; /* minor number arena */
167 static taskq_t *dtrace_taskq; /* task queue */
168 static dtrace_probe_t **dtrace_probes; /* array of all probes */
169 static int dtrace_nprobes; /* number of probes */
170 static dtrace_provider_t *dtrace_provider; /* provider list */
171 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
172 static int dtrace_opens; /* number of opens */
173 static int dtrace_helpers; /* number of helpers */
174 static int dtrace_getf; /* number of unpriv getf()s */
175 static void *dtrace_softstate; /* softstate pointer */
176 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
177 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
178 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
179 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
180 static int dtrace_toxranges; /* number of toxic ranges */
181 static int dtrace_toxranges_max; /* size of toxic range array */
182 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
183 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
184 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
185 static kthread_t *dtrace_panicked; /* panicking thread */
186 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
187 static dtrace_genid_t dtrace_probegen; /* current probe generation */
188 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
189 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
190 static dtrace_genid_t dtrace_retained_gen; /* current retained enab gen */
191 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
192 static int dtrace_dynvar_failclean; /* dynvars failed to clean */
193
194 /*
195 * DTrace Locking
196 * DTrace is protected by three (relatively coarse-grained) locks:
197 *
198 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
199 * including enabling state, probes, ECBs, consumer state, helper state,
200 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
201 * probe context is lock-free -- synchronization is handled via the
202 * dtrace_sync() cross call mechanism.
203 *
204 * (2) dtrace_provider_lock is required when manipulating provider state, or
205 * when provider state must be held constant.
206 *
207 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
208 * when meta provider state must be held constant.
209 *
210 * The lock ordering between these three locks is dtrace_meta_lock before
211 * dtrace_provider_lock before dtrace_lock. (In particular, there are
212 * several places where dtrace_provider_lock is held by the framework as it
213 * calls into the providers -- which then call back into the framework,
214 * grabbing dtrace_lock.)
215 *
216 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
217 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
218 * role as a coarse-grained lock; it is acquired before both of these locks.
219 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
220 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
221 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
222 * acquired _between_ dtrace_provider_lock and dtrace_lock.
223 */
224 static kmutex_t dtrace_lock; /* probe state lock */
225 static kmutex_t dtrace_provider_lock; /* provider state lock */
226 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
227
228 /*
229 * DTrace Provider Variables
230 *
231 * These are the variables relating to DTrace as a provider (that is, the
232 * provider of the BEGIN, END, and ERROR probes).
233 */
234 static dtrace_pattr_t dtrace_provider_attr = {
235 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
236 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
237 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
238 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
239 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
240 };
241
242 static void
243 dtrace_nullop(void)
244 {}
245
246 static int
247 dtrace_enable_nullop(void)
248 {
249 return (0);
250 }
251
252 static dtrace_pops_t dtrace_provider_ops = {
253 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
254 (void (*)(void *, struct modctl *))dtrace_nullop,
255 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop,
256 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
257 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
258 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
259 NULL,
260 NULL,
261 NULL,
262 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
263 };
264
265 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
266 static dtrace_id_t dtrace_probeid_end; /* special END probe */
267 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
268
269 /*
270 * DTrace Helper Tracing Variables
271 *
272 * These variables should be set dynamically to enable helper tracing. The
273 * only variables that should be set are dtrace_helptrace_enable (which should
274 * be set to a non-zero value to allocate helper tracing buffers on the next
275 * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
276 * non-zero value to deallocate helper tracing buffers on the next close of
277 * /dev/dtrace). When (and only when) helper tracing is disabled, the
278 * buffer size may also be set via dtrace_helptrace_bufsize.
279 */
280 int dtrace_helptrace_enable = 0;
281 int dtrace_helptrace_disable = 0;
282 int dtrace_helptrace_bufsize = 16 * 1024 * 1024;
283 uint32_t dtrace_helptrace_nlocals;
284 static dtrace_helptrace_t *dtrace_helptrace_buffer;
285 static uint32_t dtrace_helptrace_next = 0;
286 static int dtrace_helptrace_wrapped = 0;
287
288 /*
289 * DTrace Error Hashing
290 *
291 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
292 * table. This is very useful for checking coverage of tests that are
293 * expected to induce DIF or DOF processing errors, and may be useful for
294 * debugging problems in the DIF code generator or in DOF generation . The
295 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
296 */
297 #ifdef DEBUG
298 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
299 static const char *dtrace_errlast;
300 static kthread_t *dtrace_errthread;
301 static kmutex_t dtrace_errlock;
302 #endif
303
304 /*
305 * DTrace Macros and Constants
306 *
307 * These are various macros that are useful in various spots in the
308 * implementation, along with a few random constants that have no meaning
309 * outside of the implementation. There is no real structure to this cpp
310 * mishmash -- but is there ever?
311 */
312 #define DTRACE_HASHSTR(hash, probe) \
313 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
314
315 #define DTRACE_HASHNEXT(hash, probe) \
316 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
317
318 #define DTRACE_HASHPREV(hash, probe) \
319 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
320
321 #define DTRACE_HASHEQ(hash, lhs, rhs) \
322 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
323 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
324
325 #define DTRACE_AGGHASHSIZE_SLEW 17
326
327 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
328
329 /*
330 * The key for a thread-local variable consists of the lower 61 bits of the
331 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
332 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
333 * equal to a variable identifier. This is necessary (but not sufficient) to
334 * assure that global associative arrays never collide with thread-local
335 * variables. To guarantee that they cannot collide, we must also define the
336 * order for keying dynamic variables. That order is:
337 *
338 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
339 *
340 * Because the variable-key and the tls-key are in orthogonal spaces, there is
341 * no way for a global variable key signature to match a thread-local key
342 * signature.
343 */
344 #define DTRACE_TLS_THRKEY(where) { \
345 uint_t intr = 0; \
346 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
347 for (; actv; actv >>= 1) \
348 intr++; \
349 ASSERT(intr < (1 << 3)); \
350 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
351 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
352 }
353
354 #define DT_BSWAP_8(x) ((x) & 0xff)
355 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
356 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
357 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
358
359 #define DT_MASK_LO 0x00000000FFFFFFFFULL
360
361 #define DTRACE_STORE(type, tomax, offset, what) \
362 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
363
364 #ifndef __i386
365 #define DTRACE_ALIGNCHECK(addr, size, flags) \
366 if (addr & (size - 1)) { \
367 *flags |= CPU_DTRACE_BADALIGN; \
368 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
369 return (0); \
370 }
371 #else
372 #define DTRACE_ALIGNCHECK(addr, size, flags)
373 #endif
374
375 /*
376 * Test whether a range of memory starting at testaddr of size testsz falls
377 * within the range of memory described by addr, sz. We take care to avoid
378 * problems with overflow and underflow of the unsigned quantities, and
379 * disallow all negative sizes. Ranges of size 0 are allowed.
380 */
381 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
382 ((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
383 (testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
384 (testaddr) + (testsz) >= (testaddr))
385
386 /*
387 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
388 * alloc_sz on the righthand side of the comparison in order to avoid overflow
389 * or underflow in the comparison with it. This is simpler than the INRANGE
390 * check above, because we know that the dtms_scratch_ptr is valid in the
391 * range. Allocations of size zero are allowed.
392 */
393 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
394 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
395 (mstate)->dtms_scratch_ptr >= (alloc_sz))
396
397 #define DTRACE_LOADFUNC(bits) \
398 /*CSTYLED*/ \
399 uint##bits##_t \
400 dtrace_load##bits(uintptr_t addr) \
401 { \
402 size_t size = bits / NBBY; \
403 /*CSTYLED*/ \
404 uint##bits##_t rval; \
405 int i; \
406 volatile uint16_t *flags = (volatile uint16_t *) \
407 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; \
408 \
409 DTRACE_ALIGNCHECK(addr, size, flags); \
410 \
411 for (i = 0; i < dtrace_toxranges; i++) { \
412 if (addr >= dtrace_toxrange[i].dtt_limit) \
413 continue; \
414 \
415 if (addr + size <= dtrace_toxrange[i].dtt_base) \
416 continue; \
417 \
418 /* \
419 * This address falls within a toxic region; return 0. \
420 */ \
421 *flags |= CPU_DTRACE_BADADDR; \
422 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
423 return (0); \
424 } \
425 \
426 *flags |= CPU_DTRACE_NOFAULT; \
427 /*CSTYLED*/ \
428 rval = *((volatile uint##bits##_t *)addr); \
429 *flags &= ~CPU_DTRACE_NOFAULT; \
430 \
431 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
432 }
433
434 #ifdef _LP64
435 #define dtrace_loadptr dtrace_load64
436 #else
437 #define dtrace_loadptr dtrace_load32
438 #endif
439
440 #define DTRACE_DYNHASH_FREE 0
441 #define DTRACE_DYNHASH_SINK 1
442 #define DTRACE_DYNHASH_VALID 2
443
444 #define DTRACE_MATCH_FAIL -1
445 #define DTRACE_MATCH_NEXT 0
446 #define DTRACE_MATCH_DONE 1
447 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
448 #define DTRACE_STATE_ALIGN 64
449
450 #define DTRACE_FLAGS2FLT(flags) \
451 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
452 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
453 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
454 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
455 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
456 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
457 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
458 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
459 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
460 DTRACEFLT_UNKNOWN)
461
462 #define DTRACEACT_ISSTRING(act) \
463 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
464 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
465
466 static size_t dtrace_strlen(const char *, size_t);
467 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
468 static void dtrace_enabling_provide(dtrace_provider_t *);
469 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
470 static void dtrace_enabling_matchall(void);
471 static void dtrace_enabling_reap(void);
472 static dtrace_state_t *dtrace_anon_grab(void);
473 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
474 dtrace_state_t *, uint64_t, uint64_t);
475 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
476 static void dtrace_buffer_drop(dtrace_buffer_t *);
477 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
478 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
479 dtrace_state_t *, dtrace_mstate_t *);
480 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
481 dtrace_optval_t);
482 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
483 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
484 static int dtrace_priv_proc(dtrace_state_t *, dtrace_mstate_t *);
485 static void dtrace_getf_barrier(void);
486
487 /*
488 * DTrace Probe Context Functions
489 *
490 * These functions are called from probe context. Because probe context is
491 * any context in which C may be called, arbitrarily locks may be held,
492 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
493 * As a result, functions called from probe context may only call other DTrace
494 * support functions -- they may not interact at all with the system at large.
495 * (Note that the ASSERT macro is made probe-context safe by redefining it in
496 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
497 * loads are to be performed from probe context, they _must_ be in terms of
498 * the safe dtrace_load*() variants.
499 *
500 * Some functions in this block are not actually called from probe context;
501 * for these functions, there will be a comment above the function reading
502 * "Note: not called from probe context."
503 */
504 void
505 dtrace_panic(const char *format, ...)
506 {
507 va_list alist;
508
509 va_start(alist, format);
510 dtrace_vpanic(format, alist);
511 va_end(alist);
512 }
513
514 int
515 dtrace_assfail(const char *a, const char *f, int l)
516 {
517 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
518
519 /*
520 * We just need something here that even the most clever compiler
521 * cannot optimize away.
522 */
523 return (a[(uintptr_t)f]);
524 }
525
526 /*
527 * Atomically increment a specified error counter from probe context.
528 */
529 static void
530 dtrace_error(uint32_t *counter)
531 {
532 /*
533 * Most counters stored to in probe context are per-CPU counters.
534 * However, there are some error conditions that are sufficiently
535 * arcane that they don't merit per-CPU storage. If these counters
536 * are incremented concurrently on different CPUs, scalability will be
537 * adversely affected -- but we don't expect them to be white-hot in a
538 * correctly constructed enabling...
539 */
540 uint32_t oval, nval;
541
542 do {
543 oval = *counter;
544
545 if ((nval = oval + 1) == 0) {
546 /*
547 * If the counter would wrap, set it to 1 -- assuring
548 * that the counter is never zero when we have seen
549 * errors. (The counter must be 32-bits because we
550 * aren't guaranteed a 64-bit compare&swap operation.)
551 * To save this code both the infamy of being fingered
552 * by a priggish news story and the indignity of being
553 * the target of a neo-puritan witch trial, we're
554 * carefully avoiding any colorful description of the
555 * likelihood of this condition -- but suffice it to
556 * say that it is only slightly more likely than the
557 * overflow of predicate cache IDs, as discussed in
558 * dtrace_predicate_create().
559 */
560 nval = 1;
561 }
562 } while (dtrace_cas32(counter, oval, nval) != oval);
563 }
564
565 /*
566 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
567 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
568 */
569 DTRACE_LOADFUNC(8)
570 DTRACE_LOADFUNC(16)
571 DTRACE_LOADFUNC(32)
572 DTRACE_LOADFUNC(64)
573
574 static int
575 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
576 {
577 if (dest < mstate->dtms_scratch_base)
578 return (0);
579
580 if (dest + size < dest)
581 return (0);
582
583 if (dest + size > mstate->dtms_scratch_ptr)
584 return (0);
585
586 return (1);
587 }
588
589 static int
590 dtrace_canstore_statvar(uint64_t addr, size_t sz,
591 dtrace_statvar_t **svars, int nsvars)
592 {
593 int i;
594
595 for (i = 0; i < nsvars; i++) {
596 dtrace_statvar_t *svar = svars[i];
597
598 if (svar == NULL || svar->dtsv_size == 0)
599 continue;
600
601 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
602 return (1);
603 }
604
605 return (0);
606 }
607
608 /*
609 * Check to see if the address is within a memory region to which a store may
610 * be issued. This includes the DTrace scratch areas, and any DTrace variable
611 * region. The caller of dtrace_canstore() is responsible for performing any
612 * alignment checks that are needed before stores are actually executed.
613 */
614 static int
615 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
616 dtrace_vstate_t *vstate)
617 {
618 /*
619 * First, check to see if the address is in scratch space...
620 */
621 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
622 mstate->dtms_scratch_size))
623 return (1);
624
625 /*
626 * Now check to see if it's a dynamic variable. This check will pick
627 * up both thread-local variables and any global dynamically-allocated
628 * variables.
629 */
630 if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
631 vstate->dtvs_dynvars.dtds_size)) {
632 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
633 uintptr_t base = (uintptr_t)dstate->dtds_base +
634 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
635 uintptr_t chunkoffs;
636
637 /*
638 * Before we assume that we can store here, we need to make
639 * sure that it isn't in our metadata -- storing to our
640 * dynamic variable metadata would corrupt our state. For
641 * the range to not include any dynamic variable metadata,
642 * it must:
643 *
644 * (1) Start above the hash table that is at the base of
645 * the dynamic variable space
646 *
647 * (2) Have a starting chunk offset that is beyond the
648 * dtrace_dynvar_t that is at the base of every chunk
649 *
650 * (3) Not span a chunk boundary
651 *
652 */
653 if (addr < base)
654 return (0);
655
656 chunkoffs = (addr - base) % dstate->dtds_chunksize;
657
658 if (chunkoffs < sizeof (dtrace_dynvar_t))
659 return (0);
660
661 if (chunkoffs + sz > dstate->dtds_chunksize)
662 return (0);
663
664 return (1);
665 }
666
667 /*
668 * Finally, check the static local and global variables. These checks
669 * take the longest, so we perform them last.
670 */
671 if (dtrace_canstore_statvar(addr, sz,
672 vstate->dtvs_locals, vstate->dtvs_nlocals))
673 return (1);
674
675 if (dtrace_canstore_statvar(addr, sz,
676 vstate->dtvs_globals, vstate->dtvs_nglobals))
677 return (1);
678
679 return (0);
680 }
681
682
683 /*
684 * Convenience routine to check to see if the address is within a memory
685 * region in which a load may be issued given the user's privilege level;
686 * if not, it sets the appropriate error flags and loads 'addr' into the
687 * illegal value slot.
688 *
689 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
690 * appropriate memory access protection.
691 */
692 static int
693 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
694 dtrace_vstate_t *vstate)
695 {
696 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
697 file_t *fp;
698
699 /*
700 * If we hold the privilege to read from kernel memory, then
701 * everything is readable.
702 */
703 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
704 return (1);
705
706 /*
707 * You can obviously read that which you can store.
708 */
709 if (dtrace_canstore(addr, sz, mstate, vstate))
710 return (1);
711
712 /*
713 * We're allowed to read from our own string table.
714 */
715 if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
716 mstate->dtms_difo->dtdo_strlen))
717 return (1);
718
719 if (vstate->dtvs_state != NULL &&
720 dtrace_priv_proc(vstate->dtvs_state, mstate)) {
721 proc_t *p;
722
723 /*
724 * When we have privileges to the current process, there are
725 * several context-related kernel structures that are safe to
726 * read, even absent the privilege to read from kernel memory.
727 * These reads are safe because these structures contain only
728 * state that (1) we're permitted to read, (2) is harmless or
729 * (3) contains pointers to additional kernel state that we're
730 * not permitted to read (and as such, do not present an
731 * opportunity for privilege escalation). Finally (and
732 * critically), because of the nature of their relation with
733 * the current thread context, the memory associated with these
734 * structures cannot change over the duration of probe context,
735 * and it is therefore impossible for this memory to be
736 * deallocated and reallocated as something else while it's
737 * being operated upon.
738 */
739 if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t)))
740 return (1);
741
742 if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
743 sz, curthread->t_procp, sizeof (proc_t))) {
744 return (1);
745 }
746
747 if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
748 curthread->t_cred, sizeof (cred_t))) {
749 return (1);
750 }
751
752 if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
753 &(p->p_pidp->pid_id), sizeof (pid_t))) {
754 return (1);
755 }
756
757 if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
758 curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
759 return (1);
760 }
761 }
762
763 if ((fp = mstate->dtms_getf) != NULL) {
764 uintptr_t psz = sizeof (void *);
765 vnode_t *vp;
766 vnodeops_t *op;
767
768 /*
769 * When getf() returns a file_t, the enabling is implicitly
770 * granted the (transient) right to read the returned file_t
771 * as well as the v_path and v_op->vnop_name of the underlying
772 * vnode. These accesses are allowed after a successful
773 * getf() because the members that they refer to cannot change
774 * once set -- and the barrier logic in the kernel's closef()
775 * path assures that the file_t and its referenced vode_t
776 * cannot themselves be stale (that is, it impossible for
777 * either dtms_getf itself or its f_vnode member to reference
778 * freed memory).
779 */
780 if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t)))
781 return (1);
782
783 if ((vp = fp->f_vnode) != NULL) {
784 if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz))
785 return (1);
786
787 if (vp->v_path != NULL && DTRACE_INRANGE(addr, sz,
788 vp->v_path, strlen(vp->v_path) + 1)) {
789 return (1);
790 }
791
792 if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz))
793 return (1);
794
795 if ((op = vp->v_op) != NULL &&
796 DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
797 return (1);
798 }
799
800 if (op != NULL && op->vnop_name != NULL &&
801 DTRACE_INRANGE(addr, sz, op->vnop_name,
802 strlen(op->vnop_name) + 1)) {
803 return (1);
804 }
805 }
806 }
807
808 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
809 *illval = addr;
810 return (0);
811 }
812
813 /*
814 * Convenience routine to check to see if a given string is within a memory
815 * region in which a load may be issued given the user's privilege level;
816 * this exists so that we don't need to issue unnecessary dtrace_strlen()
817 * calls in the event that the user has all privileges.
818 */
819 static int
820 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
821 dtrace_vstate_t *vstate)
822 {
823 size_t strsz;
824
825 /*
826 * If we hold the privilege to read from kernel memory, then
827 * everything is readable.
828 */
829 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
830 return (1);
831
832 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
833 if (dtrace_canload(addr, strsz, mstate, vstate))
834 return (1);
835
836 return (0);
837 }
838
839 /*
840 * Convenience routine to check to see if a given variable is within a memory
841 * region in which a load may be issued given the user's privilege level.
842 */
843 static int
844 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
845 dtrace_vstate_t *vstate)
846 {
847 size_t sz, strsize;
848 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
849
850 /*
851 * If we hold the privilege to read from kernel memory, then
852 * everything is readable.
853 */
854 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
855 return (1);
856
857 if (type->dtdt_kind == DIF_TYPE_STRING) {
858 dtrace_state_t *state = vstate->dtvs_state;
859
860 if (state != NULL) {
861 strsize = state->dts_options[DTRACEOPT_STRSIZE];
862 } else {
863 /*
864 * In helper context, we have a NULL state; fall back
865 * to using the system-wide default for the string size
866 * in this case.
867 */
868 strsize = dtrace_strsize_default;
869 }
870
871 sz = dtrace_strlen(src, strsize) + 1;
872 } else {
873 sz = type->dtdt_size;
874 }
875
876 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
877 }
878
879 /*
880 * Compare two strings using safe loads.
881 */
882 static int
883 dtrace_strncmp(char *s1, char *s2, size_t limit)
884 {
885 uint8_t c1, c2;
886 volatile uint16_t *flags;
887
888 if (s1 == s2 || limit == 0)
889 return (0);
890
891 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
892
893 do {
894 if (s1 == NULL) {
895 c1 = '\0';
896 } else {
897 c1 = dtrace_load8((uintptr_t)s1++);
898 }
899
900 if (s2 == NULL) {
901 c2 = '\0';
902 } else {
903 c2 = dtrace_load8((uintptr_t)s2++);
904 }
905
906 if (c1 != c2)
907 return (c1 - c2);
908 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
909
910 return (0);
911 }
912
913 /*
914 * Compute strlen(s) for a string using safe memory accesses. The additional
915 * len parameter is used to specify a maximum length to ensure completion.
916 */
917 static size_t
918 dtrace_strlen(const char *s, size_t lim)
919 {
920 uint_t len;
921
922 for (len = 0; len != lim; len++) {
923 if (dtrace_load8((uintptr_t)s++) == '\0')
924 break;
925 }
926
927 return (len);
928 }
929
930 /*
931 * Check if an address falls within a toxic region.
932 */
933 static int
934 dtrace_istoxic(uintptr_t kaddr, size_t size)
935 {
936 uintptr_t taddr, tsize;
937 int i;
938
939 for (i = 0; i < dtrace_toxranges; i++) {
940 taddr = dtrace_toxrange[i].dtt_base;
941 tsize = dtrace_toxrange[i].dtt_limit - taddr;
942
943 if (kaddr - taddr < tsize) {
944 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
945 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
946 return (1);
947 }
948
949 if (taddr - kaddr < size) {
950 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
951 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
952 return (1);
953 }
954 }
955
956 return (0);
957 }
958
959 /*
960 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
961 * memory specified by the DIF program. The dst is assumed to be safe memory
962 * that we can store to directly because it is managed by DTrace. As with
963 * standard bcopy, overlapping copies are handled properly.
964 */
965 static void
966 dtrace_bcopy(const void *src, void *dst, size_t len)
967 {
968 if (len != 0) {
969 uint8_t *s1 = dst;
970 const uint8_t *s2 = src;
971
972 if (s1 <= s2) {
973 do {
974 *s1++ = dtrace_load8((uintptr_t)s2++);
975 } while (--len != 0);
976 } else {
977 s2 += len;
978 s1 += len;
979
980 do {
981 *--s1 = dtrace_load8((uintptr_t)--s2);
982 } while (--len != 0);
983 }
984 }
985 }
986
987 /*
988 * Copy src to dst using safe memory accesses, up to either the specified
989 * length, or the point that a nul byte is encountered. The src is assumed to
990 * be unsafe memory specified by the DIF program. The dst is assumed to be
991 * safe memory that we can store to directly because it is managed by DTrace.
992 * Unlike dtrace_bcopy(), overlapping regions are not handled.
993 */
994 static void
995 dtrace_strcpy(const void *src, void *dst, size_t len)
996 {
997 if (len != 0) {
998 uint8_t *s1 = dst, c;
999 const uint8_t *s2 = src;
1000
1001 do {
1002 *s1++ = c = dtrace_load8((uintptr_t)s2++);
1003 } while (--len != 0 && c != '\0');
1004 }
1005 }
1006
1007 /*
1008 * Copy src to dst, deriving the size and type from the specified (BYREF)
1009 * variable type. The src is assumed to be unsafe memory specified by the DIF
1010 * program. The dst is assumed to be DTrace variable memory that is of the
1011 * specified type; we assume that we can store to directly.
1012 */
1013 static void
1014 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1015 {
1016 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1017
1018 if (type->dtdt_kind == DIF_TYPE_STRING) {
1019 dtrace_strcpy(src, dst, type->dtdt_size);
1020 } else {
1021 dtrace_bcopy(src, dst, type->dtdt_size);
1022 }
1023 }
1024
1025 /*
1026 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1027 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1028 * safe memory that we can access directly because it is managed by DTrace.
1029 */
1030 static int
1031 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1032 {
1033 volatile uint16_t *flags;
1034
1035 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1036
1037 if (s1 == s2)
1038 return (0);
1039
1040 if (s1 == NULL || s2 == NULL)
1041 return (1);
1042
1043 if (s1 != s2 && len != 0) {
1044 const uint8_t *ps1 = s1;
1045 const uint8_t *ps2 = s2;
1046
1047 do {
1048 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1049 return (1);
1050 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1051 }
1052 return (0);
1053 }
1054
1055 /*
1056 * Zero the specified region using a simple byte-by-byte loop. Note that this
1057 * is for safe DTrace-managed memory only.
1058 */
1059 static void
1060 dtrace_bzero(void *dst, size_t len)
1061 {
1062 uchar_t *cp;
1063
1064 for (cp = dst; len != 0; len--)
1065 *cp++ = 0;
1066 }
1067
1068 static void
1069 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1070 {
1071 uint64_t result[2];
1072
1073 result[0] = addend1[0] + addend2[0];
1074 result[1] = addend1[1] + addend2[1] +
1075 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1076
1077 sum[0] = result[0];
1078 sum[1] = result[1];
1079 }
1080
1081 /*
1082 * Shift the 128-bit value in a by b. If b is positive, shift left.
1083 * If b is negative, shift right.
1084 */
1085 static void
1086 dtrace_shift_128(uint64_t *a, int b)
1087 {
1088 uint64_t mask;
1089
1090 if (b == 0)
1091 return;
1092
1093 if (b < 0) {
1094 b = -b;
1095 if (b >= 64) {
1096 a[0] = a[1] >> (b - 64);
1097 a[1] = 0;
1098 } else {
1099 a[0] >>= b;
1100 mask = 1LL << (64 - b);
1101 mask -= 1;
1102 a[0] |= ((a[1] & mask) << (64 - b));
1103 a[1] >>= b;
1104 }
1105 } else {
1106 if (b >= 64) {
1107 a[1] = a[0] << (b - 64);
1108 a[0] = 0;
1109 } else {
1110 a[1] <<= b;
1111 mask = a[0] >> (64 - b);
1112 a[1] |= mask;
1113 a[0] <<= b;
1114 }
1115 }
1116 }
1117
1118 /*
1119 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1120 * use native multiplication on those, and then re-combine into the
1121 * resulting 128-bit value.
1122 *
1123 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1124 * hi1 * hi2 << 64 +
1125 * hi1 * lo2 << 32 +
1126 * hi2 * lo1 << 32 +
1127 * lo1 * lo2
1128 */
1129 static void
1130 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1131 {
1132 uint64_t hi1, hi2, lo1, lo2;
1133 uint64_t tmp[2];
1134
1135 hi1 = factor1 >> 32;
1136 hi2 = factor2 >> 32;
1137
1138 lo1 = factor1 & DT_MASK_LO;
1139 lo2 = factor2 & DT_MASK_LO;
1140
1141 product[0] = lo1 * lo2;
1142 product[1] = hi1 * hi2;
1143
1144 tmp[0] = hi1 * lo2;
1145 tmp[1] = 0;
1146 dtrace_shift_128(tmp, 32);
1147 dtrace_add_128(product, tmp, product);
1148
1149 tmp[0] = hi2 * lo1;
1150 tmp[1] = 0;
1151 dtrace_shift_128(tmp, 32);
1152 dtrace_add_128(product, tmp, product);
1153 }
1154
1155 /*
1156 * This privilege check should be used by actions and subroutines to
1157 * verify that the user credentials of the process that enabled the
1158 * invoking ECB match the target credentials
1159 */
1160 static int
1161 dtrace_priv_proc_common_user(dtrace_state_t *state)
1162 {
1163 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1164
1165 /*
1166 * We should always have a non-NULL state cred here, since if cred
1167 * is null (anonymous tracing), we fast-path bypass this routine.
1168 */
1169 ASSERT(s_cr != NULL);
1170
1171 if ((cr = CRED()) != NULL &&
1172 s_cr->cr_uid == cr->cr_uid &&
1173 s_cr->cr_uid == cr->cr_ruid &&
1174 s_cr->cr_uid == cr->cr_suid &&
1175 s_cr->cr_gid == cr->cr_gid &&
1176 s_cr->cr_gid == cr->cr_rgid &&
1177 s_cr->cr_gid == cr->cr_sgid)
1178 return (1);
1179
1180 return (0);
1181 }
1182
1183 /*
1184 * This privilege check should be used by actions and subroutines to
1185 * verify that the zone of the process that enabled the invoking ECB
1186 * matches the target credentials
1187 */
1188 static int
1189 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1190 {
1191 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1192
1193 /*
1194 * We should always have a non-NULL state cred here, since if cred
1195 * is null (anonymous tracing), we fast-path bypass this routine.
1196 */
1197 ASSERT(s_cr != NULL);
1198
1199 if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1200 return (1);
1201
1202 return (0);
1203 }
1204
1205 /*
1206 * This privilege check should be used by actions and subroutines to
1207 * verify that the process has not setuid or changed credentials.
1208 */
1209 static int
1210 dtrace_priv_proc_common_nocd()
1211 {
1212 proc_t *proc;
1213
1214 if ((proc = ttoproc(curthread)) != NULL &&
1215 !(proc->p_flag & SNOCD))
1216 return (1);
1217
1218 return (0);
1219 }
1220
1221 static int
1222 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1223 {
1224 int action = state->dts_cred.dcr_action;
1225
1226 if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1227 goto bad;
1228
1229 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1230 dtrace_priv_proc_common_zone(state) == 0)
1231 goto bad;
1232
1233 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1234 dtrace_priv_proc_common_user(state) == 0)
1235 goto bad;
1236
1237 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1238 dtrace_priv_proc_common_nocd() == 0)
1239 goto bad;
1240
1241 return (1);
1242
1243 bad:
1244 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1245
1246 return (0);
1247 }
1248
1249 static int
1250 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1251 {
1252 if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1253 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1254 return (1);
1255
1256 if (dtrace_priv_proc_common_zone(state) &&
1257 dtrace_priv_proc_common_user(state) &&
1258 dtrace_priv_proc_common_nocd())
1259 return (1);
1260 }
1261
1262 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1263
1264 return (0);
1265 }
1266
1267 static int
1268 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1269 {
1270 if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1271 (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1272 return (1);
1273
1274 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1275
1276 return (0);
1277 }
1278
1279 static int
1280 dtrace_priv_kernel(dtrace_state_t *state)
1281 {
1282 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1283 return (1);
1284
1285 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1286
1287 return (0);
1288 }
1289
1290 static int
1291 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1292 {
1293 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1294 return (1);
1295
1296 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1297
1298 return (0);
1299 }
1300
1301 /*
1302 * Determine if the dte_cond of the specified ECB allows for processing of
1303 * the current probe to continue. Note that this routine may allow continued
1304 * processing, but with access(es) stripped from the mstate's dtms_access
1305 * field.
1306 */
1307 static int
1308 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1309 dtrace_ecb_t *ecb)
1310 {
1311 dtrace_probe_t *probe = ecb->dte_probe;
1312 dtrace_provider_t *prov = probe->dtpr_provider;
1313 dtrace_pops_t *pops = &prov->dtpv_pops;
1314 int mode = DTRACE_MODE_NOPRIV_DROP;
1315
1316 ASSERT(ecb->dte_cond);
1317
1318 if (pops->dtps_mode != NULL) {
1319 mode = pops->dtps_mode(prov->dtpv_arg,
1320 probe->dtpr_id, probe->dtpr_arg);
1321
1322 ASSERT(mode & (DTRACE_MODE_USER | DTRACE_MODE_KERNEL));
1323 ASSERT(mode & (DTRACE_MODE_NOPRIV_RESTRICT |
1324 DTRACE_MODE_NOPRIV_DROP));
1325 }
1326
1327 /*
1328 * If the dte_cond bits indicate that this consumer is only allowed to
1329 * see user-mode firings of this probe, check that the probe was fired
1330 * while in a user context. If that's not the case, use the policy
1331 * specified by the provider to determine if we drop the probe or
1332 * merely restrict operation.
1333 */
1334 if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1335 ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1336
1337 if (!(mode & DTRACE_MODE_USER)) {
1338 if (mode & DTRACE_MODE_NOPRIV_DROP)
1339 return (0);
1340
1341 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1342 }
1343 }
1344
1345 /*
1346 * This is more subtle than it looks. We have to be absolutely certain
1347 * that CRED() isn't going to change out from under us so it's only
1348 * legit to examine that structure if we're in constrained situations.
1349 * Currently, the only times we'll this check is if a non-super-user
1350 * has enabled the profile or syscall providers -- providers that
1351 * allow visibility of all processes. For the profile case, the check
1352 * above will ensure that we're examining a user context.
1353 */
1354 if (ecb->dte_cond & DTRACE_COND_OWNER) {
1355 cred_t *cr;
1356 cred_t *s_cr = state->dts_cred.dcr_cred;
1357 proc_t *proc;
1358
1359 ASSERT(s_cr != NULL);
1360
1361 if ((cr = CRED()) == NULL ||
1362 s_cr->cr_uid != cr->cr_uid ||
1363 s_cr->cr_uid != cr->cr_ruid ||
1364 s_cr->cr_uid != cr->cr_suid ||
1365 s_cr->cr_gid != cr->cr_gid ||
1366 s_cr->cr_gid != cr->cr_rgid ||
1367 s_cr->cr_gid != cr->cr_sgid ||
1368 (proc = ttoproc(curthread)) == NULL ||
1369 (proc->p_flag & SNOCD)) {
1370 if (mode & DTRACE_MODE_NOPRIV_DROP)
1371 return (0);
1372
1373 mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1374 }
1375 }
1376
1377 /*
1378 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1379 * in our zone, check to see if our mode policy is to restrict rather
1380 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1381 * and DTRACE_ACCESS_ARGS
1382 */
1383 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1384 cred_t *cr;
1385 cred_t *s_cr = state->dts_cred.dcr_cred;
1386
1387 ASSERT(s_cr != NULL);
1388
1389 if ((cr = CRED()) == NULL ||
1390 s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1391 if (mode & DTRACE_MODE_NOPRIV_DROP)
1392 return (0);
1393
1394 mstate->dtms_access &=
1395 ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1396 }
1397 }
1398
1399 /*
1400 * By merits of being in this code path at all, we have limited
1401 * privileges. If the provider has indicated that limited privileges
1402 * are to denote restricted operation, strip off the ability to access
1403 * arguments.
1404 */
1405 if (mode & DTRACE_MODE_LIMITEDPRIV_RESTRICT)
1406 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1407
1408 return (1);
1409 }
1410
1411 /*
1412 * Note: not called from probe context. This function is called
1413 * asynchronously (and at a regular interval) from outside of probe context to
1414 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1415 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1416 */
1417 void
1418 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1419 {
1420 dtrace_dynvar_t *dirty;
1421 dtrace_dstate_percpu_t *dcpu;
1422 dtrace_dynvar_t **rinsep;
1423 int i, j, work = 0;
1424
1425 for (i = 0; i < NCPU; i++) {
1426 dcpu = &dstate->dtds_percpu[i];
1427 rinsep = &dcpu->dtdsc_rinsing;
1428
1429 /*
1430 * If the dirty list is NULL, there is no dirty work to do.
1431 */
1432 if (dcpu->dtdsc_dirty == NULL)
1433 continue;
1434
1435 if (dcpu->dtdsc_rinsing != NULL) {
1436 /*
1437 * If the rinsing list is non-NULL, then it is because
1438 * this CPU was selected to accept another CPU's
1439 * dirty list -- and since that time, dirty buffers
1440 * have accumulated. This is a highly unlikely
1441 * condition, but we choose to ignore the dirty
1442 * buffers -- they'll be picked up a future cleanse.
1443 */
1444 continue;
1445 }
1446
1447 if (dcpu->dtdsc_clean != NULL) {
1448 /*
1449 * If the clean list is non-NULL, then we're in a
1450 * situation where a CPU has done deallocations (we
1451 * have a non-NULL dirty list) but no allocations (we
1452 * also have a non-NULL clean list). We can't simply
1453 * move the dirty list into the clean list on this
1454 * CPU, yet we also don't want to allow this condition
1455 * to persist, lest a short clean list prevent a
1456 * massive dirty list from being cleaned (which in
1457 * turn could lead to otherwise avoidable dynamic
1458 * drops). To deal with this, we look for some CPU
1459 * with a NULL clean list, NULL dirty list, and NULL
1460 * rinsing list -- and then we borrow this CPU to
1461 * rinse our dirty list.
1462 */
1463 for (j = 0; j < NCPU; j++) {
1464 dtrace_dstate_percpu_t *rinser;
1465
1466 rinser = &dstate->dtds_percpu[j];
1467
1468 if (rinser->dtdsc_rinsing != NULL)
1469 continue;
1470
1471 if (rinser->dtdsc_dirty != NULL)
1472 continue;
1473
1474 if (rinser->dtdsc_clean != NULL)
1475 continue;
1476
1477 rinsep = &rinser->dtdsc_rinsing;
1478 break;
1479 }
1480
1481 if (j == NCPU) {
1482 /*
1483 * We were unable to find another CPU that
1484 * could accept this dirty list -- we are
1485 * therefore unable to clean it now.
1486 */
1487 dtrace_dynvar_failclean++;
1488 continue;
1489 }
1490 }
1491
1492 work = 1;
1493
1494 /*
1495 * Atomically move the dirty list aside.
1496 */
1497 do {
1498 dirty = dcpu->dtdsc_dirty;
1499
1500 /*
1501 * Before we zap the dirty list, set the rinsing list.
1502 * (This allows for a potential assertion in
1503 * dtrace_dynvar(): if a free dynamic variable appears
1504 * on a hash chain, either the dirty list or the
1505 * rinsing list for some CPU must be non-NULL.)
1506 */
1507 *rinsep = dirty;
1508 dtrace_membar_producer();
1509 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1510 dirty, NULL) != dirty);
1511 }
1512
1513 if (!work) {
1514 /*
1515 * We have no work to do; we can simply return.
1516 */
1517 return;
1518 }
1519
1520 dtrace_sync();
1521
1522 for (i = 0; i < NCPU; i++) {
1523 dcpu = &dstate->dtds_percpu[i];
1524
1525 if (dcpu->dtdsc_rinsing == NULL)
1526 continue;
1527
1528 /*
1529 * We are now guaranteed that no hash chain contains a pointer
1530 * into this dirty list; we can make it clean.
1531 */
1532 ASSERT(dcpu->dtdsc_clean == NULL);
1533 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1534 dcpu->dtdsc_rinsing = NULL;
1535 }
1536
1537 /*
1538 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1539 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1540 * This prevents a race whereby a CPU incorrectly decides that
1541 * the state should be something other than DTRACE_DSTATE_CLEAN
1542 * after dtrace_dynvar_clean() has completed.
1543 */
1544 dtrace_sync();
1545
1546 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1547 }
1548
1549 /*
1550 * Depending on the value of the op parameter, this function looks-up,
1551 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1552 * allocation is requested, this function will return a pointer to a
1553 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1554 * variable can be allocated. If NULL is returned, the appropriate counter
1555 * will be incremented.
1556 */
1557 dtrace_dynvar_t *
1558 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1559 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1560 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1561 {
1562 uint64_t hashval = DTRACE_DYNHASH_VALID;
1563 dtrace_dynhash_t *hash = dstate->dtds_hash;
1564 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1565 processorid_t me = CPU->cpu_id, cpu = me;
1566 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1567 size_t bucket, ksize;
1568 size_t chunksize = dstate->dtds_chunksize;
1569 uintptr_t kdata, lock, nstate;
1570 uint_t i;
1571
1572 ASSERT(nkeys != 0);
1573
1574 /*
1575 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1576 * algorithm. For the by-value portions, we perform the algorithm in
1577 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1578 * bit, and seems to have only a minute effect on distribution. For
1579 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1580 * over each referenced byte. It's painful to do this, but it's much
1581 * better than pathological hash distribution. The efficacy of the
1582 * hashing algorithm (and a comparison with other algorithms) may be
1583 * found by running the ::dtrace_dynstat MDB dcmd.
1584 */
1585 for (i = 0; i < nkeys; i++) {
1586 if (key[i].dttk_size == 0) {
1587 uint64_t val = key[i].dttk_value;
1588
1589 hashval += (val >> 48) & 0xffff;
1590 hashval += (hashval << 10);
1591 hashval ^= (hashval >> 6);
1592
1593 hashval += (val >> 32) & 0xffff;
1594 hashval += (hashval << 10);
1595 hashval ^= (hashval >> 6);
1596
1597 hashval += (val >> 16) & 0xffff;
1598 hashval += (hashval << 10);
1599 hashval ^= (hashval >> 6);
1600
1601 hashval += val & 0xffff;
1602 hashval += (hashval << 10);
1603 hashval ^= (hashval >> 6);
1604 } else {
1605 /*
1606 * This is incredibly painful, but it beats the hell
1607 * out of the alternative.
1608 */
1609 uint64_t j, size = key[i].dttk_size;
1610 uintptr_t base = (uintptr_t)key[i].dttk_value;
1611
1612 if (!dtrace_canload(base, size, mstate, vstate))
1613 break;
1614
1615 for (j = 0; j < size; j++) {
1616 hashval += dtrace_load8(base + j);
1617 hashval += (hashval << 10);
1618 hashval ^= (hashval >> 6);
1619 }
1620 }
1621 }
1622
1623 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1624 return (NULL);
1625
1626 hashval += (hashval << 3);
1627 hashval ^= (hashval >> 11);
1628 hashval += (hashval << 15);
1629
1630 /*
1631 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1632 * comes out to be one of our two sentinel hash values. If this
1633 * actually happens, we set the hashval to be a value known to be a
1634 * non-sentinel value.
1635 */
1636 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1637 hashval = DTRACE_DYNHASH_VALID;
1638
1639 /*
1640 * Yes, it's painful to do a divide here. If the cycle count becomes
1641 * important here, tricks can be pulled to reduce it. (However, it's
1642 * critical that hash collisions be kept to an absolute minimum;
1643 * they're much more painful than a divide.) It's better to have a
1644 * solution that generates few collisions and still keeps things
1645 * relatively simple.
1646 */
1647 bucket = hashval % dstate->dtds_hashsize;
1648
1649 if (op == DTRACE_DYNVAR_DEALLOC) {
1650 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1651
1652 for (;;) {
1653 while ((lock = *lockp) & 1)
1654 continue;
1655
1656 if (dtrace_casptr((void *)lockp,
1657 (void *)lock, (void *)(lock + 1)) == (void *)lock)
1658 break;
1659 }
1660
1661 dtrace_membar_producer();
1662 }
1663
1664 top:
1665 prev = NULL;
1666 lock = hash[bucket].dtdh_lock;
1667
1668 dtrace_membar_consumer();
1669
1670 start = hash[bucket].dtdh_chain;
1671 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1672 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1673 op != DTRACE_DYNVAR_DEALLOC));
1674
1675 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1676 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1677 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1678
1679 if (dvar->dtdv_hashval != hashval) {
1680 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1681 /*
1682 * We've reached the sink, and therefore the
1683 * end of the hash chain; we can kick out of
1684 * the loop knowing that we have seen a valid
1685 * snapshot of state.
1686 */
1687 ASSERT(dvar->dtdv_next == NULL);
1688 ASSERT(dvar == &dtrace_dynhash_sink);
1689 break;
1690 }
1691
1692 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1693 /*
1694 * We've gone off the rails: somewhere along
1695 * the line, one of the members of this hash
1696 * chain was deleted. Note that we could also
1697 * detect this by simply letting this loop run
1698 * to completion, as we would eventually hit
1699 * the end of the dirty list. However, we
1700 * want to avoid running the length of the
1701 * dirty list unnecessarily (it might be quite
1702 * long), so we catch this as early as
1703 * possible by detecting the hash marker. In
1704 * this case, we simply set dvar to NULL and
1705 * break; the conditional after the loop will
1706 * send us back to top.
1707 */
1708 dvar = NULL;
1709 break;
1710 }
1711
1712 goto next;
1713 }
1714
1715 if (dtuple->dtt_nkeys != nkeys)
1716 goto next;
1717
1718 for (i = 0; i < nkeys; i++, dkey++) {
1719 if (dkey->dttk_size != key[i].dttk_size)
1720 goto next; /* size or type mismatch */
1721
1722 if (dkey->dttk_size != 0) {
1723 if (dtrace_bcmp(
1724 (void *)(uintptr_t)key[i].dttk_value,
1725 (void *)(uintptr_t)dkey->dttk_value,
1726 dkey->dttk_size))
1727 goto next;
1728 } else {
1729 if (dkey->dttk_value != key[i].dttk_value)
1730 goto next;
1731 }
1732 }
1733
1734 if (op != DTRACE_DYNVAR_DEALLOC)
1735 return (dvar);
1736
1737 ASSERT(dvar->dtdv_next == NULL ||
1738 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1739
1740 if (prev != NULL) {
1741 ASSERT(hash[bucket].dtdh_chain != dvar);
1742 ASSERT(start != dvar);
1743 ASSERT(prev->dtdv_next == dvar);
1744 prev->dtdv_next = dvar->dtdv_next;
1745 } else {
1746 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1747 start, dvar->dtdv_next) != start) {
1748 /*
1749 * We have failed to atomically swing the
1750 * hash table head pointer, presumably because
1751 * of a conflicting allocation on another CPU.
1752 * We need to reread the hash chain and try
1753 * again.
1754 */
1755 goto top;
1756 }
1757 }
1758
1759 dtrace_membar_producer();
1760
1761 /*
1762 * Now set the hash value to indicate that it's free.
1763 */
1764 ASSERT(hash[bucket].dtdh_chain != dvar);
1765 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1766
1767 dtrace_membar_producer();
1768
1769 /*
1770 * Set the next pointer to point at the dirty list, and
1771 * atomically swing the dirty pointer to the newly freed dvar.
1772 */
1773 do {
1774 next = dcpu->dtdsc_dirty;
1775 dvar->dtdv_next = next;
1776 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1777
1778 /*
1779 * Finally, unlock this hash bucket.
1780 */
1781 ASSERT(hash[bucket].dtdh_lock == lock);
1782 ASSERT(lock & 1);
1783 hash[bucket].dtdh_lock++;
1784
1785 return (NULL);
1786 next:
1787 prev = dvar;
1788 continue;
1789 }
1790
1791 if (dvar == NULL) {
1792 /*
1793 * If dvar is NULL, it is because we went off the rails:
1794 * one of the elements that we traversed in the hash chain
1795 * was deleted while we were traversing it. In this case,
1796 * we assert that we aren't doing a dealloc (deallocs lock
1797 * the hash bucket to prevent themselves from racing with
1798 * one another), and retry the hash chain traversal.
1799 */
1800 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1801 goto top;
1802 }
1803
1804 if (op != DTRACE_DYNVAR_ALLOC) {
1805 /*
1806 * If we are not to allocate a new variable, we want to
1807 * return NULL now. Before we return, check that the value
1808 * of the lock word hasn't changed. If it has, we may have
1809 * seen an inconsistent snapshot.
1810 */
1811 if (op == DTRACE_DYNVAR_NOALLOC) {
1812 if (hash[bucket].dtdh_lock != lock)
1813 goto top;
1814 } else {
1815 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1816 ASSERT(hash[bucket].dtdh_lock == lock);
1817 ASSERT(lock & 1);
1818 hash[bucket].dtdh_lock++;
1819 }
1820
1821 return (NULL);
1822 }
1823
1824 /*
1825 * We need to allocate a new dynamic variable. The size we need is the
1826 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1827 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1828 * the size of any referred-to data (dsize). We then round the final
1829 * size up to the chunksize for allocation.
1830 */
1831 for (ksize = 0, i = 0; i < nkeys; i++)
1832 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1833
1834 /*
1835 * This should be pretty much impossible, but could happen if, say,
1836 * strange DIF specified the tuple. Ideally, this should be an
1837 * assertion and not an error condition -- but that requires that the
1838 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1839 * bullet-proof. (That is, it must not be able to be fooled by
1840 * malicious DIF.) Given the lack of backwards branches in DIF,
1841 * solving this would presumably not amount to solving the Halting
1842 * Problem -- but it still seems awfully hard.
1843 */
1844 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1845 ksize + dsize > chunksize) {
1846 dcpu->dtdsc_drops++;
1847 return (NULL);
1848 }
1849
1850 nstate = DTRACE_DSTATE_EMPTY;
1851
1852 do {
1853 retry:
1854 free = dcpu->dtdsc_free;
1855
1856 if (free == NULL) {
1857 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1858 void *rval;
1859
1860 if (clean == NULL) {
1861 /*
1862 * We're out of dynamic variable space on
1863 * this CPU. Unless we have tried all CPUs,
1864 * we'll try to allocate from a different
1865 * CPU.
1866 */
1867 switch (dstate->dtds_state) {
1868 case DTRACE_DSTATE_CLEAN: {
1869 void *sp = &dstate->dtds_state;
1870
1871 if (++cpu >= NCPU)
1872 cpu = 0;
1873
1874 if (dcpu->dtdsc_dirty != NULL &&
1875 nstate == DTRACE_DSTATE_EMPTY)
1876 nstate = DTRACE_DSTATE_DIRTY;
1877
1878 if (dcpu->dtdsc_rinsing != NULL)
1879 nstate = DTRACE_DSTATE_RINSING;
1880
1881 dcpu = &dstate->dtds_percpu[cpu];
1882
1883 if (cpu != me)
1884 goto retry;
1885
1886 (void) dtrace_cas32(sp,
1887 DTRACE_DSTATE_CLEAN, nstate);
1888
1889 /*
1890 * To increment the correct bean
1891 * counter, take another lap.
1892 */
1893 goto retry;
1894 }
1895
1896 case DTRACE_DSTATE_DIRTY:
1897 dcpu->dtdsc_dirty_drops++;
1898 break;
1899
1900 case DTRACE_DSTATE_RINSING:
1901 dcpu->dtdsc_rinsing_drops++;
1902 break;
1903
1904 case DTRACE_DSTATE_EMPTY:
1905 dcpu->dtdsc_drops++;
1906 break;
1907 }
1908
1909 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1910 return (NULL);
1911 }
1912
1913 /*
1914 * The clean list appears to be non-empty. We want to
1915 * move the clean list to the free list; we start by
1916 * moving the clean pointer aside.
1917 */
1918 if (dtrace_casptr(&dcpu->dtdsc_clean,
1919 clean, NULL) != clean) {
1920 /*
1921 * We are in one of two situations:
1922 *
1923 * (a) The clean list was switched to the
1924 * free list by another CPU.
1925 *
1926 * (b) The clean list was added to by the
1927 * cleansing cyclic.
1928 *
1929 * In either of these situations, we can
1930 * just reattempt the free list allocation.
1931 */
1932 goto retry;
1933 }
1934
1935 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1936
1937 /*
1938 * Now we'll move the clean list to our free list.
1939 * It's impossible for this to fail: the only way
1940 * the free list can be updated is through this
1941 * code path, and only one CPU can own the clean list.
1942 * Thus, it would only be possible for this to fail if
1943 * this code were racing with dtrace_dynvar_clean().
1944 * (That is, if dtrace_dynvar_clean() updated the clean
1945 * list, and we ended up racing to update the free
1946 * list.) This race is prevented by the dtrace_sync()
1947 * in dtrace_dynvar_clean() -- which flushes the
1948 * owners of the clean lists out before resetting
1949 * the clean lists.
1950 */
1951 dcpu = &dstate->dtds_percpu[me];
1952 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1953 ASSERT(rval == NULL);
1954 goto retry;
1955 }
1956
1957 dvar = free;
1958 new_free = dvar->dtdv_next;
1959 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1960
1961 /*
1962 * We have now allocated a new chunk. We copy the tuple keys into the
1963 * tuple array and copy any referenced key data into the data space
1964 * following the tuple array. As we do this, we relocate dttk_value
1965 * in the final tuple to point to the key data address in the chunk.
1966 */
1967 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1968 dvar->dtdv_data = (void *)(kdata + ksize);
1969 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1970
1971 for (i = 0; i < nkeys; i++) {
1972 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1973 size_t kesize = key[i].dttk_size;
1974
1975 if (kesize != 0) {
1976 dtrace_bcopy(
1977 (const void *)(uintptr_t)key[i].dttk_value,
1978 (void *)kdata, kesize);
1979 dkey->dttk_value = kdata;
1980 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1981 } else {
1982 dkey->dttk_value = key[i].dttk_value;
1983 }
1984
1985 dkey->dttk_size = kesize;
1986 }
1987
1988 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1989 dvar->dtdv_hashval = hashval;
1990 dvar->dtdv_next = start;
1991
1992 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1993 return (dvar);
1994
1995 /*
1996 * The cas has failed. Either another CPU is adding an element to
1997 * this hash chain, or another CPU is deleting an element from this
1998 * hash chain. The simplest way to deal with both of these cases
1999 * (though not necessarily the most efficient) is to free our
2000 * allocated block and tail-call ourselves. Note that the free is
2001 * to the dirty list and _not_ to the free list. This is to prevent
2002 * races with allocators, above.
2003 */
2004 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2005
2006 dtrace_membar_producer();
2007
2008 do {
2009 free = dcpu->dtdsc_dirty;
2010 dvar->dtdv_next = free;
2011 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2012
2013 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
2014 }
2015
2016 /*ARGSUSED*/
2017 static void
2018 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2019 {
2020 if ((int64_t)nval < (int64_t)*oval)
2021 *oval = nval;
2022 }
2023
2024 /*ARGSUSED*/
2025 static void
2026 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2027 {
2028 if ((int64_t)nval > (int64_t)*oval)
2029 *oval = nval;
2030 }
2031
2032 static void
2033 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2034 {
2035 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2036 int64_t val = (int64_t)nval;
2037
2038 if (val < 0) {
2039 for (i = 0; i < zero; i++) {
2040 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2041 quanta[i] += incr;
2042 return;
2043 }
2044 }
2045 } else {
2046 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2047 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2048 quanta[i - 1] += incr;
2049 return;
2050 }
2051 }
2052
2053 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2054 return;
2055 }
2056
2057 ASSERT(0);
2058 }
2059
2060 static void
2061 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2062 {
2063 uint64_t arg = *lquanta++;
2064 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2065 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2066 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2067 int32_t val = (int32_t)nval, level;
2068
2069 ASSERT(step != 0);
2070 ASSERT(levels != 0);
2071
2072 if (val < base) {
2073 /*
2074 * This is an underflow.
2075 */
2076 lquanta[0] += incr;
2077 return;
2078 }
2079
2080 level = (val - base) / step;
2081
2082 if (level < levels) {
2083 lquanta[level + 1] += incr;
2084 return;
2085 }
2086
2087 /*
2088 * This is an overflow.
2089 */
2090 lquanta[levels + 1] += incr;
2091 }
2092
2093 static int
2094 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2095 uint16_t high, uint16_t nsteps, int64_t value)
2096 {
2097 int64_t this = 1, last, next;
2098 int base = 1, order;
2099
2100 ASSERT(factor <= nsteps);
2101 ASSERT(nsteps % factor == 0);
2102
2103 for (order = 0; order < low; order++)
2104 this *= factor;
2105
2106 /*
2107 * If our value is less than our factor taken to the power of the
2108 * low order of magnitude, it goes into the zeroth bucket.
2109 */
2110 if (value < (last = this))
2111 return (0);
2112
2113 for (this *= factor; order <= high; order++) {
2114 int nbuckets = this > nsteps ? nsteps : this;
2115
2116 if ((next = this * factor) < this) {
2117 /*
2118 * We should not generally get log/linear quantizations
2119 * with a high magnitude that allows 64-bits to
2120 * overflow, but we nonetheless protect against this
2121 * by explicitly checking for overflow, and clamping
2122 * our value accordingly.
2123 */
2124 value = this - 1;
2125 }
2126
2127 if (value < this) {
2128 /*
2129 * If our value lies within this order of magnitude,
2130 * determine its position by taking the offset within
2131 * the order of magnitude, dividing by the bucket
2132 * width, and adding to our (accumulated) base.
2133 */
2134 return (base + (value - last) / (this / nbuckets));
2135 }
2136
2137 base += nbuckets - (nbuckets / factor);
2138 last = this;
2139 this = next;
2140 }
2141
2142 /*
2143 * Our value is greater than or equal to our factor taken to the
2144 * power of one plus the high magnitude -- return the top bucket.
2145 */
2146 return (base);
2147 }
2148
2149 static void
2150 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2151 {
2152 uint64_t arg = *llquanta++;
2153 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2154 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2155 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2156 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2157
2158 llquanta[dtrace_aggregate_llquantize_bucket(factor,
2159 low, high, nsteps, nval)] += incr;
2160 }
2161
2162 /*ARGSUSED*/
2163 static void
2164 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2165 {
2166 data[0]++;
2167 data[1] += nval;
2168 }
2169
2170 /*ARGSUSED*/
2171 static void
2172 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2173 {
2174 int64_t snval = (int64_t)nval;
2175 uint64_t tmp[2];
2176
2177 data[0]++;
2178 data[1] += nval;
2179
2180 /*
2181 * What we want to say here is:
2182 *
2183 * data[2] += nval * nval;
2184 *
2185 * But given that nval is 64-bit, we could easily overflow, so
2186 * we do this as 128-bit arithmetic.
2187 */
2188 if (snval < 0)
2189 snval = -snval;
2190
2191 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2192 dtrace_add_128(data + 2, tmp, data + 2);
2193 }
2194
2195 /*ARGSUSED*/
2196 static void
2197 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2198 {
2199 *oval = *oval + 1;
2200 }
2201
2202 /*ARGSUSED*/
2203 static void
2204 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2205 {
2206 *oval += nval;
2207 }
2208
2209 /*
2210 * Aggregate given the tuple in the principal data buffer, and the aggregating
2211 * action denoted by the specified dtrace_aggregation_t. The aggregation
2212 * buffer is specified as the buf parameter. This routine does not return
2213 * failure; if there is no space in the aggregation buffer, the data will be
2214 * dropped, and a corresponding counter incremented.
2215 */
2216 static void
2217 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2218 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2219 {
2220 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2221 uint32_t i, ndx, size, fsize;
2222 uint32_t align = sizeof (uint64_t) - 1;
2223 dtrace_aggbuffer_t *agb;
2224 dtrace_aggkey_t *key;
2225 uint32_t hashval = 0, limit, isstr;
2226 caddr_t tomax, data, kdata;
2227 dtrace_actkind_t action;
2228 dtrace_action_t *act;
2229 uintptr_t offs;
2230
2231 if (buf == NULL)
2232 return;
2233
2234 if (!agg->dtag_hasarg) {
2235 /*
2236 * Currently, only quantize() and lquantize() take additional
2237 * arguments, and they have the same semantics: an increment
2238 * value that defaults to 1 when not present. If additional
2239 * aggregating actions take arguments, the setting of the
2240 * default argument value will presumably have to become more
2241 * sophisticated...
2242 */
2243 arg = 1;
2244 }
2245
2246 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2247 size = rec->dtrd_offset - agg->dtag_base;
2248 fsize = size + rec->dtrd_size;
2249
2250 ASSERT(dbuf->dtb_tomax != NULL);
2251 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2252
2253 if ((tomax = buf->dtb_tomax) == NULL) {
2254 dtrace_buffer_drop(buf);
2255 return;
2256 }
2257
2258 /*
2259 * The metastructure is always at the bottom of the buffer.
2260 */
2261 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2262 sizeof (dtrace_aggbuffer_t));
2263
2264 if (buf->dtb_offset == 0) {
2265 /*
2266 * We just kludge up approximately 1/8th of the size to be
2267 * buckets. If this guess ends up being routinely
2268 * off-the-mark, we may need to dynamically readjust this
2269 * based on past performance.
2270 */
2271 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2272
2273 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2274 (uintptr_t)tomax || hashsize == 0) {
2275 /*
2276 * We've been given a ludicrously small buffer;
2277 * increment our drop count and leave.
2278 */
2279 dtrace_buffer_drop(buf);
2280 return;
2281 }
2282
2283 /*
2284 * And now, a pathetic attempt to try to get a an odd (or
2285 * perchance, a prime) hash size for better hash distribution.
2286 */
2287 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2288 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2289
2290 agb->dtagb_hashsize = hashsize;
2291 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2292 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2293 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2294
2295 for (i = 0; i < agb->dtagb_hashsize; i++)
2296 agb->dtagb_hash[i] = NULL;
2297 }
2298
2299 ASSERT(agg->dtag_first != NULL);
2300 ASSERT(agg->dtag_first->dta_intuple);
2301
2302 /*
2303 * Calculate the hash value based on the key. Note that we _don't_
2304 * include the aggid in the hashing (but we will store it as part of
2305 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2306 * algorithm: a simple, quick algorithm that has no known funnels, and
2307 * gets good distribution in practice. The efficacy of the hashing
2308 * algorithm (and a comparison with other algorithms) may be found by
2309 * running the ::dtrace_aggstat MDB dcmd.
2310 */
2311 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2312 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2313 limit = i + act->dta_rec.dtrd_size;
2314 ASSERT(limit <= size);
2315 isstr = DTRACEACT_ISSTRING(act);
2316
2317 for (; i < limit; i++) {
2318 hashval += data[i];
2319 hashval += (hashval << 10);
2320 hashval ^= (hashval >> 6);
2321
2322 if (isstr && data[i] == '\0')
2323 break;
2324 }
2325 }
2326
2327 hashval += (hashval << 3);
2328 hashval ^= (hashval >> 11);
2329 hashval += (hashval << 15);
2330
2331 /*
2332 * Yes, the divide here is expensive -- but it's generally the least
2333 * of the performance issues given the amount of data that we iterate
2334 * over to compute hash values, compare data, etc.
2335 */
2336 ndx = hashval % agb->dtagb_hashsize;
2337
2338 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2339 ASSERT((caddr_t)key >= tomax);
2340 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2341
2342 if (hashval != key->dtak_hashval || key->dtak_size != size)
2343 continue;
2344
2345 kdata = key->dtak_data;
2346 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2347
2348 for (act = agg->dtag_first; act->dta_intuple;
2349 act = act->dta_next) {
2350 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2351 limit = i + act->dta_rec.dtrd_size;
2352 ASSERT(limit <= size);
2353 isstr = DTRACEACT_ISSTRING(act);
2354
2355 for (; i < limit; i++) {
2356 if (kdata[i] != data[i])
2357 goto next;
2358
2359 if (isstr && data[i] == '\0')
2360 break;
2361 }
2362 }
2363
2364 if (action != key->dtak_action) {
2365 /*
2366 * We are aggregating on the same value in the same
2367 * aggregation with two different aggregating actions.
2368 * (This should have been picked up in the compiler,
2369 * so we may be dealing with errant or devious DIF.)
2370 * This is an error condition; we indicate as much,
2371 * and return.
2372 */
2373 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2374 return;
2375 }
2376
2377 /*
2378 * This is a hit: we need to apply the aggregator to
2379 * the value at this key.
2380 */
2381 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2382 return;
2383 next:
2384 continue;
2385 }
2386
2387 /*
2388 * We didn't find it. We need to allocate some zero-filled space,
2389 * link it into the hash table appropriately, and apply the aggregator
2390 * to the (zero-filled) value.
2391 */
2392 offs = buf->dtb_offset;
2393 while (offs & (align - 1))
2394 offs += sizeof (uint32_t);
2395
2396 /*
2397 * If we don't have enough room to both allocate a new key _and_
2398 * its associated data, increment the drop count and return.
2399 */
2400 if ((uintptr_t)tomax + offs + fsize >
2401 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2402 dtrace_buffer_drop(buf);
2403 return;
2404 }
2405
2406 /*CONSTCOND*/
2407 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2408 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2409 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2410
2411 key->dtak_data = kdata = tomax + offs;
2412 buf->dtb_offset = offs + fsize;
2413
2414 /*
2415 * Now copy the data across.
2416 */
2417 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2418
2419 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2420 kdata[i] = data[i];
2421
2422 /*
2423 * Because strings are not zeroed out by default, we need to iterate
2424 * looking for actions that store strings, and we need to explicitly
2425 * pad these strings out with zeroes.
2426 */
2427 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2428 int nul;
2429
2430 if (!DTRACEACT_ISSTRING(act))
2431 continue;
2432
2433 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2434 limit = i + act->dta_rec.dtrd_size;
2435 ASSERT(limit <= size);
2436
2437 for (nul = 0; i < limit; i++) {
2438 if (nul) {
2439 kdata[i] = '\0';
2440 continue;
2441 }
2442
2443 if (data[i] != '\0')
2444 continue;
2445
2446 nul = 1;
2447 }
2448 }
2449
2450 for (i = size; i < fsize; i++)
2451 kdata[i] = 0;
2452
2453 key->dtak_hashval = hashval;
2454 key->dtak_size = size;
2455 key->dtak_action = action;
2456 key->dtak_next = agb->dtagb_hash[ndx];
2457 agb->dtagb_hash[ndx] = key;
2458
2459 /*
2460 * Finally, apply the aggregator.
2461 */
2462 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2463 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2464 }
2465
2466 /*
2467 * Given consumer state, this routine finds a speculation in the INACTIVE
2468 * state and transitions it into the ACTIVE state. If there is no speculation
2469 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2470 * incremented -- it is up to the caller to take appropriate action.
2471 */
2472 static int
2473 dtrace_speculation(dtrace_state_t *state)
2474 {
2475 int i = 0;
2476 dtrace_speculation_state_t current;
2477 uint32_t *stat = &state->dts_speculations_unavail, count;
2478
2479 while (i < state->dts_nspeculations) {
2480 dtrace_speculation_t *spec = &state->dts_speculations[i];
2481
2482 current = spec->dtsp_state;
2483
2484 if (current != DTRACESPEC_INACTIVE) {
2485 if (current == DTRACESPEC_COMMITTINGMANY ||
2486 current == DTRACESPEC_COMMITTING ||
2487 current == DTRACESPEC_DISCARDING)
2488 stat = &state->dts_speculations_busy;
2489 i++;
2490 continue;
2491 }
2492
2493 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2494 current, DTRACESPEC_ACTIVE) == current)
2495 return (i + 1);
2496 }
2497
2498 /*
2499 * We couldn't find a speculation. If we found as much as a single
2500 * busy speculation buffer, we'll attribute this failure as "busy"
2501 * instead of "unavail".
2502 */
2503 do {
2504 count = *stat;
2505 } while (dtrace_cas32(stat, count, count + 1) != count);
2506
2507 return (0);
2508 }
2509
2510 /*
2511 * This routine commits an active speculation. If the specified speculation
2512 * is not in a valid state to perform a commit(), this routine will silently do
2513 * nothing. The state of the specified speculation is transitioned according
2514 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2515 */
2516 static void
2517 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2518 dtrace_specid_t which)
2519 {
2520 dtrace_speculation_t *spec;
2521 dtrace_buffer_t *src, *dest;
2522 uintptr_t daddr, saddr, dlimit, slimit;
2523 dtrace_speculation_state_t current, new;
2524 intptr_t offs;
2525 uint64_t timestamp;
2526
2527 if (which == 0)
2528 return;
2529
2530 if (which > state->dts_nspeculations) {
2531 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2532 return;
2533 }
2534
2535 spec = &state->dts_speculations[which - 1];
2536 src = &spec->dtsp_buffer[cpu];
2537 dest = &state->dts_buffer[cpu];
2538
2539 do {
2540 current = spec->dtsp_state;
2541
2542 if (current == DTRACESPEC_COMMITTINGMANY)
2543 break;
2544
2545 switch (current) {
2546 case DTRACESPEC_INACTIVE:
2547 case DTRACESPEC_DISCARDING:
2548 return;
2549
2550 case DTRACESPEC_COMMITTING:
2551 /*
2552 * This is only possible if we are (a) commit()'ing
2553 * without having done a prior speculate() on this CPU
2554 * and (b) racing with another commit() on a different
2555 * CPU. There's nothing to do -- we just assert that
2556 * our offset is 0.
2557 */
2558 ASSERT(src->dtb_offset == 0);
2559 return;
2560
2561 case DTRACESPEC_ACTIVE:
2562 new = DTRACESPEC_COMMITTING;
2563 break;
2564
2565 case DTRACESPEC_ACTIVEONE:
2566 /*
2567 * This speculation is active on one CPU. If our
2568 * buffer offset is non-zero, we know that the one CPU
2569 * must be us. Otherwise, we are committing on a
2570 * different CPU from the speculate(), and we must
2571 * rely on being asynchronously cleaned.
2572 */
2573 if (src->dtb_offset != 0) {
2574 new = DTRACESPEC_COMMITTING;
2575 break;
2576 }
2577 /*FALLTHROUGH*/
2578
2579 case DTRACESPEC_ACTIVEMANY:
2580 new = DTRACESPEC_COMMITTINGMANY;
2581 break;
2582
2583 default:
2584 ASSERT(0);
2585 }
2586 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2587 current, new) != current);
2588
2589 /*
2590 * We have set the state to indicate that we are committing this
2591 * speculation. Now reserve the necessary space in the destination
2592 * buffer.
2593 */
2594 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2595 sizeof (uint64_t), state, NULL)) < 0) {
2596 dtrace_buffer_drop(dest);
2597 goto out;
2598 }
2599
2600 /*
2601 * We have sufficient space to copy the speculative buffer into the
2602 * primary buffer. First, modify the speculative buffer, filling
2603 * in the timestamp of all entries with the current time. The data
2604 * must have the commit() time rather than the time it was traced,
2605 * so that all entries in the primary buffer are in timestamp order.
2606 */
2607 timestamp = dtrace_gethrtime();
2608 saddr = (uintptr_t)src->dtb_tomax;
2609 slimit = saddr + src->dtb_offset;
2610 while (saddr < slimit) {
2611 size_t size;
2612 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2613
2614 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2615 saddr += sizeof (dtrace_epid_t);
2616 continue;
2617 }
2618 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2619 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2620
2621 ASSERT3U(saddr + size, <=, slimit);
2622 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2623 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2624
2625 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2626
2627 saddr += size;
2628 }
2629
2630 /*
2631 * Copy the buffer across. (Note that this is a
2632 * highly subobtimal bcopy(); in the unlikely event that this becomes
2633 * a serious performance issue, a high-performance DTrace-specific
2634 * bcopy() should obviously be invented.)
2635 */
2636 daddr = (uintptr_t)dest->dtb_tomax + offs;
2637 dlimit = daddr + src->dtb_offset;
2638 saddr = (uintptr_t)src->dtb_tomax;
2639
2640 /*
2641 * First, the aligned portion.
2642 */
2643 while (dlimit - daddr >= sizeof (uint64_t)) {
2644 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2645
2646 daddr += sizeof (uint64_t);
2647 saddr += sizeof (uint64_t);
2648 }
2649
2650 /*
2651 * Now any left-over bit...
2652 */
2653 while (dlimit - daddr)
2654 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2655
2656 /*
2657 * Finally, commit the reserved space in the destination buffer.
2658 */
2659 dest->dtb_offset = offs + src->dtb_offset;
2660
2661 out:
2662 /*
2663 * If we're lucky enough to be the only active CPU on this speculation
2664 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2665 */
2666 if (current == DTRACESPEC_ACTIVE ||
2667 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2668 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2669 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2670
2671 ASSERT(rval == DTRACESPEC_COMMITTING);
2672 }
2673
2674 src->dtb_offset = 0;
2675 src->dtb_xamot_drops += src->dtb_drops;
2676 src->dtb_drops = 0;
2677 }
2678
2679 /*
2680 * This routine discards an active speculation. If the specified speculation
2681 * is not in a valid state to perform a discard(), this routine will silently
2682 * do nothing. The state of the specified speculation is transitioned
2683 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2684 */
2685 static void
2686 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2687 dtrace_specid_t which)
2688 {
2689 dtrace_speculation_t *spec;
2690 dtrace_speculation_state_t current, new;
2691 dtrace_buffer_t *buf;
2692
2693 if (which == 0)
2694 return;
2695
2696 if (which > state->dts_nspeculations) {
2697 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2698 return;
2699 }
2700
2701 spec = &state->dts_speculations[which - 1];
2702 buf = &spec->dtsp_buffer[cpu];
2703
2704 do {
2705 current = spec->dtsp_state;
2706
2707 switch (current) {
2708 case DTRACESPEC_INACTIVE:
2709 case DTRACESPEC_COMMITTINGMANY:
2710 case DTRACESPEC_COMMITTING:
2711 case DTRACESPEC_DISCARDING:
2712 return;
2713
2714 case DTRACESPEC_ACTIVE:
2715 case DTRACESPEC_ACTIVEMANY:
2716 new = DTRACESPEC_DISCARDING;
2717 break;
2718
2719 case DTRACESPEC_ACTIVEONE:
2720 if (buf->dtb_offset != 0) {
2721 new = DTRACESPEC_INACTIVE;
2722 } else {
2723 new = DTRACESPEC_DISCARDING;
2724 }
2725 break;
2726
2727 default:
2728 ASSERT(0);
2729 }
2730 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2731 current, new) != current);
2732
2733 buf->dtb_offset = 0;
2734 buf->dtb_drops = 0;
2735 }
2736
2737 /*
2738 * Note: not called from probe context. This function is called
2739 * asynchronously from cross call context to clean any speculations that are
2740 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2741 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2742 * speculation.
2743 */
2744 static void
2745 dtrace_speculation_clean_here(dtrace_state_t *state)
2746 {
2747 dtrace_icookie_t cookie;
2748 processorid_t cpu = CPU->cpu_id;
2749 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2750 dtrace_specid_t i;
2751
2752 cookie = dtrace_interrupt_disable();
2753
2754 if (dest->dtb_tomax == NULL) {
2755 dtrace_interrupt_enable(cookie);
2756 return;
2757 }
2758
2759 for (i = 0; i < state->dts_nspeculations; i++) {
2760 dtrace_speculation_t *spec = &state->dts_speculations[i];
2761 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2762
2763 if (src->dtb_tomax == NULL)
2764 continue;
2765
2766 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2767 src->dtb_offset = 0;
2768 continue;
2769 }
2770
2771 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2772 continue;
2773
2774 if (src->dtb_offset == 0)
2775 continue;
2776
2777 dtrace_speculation_commit(state, cpu, i + 1);
2778 }
2779
2780 dtrace_interrupt_enable(cookie);
2781 }
2782
2783 /*
2784 * Note: not called from probe context. This function is called
2785 * asynchronously (and at a regular interval) to clean any speculations that
2786 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2787 * is work to be done, it cross calls all CPUs to perform that work;
2788 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2789 * INACTIVE state until they have been cleaned by all CPUs.
2790 */
2791 static void
2792 dtrace_speculation_clean(dtrace_state_t *state)
2793 {
2794 int work = 0, rv;
2795 dtrace_specid_t i;
2796
2797 for (i = 0; i < state->dts_nspeculations; i++) {
2798 dtrace_speculation_t *spec = &state->dts_speculations[i];
2799
2800 ASSERT(!spec->dtsp_cleaning);
2801
2802 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2803 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2804 continue;
2805
2806 work++;
2807 spec->dtsp_cleaning = 1;
2808 }
2809
2810 if (!work)
2811 return;
2812
2813 dtrace_xcall(DTRACE_CPUALL,
2814 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2815
2816 /*
2817 * We now know that all CPUs have committed or discarded their
2818 * speculation buffers, as appropriate. We can now set the state
2819 * to inactive.
2820 */
2821 for (i = 0; i < state->dts_nspeculations; i++) {
2822 dtrace_speculation_t *spec = &state->dts_speculations[i];
2823 dtrace_speculation_state_t current, new;
2824
2825 if (!spec->dtsp_cleaning)
2826 continue;
2827
2828 current = spec->dtsp_state;
2829 ASSERT(current == DTRACESPEC_DISCARDING ||
2830 current == DTRACESPEC_COMMITTINGMANY);
2831
2832 new = DTRACESPEC_INACTIVE;
2833
2834 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2835 ASSERT(rv == current);
2836 spec->dtsp_cleaning = 0;
2837 }
2838 }
2839
2840 /*
2841 * Called as part of a speculate() to get the speculative buffer associated
2842 * with a given speculation. Returns NULL if the specified speculation is not
2843 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2844 * the active CPU is not the specified CPU -- the speculation will be
2845 * atomically transitioned into the ACTIVEMANY state.
2846 */
2847 static dtrace_buffer_t *
2848 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2849 dtrace_specid_t which)
2850 {
2851 dtrace_speculation_t *spec;
2852 dtrace_speculation_state_t current, new;
2853 dtrace_buffer_t *buf;
2854
2855 if (which == 0)
2856 return (NULL);
2857
2858 if (which > state->dts_nspeculations) {
2859 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2860 return (NULL);
2861 }
2862
2863 spec = &state->dts_speculations[which - 1];
2864 buf = &spec->dtsp_buffer[cpuid];
2865
2866 do {
2867 current = spec->dtsp_state;
2868
2869 switch (current) {
2870 case DTRACESPEC_INACTIVE:
2871 case DTRACESPEC_COMMITTINGMANY:
2872 case DTRACESPEC_DISCARDING:
2873 return (NULL);
2874
2875 case DTRACESPEC_COMMITTING:
2876 ASSERT(buf->dtb_offset == 0);
2877 return (NULL);
2878
2879 case DTRACESPEC_ACTIVEONE:
2880 /*
2881 * This speculation is currently active on one CPU.
2882 * Check the offset in the buffer; if it's non-zero,
2883 * that CPU must be us (and we leave the state alone).
2884 * If it's zero, assume that we're starting on a new
2885 * CPU -- and change the state to indicate that the
2886 * speculation is active on more than one CPU.
2887 */
2888 if (buf->dtb_offset != 0)
2889 return (buf);
2890
2891 new = DTRACESPEC_ACTIVEMANY;
2892 break;
2893
2894 case DTRACESPEC_ACTIVEMANY:
2895 return (buf);
2896
2897 case DTRACESPEC_ACTIVE:
2898 new = DTRACESPEC_ACTIVEONE;
2899 break;
2900
2901 default:
2902 ASSERT(0);
2903 }
2904 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2905 current, new) != current);
2906
2907 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2908 return (buf);
2909 }
2910
2911 /*
2912 * Return a string. In the event that the user lacks the privilege to access
2913 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2914 * don't fail access checking.
2915 *
2916 * dtrace_dif_variable() uses this routine as a helper for various
2917 * builtin values such as 'execname' and 'probefunc.'
2918 */
2919 uintptr_t
2920 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2921 dtrace_mstate_t *mstate)
2922 {
2923 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2924 uintptr_t ret;
2925 size_t strsz;
2926
2927 /*
2928 * The easy case: this probe is allowed to read all of memory, so
2929 * we can just return this as a vanilla pointer.
2930 */
2931 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2932 return (addr);
2933
2934 /*
2935 * This is the tougher case: we copy the string in question from
2936 * kernel memory into scratch memory and return it that way: this
2937 * ensures that we won't trip up when access checking tests the
2938 * BYREF return value.
2939 */
2940 strsz = dtrace_strlen((char *)addr, size) + 1;
2941
2942 if (mstate->dtms_scratch_ptr + strsz >
2943 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2944 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2945 return (NULL);
2946 }
2947
2948 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2949 strsz);
2950 ret = mstate->dtms_scratch_ptr;
2951 mstate->dtms_scratch_ptr += strsz;
2952 return (ret);
2953 }
2954
2955 /*
2956 * This function implements the DIF emulator's variable lookups. The emulator
2957 * passes a reserved variable identifier and optional built-in array index.
2958 */
2959 static uint64_t
2960 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2961 uint64_t ndx)
2962 {
2963 /*
2964 * If we're accessing one of the uncached arguments, we'll turn this
2965 * into a reference in the args array.
2966 */
2967 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2968 ndx = v - DIF_VAR_ARG0;
2969 v = DIF_VAR_ARGS;
2970 }
2971
2972 switch (v) {
2973 case DIF_VAR_ARGS:
2974 if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
2975 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
2976 CPU_DTRACE_KPRIV;
2977 return (0);
2978 }
2979
2980 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2981 if (ndx >= sizeof (mstate->dtms_arg) /
2982 sizeof (mstate->dtms_arg[0])) {
2983 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2984 dtrace_provider_t *pv;
2985 uint64_t val;
2986
2987 pv = mstate->dtms_probe->dtpr_provider;
2988 if (pv->dtpv_pops.dtps_getargval != NULL)
2989 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2990 mstate->dtms_probe->dtpr_id,
2991 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2992 else
2993 val = dtrace_getarg(ndx, aframes);
2994
2995 /*
2996 * This is regrettably required to keep the compiler
2997 * from tail-optimizing the call to dtrace_getarg().
2998 * The condition always evaluates to true, but the
2999 * compiler has no way of figuring that out a priori.
3000 * (None of this would be necessary if the compiler
3001 * could be relied upon to _always_ tail-optimize
3002 * the call to dtrace_getarg() -- but it can't.)
3003 */
3004 if (mstate->dtms_probe != NULL)
3005 return (val);
3006
3007 ASSERT(0);
3008 }
3009
3010 return (mstate->dtms_arg[ndx]);
3011
3012 case DIF_VAR_UREGS: {
3013 klwp_t *lwp;
3014
3015 if (!dtrace_priv_proc(state, mstate))
3016 return (0);
3017
3018 if ((lwp = curthread->t_lwp) == NULL) {
3019 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3020 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
3021 return (0);
3022 }
3023
3024 return (dtrace_getreg(lwp->lwp_regs, ndx));
3025 }
3026
3027 case DIF_VAR_VMREGS: {
3028 uint64_t rval;
3029
3030 if (!dtrace_priv_kernel(state))
3031 return (0);
3032
3033 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3034
3035 rval = dtrace_getvmreg(ndx,
3036 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
3037
3038 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3039
3040 return (rval);
3041 }
3042
3043 case DIF_VAR_CURTHREAD:
3044 if (!dtrace_priv_proc(state, mstate))
3045 return (0);
3046 return ((uint64_t)(uintptr_t)curthread);
3047
3048 case DIF_VAR_TIMESTAMP:
3049 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3050 mstate->dtms_timestamp = dtrace_gethrtime();
3051 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3052 }
3053 return (mstate->dtms_timestamp);
3054
3055 case DIF_VAR_VTIMESTAMP:
3056 ASSERT(dtrace_vtime_references != 0);
3057 return (curthread->t_dtrace_vtime);
3058
3059 case DIF_VAR_WALLTIMESTAMP:
3060 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3061 mstate->dtms_walltimestamp = dtrace_gethrestime();
3062 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3063 }
3064 return (mstate->dtms_walltimestamp);
3065
3066 case DIF_VAR_IPL:
3067 if (!dtrace_priv_kernel(state))
3068 return (0);
3069 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3070 mstate->dtms_ipl = dtrace_getipl();
3071 mstate->dtms_present |= DTRACE_MSTATE_IPL;
3072 }
3073 return (mstate->dtms_ipl);
3074
3075 case DIF_VAR_EPID:
3076 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3077 return (mstate->dtms_epid);
3078
3079 case DIF_VAR_ID:
3080 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3081 return (mstate->dtms_probe->dtpr_id);
3082
3083 case DIF_VAR_STACKDEPTH:
3084 if (!dtrace_priv_kernel(state))
3085 return (0);
3086 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3087 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3088
3089 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3090 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3091 }
3092 return (mstate->dtms_stackdepth);
3093
3094 case DIF_VAR_USTACKDEPTH:
3095 if (!dtrace_priv_proc(state, mstate))
3096 return (0);
3097 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3098 /*
3099 * See comment in DIF_VAR_PID.
3100 */
3101 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3102 CPU_ON_INTR(CPU)) {
3103 mstate->dtms_ustackdepth = 0;
3104 } else {
3105 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3106 mstate->dtms_ustackdepth =
3107 dtrace_getustackdepth();
3108 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3109 }
3110 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3111 }
3112 return (mstate->dtms_ustackdepth);
3113
3114 case DIF_VAR_CALLER:
3115 if (!dtrace_priv_kernel(state))
3116 return (0);
3117 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3118 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3119
3120 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3121 /*
3122 * If this is an unanchored probe, we are
3123 * required to go through the slow path:
3124 * dtrace_caller() only guarantees correct
3125 * results for anchored probes.
3126 */
3127 pc_t caller[2];
3128
3129 dtrace_getpcstack(caller, 2, aframes,
3130 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3131 mstate->dtms_caller = caller[1];
3132 } else if ((mstate->dtms_caller =
3133 dtrace_caller(aframes)) == -1) {
3134 /*
3135 * We have failed to do this the quick way;
3136 * we must resort to the slower approach of
3137 * calling dtrace_getpcstack().
3138 */
3139 pc_t caller;
3140
3141 dtrace_getpcstack(&caller, 1, aframes, NULL);
3142 mstate->dtms_caller = caller;
3143 }
3144
3145 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3146 }
3147 return (mstate->dtms_caller);
3148
3149 case DIF_VAR_UCALLER:
3150 if (!dtrace_priv_proc(state, mstate))
3151 return (0);
3152
3153 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3154 uint64_t ustack[3];
3155
3156 /*
3157 * dtrace_getupcstack() fills in the first uint64_t
3158 * with the current PID. The second uint64_t will
3159 * be the program counter at user-level. The third
3160 * uint64_t will contain the caller, which is what
3161 * we're after.
3162 */
3163 ustack[2] = NULL;
3164 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3165 dtrace_getupcstack(ustack, 3);
3166 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3167 mstate->dtms_ucaller = ustack[2];
3168 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3169 }
3170
3171 return (mstate->dtms_ucaller);
3172
3173 case DIF_VAR_PROBEPROV:
3174 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3175 return (dtrace_dif_varstr(
3176 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3177 state, mstate));
3178
3179 case DIF_VAR_PROBEMOD:
3180 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3181 return (dtrace_dif_varstr(
3182 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3183 state, mstate));
3184
3185 case DIF_VAR_PROBEFUNC:
3186 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3187 return (dtrace_dif_varstr(
3188 (uintptr_t)mstate->dtms_probe->dtpr_func,
3189 state, mstate));
3190
3191 case DIF_VAR_PROBENAME:
3192 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3193 return (dtrace_dif_varstr(
3194 (uintptr_t)mstate->dtms_probe->dtpr_name,
3195 state, mstate));
3196
3197 case DIF_VAR_PID:
3198 if (!dtrace_priv_proc(state, mstate))
3199 return (0);
3200
3201 /*
3202 * Note that we are assuming that an unanchored probe is
3203 * always due to a high-level interrupt. (And we're assuming
3204 * that there is only a single high level interrupt.)
3205 */
3206 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3207 return (pid0.pid_id);
3208
3209 /*
3210 * It is always safe to dereference one's own t_procp pointer:
3211 * it always points to a valid, allocated proc structure.
3212 * Further, it is always safe to dereference the p_pidp member
3213 * of one's own proc structure. (These are truisms becuase
3214 * threads and processes don't clean up their own state --
3215 * they leave that task to whomever reaps them.)
3216 */
3217 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3218
3219 case DIF_VAR_PPID:
3220 if (!dtrace_priv_proc(state, mstate))
3221 return (0);
3222
3223 /*
3224 * See comment in DIF_VAR_PID.
3225 */
3226 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3227 return (pid0.pid_id);
3228
3229 /*
3230 * It is always safe to dereference one's own t_procp pointer:
3231 * it always points to a valid, allocated proc structure.
3232 * (This is true because threads don't clean up their own
3233 * state -- they leave that task to whomever reaps them.)
3234 */
3235 return ((uint64_t)curthread->t_procp->p_ppid);
3236
3237 case DIF_VAR_TID:
3238 /*
3239 * See comment in DIF_VAR_PID.
3240 */
3241 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3242 return (0);
3243
3244 return ((uint64_t)curthread->t_tid);
3245
3246 case DIF_VAR_EXECNAME:
3247 if (!dtrace_priv_proc(state, mstate))
3248 return (0);
3249
3250 /*
3251 * See comment in DIF_VAR_PID.
3252 */
3253 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3254 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3255
3256 /*
3257 * It is always safe to dereference one's own t_procp pointer:
3258 * it always points to a valid, allocated proc structure.
3259 * (This is true because threads don't clean up their own
3260 * state -- they leave that task to whomever reaps them.)
3261 */
3262 return (dtrace_dif_varstr(
3263 (uintptr_t)curthread->t_procp->p_user.u_comm,
3264 state, mstate));
3265
3266 case DIF_VAR_ZONENAME:
3267 if (!dtrace_priv_proc(state, mstate))
3268 return (0);
3269
3270 /*
3271 * See comment in DIF_VAR_PID.
3272 */
3273 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3274 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3275
3276 /*
3277 * It is always safe to dereference one's own t_procp pointer:
3278 * it always points to a valid, allocated proc structure.
3279 * (This is true because threads don't clean up their own
3280 * state -- they leave that task to whomever reaps them.)
3281 */
3282 return (dtrace_dif_varstr(
3283 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3284 state, mstate));
3285
3286 case DIF_VAR_UID:
3287 if (!dtrace_priv_proc(state, mstate))
3288 return (0);
3289
3290 /*
3291 * See comment in DIF_VAR_PID.
3292 */
3293 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3294 return ((uint64_t)p0.p_cred->cr_uid);
3295
3296 /*
3297 * It is always safe to dereference one's own t_procp pointer:
3298 * it always points to a valid, allocated proc structure.
3299 * (This is true because threads don't clean up their own
3300 * state -- they leave that task to whomever reaps them.)
3301 *
3302 * Additionally, it is safe to dereference one's own process
3303 * credential, since this is never NULL after process birth.
3304 */
3305 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3306
3307 case DIF_VAR_GID:
3308 if (!dtrace_priv_proc(state, mstate))
3309 return (0);
3310
3311 /*
3312 * See comment in DIF_VAR_PID.
3313 */
3314 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3315 return ((uint64_t)p0.p_cred->cr_gid);
3316
3317 /*
3318 * It is always safe to dereference one's own t_procp pointer:
3319 * it always points to a valid, allocated proc structure.
3320 * (This is true because threads don't clean up their own
3321 * state -- they leave that task to whomever reaps them.)
3322 *
3323 * Additionally, it is safe to dereference one's own process
3324 * credential, since this is never NULL after process birth.
3325 */
3326 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3327
3328 case DIF_VAR_ERRNO: {
3329 klwp_t *lwp;
3330 if (!dtrace_priv_proc(state, mstate))
3331 return (0);
3332
3333 /*
3334 * See comment in DIF_VAR_PID.
3335 */
3336 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3337 return (0);
3338
3339 /*
3340 * It is always safe to dereference one's own t_lwp pointer in
3341 * the event that this pointer is non-NULL. (This is true
3342 * because threads and lwps don't clean up their own state --
3343 * they leave that task to whomever reaps them.)
3344 */
3345 if ((lwp = curthread->t_lwp) == NULL)
3346 return (0);
3347
3348 return ((uint64_t)lwp->lwp_errno);
3349 }
3350 default:
3351 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3352 return (0);
3353 }
3354 }
3355
3356 /*
3357 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3358 * Notice that we don't bother validating the proper number of arguments or
3359 * their types in the tuple stack. This isn't needed because all argument
3360 * interpretation is safe because of our load safety -- the worst that can
3361 * happen is that a bogus program can obtain bogus results.
3362 */
3363 static void
3364 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3365 dtrace_key_t *tupregs, int nargs,
3366 dtrace_mstate_t *mstate, dtrace_state_t *state)
3367 {
3368 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
3369 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
3370 dtrace_vstate_t *vstate = &state->dts_vstate;
3371
3372 union {
3373 mutex_impl_t mi;
3374 uint64_t mx;
3375 } m;
3376
3377 union {
3378 krwlock_t ri;
3379 uintptr_t rw;
3380 } r;
3381
3382 switch (subr) {
3383 case DIF_SUBR_RAND:
3384 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3385 break;
3386
3387 case DIF_SUBR_MUTEX_OWNED:
3388 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3389 mstate, vstate)) {
3390 regs[rd] = NULL;
3391 break;
3392 }
3393
3394 m.mx = dtrace_load64(tupregs[0].dttk_value);
3395 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3396 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3397 else
3398 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3399 break;
3400
3401 case DIF_SUBR_MUTEX_OWNER:
3402 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3403 mstate, vstate)) {
3404 regs[rd] = NULL;
3405 break;
3406 }
3407
3408 m.mx = dtrace_load64(tupregs[0].dttk_value);
3409 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3410 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3411 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3412 else
3413 regs[rd] = 0;
3414 break;
3415
3416 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3417 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3418 mstate, vstate)) {
3419 regs[rd] = NULL;
3420 break;
3421 }
3422
3423 m.mx = dtrace_load64(tupregs[0].dttk_value);
3424 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3425 break;
3426
3427 case DIF_SUBR_MUTEX_TYPE_SPIN:
3428 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3429 mstate, vstate)) {
3430 regs[rd] = NULL;
3431 break;
3432 }
3433
3434 m.mx = dtrace_load64(tupregs[0].dttk_value);
3435 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3436 break;
3437
3438 case DIF_SUBR_RW_READ_HELD: {
3439 uintptr_t tmp;
3440
3441 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3442 mstate, vstate)) {
3443 regs[rd] = NULL;
3444 break;
3445 }
3446
3447 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3448 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3449 break;
3450 }
3451
3452 case DIF_SUBR_RW_WRITE_HELD:
3453 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3454 mstate, vstate)) {
3455 regs[rd] = NULL;
3456 break;
3457 }
3458
3459 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3460 regs[rd] = _RW_WRITE_HELD(&r.ri);
3461 break;
3462
3463 case DIF_SUBR_RW_ISWRITER:
3464 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3465 mstate, vstate)) {
3466 regs[rd] = NULL;
3467 break;
3468 }
3469
3470 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3471 regs[rd] = _RW_ISWRITER(&r.ri);
3472 break;
3473
3474 case DIF_SUBR_BCOPY: {
3475 /*
3476 * We need to be sure that the destination is in the scratch
3477 * region -- no other region is allowed.
3478 */
3479 uintptr_t src = tupregs[0].dttk_value;
3480 uintptr_t dest = tupregs[1].dttk_value;
3481 size_t size = tupregs[2].dttk_value;
3482
3483 if (!dtrace_inscratch(dest, size, mstate)) {
3484 *flags |= CPU_DTRACE_BADADDR;
3485 *illval = regs[rd];
3486 break;
3487 }
3488
3489 if (!dtrace_canload(src, size, mstate, vstate)) {
3490 regs[rd] = NULL;
3491 break;
3492 }
3493
3494 dtrace_bcopy((void *)src, (void *)dest, size);
3495 break;
3496 }
3497
3498 case DIF_SUBR_ALLOCA:
3499 case DIF_SUBR_COPYIN: {
3500 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3501 uint64_t size =
3502 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3503 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3504
3505 /*
3506 * This action doesn't require any credential checks since
3507 * probes will not activate in user contexts to which the
3508 * enabling user does not have permissions.
3509 */
3510
3511 /*
3512 * Rounding up the user allocation size could have overflowed
3513 * a large, bogus allocation (like -1ULL) to 0.
3514 */
3515 if (scratch_size < size ||
3516 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3517 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3518 regs[rd] = NULL;
3519 break;
3520 }
3521
3522 if (subr == DIF_SUBR_COPYIN) {
3523 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3524 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3525 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3526 }
3527
3528 mstate->dtms_scratch_ptr += scratch_size;
3529 regs[rd] = dest;
3530 break;
3531 }
3532
3533 case DIF_SUBR_COPYINTO: {
3534 uint64_t size = tupregs[1].dttk_value;
3535 uintptr_t dest = tupregs[2].dttk_value;
3536
3537 /*
3538 * This action doesn't require any credential checks since
3539 * probes will not activate in user contexts to which the
3540 * enabling user does not have permissions.
3541 */
3542 if (!dtrace_inscratch(dest, size, mstate)) {
3543 *flags |= CPU_DTRACE_BADADDR;
3544 *illval = regs[rd];
3545 break;
3546 }
3547
3548 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3549 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3550 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3551 break;
3552 }
3553
3554 case DIF_SUBR_COPYINSTR: {
3555 uintptr_t dest = mstate->dtms_scratch_ptr;
3556 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3557
3558 if (nargs > 1 && tupregs[1].dttk_value < size)
3559 size = tupregs[1].dttk_value + 1;
3560
3561 /*
3562 * This action doesn't require any credential checks since
3563 * probes will not activate in user contexts to which the
3564 * enabling user does not have permissions.
3565 */
3566 if (!DTRACE_INSCRATCH(mstate, size)) {
3567 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3568 regs[rd] = NULL;
3569 break;
3570 }
3571
3572 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3573 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3574 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3575
3576 ((char *)dest)[size - 1] = '\0';
3577 mstate->dtms_scratch_ptr += size;
3578 regs[rd] = dest;
3579 break;
3580 }
3581
3582 case DIF_SUBR_MSGSIZE:
3583 case DIF_SUBR_MSGDSIZE: {
3584 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3585 uintptr_t wptr, rptr;
3586 size_t count = 0;
3587 int cont = 0;
3588
3589 while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3590
3591 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3592 vstate)) {
3593 regs[rd] = NULL;
3594 break;
3595 }
3596
3597 wptr = dtrace_loadptr(baddr +
3598 offsetof(mblk_t, b_wptr));
3599
3600 rptr = dtrace_loadptr(baddr +
3601 offsetof(mblk_t, b_rptr));
3602
3603 if (wptr < rptr) {
3604 *flags |= CPU_DTRACE_BADADDR;
3605 *illval = tupregs[0].dttk_value;
3606 break;
3607 }
3608
3609 daddr = dtrace_loadptr(baddr +
3610 offsetof(mblk_t, b_datap));
3611
3612 baddr = dtrace_loadptr(baddr +
3613 offsetof(mblk_t, b_cont));
3614
3615 /*
3616 * We want to prevent against denial-of-service here,
3617 * so we're only going to search the list for
3618 * dtrace_msgdsize_max mblks.
3619 */
3620 if (cont++ > dtrace_msgdsize_max) {
3621 *flags |= CPU_DTRACE_ILLOP;
3622 break;
3623 }
3624
3625 if (subr == DIF_SUBR_MSGDSIZE) {
3626 if (dtrace_load8(daddr +
3627 offsetof(dblk_t, db_type)) != M_DATA)
3628 continue;
3629 }
3630
3631 count += wptr - rptr;
3632 }
3633
3634 if (!(*flags & CPU_DTRACE_FAULT))
3635 regs[rd] = count;
3636
3637 break;
3638 }
3639
3640 case DIF_SUBR_PROGENYOF: {
3641 pid_t pid = tupregs[0].dttk_value;
3642 proc_t *p;
3643 int rval = 0;
3644
3645 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3646
3647 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3648 if (p->p_pidp->pid_id == pid) {
3649 rval = 1;
3650 break;
3651 }
3652 }
3653
3654 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3655
3656 regs[rd] = rval;
3657 break;
3658 }
3659
3660 case DIF_SUBR_SPECULATION:
3661 regs[rd] = dtrace_speculation(state);
3662 break;
3663
3664 case DIF_SUBR_COPYOUT: {
3665 uintptr_t kaddr = tupregs[0].dttk_value;
3666 uintptr_t uaddr = tupregs[1].dttk_value;
3667 uint64_t size = tupregs[2].dttk_value;
3668
3669 if (!dtrace_destructive_disallow &&
3670 dtrace_priv_proc_control(state, mstate) &&
3671 !dtrace_istoxic(kaddr, size)) {
3672 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3673 dtrace_copyout(kaddr, uaddr, size, flags);
3674 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3675 }
3676 break;
3677 }
3678
3679 case DIF_SUBR_COPYOUTSTR: {
3680 uintptr_t kaddr = tupregs[0].dttk_value;
3681 uintptr_t uaddr = tupregs[1].dttk_value;
3682 uint64_t size = tupregs[2].dttk_value;
3683
3684 if (!dtrace_destructive_disallow &&
3685 dtrace_priv_proc_control(state, mstate) &&
3686 !dtrace_istoxic(kaddr, size)) {
3687 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3688 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3689 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3690 }
3691 break;
3692 }
3693
3694 case DIF_SUBR_STRLEN: {
3695 size_t sz;
3696 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3697 sz = dtrace_strlen((char *)addr,
3698 state->dts_options[DTRACEOPT_STRSIZE]);
3699
3700 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3701 regs[rd] = NULL;
3702 break;
3703 }
3704
3705 regs[rd] = sz;
3706
3707 break;
3708 }
3709
3710 case DIF_SUBR_STRCHR:
3711 case DIF_SUBR_STRRCHR: {
3712 /*
3713 * We're going to iterate over the string looking for the
3714 * specified character. We will iterate until we have reached
3715 * the string length or we have found the character. If this
3716 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3717 * of the specified character instead of the first.
3718 */
3719 uintptr_t saddr = tupregs[0].dttk_value;
3720 uintptr_t addr = tupregs[0].dttk_value;
3721 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3722 char c, target = (char)tupregs[1].dttk_value;
3723
3724 for (regs[rd] = NULL; addr < limit; addr++) {
3725 if ((c = dtrace_load8(addr)) == target) {
3726 regs[rd] = addr;
3727
3728 if (subr == DIF_SUBR_STRCHR)
3729 break;
3730 }
3731
3732 if (c == '\0')
3733 break;
3734 }
3735
3736 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3737 regs[rd] = NULL;
3738 break;
3739 }
3740
3741 break;
3742 }
3743
3744 case DIF_SUBR_STRSTR:
3745 case DIF_SUBR_INDEX:
3746 case DIF_SUBR_RINDEX: {
3747 /*
3748 * We're going to iterate over the string looking for the
3749 * specified string. We will iterate until we have reached
3750 * the string length or we have found the string. (Yes, this
3751 * is done in the most naive way possible -- but considering
3752 * that the string we're searching for is likely to be
3753 * relatively short, the complexity of Rabin-Karp or similar
3754 * hardly seems merited.)
3755 */
3756 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3757 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3758 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3759 size_t len = dtrace_strlen(addr, size);
3760 size_t sublen = dtrace_strlen(substr, size);
3761 char *limit = addr + len, *orig = addr;
3762 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3763 int inc = 1;
3764
3765 regs[rd] = notfound;
3766
3767 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3768 regs[rd] = NULL;
3769 break;
3770 }
3771
3772 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3773 vstate)) {
3774 regs[rd] = NULL;
3775 break;
3776 }
3777
3778 /*
3779 * strstr() and index()/rindex() have similar semantics if
3780 * both strings are the empty string: strstr() returns a
3781 * pointer to the (empty) string, and index() and rindex()
3782 * both return index 0 (regardless of any position argument).
3783 */
3784 if (sublen == 0 && len == 0) {
3785 if (subr == DIF_SUBR_STRSTR)
3786 regs[rd] = (uintptr_t)addr;
3787 else
3788 regs[rd] = 0;
3789 break;
3790 }
3791
3792 if (subr != DIF_SUBR_STRSTR) {
3793 if (subr == DIF_SUBR_RINDEX) {
3794 limit = orig - 1;
3795 addr += len;
3796 inc = -1;
3797 }
3798
3799 /*
3800 * Both index() and rindex() take an optional position
3801 * argument that denotes the starting position.
3802 */
3803 if (nargs == 3) {
3804 int64_t pos = (int64_t)tupregs[2].dttk_value;
3805
3806 /*
3807 * If the position argument to index() is
3808 * negative, Perl implicitly clamps it at
3809 * zero. This semantic is a little surprising
3810 * given the special meaning of negative
3811 * positions to similar Perl functions like
3812 * substr(), but it appears to reflect a
3813 * notion that index() can start from a
3814 * negative index and increment its way up to
3815 * the string. Given this notion, Perl's
3816 * rindex() is at least self-consistent in
3817 * that it implicitly clamps positions greater
3818 * than the string length to be the string
3819 * length. Where Perl completely loses
3820 * coherence, however, is when the specified
3821 * substring is the empty string (""). In
3822 * this case, even if the position is
3823 * negative, rindex() returns 0 -- and even if
3824 * the position is greater than the length,
3825 * index() returns the string length. These
3826 * semantics violate the notion that index()
3827 * should never return a value less than the
3828 * specified position and that rindex() should
3829 * never return a value greater than the
3830 * specified position. (One assumes that
3831 * these semantics are artifacts of Perl's
3832 * implementation and not the results of
3833 * deliberate design -- it beggars belief that
3834 * even Larry Wall could desire such oddness.)
3835 * While in the abstract one would wish for
3836 * consistent position semantics across
3837 * substr(), index() and rindex() -- or at the
3838 * very least self-consistent position
3839 * semantics for index() and rindex() -- we
3840 * instead opt to keep with the extant Perl
3841 * semantics, in all their broken glory. (Do
3842 * we have more desire to maintain Perl's
3843 * semantics than Perl does? Probably.)
3844 */
3845 if (subr == DIF_SUBR_RINDEX) {
3846 if (pos < 0) {
3847 if (sublen == 0)
3848 regs[rd] = 0;
3849 break;
3850 }
3851
3852 if (pos > len)
3853 pos = len;
3854 } else {
3855 if (pos < 0)
3856 pos = 0;
3857
3858 if (pos >= len) {
3859 if (sublen == 0)
3860 regs[rd] = len;
3861 break;
3862 }
3863 }
3864
3865 addr = orig + pos;
3866 }
3867 }
3868
3869 for (regs[rd] = notfound; addr != limit; addr += inc) {
3870 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3871 if (subr != DIF_SUBR_STRSTR) {
3872 /*
3873 * As D index() and rindex() are
3874 * modeled on Perl (and not on awk),
3875 * we return a zero-based (and not a
3876 * one-based) index. (For you Perl
3877 * weenies: no, we're not going to add
3878 * $[ -- and shouldn't you be at a con
3879 * or something?)
3880 */
3881 regs[rd] = (uintptr_t)(addr - orig);
3882 break;
3883 }
3884
3885 ASSERT(subr == DIF_SUBR_STRSTR);
3886 regs[rd] = (uintptr_t)addr;
3887 break;
3888 }
3889 }
3890
3891 break;
3892 }
3893
3894 case DIF_SUBR_STRTOK: {
3895 uintptr_t addr = tupregs[0].dttk_value;
3896 uintptr_t tokaddr = tupregs[1].dttk_value;
3897 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3898 uintptr_t limit, toklimit = tokaddr + size;
3899 uint8_t c, tokmap[32]; /* 256 / 8 */
3900 char *dest = (char *)mstate->dtms_scratch_ptr;
3901 int i;
3902
3903 /*
3904 * Check both the token buffer and (later) the input buffer,
3905 * since both could be non-scratch addresses.
3906 */
3907 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3908 regs[rd] = NULL;
3909 break;
3910 }
3911
3912 if (!DTRACE_INSCRATCH(mstate, size)) {
3913 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3914 regs[rd] = NULL;
3915 break;
3916 }
3917
3918 if (addr == NULL) {
3919 /*
3920 * If the address specified is NULL, we use our saved
3921 * strtok pointer from the mstate. Note that this
3922 * means that the saved strtok pointer is _only_
3923 * valid within multiple enablings of the same probe --
3924 * it behaves like an implicit clause-local variable.
3925 */
3926 addr = mstate->dtms_strtok;
3927 } else {
3928 /*
3929 * If the user-specified address is non-NULL we must
3930 * access check it. This is the only time we have
3931 * a chance to do so, since this address may reside
3932 * in the string table of this clause-- future calls
3933 * (when we fetch addr from mstate->dtms_strtok)
3934 * would fail this access check.
3935 */
3936 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3937 regs[rd] = NULL;
3938 break;
3939 }
3940 }
3941
3942 /*
3943 * First, zero the token map, and then process the token
3944 * string -- setting a bit in the map for every character
3945 * found in the token string.
3946 */
3947 for (i = 0; i < sizeof (tokmap); i++)
3948 tokmap[i] = 0;
3949
3950 for (; tokaddr < toklimit; tokaddr++) {
3951 if ((c = dtrace_load8(tokaddr)) == '\0')
3952 break;
3953
3954 ASSERT((c >> 3) < sizeof (tokmap));
3955 tokmap[c >> 3] |= (1 << (c & 0x7));
3956 }
3957
3958 for (limit = addr + size; addr < limit; addr++) {
3959 /*
3960 * We're looking for a character that is _not_ contained
3961 * in the token string.
3962 */
3963 if ((c = dtrace_load8(addr)) == '\0')
3964 break;
3965
3966 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3967 break;
3968 }
3969
3970 if (c == '\0') {
3971 /*
3972 * We reached the end of the string without finding
3973 * any character that was not in the token string.
3974 * We return NULL in this case, and we set the saved
3975 * address to NULL as well.
3976 */
3977 regs[rd] = NULL;
3978 mstate->dtms_strtok = NULL;
3979 break;
3980 }
3981
3982 /*
3983 * From here on, we're copying into the destination string.
3984 */
3985 for (i = 0; addr < limit && i < size - 1; addr++) {
3986 if ((c = dtrace_load8(addr)) == '\0')
3987 break;
3988
3989 if (tokmap[c >> 3] & (1 << (c & 0x7)))
3990 break;
3991
3992 ASSERT(i < size);
3993 dest[i++] = c;
3994 }
3995
3996 ASSERT(i < size);
3997 dest[i] = '\0';
3998 regs[rd] = (uintptr_t)dest;
3999 mstate->dtms_scratch_ptr += size;
4000 mstate->dtms_strtok = addr;
4001 break;
4002 }
4003
4004 case DIF_SUBR_SUBSTR: {
4005 uintptr_t s = tupregs[0].dttk_value;
4006 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4007 char *d = (char *)mstate->dtms_scratch_ptr;
4008 int64_t index = (int64_t)tupregs[1].dttk_value;
4009 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4010 size_t len = dtrace_strlen((char *)s, size);
4011 int64_t i;
4012
4013 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4014 regs[rd] = NULL;
4015 break;
4016 }
4017
4018 if (!DTRACE_INSCRATCH(mstate, size)) {
4019 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4020 regs[rd] = NULL;
4021 break;
4022 }
4023
4024 if (nargs <= 2)
4025 remaining = (int64_t)size;
4026
4027 if (index < 0) {
4028 index += len;
4029
4030 if (index < 0 && index + remaining > 0) {
4031 remaining += index;
4032 index = 0;
4033 }
4034 }
4035
4036 if (index >= len || index < 0) {
4037 remaining = 0;
4038 } else if (remaining < 0) {
4039 remaining += len - index;
4040 } else if (index + remaining > size) {
4041 remaining = size - index;
4042 }
4043
4044 for (i = 0; i < remaining; i++) {
4045 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4046 break;
4047 }
4048
4049 d[i] = '\0';
4050
4051 mstate->dtms_scratch_ptr += size;
4052 regs[rd] = (uintptr_t)d;
4053 break;
4054 }
4055
4056 case DIF_SUBR_TOUPPER:
4057 case DIF_SUBR_TOLOWER: {
4058 uintptr_t s = tupregs[0].dttk_value;
4059 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4060 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4061 size_t len = dtrace_strlen((char *)s, size);
4062 char lower, upper, convert;
4063 int64_t i;
4064
4065 if (subr == DIF_SUBR_TOUPPER) {
4066 lower = 'a';
4067 upper = 'z';
4068 convert = 'A';
4069 } else {
4070 lower = 'A';
4071 upper = 'Z';
4072 convert = 'a';
4073 }
4074
4075 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4076 regs[rd] = NULL;
4077 break;
4078 }
4079
4080 if (!DTRACE_INSCRATCH(mstate, size)) {
4081 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4082 regs[rd] = NULL;
4083 break;
4084 }
4085
4086 for (i = 0; i < size - 1; i++) {
4087 if ((c = dtrace_load8(s + i)) == '\0')
4088 break;
4089
4090 if (c >= lower && c <= upper)
4091 c = convert + (c - lower);
4092
4093 dest[i] = c;
4094 }
4095
4096 ASSERT(i < size);
4097 dest[i] = '\0';
4098 regs[rd] = (uintptr_t)dest;
4099 mstate->dtms_scratch_ptr += size;
4100 break;
4101 }
4102
4103 case DIF_SUBR_GETMAJOR:
4104 #ifdef _LP64
4105 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4106 #else
4107 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4108 #endif
4109 break;
4110
4111 case DIF_SUBR_GETMINOR:
4112 #ifdef _LP64
4113 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4114 #else
4115 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4116 #endif
4117 break;
4118
4119 case DIF_SUBR_DDI_PATHNAME: {
4120 /*
4121 * This one is a galactic mess. We are going to roughly
4122 * emulate ddi_pathname(), but it's made more complicated
4123 * by the fact that we (a) want to include the minor name and
4124 * (b) must proceed iteratively instead of recursively.
4125 */
4126 uintptr_t dest = mstate->dtms_scratch_ptr;
4127 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4128 char *start = (char *)dest, *end = start + size - 1;
4129 uintptr_t daddr = tupregs[0].dttk_value;
4130 int64_t minor = (int64_t)tupregs[1].dttk_value;
4131 char *s;
4132 int i, len, depth = 0;
4133
4134 /*
4135 * Due to all the pointer jumping we do and context we must
4136 * rely upon, we just mandate that the user must have kernel
4137 * read privileges to use this routine.
4138 */
4139 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4140 *flags |= CPU_DTRACE_KPRIV;
4141 *illval = daddr;
4142 regs[rd] = NULL;
4143 }
4144
4145 if (!DTRACE_INSCRATCH(mstate, size)) {
4146 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4147 regs[rd] = NULL;
4148 break;
4149 }
4150
4151 *end = '\0';
4152
4153 /*
4154 * We want to have a name for the minor. In order to do this,
4155 * we need to walk the minor list from the devinfo. We want
4156 * to be sure that we don't infinitely walk a circular list,
4157 * so we check for circularity by sending a scout pointer
4158 * ahead two elements for every element that we iterate over;
4159 * if the list is circular, these will ultimately point to the
4160 * same element. You may recognize this little trick as the
4161 * answer to a stupid interview question -- one that always
4162 * seems to be asked by those who had to have it laboriously
4163 * explained to them, and who can't even concisely describe
4164 * the conditions under which one would be forced to resort to
4165 * this technique. Needless to say, those conditions are
4166 * found here -- and probably only here. Is this the only use
4167 * of this infamous trick in shipping, production code? If it
4168 * isn't, it probably should be...
4169 */
4170 if (minor != -1) {
4171 uintptr_t maddr = dtrace_loadptr(daddr +
4172 offsetof(struct dev_info, devi_minor));
4173
4174 uintptr_t next = offsetof(struct ddi_minor_data, next);
4175 uintptr_t name = offsetof(struct ddi_minor_data,
4176 d_minor) + offsetof(struct ddi_minor, name);
4177 uintptr_t dev = offsetof(struct ddi_minor_data,
4178 d_minor) + offsetof(struct ddi_minor, dev);
4179 uintptr_t scout;
4180
4181 if (maddr != NULL)
4182 scout = dtrace_loadptr(maddr + next);
4183
4184 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4185 uint64_t m;
4186 #ifdef _LP64
4187 m = dtrace_load64(maddr + dev) & MAXMIN64;
4188 #else
4189 m = dtrace_load32(maddr + dev) & MAXMIN;
4190 #endif
4191 if (m != minor) {
4192 maddr = dtrace_loadptr(maddr + next);
4193
4194 if (scout == NULL)
4195 continue;
4196
4197 scout = dtrace_loadptr(scout + next);
4198
4199 if (scout == NULL)
4200 continue;
4201
4202 scout = dtrace_loadptr(scout + next);
4203
4204 if (scout == NULL)
4205 continue;
4206
4207 if (scout == maddr) {
4208 *flags |= CPU_DTRACE_ILLOP;
4209 break;
4210 }
4211
4212 continue;
4213 }
4214
4215 /*
4216 * We have the minor data. Now we need to
4217 * copy the minor's name into the end of the
4218 * pathname.
4219 */
4220 s = (char *)dtrace_loadptr(maddr + name);
4221 len = dtrace_strlen(s, size);
4222
4223 if (*flags & CPU_DTRACE_FAULT)
4224 break;
4225
4226 if (len != 0) {
4227 if ((end -= (len + 1)) < start)
4228 break;
4229
4230 *end = ':';
4231 }
4232
4233 for (i = 1; i <= len; i++)
4234 end[i] = dtrace_load8((uintptr_t)s++);
4235 break;
4236 }
4237 }
4238
4239 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4240 ddi_node_state_t devi_state;
4241
4242 devi_state = dtrace_load32(daddr +
4243 offsetof(struct dev_info, devi_node_state));
4244
4245 if (*flags & CPU_DTRACE_FAULT)
4246 break;
4247
4248 if (devi_state >= DS_INITIALIZED) {
4249 s = (char *)dtrace_loadptr(daddr +
4250 offsetof(struct dev_info, devi_addr));
4251 len = dtrace_strlen(s, size);
4252
4253 if (*flags & CPU_DTRACE_FAULT)
4254 break;
4255
4256 if (len != 0) {
4257 if ((end -= (len + 1)) < start)
4258 break;
4259
4260 *end = '@';
4261 }
4262
4263 for (i = 1; i <= len; i++)
4264 end[i] = dtrace_load8((uintptr_t)s++);
4265 }
4266
4267 /*
4268 * Now for the node name...
4269 */
4270 s = (char *)dtrace_loadptr(daddr +
4271 offsetof(struct dev_info, devi_node_name));
4272
4273 daddr = dtrace_loadptr(daddr +
4274 offsetof(struct dev_info, devi_parent));
4275
4276 /*
4277 * If our parent is NULL (that is, if we're the root
4278 * node), we're going to use the special path
4279 * "devices".
4280 */
4281 if (daddr == NULL)
4282 s = "devices";
4283
4284 len = dtrace_strlen(s, size);
4285 if (*flags & CPU_DTRACE_FAULT)
4286 break;
4287
4288 if ((end -= (len + 1)) < start)
4289 break;
4290
4291 for (i = 1; i <= len; i++)
4292 end[i] = dtrace_load8((uintptr_t)s++);
4293 *end = '/';
4294
4295 if (depth++ > dtrace_devdepth_max) {
4296 *flags |= CPU_DTRACE_ILLOP;
4297 break;
4298 }
4299 }
4300
4301 if (end < start)
4302 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4303
4304 if (daddr == NULL) {
4305 regs[rd] = (uintptr_t)end;
4306 mstate->dtms_scratch_ptr += size;
4307 }
4308
4309 break;
4310 }
4311
4312 case DIF_SUBR_STRJOIN: {
4313 char *d = (char *)mstate->dtms_scratch_ptr;
4314 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4315 uintptr_t s1 = tupregs[0].dttk_value;
4316 uintptr_t s2 = tupregs[1].dttk_value;
4317 int i = 0;
4318
4319 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4320 !dtrace_strcanload(s2, size, mstate, vstate)) {
4321 regs[rd] = NULL;
4322 break;
4323 }
4324
4325 if (!DTRACE_INSCRATCH(mstate, size)) {
4326 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4327 regs[rd] = NULL;
4328 break;
4329 }
4330
4331 for (;;) {
4332 if (i >= size) {
4333 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4334 regs[rd] = NULL;
4335 break;
4336 }
4337
4338 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4339 i--;
4340 break;
4341 }
4342 }
4343
4344 for (;;) {
4345 if (i >= size) {
4346 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4347 regs[rd] = NULL;
4348 break;
4349 }
4350
4351 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4352 break;
4353 }
4354
4355 if (i < size) {
4356 mstate->dtms_scratch_ptr += i;
4357 regs[rd] = (uintptr_t)d;
4358 }
4359
4360 break;
4361 }
4362
4363 case DIF_SUBR_LLTOSTR: {
4364 int64_t i = (int64_t)tupregs[0].dttk_value;
4365 uint64_t val, digit;
4366 uint64_t size = 65; /* enough room for 2^64 in binary */
4367 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4368 int base = 10;
4369
4370 if (nargs > 1) {
4371 if ((base = tupregs[1].dttk_value) <= 1 ||
4372 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4373 *flags |= CPU_DTRACE_ILLOP;
4374 break;
4375 }
4376 }
4377
4378 val = (base == 10 && i < 0) ? i * -1 : i;
4379
4380 if (!DTRACE_INSCRATCH(mstate, size)) {
4381 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4382 regs[rd] = NULL;
4383 break;
4384 }
4385
4386 for (*end-- = '\0'; val; val /= base) {
4387 if ((digit = val % base) <= '9' - '0') {
4388 *end-- = '0' + digit;
4389 } else {
4390 *end-- = 'a' + (digit - ('9' - '0') - 1);
4391 }
4392 }
4393
4394 if (i == 0 && base == 16)
4395 *end-- = '0';
4396
4397 if (base == 16)
4398 *end-- = 'x';
4399
4400 if (i == 0 || base == 8 || base == 16)
4401 *end-- = '0';
4402
4403 if (i < 0 && base == 10)
4404 *end-- = '-';
4405
4406 regs[rd] = (uintptr_t)end + 1;
4407 mstate->dtms_scratch_ptr += size;
4408 break;
4409 }
4410
4411 case DIF_SUBR_HTONS:
4412 case DIF_SUBR_NTOHS:
4413 #ifdef _BIG_ENDIAN
4414 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4415 #else
4416 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4417 #endif
4418 break;
4419
4420
4421 case DIF_SUBR_HTONL:
4422 case DIF_SUBR_NTOHL:
4423 #ifdef _BIG_ENDIAN
4424 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4425 #else
4426 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4427 #endif
4428 break;
4429
4430
4431 case DIF_SUBR_HTONLL:
4432 case DIF_SUBR_NTOHLL:
4433 #ifdef _BIG_ENDIAN
4434 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4435 #else
4436 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4437 #endif
4438 break;
4439
4440
4441 case DIF_SUBR_DIRNAME:
4442 case DIF_SUBR_BASENAME: {
4443 char *dest = (char *)mstate->dtms_scratch_ptr;
4444 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4445 uintptr_t src = tupregs[0].dttk_value;
4446 int i, j, len = dtrace_strlen((char *)src, size);
4447 int lastbase = -1, firstbase = -1, lastdir = -1;
4448 int start, end;
4449
4450 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4451 regs[rd] = NULL;
4452 break;
4453 }
4454
4455 if (!DTRACE_INSCRATCH(mstate, size)) {
4456 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4457 regs[rd] = NULL;
4458 break;
4459 }
4460
4461 /*
4462 * The basename and dirname for a zero-length string is
4463 * defined to be "."
4464 */
4465 if (len == 0) {
4466 len = 1;
4467 src = (uintptr_t)".";
4468 }
4469
4470 /*
4471 * Start from the back of the string, moving back toward the
4472 * front until we see a character that isn't a slash. That
4473 * character is the last character in the basename.
4474 */
4475 for (i = len - 1; i >= 0; i--) {
4476 if (dtrace_load8(src + i) != '/')
4477 break;
4478 }
4479
4480 if (i >= 0)
4481 lastbase = i;
4482
4483 /*
4484 * Starting from the last character in the basename, move
4485 * towards the front until we find a slash. The character
4486 * that we processed immediately before that is the first
4487 * character in the basename.
4488 */
4489 for (; i >= 0; i--) {
4490 if (dtrace_load8(src + i) == '/')
4491 break;
4492 }
4493
4494 if (i >= 0)
4495 firstbase = i + 1;
4496
4497 /*
4498 * Now keep going until we find a non-slash character. That
4499 * character is the last character in the dirname.
4500 */
4501 for (; i >= 0; i--) {
4502 if (dtrace_load8(src + i) != '/')
4503 break;
4504 }
4505
4506 if (i >= 0)
4507 lastdir = i;
4508
4509 ASSERT(!(lastbase == -1 && firstbase != -1));
4510 ASSERT(!(firstbase == -1 && lastdir != -1));
4511
4512 if (lastbase == -1) {
4513 /*
4514 * We didn't find a non-slash character. We know that
4515 * the length is non-zero, so the whole string must be
4516 * slashes. In either the dirname or the basename
4517 * case, we return '/'.
4518 */
4519 ASSERT(firstbase == -1);
4520 firstbase = lastbase = lastdir = 0;
4521 }
4522
4523 if (firstbase == -1) {
4524 /*
4525 * The entire string consists only of a basename
4526 * component. If we're looking for dirname, we need
4527 * to change our string to be just "."; if we're
4528 * looking for a basename, we'll just set the first
4529 * character of the basename to be 0.
4530 */
4531 if (subr == DIF_SUBR_DIRNAME) {
4532 ASSERT(lastdir == -1);
4533 src = (uintptr_t)".";
4534 lastdir = 0;
4535 } else {
4536 firstbase = 0;
4537 }
4538 }
4539
4540 if (subr == DIF_SUBR_DIRNAME) {
4541 if (lastdir == -1) {
4542 /*
4543 * We know that we have a slash in the name --
4544 * or lastdir would be set to 0, above. And
4545 * because lastdir is -1, we know that this
4546 * slash must be the first character. (That
4547 * is, the full string must be of the form
4548 * "/basename".) In this case, the last
4549 * character of the directory name is 0.
4550 */
4551 lastdir = 0;
4552 }
4553
4554 start = 0;
4555 end = lastdir;
4556 } else {
4557 ASSERT(subr == DIF_SUBR_BASENAME);
4558 ASSERT(firstbase != -1 && lastbase != -1);
4559 start = firstbase;
4560 end = lastbase;
4561 }
4562
4563 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4564 dest[j] = dtrace_load8(src + i);
4565
4566 dest[j] = '\0';
4567 regs[rd] = (uintptr_t)dest;
4568 mstate->dtms_scratch_ptr += size;
4569 break;
4570 }
4571
4572 case DIF_SUBR_GETF: {
4573 uintptr_t fd = tupregs[0].dttk_value;
4574 uf_info_t *finfo = &curthread->t_procp->p_user.u_finfo;
4575 file_t *fp;
4576
4577 if (!dtrace_priv_proc(state, mstate)) {
4578 regs[rd] = NULL;
4579 break;
4580 }
4581
4582 /*
4583 * This is safe because fi_nfiles only increases, and the
4584 * fi_list array is not freed when the array size doubles.
4585 * (See the comment in flist_grow() for details on the
4586 * management of the u_finfo structure.)
4587 */
4588 fp = fd < finfo->fi_nfiles ? finfo->fi_list[fd].uf_file : NULL;
4589
4590 mstate->dtms_getf = fp;
4591 regs[rd] = (uintptr_t)fp;
4592 break;
4593 }
4594
4595 case DIF_SUBR_CLEANPATH: {
4596 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4597 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4598 uintptr_t src = tupregs[0].dttk_value;
4599 int i = 0, j = 0;
4600 zone_t *z;
4601
4602 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4603 regs[rd] = NULL;
4604 break;
4605 }
4606
4607 if (!DTRACE_INSCRATCH(mstate, size)) {
4608 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4609 regs[rd] = NULL;
4610 break;
4611 }
4612
4613 /*
4614 * Move forward, loading each character.
4615 */
4616 do {
4617 c = dtrace_load8(src + i++);
4618 next:
4619 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4620 break;
4621
4622 if (c != '/') {
4623 dest[j++] = c;
4624 continue;
4625 }
4626
4627 c = dtrace_load8(src + i++);
4628
4629 if (c == '/') {
4630 /*
4631 * We have two slashes -- we can just advance
4632 * to the next character.
4633 */
4634 goto next;
4635 }
4636
4637 if (c != '.') {
4638 /*
4639 * This is not "." and it's not ".." -- we can
4640 * just store the "/" and this character and
4641 * drive on.
4642 */
4643 dest[j++] = '/';
4644 dest[j++] = c;
4645 continue;
4646 }
4647
4648 c = dtrace_load8(src + i++);
4649
4650 if (c == '/') {
4651 /*
4652 * This is a "/./" component. We're not going
4653 * to store anything in the destination buffer;
4654 * we're just going to go to the next component.
4655 */
4656 goto next;
4657 }
4658
4659 if (c != '.') {
4660 /*
4661 * This is not ".." -- we can just store the
4662 * "/." and this character and continue
4663 * processing.
4664 */
4665 dest[j++] = '/';
4666 dest[j++] = '.';
4667 dest[j++] = c;
4668 continue;
4669 }
4670
4671 c = dtrace_load8(src + i++);
4672
4673 if (c != '/' && c != '\0') {
4674 /*
4675 * This is not ".." -- it's "..[mumble]".
4676 * We'll store the "/.." and this character
4677 * and continue processing.
4678 */
4679 dest[j++] = '/';
4680 dest[j++] = '.';
4681 dest[j++] = '.';
4682 dest[j++] = c;
4683 continue;
4684 }
4685
4686 /*
4687 * This is "/../" or "/..\0". We need to back up
4688 * our destination pointer until we find a "/".
4689 */
4690 i--;
4691 while (j != 0 && dest[--j] != '/')
4692 continue;
4693
4694 if (c == '\0')
4695 dest[++j] = '/';
4696 } while (c != '\0');
4697
4698 dest[j] = '\0';
4699
4700 if (mstate->dtms_getf != NULL &&
4701 !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
4702 (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
4703 /*
4704 * If we've done a getf() as a part of this ECB and we
4705 * don't have kernel access (and we're not in the global
4706 * zone), check if the path we cleaned up begins with
4707 * the zone's root path, and trim it off if so. Note
4708 * that this is an output cleanliness issue, not a
4709 * security issue: knowing one's zone root path does
4710 * not enable privilege escalation.
4711 */
4712 if (strstr(dest, z->zone_rootpath) == dest)
4713 dest += strlen(z->zone_rootpath) - 1;
4714 }
4715
4716 regs[rd] = (uintptr_t)dest;
4717 mstate->dtms_scratch_ptr += size;
4718 break;
4719 }
4720
4721 case DIF_SUBR_INET_NTOA:
4722 case DIF_SUBR_INET_NTOA6:
4723 case DIF_SUBR_INET_NTOP: {
4724 size_t size;
4725 int af, argi, i;
4726 char *base, *end;
4727
4728 if (subr == DIF_SUBR_INET_NTOP) {
4729 af = (int)tupregs[0].dttk_value;
4730 argi = 1;
4731 } else {
4732 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4733 argi = 0;
4734 }
4735
4736 if (af == AF_INET) {
4737 ipaddr_t ip4;
4738 uint8_t *ptr8, val;
4739
4740 /*
4741 * Safely load the IPv4 address.
4742 */
4743 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4744
4745 /*
4746 * Check an IPv4 string will fit in scratch.
4747 */
4748 size = INET_ADDRSTRLEN;
4749 if (!DTRACE_INSCRATCH(mstate, size)) {
4750 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4751 regs[rd] = NULL;
4752 break;
4753 }
4754 base = (char *)mstate->dtms_scratch_ptr;
4755 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4756
4757 /*
4758 * Stringify as a dotted decimal quad.
4759 */
4760 *end-- = '\0';
4761 ptr8 = (uint8_t *)&ip4;
4762 for (i = 3; i >= 0; i--) {
4763 val = ptr8[i];
4764
4765 if (val == 0) {
4766 *end-- = '0';
4767 } else {
4768 for (; val; val /= 10) {
4769 *end-- = '0' + (val % 10);
4770 }
4771 }
4772
4773 if (i > 0)
4774 *end-- = '.';
4775 }
4776 ASSERT(end + 1 >= base);
4777
4778 } else if (af == AF_INET6) {
4779 struct in6_addr ip6;
4780 int firstzero, tryzero, numzero, v6end;
4781 uint16_t val;
4782 const char digits[] = "0123456789abcdef";
4783
4784 /*
4785 * Stringify using RFC 1884 convention 2 - 16 bit
4786 * hexadecimal values with a zero-run compression.
4787 * Lower case hexadecimal digits are used.
4788 * eg, fe80::214:4fff:fe0b:76c8.
4789 * The IPv4 embedded form is returned for inet_ntop,
4790 * just the IPv4 string is returned for inet_ntoa6.
4791 */
4792
4793 /*
4794 * Safely load the IPv6 address.
4795 */
4796 dtrace_bcopy(
4797 (void *)(uintptr_t)tupregs[argi].dttk_value,
4798 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4799
4800 /*
4801 * Check an IPv6 string will fit in scratch.
4802 */
4803 size = INET6_ADDRSTRLEN;
4804 if (!DTRACE_INSCRATCH(mstate, size)) {
4805 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4806 regs[rd] = NULL;
4807 break;
4808 }
4809 base = (char *)mstate->dtms_scratch_ptr;
4810 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4811 *end-- = '\0';
4812
4813 /*
4814 * Find the longest run of 16 bit zero values
4815 * for the single allowed zero compression - "::".
4816 */
4817 firstzero = -1;
4818 tryzero = -1;
4819 numzero = 1;
4820 for (i = 0; i < sizeof (struct in6_addr); i++) {
4821 if (ip6._S6_un._S6_u8[i] == 0 &&
4822 tryzero == -1 && i % 2 == 0) {
4823 tryzero = i;
4824 continue;
4825 }
4826
4827 if (tryzero != -1 &&
4828 (ip6._S6_un._S6_u8[i] != 0 ||
4829 i == sizeof (struct in6_addr) - 1)) {
4830
4831 if (i - tryzero <= numzero) {
4832 tryzero = -1;
4833 continue;
4834 }
4835
4836 firstzero = tryzero;
4837 numzero = i - i % 2 - tryzero;
4838 tryzero = -1;
4839
4840 if (ip6._S6_un._S6_u8[i] == 0 &&
4841 i == sizeof (struct in6_addr) - 1)
4842 numzero += 2;
4843 }
4844 }
4845 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4846
4847 /*
4848 * Check for an IPv4 embedded address.
4849 */
4850 v6end = sizeof (struct in6_addr) - 2;
4851 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4852 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4853 for (i = sizeof (struct in6_addr) - 1;
4854 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4855 ASSERT(end >= base);
4856
4857 val = ip6._S6_un._S6_u8[i];
4858
4859 if (val == 0) {
4860 *end-- = '0';
4861 } else {
4862 for (; val; val /= 10) {
4863 *end-- = '0' + val % 10;
4864 }
4865 }
4866
4867 if (i > DTRACE_V4MAPPED_OFFSET)
4868 *end-- = '.';
4869 }
4870
4871 if (subr == DIF_SUBR_INET_NTOA6)
4872 goto inetout;
4873
4874 /*
4875 * Set v6end to skip the IPv4 address that
4876 * we have already stringified.
4877 */
4878 v6end = 10;
4879 }
4880
4881 /*
4882 * Build the IPv6 string by working through the
4883 * address in reverse.
4884 */
4885 for (i = v6end; i >= 0; i -= 2) {
4886 ASSERT(end >= base);
4887
4888 if (i == firstzero + numzero - 2) {
4889 *end-- = ':';
4890 *end-- = ':';
4891 i -= numzero - 2;
4892 continue;
4893 }
4894
4895 if (i < 14 && i != firstzero - 2)
4896 *end-- = ':';
4897
4898 val = (ip6._S6_un._S6_u8[i] << 8) +
4899 ip6._S6_un._S6_u8[i + 1];
4900
4901 if (val == 0) {
4902 *end-- = '0';
4903 } else {
4904 for (; val; val /= 16) {
4905 *end-- = digits[val % 16];
4906 }
4907 }
4908 }
4909 ASSERT(end + 1 >= base);
4910
4911 } else {
4912 /*
4913 * The user didn't use AH_INET or AH_INET6.
4914 */
4915 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4916 regs[rd] = NULL;
4917 break;
4918 }
4919
4920 inetout: regs[rd] = (uintptr_t)end + 1;
4921 mstate->dtms_scratch_ptr += size;
4922 break;
4923 }
4924
4925 }
4926 }
4927
4928 /*
4929 * Emulate the execution of DTrace IR instructions specified by the given
4930 * DIF object. This function is deliberately void of assertions as all of
4931 * the necessary checks are handled by a call to dtrace_difo_validate().
4932 */
4933 static uint64_t
4934 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4935 dtrace_vstate_t *vstate, dtrace_state_t *state)
4936 {
4937 const dif_instr_t *text = difo->dtdo_buf;
4938 const uint_t textlen = difo->dtdo_len;
4939 const char *strtab = difo->dtdo_strtab;
4940 const uint64_t *inttab = difo->dtdo_inttab;
4941
4942 uint64_t rval = 0;
4943 dtrace_statvar_t *svar;
4944 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4945 dtrace_difv_t *v;
4946 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4947 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
4948
4949 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4950 uint64_t regs[DIF_DIR_NREGS];
4951 uint64_t *tmp;
4952
4953 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4954 int64_t cc_r;
4955 uint_t pc = 0, id, opc;
4956 uint8_t ttop = 0;
4957 dif_instr_t instr;
4958 uint_t r1, r2, rd;
4959
4960 /*
4961 * We stash the current DIF object into the machine state: we need it
4962 * for subsequent access checking.
4963 */
4964 mstate->dtms_difo = difo;
4965
4966 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
4967
4968 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4969 opc = pc;
4970
4971 instr = text[pc++];
4972 r1 = DIF_INSTR_R1(instr);
4973 r2 = DIF_INSTR_R2(instr);
4974 rd = DIF_INSTR_RD(instr);
4975
4976 switch (DIF_INSTR_OP(instr)) {
4977 case DIF_OP_OR:
4978 regs[rd] = regs[r1] | regs[r2];
4979 break;
4980 case DIF_OP_XOR:
4981 regs[rd] = regs[r1] ^ regs[r2];
4982 break;
4983 case DIF_OP_AND:
4984 regs[rd] = regs[r1] & regs[r2];
4985 break;
4986 case DIF_OP_SLL:
4987 regs[rd] = regs[r1] << regs[r2];
4988 break;
4989 case DIF_OP_SRL:
4990 regs[rd] = regs[r1] >> regs[r2];
4991 break;
4992 case DIF_OP_SUB:
4993 regs[rd] = regs[r1] - regs[r2];
4994 break;
4995 case DIF_OP_ADD:
4996 regs[rd] = regs[r1] + regs[r2];
4997 break;
4998 case DIF_OP_MUL:
4999 regs[rd] = regs[r1] * regs[r2];
5000 break;
5001 case DIF_OP_SDIV:
5002 if (regs[r2] == 0) {
5003 regs[rd] = 0;
5004 *flags |= CPU_DTRACE_DIVZERO;
5005 } else {
5006 regs[rd] = (int64_t)regs[r1] /
5007 (int64_t)regs[r2];
5008 }
5009 break;
5010
5011 case DIF_OP_UDIV:
5012 if (regs[r2] == 0) {
5013 regs[rd] = 0;
5014 *flags |= CPU_DTRACE_DIVZERO;
5015 } else {
5016 regs[rd] = regs[r1] / regs[r2];
5017 }
5018 break;
5019
5020 case DIF_OP_SREM:
5021 if (regs[r2] == 0) {
5022 regs[rd] = 0;
5023 *flags |= CPU_DTRACE_DIVZERO;
5024 } else {
5025 regs[rd] = (int64_t)regs[r1] %
5026 (int64_t)regs[r2];
5027 }
5028 break;
5029
5030 case DIF_OP_UREM:
5031 if (regs[r2] == 0) {
5032 regs[rd] = 0;
5033 *flags |= CPU_DTRACE_DIVZERO;
5034 } else {
5035 regs[rd] = regs[r1] % regs[r2];
5036 }
5037 break;
5038
5039 case DIF_OP_NOT:
5040 regs[rd] = ~regs[r1];
5041 break;
5042 case DIF_OP_MOV:
5043 regs[rd] = regs[r1];
5044 break;
5045 case DIF_OP_CMP:
5046 cc_r = regs[r1] - regs[r2];
5047 cc_n = cc_r < 0;
5048 cc_z = cc_r == 0;
5049 cc_v = 0;
5050 cc_c = regs[r1] < regs[r2];
5051 break;
5052 case DIF_OP_TST:
5053 cc_n = cc_v = cc_c = 0;
5054 cc_z = regs[r1] == 0;
5055 break;
5056 case DIF_OP_BA:
5057 pc = DIF_INSTR_LABEL(instr);
5058 break;
5059 case DIF_OP_BE:
5060 if (cc_z)
5061 pc = DIF_INSTR_LABEL(instr);
5062 break;
5063 case DIF_OP_BNE:
5064 if (cc_z == 0)
5065 pc = DIF_INSTR_LABEL(instr);
5066 break;
5067 case DIF_OP_BG:
5068 if ((cc_z | (cc_n ^ cc_v)) == 0)
5069 pc = DIF_INSTR_LABEL(instr);
5070 break;
5071 case DIF_OP_BGU:
5072 if ((cc_c | cc_z) == 0)
5073 pc = DIF_INSTR_LABEL(instr);
5074 break;
5075 case DIF_OP_BGE:
5076 if ((cc_n ^ cc_v) == 0)
5077 pc = DIF_INSTR_LABEL(instr);
5078 break;
5079 case DIF_OP_BGEU:
5080 if (cc_c == 0)
5081 pc = DIF_INSTR_LABEL(instr);
5082 break;
5083 case DIF_OP_BL:
5084 if (cc_n ^ cc_v)
5085 pc = DIF_INSTR_LABEL(instr);
5086 break;
5087 case DIF_OP_BLU:
5088 if (cc_c)
5089 pc = DIF_INSTR_LABEL(instr);
5090 break;
5091 case DIF_OP_BLE:
5092 if (cc_z | (cc_n ^ cc_v))
5093 pc = DIF_INSTR_LABEL(instr);
5094 break;
5095 case DIF_OP_BLEU:
5096 if (cc_c | cc_z)
5097 pc = DIF_INSTR_LABEL(instr);
5098 break;
5099 case DIF_OP_RLDSB:
5100 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5101 break;
5102 /*FALLTHROUGH*/
5103 case DIF_OP_LDSB:
5104 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5105 break;
5106 case DIF_OP_RLDSH:
5107 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5108 break;
5109 /*FALLTHROUGH*/
5110 case DIF_OP_LDSH:
5111 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5112 break;
5113 case DIF_OP_RLDSW:
5114 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5115 break;
5116 /*FALLTHROUGH*/
5117 case DIF_OP_LDSW:
5118 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5119 break;
5120 case DIF_OP_RLDUB:
5121 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5122 break;
5123 /*FALLTHROUGH*/
5124 case DIF_OP_LDUB:
5125 regs[rd] = dtrace_load8(regs[r1]);
5126 break;
5127 case DIF_OP_RLDUH:
5128 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5129 break;
5130 /*FALLTHROUGH*/
5131 case DIF_OP_LDUH:
5132 regs[rd] = dtrace_load16(regs[r1]);
5133 break;
5134 case DIF_OP_RLDUW:
5135 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5136 break;
5137 /*FALLTHROUGH*/
5138 case DIF_OP_LDUW:
5139 regs[rd] = dtrace_load32(regs[r1]);
5140 break;
5141 case DIF_OP_RLDX:
5142 if (!dtrace_canload(regs[r1], 8, mstate, vstate))
5143 break;
5144 /*FALLTHROUGH*/
5145 case DIF_OP_LDX:
5146 regs[rd] = dtrace_load64(regs[r1]);
5147 break;
5148 case DIF_OP_ULDSB:
5149 regs[rd] = (int8_t)
5150 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5151 break;
5152 case DIF_OP_ULDSH:
5153 regs[rd] = (int16_t)
5154 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5155 break;
5156 case DIF_OP_ULDSW:
5157 regs[rd] = (int32_t)
5158 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5159 break;
5160 case DIF_OP_ULDUB:
5161 regs[rd] =
5162 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5163 break;
5164 case DIF_OP_ULDUH:
5165 regs[rd] =
5166 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5167 break;
5168 case DIF_OP_ULDUW:
5169 regs[rd] =
5170 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5171 break;
5172 case DIF_OP_ULDX:
5173 regs[rd] =
5174 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5175 break;
5176 case DIF_OP_RET:
5177 rval = regs[rd];
5178 pc = textlen;
5179 break;
5180 case DIF_OP_NOP:
5181 break;
5182 case DIF_OP_SETX:
5183 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5184 break;
5185 case DIF_OP_SETS:
5186 regs[rd] = (uint64_t)(uintptr_t)
5187 (strtab + DIF_INSTR_STRING(instr));
5188 break;
5189 case DIF_OP_SCMP: {
5190 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5191 uintptr_t s1 = regs[r1];
5192 uintptr_t s2 = regs[r2];
5193
5194 if (s1 != NULL &&
5195 !dtrace_strcanload(s1, sz, mstate, vstate))
5196 break;
5197 if (s2 != NULL &&
5198 !dtrace_strcanload(s2, sz, mstate, vstate))
5199 break;
5200
5201 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5202
5203 cc_n = cc_r < 0;
5204 cc_z = cc_r == 0;
5205 cc_v = cc_c = 0;
5206 break;
5207 }
5208 case DIF_OP_LDGA:
5209 regs[rd] = dtrace_dif_variable(mstate, state,
5210 r1, regs[r2]);
5211 break;
5212 case DIF_OP_LDGS:
5213 id = DIF_INSTR_VAR(instr);
5214
5215 if (id >= DIF_VAR_OTHER_UBASE) {
5216 uintptr_t a;
5217
5218 id -= DIF_VAR_OTHER_UBASE;
5219 svar = vstate->dtvs_globals[id];
5220 ASSERT(svar != NULL);
5221 v = &svar->dtsv_var;
5222
5223 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5224 regs[rd] = svar->dtsv_data;
5225 break;
5226 }
5227
5228 a = (uintptr_t)svar->dtsv_data;
5229
5230 if (*(uint8_t *)a == UINT8_MAX) {
5231 /*
5232 * If the 0th byte is set to UINT8_MAX
5233 * then this is to be treated as a
5234 * reference to a NULL variable.
5235 */
5236 regs[rd] = NULL;
5237 } else {
5238 regs[rd] = a + sizeof (uint64_t);
5239 }
5240
5241 break;
5242 }
5243
5244 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5245 break;
5246
5247 case DIF_OP_STGS:
5248 id = DIF_INSTR_VAR(instr);
5249
5250 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5251 id -= DIF_VAR_OTHER_UBASE;
5252
5253 svar = vstate->dtvs_globals[id];
5254 ASSERT(svar != NULL);
5255 v = &svar->dtsv_var;
5256
5257 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5258 uintptr_t a = (uintptr_t)svar->dtsv_data;
5259
5260 ASSERT(a != NULL);
5261 ASSERT(svar->dtsv_size != 0);
5262
5263 if (regs[rd] == NULL) {
5264 *(uint8_t *)a = UINT8_MAX;
5265 break;
5266 } else {
5267 *(uint8_t *)a = 0;
5268 a += sizeof (uint64_t);
5269 }
5270 if (!dtrace_vcanload(
5271 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5272 mstate, vstate))
5273 break;
5274
5275 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5276 (void *)a, &v->dtdv_type);
5277 break;
5278 }
5279
5280 svar->dtsv_data = regs[rd];
5281 break;
5282
5283 case DIF_OP_LDTA:
5284 /*
5285 * There are no DTrace built-in thread-local arrays at
5286 * present. This opcode is saved for future work.
5287 */
5288 *flags |= CPU_DTRACE_ILLOP;
5289 regs[rd] = 0;
5290 break;
5291
5292 case DIF_OP_LDLS:
5293 id = DIF_INSTR_VAR(instr);
5294
5295 if (id < DIF_VAR_OTHER_UBASE) {
5296 /*
5297 * For now, this has no meaning.
5298 */
5299 regs[rd] = 0;
5300 break;
5301 }
5302
5303 id -= DIF_VAR_OTHER_UBASE;
5304
5305 ASSERT(id < vstate->dtvs_nlocals);
5306 ASSERT(vstate->dtvs_locals != NULL);
5307
5308 svar = vstate->dtvs_locals[id];
5309 ASSERT(svar != NULL);
5310 v = &svar->dtsv_var;
5311
5312 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5313 uintptr_t a = (uintptr_t)svar->dtsv_data;
5314 size_t sz = v->dtdv_type.dtdt_size;
5315
5316 sz += sizeof (uint64_t);
5317 ASSERT(svar->dtsv_size == NCPU * sz);
5318 a += CPU->cpu_id * sz;
5319
5320 if (*(uint8_t *)a == UINT8_MAX) {
5321 /*
5322 * If the 0th byte is set to UINT8_MAX
5323 * then this is to be treated as a
5324 * reference to a NULL variable.
5325 */
5326 regs[rd] = NULL;
5327 } else {
5328 regs[rd] = a + sizeof (uint64_t);
5329 }
5330
5331 break;
5332 }
5333
5334 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5335 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5336 regs[rd] = tmp[CPU->cpu_id];
5337 break;
5338
5339 case DIF_OP_STLS:
5340 id = DIF_INSTR_VAR(instr);
5341
5342 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5343 id -= DIF_VAR_OTHER_UBASE;
5344 ASSERT(id < vstate->dtvs_nlocals);
5345
5346 ASSERT(vstate->dtvs_locals != NULL);
5347 svar = vstate->dtvs_locals[id];
5348 ASSERT(svar != NULL);
5349 v = &svar->dtsv_var;
5350
5351 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5352 uintptr_t a = (uintptr_t)svar->dtsv_data;
5353 size_t sz = v->dtdv_type.dtdt_size;
5354
5355 sz += sizeof (uint64_t);
5356 ASSERT(svar->dtsv_size == NCPU * sz);
5357 a += CPU->cpu_id * sz;
5358
5359 if (regs[rd] == NULL) {
5360 *(uint8_t *)a = UINT8_MAX;
5361 break;
5362 } else {
5363 *(uint8_t *)a = 0;
5364 a += sizeof (uint64_t);
5365 }
5366
5367 if (!dtrace_vcanload(
5368 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5369 mstate, vstate))
5370 break;
5371
5372 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5373 (void *)a, &v->dtdv_type);
5374 break;
5375 }
5376
5377 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5378 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5379 tmp[CPU->cpu_id] = regs[rd];
5380 break;
5381
5382 case DIF_OP_LDTS: {
5383 dtrace_dynvar_t *dvar;
5384 dtrace_key_t *key;
5385
5386 id = DIF_INSTR_VAR(instr);
5387 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5388 id -= DIF_VAR_OTHER_UBASE;
5389 v = &vstate->dtvs_tlocals[id];
5390
5391 key = &tupregs[DIF_DTR_NREGS];
5392 key[0].dttk_value = (uint64_t)id;
5393 key[0].dttk_size = 0;
5394 DTRACE_TLS_THRKEY(key[1].dttk_value);
5395 key[1].dttk_size = 0;
5396
5397 dvar = dtrace_dynvar(dstate, 2, key,
5398 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5399 mstate, vstate);
5400
5401 if (dvar == NULL) {
5402 regs[rd] = 0;
5403 break;
5404 }
5405
5406 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5407 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5408 } else {
5409 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5410 }
5411
5412 break;
5413 }
5414
5415 case DIF_OP_STTS: {
5416 dtrace_dynvar_t *dvar;
5417 dtrace_key_t *key;
5418
5419 id = DIF_INSTR_VAR(instr);
5420 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5421 id -= DIF_VAR_OTHER_UBASE;
5422
5423 key = &tupregs[DIF_DTR_NREGS];
5424 key[0].dttk_value = (uint64_t)id;
5425 key[0].dttk_size = 0;
5426 DTRACE_TLS_THRKEY(key[1].dttk_value);
5427 key[1].dttk_size = 0;
5428 v = &vstate->dtvs_tlocals[id];
5429
5430 dvar = dtrace_dynvar(dstate, 2, key,
5431 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5432 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5433 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5434 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5435
5436 /*
5437 * Given that we're storing to thread-local data,
5438 * we need to flush our predicate cache.
5439 */
5440 curthread->t_predcache = NULL;
5441
5442 if (dvar == NULL)
5443 break;
5444
5445 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5446 if (!dtrace_vcanload(
5447 (void *)(uintptr_t)regs[rd],
5448 &v->dtdv_type, mstate, vstate))
5449 break;
5450
5451 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5452 dvar->dtdv_data, &v->dtdv_type);
5453 } else {
5454 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5455 }
5456
5457 break;
5458 }
5459
5460 case DIF_OP_SRA:
5461 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5462 break;
5463
5464 case DIF_OP_CALL:
5465 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5466 regs, tupregs, ttop, mstate, state);
5467 break;
5468
5469 case DIF_OP_PUSHTR:
5470 if (ttop == DIF_DTR_NREGS) {
5471 *flags |= CPU_DTRACE_TUPOFLOW;
5472 break;
5473 }
5474
5475 if (r1 == DIF_TYPE_STRING) {
5476 /*
5477 * If this is a string type and the size is 0,
5478 * we'll use the system-wide default string
5479 * size. Note that we are _not_ looking at
5480 * the value of the DTRACEOPT_STRSIZE option;
5481 * had this been set, we would expect to have
5482 * a non-zero size value in the "pushtr".
5483 */
5484 tupregs[ttop].dttk_size =
5485 dtrace_strlen((char *)(uintptr_t)regs[rd],
5486 regs[r2] ? regs[r2] :
5487 dtrace_strsize_default) + 1;
5488 } else {
5489 tupregs[ttop].dttk_size = regs[r2];
5490 }
5491
5492 tupregs[ttop++].dttk_value = regs[rd];
5493 break;
5494
5495 case DIF_OP_PUSHTV:
5496 if (ttop == DIF_DTR_NREGS) {
5497 *flags |= CPU_DTRACE_TUPOFLOW;
5498 break;
5499 }
5500
5501 tupregs[ttop].dttk_value = regs[rd];
5502 tupregs[ttop++].dttk_size = 0;
5503 break;
5504
5505 case DIF_OP_POPTS:
5506 if (ttop != 0)
5507 ttop--;
5508 break;
5509
5510 case DIF_OP_FLUSHTS:
5511 ttop = 0;
5512 break;
5513
5514 case DIF_OP_LDGAA:
5515 case DIF_OP_LDTAA: {
5516 dtrace_dynvar_t *dvar;
5517 dtrace_key_t *key = tupregs;
5518 uint_t nkeys = ttop;
5519
5520 id = DIF_INSTR_VAR(instr);
5521 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5522 id -= DIF_VAR_OTHER_UBASE;
5523
5524 key[nkeys].dttk_value = (uint64_t)id;
5525 key[nkeys++].dttk_size = 0;
5526
5527 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5528 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5529 key[nkeys++].dttk_size = 0;
5530 v = &vstate->dtvs_tlocals[id];
5531 } else {
5532 v = &vstate->dtvs_globals[id]->dtsv_var;
5533 }
5534
5535 dvar = dtrace_dynvar(dstate, nkeys, key,
5536 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5537 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5538 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5539
5540 if (dvar == NULL) {
5541 regs[rd] = 0;
5542 break;
5543 }
5544
5545 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5546 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5547 } else {
5548 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5549 }
5550
5551 break;
5552 }
5553
5554 case DIF_OP_STGAA:
5555 case DIF_OP_STTAA: {
5556 dtrace_dynvar_t *dvar;
5557 dtrace_key_t *key = tupregs;
5558 uint_t nkeys = ttop;
5559
5560 id = DIF_INSTR_VAR(instr);
5561 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5562 id -= DIF_VAR_OTHER_UBASE;
5563
5564 key[nkeys].dttk_value = (uint64_t)id;
5565 key[nkeys++].dttk_size = 0;
5566
5567 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5568 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5569 key[nkeys++].dttk_size = 0;
5570 v = &vstate->dtvs_tlocals[id];
5571 } else {
5572 v = &vstate->dtvs_globals[id]->dtsv_var;
5573 }
5574
5575 dvar = dtrace_dynvar(dstate, nkeys, key,
5576 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5577 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5578 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5579 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5580
5581 if (dvar == NULL)
5582 break;
5583
5584 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5585 if (!dtrace_vcanload(
5586 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5587 mstate, vstate))
5588 break;
5589
5590 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5591 dvar->dtdv_data, &v->dtdv_type);
5592 } else {
5593 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5594 }
5595
5596 break;
5597 }
5598
5599 case DIF_OP_ALLOCS: {
5600 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5601 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5602
5603 /*
5604 * Rounding up the user allocation size could have
5605 * overflowed large, bogus allocations (like -1ULL) to
5606 * 0.
5607 */
5608 if (size < regs[r1] ||
5609 !DTRACE_INSCRATCH(mstate, size)) {
5610 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5611 regs[rd] = NULL;
5612 break;
5613 }
5614
5615 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5616 mstate->dtms_scratch_ptr += size;
5617 regs[rd] = ptr;
5618 break;
5619 }
5620
5621 case DIF_OP_COPYS:
5622 if (!dtrace_canstore(regs[rd], regs[r2],
5623 mstate, vstate)) {
5624 *flags |= CPU_DTRACE_BADADDR;
5625 *illval = regs[rd];
5626 break;
5627 }
5628
5629 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5630 break;
5631
5632 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5633 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5634 break;
5635
5636 case DIF_OP_STB:
5637 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5638 *flags |= CPU_DTRACE_BADADDR;
5639 *illval = regs[rd];
5640 break;
5641 }
5642 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5643 break;
5644
5645 case DIF_OP_STH:
5646 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5647 *flags |= CPU_DTRACE_BADADDR;
5648 *illval = regs[rd];
5649 break;
5650 }
5651 if (regs[rd] & 1) {
5652 *flags |= CPU_DTRACE_BADALIGN;
5653 *illval = regs[rd];
5654 break;
5655 }
5656 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5657 break;
5658
5659 case DIF_OP_STW:
5660 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5661 *flags |= CPU_DTRACE_BADADDR;
5662 *illval = regs[rd];
5663 break;
5664 }
5665 if (regs[rd] & 3) {
5666 *flags |= CPU_DTRACE_BADALIGN;
5667 *illval = regs[rd];
5668 break;
5669 }
5670 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5671 break;
5672
5673 case DIF_OP_STX:
5674 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5675 *flags |= CPU_DTRACE_BADADDR;
5676 *illval = regs[rd];
5677 break;
5678 }
5679 if (regs[rd] & 7) {
5680 *flags |= CPU_DTRACE_BADALIGN;
5681 *illval = regs[rd];
5682 break;
5683 }
5684 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5685 break;
5686 }
5687 }
5688
5689 if (!(*flags & CPU_DTRACE_FAULT))
5690 return (rval);
5691
5692 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5693 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5694
5695 return (0);
5696 }
5697
5698 static void
5699 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5700 {
5701 dtrace_probe_t *probe = ecb->dte_probe;
5702 dtrace_provider_t *prov = probe->dtpr_provider;
5703 char c[DTRACE_FULLNAMELEN + 80], *str;
5704 char *msg = "dtrace: breakpoint action at probe ";
5705 char *ecbmsg = " (ecb ";
5706 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5707 uintptr_t val = (uintptr_t)ecb;
5708 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5709
5710 if (dtrace_destructive_disallow)
5711 return;
5712
5713 /*
5714 * It's impossible to be taking action on the NULL probe.
5715 */
5716 ASSERT(probe != NULL);
5717
5718 /*
5719 * This is a poor man's (destitute man's?) sprintf(): we want to
5720 * print the provider name, module name, function name and name of
5721 * the probe, along with the hex address of the ECB with the breakpoint
5722 * action -- all of which we must place in the character buffer by
5723 * hand.
5724 */
5725 while (*msg != '\0')
5726 c[i++] = *msg++;
5727
5728 for (str = prov->dtpv_name; *str != '\0'; str++)
5729 c[i++] = *str;
5730 c[i++] = ':';
5731
5732 for (str = probe->dtpr_mod; *str != '\0'; str++)
5733 c[i++] = *str;
5734 c[i++] = ':';
5735
5736 for (str = probe->dtpr_func; *str != '\0'; str++)
5737 c[i++] = *str;
5738 c[i++] = ':';
5739
5740 for (str = probe->dtpr_name; *str != '\0'; str++)
5741 c[i++] = *str;
5742
5743 while (*ecbmsg != '\0')
5744 c[i++] = *ecbmsg++;
5745
5746 while (shift >= 0) {
5747 mask = (uintptr_t)0xf << shift;
5748
5749 if (val >= ((uintptr_t)1 << shift))
5750 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5751 shift -= 4;
5752 }
5753
5754 c[i++] = ')';
5755 c[i] = '\0';
5756
5757 debug_enter(c);
5758 }
5759
5760 static void
5761 dtrace_action_panic(dtrace_ecb_t *ecb)
5762 {
5763 dtrace_probe_t *probe = ecb->dte_probe;
5764
5765 /*
5766 * It's impossible to be taking action on the NULL probe.
5767 */
5768 ASSERT(probe != NULL);
5769
5770 if (dtrace_destructive_disallow)
5771 return;
5772
5773 if (dtrace_panicked != NULL)
5774 return;
5775
5776 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5777 return;
5778
5779 /*
5780 * We won the right to panic. (We want to be sure that only one
5781 * thread calls panic() from dtrace_probe(), and that panic() is
5782 * called exactly once.)
5783 */
5784 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5785 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5786 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5787 }
5788
5789 static void
5790 dtrace_action_raise(uint64_t sig)
5791 {
5792 if (dtrace_destructive_disallow)
5793 return;
5794
5795 if (sig >= NSIG) {
5796 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5797 return;
5798 }
5799
5800 /*
5801 * raise() has a queue depth of 1 -- we ignore all subsequent
5802 * invocations of the raise() action.
5803 */
5804 if (curthread->t_dtrace_sig == 0)
5805 curthread->t_dtrace_sig = (uint8_t)sig;
5806
5807 curthread->t_sig_check = 1;
5808 aston(curthread);
5809 }
5810
5811 static void
5812 dtrace_action_stop(void)
5813 {
5814 if (dtrace_destructive_disallow)
5815 return;
5816
5817 if (!curthread->t_dtrace_stop) {
5818 curthread->t_dtrace_stop = 1;
5819 curthread->t_sig_check = 1;
5820 aston(curthread);
5821 }
5822 }
5823
5824 static void
5825 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5826 {
5827 hrtime_t now;
5828 volatile uint16_t *flags;
5829 cpu_t *cpu = CPU;
5830
5831 if (dtrace_destructive_disallow)
5832 return;
5833
5834 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5835
5836 now = dtrace_gethrtime();
5837
5838 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5839 /*
5840 * We need to advance the mark to the current time.
5841 */
5842 cpu->cpu_dtrace_chillmark = now;
5843 cpu->cpu_dtrace_chilled = 0;
5844 }
5845
5846 /*
5847 * Now check to see if the requested chill time would take us over
5848 * the maximum amount of time allowed in the chill interval. (Or
5849 * worse, if the calculation itself induces overflow.)
5850 */
5851 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5852 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5853 *flags |= CPU_DTRACE_ILLOP;
5854 return;
5855 }
5856
5857 while (dtrace_gethrtime() - now < val)
5858 continue;
5859
5860 /*
5861 * Normally, we assure that the value of the variable "timestamp" does
5862 * not change within an ECB. The presence of chill() represents an
5863 * exception to this rule, however.
5864 */
5865 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5866 cpu->cpu_dtrace_chilled += val;
5867 }
5868
5869 static void
5870 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5871 uint64_t *buf, uint64_t arg)
5872 {
5873 int nframes = DTRACE_USTACK_NFRAMES(arg);
5874 int strsize = DTRACE_USTACK_STRSIZE(arg);
5875 uint64_t *pcs = &buf[1], *fps;
5876 char *str = (char *)&pcs[nframes];
5877 int size, offs = 0, i, j;
5878 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5879 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5880 char *sym;
5881
5882 /*
5883 * Should be taking a faster path if string space has not been
5884 * allocated.
5885 */
5886 ASSERT(strsize != 0);
5887
5888 /*
5889 * We will first allocate some temporary space for the frame pointers.
5890 */
5891 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5892 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5893 (nframes * sizeof (uint64_t));
5894
5895 if (!DTRACE_INSCRATCH(mstate, size)) {
5896 /*
5897 * Not enough room for our frame pointers -- need to indicate
5898 * that we ran out of scratch space.
5899 */
5900 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5901 return;
5902 }
5903
5904 mstate->dtms_scratch_ptr += size;
5905 saved = mstate->dtms_scratch_ptr;
5906
5907 /*
5908 * Now get a stack with both program counters and frame pointers.
5909 */
5910 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5911 dtrace_getufpstack(buf, fps, nframes + 1);
5912 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5913
5914 /*
5915 * If that faulted, we're cooked.
5916 */
5917 if (*flags & CPU_DTRACE_FAULT)
5918 goto out;
5919
5920 /*
5921 * Now we want to walk up the stack, calling the USTACK helper. For
5922 * each iteration, we restore the scratch pointer.
5923 */
5924 for (i = 0; i < nframes; i++) {
5925 mstate->dtms_scratch_ptr = saved;
5926
5927 if (offs >= strsize)
5928 break;
5929
5930 sym = (char *)(uintptr_t)dtrace_helper(
5931 DTRACE_HELPER_ACTION_USTACK,
5932 mstate, state, pcs[i], fps[i]);
5933
5934 /*
5935 * If we faulted while running the helper, we're going to
5936 * clear the fault and null out the corresponding string.
5937 */
5938 if (*flags & CPU_DTRACE_FAULT) {
5939 *flags &= ~CPU_DTRACE_FAULT;
5940 str[offs++] = '\0';
5941 continue;
5942 }
5943
5944 if (sym == NULL) {
5945 str[offs++] = '\0';
5946 continue;
5947 }
5948
5949 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5950
5951 /*
5952 * Now copy in the string that the helper returned to us.
5953 */
5954 for (j = 0; offs + j < strsize; j++) {
5955 if ((str[offs + j] = sym[j]) == '\0')
5956 break;
5957 }
5958
5959 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5960
5961 offs += j + 1;
5962 }
5963
5964 if (offs >= strsize) {
5965 /*
5966 * If we didn't have room for all of the strings, we don't
5967 * abort processing -- this needn't be a fatal error -- but we
5968 * still want to increment a counter (dts_stkstroverflows) to
5969 * allow this condition to be warned about. (If this is from
5970 * a jstack() action, it is easily tuned via jstackstrsize.)
5971 */
5972 dtrace_error(&state->dts_stkstroverflows);
5973 }
5974
5975 while (offs < strsize)
5976 str[offs++] = '\0';
5977
5978 out:
5979 mstate->dtms_scratch_ptr = old;
5980 }
5981
5982 /*
5983 * If you're looking for the epicenter of DTrace, you just found it. This
5984 * is the function called by the provider to fire a probe -- from which all
5985 * subsequent probe-context DTrace activity emanates.
5986 */
5987 void
5988 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5989 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5990 {
5991 processorid_t cpuid;
5992 dtrace_icookie_t cookie;
5993 dtrace_probe_t *probe;
5994 dtrace_mstate_t mstate;
5995 dtrace_ecb_t *ecb;
5996 dtrace_action_t *act;
5997 intptr_t offs;
5998 size_t size;
5999 int vtime, onintr;
6000 volatile uint16_t *flags;
6001 hrtime_t now;
6002
6003 /*
6004 * Kick out immediately if this CPU is still being born (in which case
6005 * curthread will be set to -1) or the current thread can't allow
6006 * probes in its current context.
6007 */
6008 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6009 return;
6010
6011 cookie = dtrace_interrupt_disable();
6012 probe = dtrace_probes[id - 1];
6013 cpuid = CPU->cpu_id;
6014 onintr = CPU_ON_INTR(CPU);
6015
6016 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6017 probe->dtpr_predcache == curthread->t_predcache) {
6018 /*
6019 * We have hit in the predicate cache; we know that
6020 * this predicate would evaluate to be false.
6021 */
6022 dtrace_interrupt_enable(cookie);
6023 return;
6024 }
6025
6026 if (panic_quiesce) {
6027 /*
6028 * We don't trace anything if we're panicking.
6029 */
6030 dtrace_interrupt_enable(cookie);
6031 return;
6032 }
6033
6034 now = dtrace_gethrtime();
6035 vtime = dtrace_vtime_references != 0;
6036
6037 if (vtime && curthread->t_dtrace_start)
6038 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6039
6040 mstate.dtms_difo = NULL;
6041 mstate.dtms_probe = probe;
6042 mstate.dtms_strtok = NULL;
6043 mstate.dtms_arg[0] = arg0;
6044 mstate.dtms_arg[1] = arg1;
6045 mstate.dtms_arg[2] = arg2;
6046 mstate.dtms_arg[3] = arg3;
6047 mstate.dtms_arg[4] = arg4;
6048
6049 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6050
6051 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6052 dtrace_predicate_t *pred = ecb->dte_predicate;
6053 dtrace_state_t *state = ecb->dte_state;
6054 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6055 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6056 dtrace_vstate_t *vstate = &state->dts_vstate;
6057 dtrace_provider_t *prov = probe->dtpr_provider;
6058 uint64_t tracememsize = 0;
6059 int committed = 0;
6060 caddr_t tomax;
6061
6062 /*
6063 * A little subtlety with the following (seemingly innocuous)
6064 * declaration of the automatic 'val': by looking at the
6065 * code, you might think that it could be declared in the
6066 * action processing loop, below. (That is, it's only used in
6067 * the action processing loop.) However, it must be declared
6068 * out of that scope because in the case of DIF expression
6069 * arguments to aggregating actions, one iteration of the
6070 * action loop will use the last iteration's value.
6071 */
6072 #ifdef lint
6073 uint64_t val = 0;
6074 #else
6075 uint64_t val;
6076 #endif
6077
6078 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6079 mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
6080 mstate.dtms_getf = NULL;
6081
6082 *flags &= ~CPU_DTRACE_ERROR;
6083
6084 if (prov == dtrace_provider) {
6085 /*
6086 * If dtrace itself is the provider of this probe,
6087 * we're only going to continue processing the ECB if
6088 * arg0 (the dtrace_state_t) is equal to the ECB's
6089 * creating state. (This prevents disjoint consumers
6090 * from seeing one another's metaprobes.)
6091 */
6092 if (arg0 != (uint64_t)(uintptr_t)state)
6093 continue;
6094 }
6095
6096 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6097 /*
6098 * We're not currently active. If our provider isn't
6099 * the dtrace pseudo provider, we're not interested.
6100 */
6101 if (prov != dtrace_provider)
6102 continue;
6103
6104 /*
6105 * Now we must further check if we are in the BEGIN
6106 * probe. If we are, we will only continue processing
6107 * if we're still in WARMUP -- if one BEGIN enabling
6108 * has invoked the exit() action, we don't want to
6109 * evaluate subsequent BEGIN enablings.
6110 */
6111 if (probe->dtpr_id == dtrace_probeid_begin &&
6112 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6113 ASSERT(state->dts_activity ==
6114 DTRACE_ACTIVITY_DRAINING);
6115 continue;
6116 }
6117 }
6118
6119 if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
6120 continue;
6121
6122 if (now - state->dts_alive > dtrace_deadman_timeout) {
6123 /*
6124 * We seem to be dead. Unless we (a) have kernel
6125 * destructive permissions (b) have explicitly enabled
6126 * destructive actions and (c) destructive actions have
6127 * not been disabled, we're going to transition into
6128 * the KILLED state, from which no further processing
6129 * on this state will be performed.
6130 */
6131 if (!dtrace_priv_kernel_destructive(state) ||
6132 !state->dts_cred.dcr_destructive ||
6133 dtrace_destructive_disallow) {
6134 void *activity = &state->dts_activity;
6135 dtrace_activity_t current;
6136
6137 do {
6138 current = state->dts_activity;
6139 } while (dtrace_cas32(activity, current,
6140 DTRACE_ACTIVITY_KILLED) != current);
6141
6142 continue;
6143 }
6144 }
6145
6146 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6147 ecb->dte_alignment, state, &mstate)) < 0)
6148 continue;
6149
6150 tomax = buf->dtb_tomax;
6151 ASSERT(tomax != NULL);
6152
6153 if (ecb->dte_size != 0) {
6154 dtrace_rechdr_t dtrh;
6155 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6156 mstate.dtms_timestamp = dtrace_gethrtime();
6157 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6158 }
6159 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6160 dtrh.dtrh_epid = ecb->dte_epid;
6161 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6162 mstate.dtms_timestamp);
6163 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6164 }
6165
6166 mstate.dtms_epid = ecb->dte_epid;
6167 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6168
6169 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6170 mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
6171
6172 if (pred != NULL) {
6173 dtrace_difo_t *dp = pred->dtp_difo;
6174 int rval;
6175
6176 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6177
6178 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6179 dtrace_cacheid_t cid = probe->dtpr_predcache;
6180
6181 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6182 /*
6183 * Update the predicate cache...
6184 */
6185 ASSERT(cid == pred->dtp_cacheid);
6186 curthread->t_predcache = cid;
6187 }
6188
6189 continue;
6190 }
6191 }
6192
6193 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6194 act != NULL; act = act->dta_next) {
6195 size_t valoffs;
6196 dtrace_difo_t *dp;
6197 dtrace_recdesc_t *rec = &act->dta_rec;
6198
6199 size = rec->dtrd_size;
6200 valoffs = offs + rec->dtrd_offset;
6201
6202 if (DTRACEACT_ISAGG(act->dta_kind)) {
6203 uint64_t v = 0xbad;
6204 dtrace_aggregation_t *agg;
6205
6206 agg = (dtrace_aggregation_t *)act;
6207
6208 if ((dp = act->dta_difo) != NULL)
6209 v = dtrace_dif_emulate(dp,
6210 &mstate, vstate, state);
6211
6212 if (*flags & CPU_DTRACE_ERROR)
6213 continue;
6214
6215 /*
6216 * Note that we always pass the expression
6217 * value from the previous iteration of the
6218 * action loop. This value will only be used
6219 * if there is an expression argument to the
6220 * aggregating action, denoted by the
6221 * dtag_hasarg field.
6222 */
6223 dtrace_aggregate(agg, buf,
6224 offs, aggbuf, v, val);
6225 continue;
6226 }
6227
6228 switch (act->dta_kind) {
6229 case DTRACEACT_STOP:
6230 if (dtrace_priv_proc_destructive(state,
6231 &mstate))
6232 dtrace_action_stop();
6233 continue;
6234
6235 case DTRACEACT_BREAKPOINT:
6236 if (dtrace_priv_kernel_destructive(state))
6237 dtrace_action_breakpoint(ecb);
6238 continue;
6239
6240 case DTRACEACT_PANIC:
6241 if (dtrace_priv_kernel_destructive(state))
6242 dtrace_action_panic(ecb);
6243 continue;
6244
6245 case DTRACEACT_STACK:
6246 if (!dtrace_priv_kernel(state))
6247 continue;
6248
6249 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6250 size / sizeof (pc_t), probe->dtpr_aframes,
6251 DTRACE_ANCHORED(probe) ? NULL :
6252 (uint32_t *)arg0);
6253
6254 continue;
6255
6256 case DTRACEACT_JSTACK:
6257 case DTRACEACT_USTACK:
6258 if (!dtrace_priv_proc(state, &mstate))
6259 continue;
6260
6261 /*
6262 * See comment in DIF_VAR_PID.
6263 */
6264 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6265 CPU_ON_INTR(CPU)) {
6266 int depth = DTRACE_USTACK_NFRAMES(
6267 rec->dtrd_arg) + 1;
6268
6269 dtrace_bzero((void *)(tomax + valoffs),
6270 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6271 + depth * sizeof (uint64_t));
6272
6273 continue;
6274 }
6275
6276 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6277 curproc->p_dtrace_helpers != NULL) {
6278 /*
6279 * This is the slow path -- we have
6280 * allocated string space, and we're
6281 * getting the stack of a process that
6282 * has helpers. Call into a separate
6283 * routine to perform this processing.
6284 */
6285 dtrace_action_ustack(&mstate, state,
6286 (uint64_t *)(tomax + valoffs),
6287 rec->dtrd_arg);
6288 continue;
6289 }
6290
6291 /*
6292 * Clear the string space, since there's no
6293 * helper to do it for us.
6294 */
6295 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
6296 int depth = DTRACE_USTACK_NFRAMES(
6297 rec->dtrd_arg);
6298 size_t strsize = DTRACE_USTACK_STRSIZE(
6299 rec->dtrd_arg);
6300 uint64_t *buf = (uint64_t *)(tomax +
6301 valoffs);
6302 void *strspace = &buf[depth + 1];
6303
6304 dtrace_bzero(strspace,
6305 MIN(depth, strsize));
6306 }
6307
6308 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6309 dtrace_getupcstack((uint64_t *)
6310 (tomax + valoffs),
6311 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6312 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6313 continue;
6314
6315 default:
6316 break;
6317 }
6318
6319 dp = act->dta_difo;
6320 ASSERT(dp != NULL);
6321
6322 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6323
6324 if (*flags & CPU_DTRACE_ERROR)
6325 continue;
6326
6327 switch (act->dta_kind) {
6328 case DTRACEACT_SPECULATE: {
6329 dtrace_rechdr_t *dtrh;
6330
6331 ASSERT(buf == &state->dts_buffer[cpuid]);
6332 buf = dtrace_speculation_buffer(state,
6333 cpuid, val);
6334
6335 if (buf == NULL) {
6336 *flags |= CPU_DTRACE_DROP;
6337 continue;
6338 }
6339
6340 offs = dtrace_buffer_reserve(buf,
6341 ecb->dte_needed, ecb->dte_alignment,
6342 state, NULL);
6343
6344 if (offs < 0) {
6345 *flags |= CPU_DTRACE_DROP;
6346 continue;
6347 }
6348
6349 tomax = buf->dtb_tomax;
6350 ASSERT(tomax != NULL);
6351
6352 if (ecb->dte_size == 0)
6353 continue;
6354
6355 ASSERT3U(ecb->dte_size, >=,
6356 sizeof (dtrace_rechdr_t));
6357 dtrh = ((void *)(tomax + offs));
6358 dtrh->dtrh_epid = ecb->dte_epid;
6359 /*
6360 * When the speculation is committed, all of
6361 * the records in the speculative buffer will
6362 * have their timestamps set to the commit
6363 * time. Until then, it is set to a sentinel
6364 * value, for debugability.
6365 */
6366 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6367 continue;
6368 }
6369
6370 case DTRACEACT_CHILL:
6371 if (dtrace_priv_kernel_destructive(state))
6372 dtrace_action_chill(&mstate, val);
6373 continue;
6374
6375 case DTRACEACT_RAISE:
6376 if (dtrace_priv_proc_destructive(state,
6377 &mstate))
6378 dtrace_action_raise(val);
6379 continue;
6380
6381 case DTRACEACT_COMMIT:
6382 ASSERT(!committed);
6383
6384 /*
6385 * We need to commit our buffer state.
6386 */
6387 if (ecb->dte_size)
6388 buf->dtb_offset = offs + ecb->dte_size;
6389 buf = &state->dts_buffer[cpuid];
6390 dtrace_speculation_commit(state, cpuid, val);
6391 committed = 1;
6392 continue;
6393
6394 case DTRACEACT_DISCARD:
6395 dtrace_speculation_discard(state, cpuid, val);
6396 continue;
6397
6398 case DTRACEACT_DIFEXPR:
6399 case DTRACEACT_LIBACT:
6400 case DTRACEACT_PRINTF:
6401 case DTRACEACT_PRINTA:
6402 case DTRACEACT_SYSTEM:
6403 case DTRACEACT_FREOPEN:
6404 case DTRACEACT_TRACEMEM:
6405 break;
6406
6407 case DTRACEACT_TRACEMEM_DYNSIZE:
6408 tracememsize = val;
6409 break;
6410
6411 case DTRACEACT_SYM:
6412 case DTRACEACT_MOD:
6413 if (!dtrace_priv_kernel(state))
6414 continue;
6415 break;
6416
6417 case DTRACEACT_USYM:
6418 case DTRACEACT_UMOD:
6419 case DTRACEACT_UADDR: {
6420 struct pid *pid = curthread->t_procp->p_pidp;
6421
6422 if (!dtrace_priv_proc(state, &mstate))
6423 continue;
6424
6425 DTRACE_STORE(uint64_t, tomax,
6426 valoffs, (uint64_t)pid->pid_id);
6427 DTRACE_STORE(uint64_t, tomax,
6428 valoffs + sizeof (uint64_t), val);
6429
6430 continue;
6431 }
6432
6433 case DTRACEACT_EXIT: {
6434 /*
6435 * For the exit action, we are going to attempt
6436 * to atomically set our activity to be
6437 * draining. If this fails (either because
6438 * another CPU has beat us to the exit action,
6439 * or because our current activity is something
6440 * other than ACTIVE or WARMUP), we will
6441 * continue. This assures that the exit action
6442 * can be successfully recorded at most once
6443 * when we're in the ACTIVE state. If we're
6444 * encountering the exit() action while in
6445 * COOLDOWN, however, we want to honor the new
6446 * status code. (We know that we're the only
6447 * thread in COOLDOWN, so there is no race.)
6448 */
6449 void *activity = &state->dts_activity;
6450 dtrace_activity_t current = state->dts_activity;
6451
6452 if (current == DTRACE_ACTIVITY_COOLDOWN)
6453 break;
6454
6455 if (current != DTRACE_ACTIVITY_WARMUP)
6456 current = DTRACE_ACTIVITY_ACTIVE;
6457
6458 if (dtrace_cas32(activity, current,
6459 DTRACE_ACTIVITY_DRAINING) != current) {
6460 *flags |= CPU_DTRACE_DROP;
6461 continue;
6462 }
6463
6464 break;
6465 }
6466
6467 default:
6468 ASSERT(0);
6469 }
6470
6471 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6472 uintptr_t end = valoffs + size;
6473
6474 if (tracememsize != 0 &&
6475 valoffs + tracememsize < end) {
6476 end = valoffs + tracememsize;
6477 tracememsize = 0;
6478 }
6479
6480 if (!dtrace_vcanload((void *)(uintptr_t)val,
6481 &dp->dtdo_rtype, &mstate, vstate))
6482 continue;
6483
6484 /*
6485 * If this is a string, we're going to only
6486 * load until we find the zero byte -- after
6487 * which we'll store zero bytes.
6488 */
6489 if (dp->dtdo_rtype.dtdt_kind ==
6490 DIF_TYPE_STRING) {
6491 char c = '\0' + 1;
6492 int intuple = act->dta_intuple;
6493 size_t s;
6494
6495 for (s = 0; s < size; s++) {
6496 if (c != '\0')
6497 c = dtrace_load8(val++);
6498
6499 DTRACE_STORE(uint8_t, tomax,
6500 valoffs++, c);
6501
6502 if (c == '\0' && intuple)
6503 break;
6504 }
6505
6506 continue;
6507 }
6508
6509 while (valoffs < end) {
6510 DTRACE_STORE(uint8_t, tomax, valoffs++,
6511 dtrace_load8(val++));
6512 }
6513
6514 continue;
6515 }
6516
6517 switch (size) {
6518 case 0:
6519 break;
6520
6521 case sizeof (uint8_t):
6522 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6523 break;
6524 case sizeof (uint16_t):
6525 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6526 break;
6527 case sizeof (uint32_t):
6528 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6529 break;
6530 case sizeof (uint64_t):
6531 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6532 break;
6533 default:
6534 /*
6535 * Any other size should have been returned by
6536 * reference, not by value.
6537 */
6538 ASSERT(0);
6539 break;
6540 }
6541 }
6542
6543 if (*flags & CPU_DTRACE_DROP)
6544 continue;
6545
6546 if (*flags & CPU_DTRACE_FAULT) {
6547 int ndx;
6548 dtrace_action_t *err;
6549
6550 buf->dtb_errors++;
6551
6552 if (probe->dtpr_id == dtrace_probeid_error) {
6553 /*
6554 * There's nothing we can do -- we had an
6555 * error on the error probe. We bump an
6556 * error counter to at least indicate that
6557 * this condition happened.
6558 */
6559 dtrace_error(&state->dts_dblerrors);
6560 continue;
6561 }
6562
6563 if (vtime) {
6564 /*
6565 * Before recursing on dtrace_probe(), we
6566 * need to explicitly clear out our start
6567 * time to prevent it from being accumulated
6568 * into t_dtrace_vtime.
6569 */
6570 curthread->t_dtrace_start = 0;
6571 }
6572
6573 /*
6574 * Iterate over the actions to figure out which action
6575 * we were processing when we experienced the error.
6576 * Note that act points _past_ the faulting action; if
6577 * act is ecb->dte_action, the fault was in the
6578 * predicate, if it's ecb->dte_action->dta_next it's
6579 * in action #1, and so on.
6580 */
6581 for (err = ecb->dte_action, ndx = 0;
6582 err != act; err = err->dta_next, ndx++)
6583 continue;
6584
6585 dtrace_probe_error(state, ecb->dte_epid, ndx,
6586 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6587 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6588 cpu_core[cpuid].cpuc_dtrace_illval);
6589
6590 continue;
6591 }
6592
6593 if (!committed)
6594 buf->dtb_offset = offs + ecb->dte_size;
6595 }
6596
6597 if (vtime)
6598 curthread->t_dtrace_start = dtrace_gethrtime();
6599
6600 dtrace_interrupt_enable(cookie);
6601 }
6602
6603 /*
6604 * DTrace Probe Hashing Functions
6605 *
6606 * The functions in this section (and indeed, the functions in remaining
6607 * sections) are not _called_ from probe context. (Any exceptions to this are
6608 * marked with a "Note:".) Rather, they are called from elsewhere in the
6609 * DTrace framework to look-up probes in, add probes to and remove probes from
6610 * the DTrace probe hashes. (Each probe is hashed by each element of the
6611 * probe tuple -- allowing for fast lookups, regardless of what was
6612 * specified.)
6613 */
6614 static uint_t
6615 dtrace_hash_str(char *p)
6616 {
6617 unsigned int g;
6618 uint_t hval = 0;
6619
6620 while (*p) {
6621 hval = (hval << 4) + *p++;
6622 if ((g = (hval & 0xf0000000)) != 0)
6623 hval ^= g >> 24;
6624 hval &= ~g;
6625 }
6626 return (hval);
6627 }
6628
6629 static dtrace_hash_t *
6630 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6631 {
6632 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6633
6634 hash->dth_stroffs = stroffs;
6635 hash->dth_nextoffs = nextoffs;
6636 hash->dth_prevoffs = prevoffs;
6637
6638 hash->dth_size = 1;
6639 hash->dth_mask = hash->dth_size - 1;
6640
6641 hash->dth_tab = kmem_zalloc(hash->dth_size *
6642 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6643
6644 return (hash);
6645 }
6646
6647 static void
6648 dtrace_hash_destroy(dtrace_hash_t *hash)
6649 {
6650 #ifdef DEBUG
6651 int i;
6652
6653 for (i = 0; i < hash->dth_size; i++)
6654 ASSERT(hash->dth_tab[i] == NULL);
6655 #endif
6656
6657 kmem_free(hash->dth_tab,
6658 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6659 kmem_free(hash, sizeof (dtrace_hash_t));
6660 }
6661
6662 static void
6663 dtrace_hash_resize(dtrace_hash_t *hash)
6664 {
6665 int size = hash->dth_size, i, ndx;
6666 int new_size = hash->dth_size << 1;
6667 int new_mask = new_size - 1;
6668 dtrace_hashbucket_t **new_tab, *bucket, *next;
6669
6670 ASSERT((new_size & new_mask) == 0);
6671
6672 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6673
6674 for (i = 0; i < size; i++) {
6675 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6676 dtrace_probe_t *probe = bucket->dthb_chain;
6677
6678 ASSERT(probe != NULL);
6679 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6680
6681 next = bucket->dthb_next;
6682 bucket->dthb_next = new_tab[ndx];
6683 new_tab[ndx] = bucket;
6684 }
6685 }
6686
6687 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6688 hash->dth_tab = new_tab;
6689 hash->dth_size = new_size;
6690 hash->dth_mask = new_mask;
6691 }
6692
6693 static void
6694 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6695 {
6696 int hashval = DTRACE_HASHSTR(hash, new);
6697 int ndx = hashval & hash->dth_mask;
6698 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6699 dtrace_probe_t **nextp, **prevp;
6700
6701 for (; bucket != NULL; bucket = bucket->dthb_next) {
6702 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6703 goto add;
6704 }
6705
6706 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6707 dtrace_hash_resize(hash);
6708 dtrace_hash_add(hash, new);
6709 return;
6710 }
6711
6712 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6713 bucket->dthb_next = hash->dth_tab[ndx];
6714 hash->dth_tab[ndx] = bucket;
6715 hash->dth_nbuckets++;
6716
6717 add:
6718 nextp = DTRACE_HASHNEXT(hash, new);
6719 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6720 *nextp = bucket->dthb_chain;
6721
6722 if (bucket->dthb_chain != NULL) {
6723 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6724 ASSERT(*prevp == NULL);
6725 *prevp = new;
6726 }
6727
6728 bucket->dthb_chain = new;
6729 bucket->dthb_len++;
6730 }
6731
6732 static dtrace_probe_t *
6733 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6734 {
6735 int hashval = DTRACE_HASHSTR(hash, template);
6736 int ndx = hashval & hash->dth_mask;
6737 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6738
6739 for (; bucket != NULL; bucket = bucket->dthb_next) {
6740 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6741 return (bucket->dthb_chain);
6742 }
6743
6744 return (NULL);
6745 }
6746
6747 static int
6748 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6749 {
6750 int hashval = DTRACE_HASHSTR(hash, template);
6751 int ndx = hashval & hash->dth_mask;
6752 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6753
6754 for (; bucket != NULL; bucket = bucket->dthb_next) {
6755 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6756 return (bucket->dthb_len);
6757 }
6758
6759 return (NULL);
6760 }
6761
6762 static void
6763 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6764 {
6765 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6766 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6767
6768 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6769 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6770
6771 /*
6772 * Find the bucket that we're removing this probe from.
6773 */
6774 for (; bucket != NULL; bucket = bucket->dthb_next) {
6775 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6776 break;
6777 }
6778
6779 ASSERT(bucket != NULL);
6780
6781 if (*prevp == NULL) {
6782 if (*nextp == NULL) {
6783 /*
6784 * The removed probe was the only probe on this
6785 * bucket; we need to remove the bucket.
6786 */
6787 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6788
6789 ASSERT(bucket->dthb_chain == probe);
6790 ASSERT(b != NULL);
6791
6792 if (b == bucket) {
6793 hash->dth_tab[ndx] = bucket->dthb_next;
6794 } else {
6795 while (b->dthb_next != bucket)
6796 b = b->dthb_next;
6797 b->dthb_next = bucket->dthb_next;
6798 }
6799
6800 ASSERT(hash->dth_nbuckets > 0);
6801 hash->dth_nbuckets--;
6802 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6803 return;
6804 }
6805
6806 bucket->dthb_chain = *nextp;
6807 } else {
6808 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6809 }
6810
6811 if (*nextp != NULL)
6812 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6813 }
6814
6815 /*
6816 * DTrace Utility Functions
6817 *
6818 * These are random utility functions that are _not_ called from probe context.
6819 */
6820 static int
6821 dtrace_badattr(const dtrace_attribute_t *a)
6822 {
6823 return (a->dtat_name > DTRACE_STABILITY_MAX ||
6824 a->dtat_data > DTRACE_STABILITY_MAX ||
6825 a->dtat_class > DTRACE_CLASS_MAX);
6826 }
6827
6828 /*
6829 * Return a duplicate copy of a string. If the specified string is NULL,
6830 * this function returns a zero-length string.
6831 */
6832 static char *
6833 dtrace_strdup(const char *str)
6834 {
6835 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6836
6837 if (str != NULL)
6838 (void) strcpy(new, str);
6839
6840 return (new);
6841 }
6842
6843 #define DTRACE_ISALPHA(c) \
6844 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6845
6846 static int
6847 dtrace_badname(const char *s)
6848 {
6849 char c;
6850
6851 if (s == NULL || (c = *s++) == '\0')
6852 return (0);
6853
6854 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6855 return (1);
6856
6857 while ((c = *s++) != '\0') {
6858 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6859 c != '-' && c != '_' && c != '.' && c != '`')
6860 return (1);
6861 }
6862
6863 return (0);
6864 }
6865
6866 static void
6867 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6868 {
6869 uint32_t priv;
6870
6871 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6872 /*
6873 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6874 */
6875 priv = DTRACE_PRIV_ALL;
6876 } else {
6877 *uidp = crgetuid(cr);
6878 *zoneidp = crgetzonedid(cr);
6879
6880 priv = 0;
6881 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6882 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6883 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6884 priv |= DTRACE_PRIV_USER;
6885 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6886 priv |= DTRACE_PRIV_PROC;
6887 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6888 priv |= DTRACE_PRIV_OWNER;
6889 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6890 priv |= DTRACE_PRIV_ZONEOWNER;
6891 }
6892
6893 *privp = priv;
6894 }
6895
6896 #ifdef DTRACE_ERRDEBUG
6897 static void
6898 dtrace_errdebug(const char *str)
6899 {
6900 int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
6901 int occupied = 0;
6902
6903 mutex_enter(&dtrace_errlock);
6904 dtrace_errlast = str;
6905 dtrace_errthread = curthread;
6906
6907 while (occupied++ < DTRACE_ERRHASHSZ) {
6908 if (dtrace_errhash[hval].dter_msg == str) {
6909 dtrace_errhash[hval].dter_count++;
6910 goto out;
6911 }
6912
6913 if (dtrace_errhash[hval].dter_msg != NULL) {
6914 hval = (hval + 1) % DTRACE_ERRHASHSZ;
6915 continue;
6916 }
6917
6918 dtrace_errhash[hval].dter_msg = str;
6919 dtrace_errhash[hval].dter_count = 1;
6920 goto out;
6921 }
6922
6923 panic("dtrace: undersized error hash");
6924 out:
6925 mutex_exit(&dtrace_errlock);
6926 }
6927 #endif
6928
6929 /*
6930 * DTrace Matching Functions
6931 *
6932 * These functions are used to match groups of probes, given some elements of
6933 * a probe tuple, or some globbed expressions for elements of a probe tuple.
6934 */
6935 static int
6936 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
6937 zoneid_t zoneid)
6938 {
6939 if (priv != DTRACE_PRIV_ALL) {
6940 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
6941 uint32_t match = priv & ppriv;
6942
6943 /*
6944 * No PRIV_DTRACE_* privileges...
6945 */
6946 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
6947 DTRACE_PRIV_KERNEL)) == 0)
6948 return (0);
6949
6950 /*
6951 * No matching bits, but there were bits to match...
6952 */
6953 if (match == 0 && ppriv != 0)
6954 return (0);
6955
6956 /*
6957 * Need to have permissions to the process, but don't...
6958 */
6959 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
6960 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
6961 return (0);
6962 }
6963
6964 /*
6965 * Need to be in the same zone unless we possess the
6966 * privilege to examine all zones.
6967 */
6968 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
6969 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
6970 return (0);
6971 }
6972 }
6973
6974 return (1);
6975 }
6976
6977 /*
6978 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
6979 * consists of input pattern strings and an ops-vector to evaluate them.
6980 * This function returns >0 for match, 0 for no match, and <0 for error.
6981 */
6982 static int
6983 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
6984 uint32_t priv, uid_t uid, zoneid_t zoneid)
6985 {
6986 dtrace_provider_t *pvp = prp->dtpr_provider;
6987 int rv;
6988
6989 if (pvp->dtpv_defunct)
6990 return (0);
6991
6992 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
6993 return (rv);
6994
6995 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
6996 return (rv);
6997
6998 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
6999 return (rv);
7000
7001 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7002 return (rv);
7003
7004 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7005 return (0);
7006
7007 return (rv);
7008 }
7009
7010 /*
7011 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7012 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7013 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7014 * In addition, all of the recursion cases except for '*' matching have been
7015 * unwound. For '*', we still implement recursive evaluation, but a depth
7016 * counter is maintained and matching is aborted if we recurse too deep.
7017 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7018 */
7019 static int
7020 dtrace_match_glob(const char *s, const char *p, int depth)
7021 {
7022 const char *olds;
7023 char s1, c;
7024 int gs;
7025
7026 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7027 return (-1);
7028
7029 if (s == NULL)
7030 s = ""; /* treat NULL as empty string */
7031
7032 top:
7033 olds = s;
7034 s1 = *s++;
7035
7036 if (p == NULL)
7037 return (0);
7038
7039 if ((c = *p++) == '\0')
7040 return (s1 == '\0');
7041
7042 switch (c) {
7043 case '[': {
7044 int ok = 0, notflag = 0;
7045 char lc = '\0';
7046
7047 if (s1 == '\0')
7048 return (0);
7049
7050 if (*p == '!') {
7051 notflag = 1;
7052 p++;
7053 }
7054
7055 if ((c = *p++) == '\0')
7056 return (0);
7057
7058 do {
7059 if (c == '-' && lc != '\0' && *p != ']') {
7060 if ((c = *p++) == '\0')
7061 return (0);
7062 if (c == '\\' && (c = *p++) == '\0')
7063 return (0);
7064
7065 if (notflag) {
7066 if (s1 < lc || s1 > c)
7067 ok++;
7068 else
7069 return (0);
7070 } else if (lc <= s1 && s1 <= c)
7071 ok++;
7072
7073 } else if (c == '\\' && (c = *p++) == '\0')
7074 return (0);
7075
7076 lc = c; /* save left-hand 'c' for next iteration */
7077
7078 if (notflag) {
7079 if (s1 != c)
7080 ok++;
7081 else
7082 return (0);
7083 } else if (s1 == c)
7084 ok++;
7085
7086 if ((c = *p++) == '\0')
7087 return (0);
7088
7089 } while (c != ']');
7090
7091 if (ok)
7092 goto top;
7093
7094 return (0);
7095 }
7096
7097 case '\\':
7098 if ((c = *p++) == '\0')
7099 return (0);
7100 /*FALLTHRU*/
7101
7102 default:
7103 if (c != s1)
7104 return (0);
7105 /*FALLTHRU*/
7106
7107 case '?':
7108 if (s1 != '\0')
7109 goto top;
7110 return (0);
7111
7112 case '*':
7113 while (*p == '*')
7114 p++; /* consecutive *'s are identical to a single one */
7115
7116 if (*p == '\0')
7117 return (1);
7118
7119 for (s = olds; *s != '\0'; s++) {
7120 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7121 return (gs);
7122 }
7123
7124 return (0);
7125 }
7126 }
7127
7128 /*ARGSUSED*/
7129 static int
7130 dtrace_match_string(const char *s, const char *p, int depth)
7131 {
7132 return (s != NULL && strcmp(s, p) == 0);
7133 }
7134
7135 /*ARGSUSED*/
7136 static int
7137 dtrace_match_nul(const char *s, const char *p, int depth)
7138 {
7139 return (1); /* always match the empty pattern */
7140 }
7141
7142 /*ARGSUSED*/
7143 static int
7144 dtrace_match_nonzero(const char *s, const char *p, int depth)
7145 {
7146 return (s != NULL && s[0] != '\0');
7147 }
7148
7149 static int
7150 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7151 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7152 {
7153 dtrace_probe_t template, *probe;
7154 dtrace_hash_t *hash = NULL;
7155 int len, rc, best = INT_MAX, nmatched = 0;
7156 dtrace_id_t i;
7157
7158 ASSERT(MUTEX_HELD(&dtrace_lock));
7159
7160 /*
7161 * If the probe ID is specified in the key, just lookup by ID and
7162 * invoke the match callback once if a matching probe is found.
7163 */
7164 if (pkp->dtpk_id != DTRACE_IDNONE) {
7165 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7166 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7167 if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
7168 return (DTRACE_MATCH_FAIL);
7169 nmatched++;
7170 }
7171 return (nmatched);
7172 }
7173
7174 template.dtpr_mod = (char *)pkp->dtpk_mod;
7175 template.dtpr_func = (char *)pkp->dtpk_func;
7176 template.dtpr_name = (char *)pkp->dtpk_name;
7177
7178 /*
7179 * We want to find the most distinct of the module name, function
7180 * name, and name. So for each one that is not a glob pattern or
7181 * empty string, we perform a lookup in the corresponding hash and
7182 * use the hash table with the fewest collisions to do our search.
7183 */
7184 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7185 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7186 best = len;
7187 hash = dtrace_bymod;
7188 }
7189
7190 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7191 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7192 best = len;
7193 hash = dtrace_byfunc;
7194 }
7195
7196 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7197 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7198 best = len;
7199 hash = dtrace_byname;
7200 }
7201
7202 /*
7203 * If we did not select a hash table, iterate over every probe and
7204 * invoke our callback for each one that matches our input probe key.
7205 */
7206 if (hash == NULL) {
7207 for (i = 0; i < dtrace_nprobes; i++) {
7208 if ((probe = dtrace_probes[i]) == NULL ||
7209 dtrace_match_probe(probe, pkp, priv, uid,
7210 zoneid) <= 0)
7211 continue;
7212
7213 nmatched++;
7214
7215 if ((rc = (*matched)(probe, arg)) !=
7216 DTRACE_MATCH_NEXT) {
7217 if (rc == DTRACE_MATCH_FAIL)
7218 return (DTRACE_MATCH_FAIL);
7219 break;
7220 }
7221 }
7222
7223 return (nmatched);
7224 }
7225
7226 /*
7227 * If we selected a hash table, iterate over each probe of the same key
7228 * name and invoke the callback for every probe that matches the other
7229 * attributes of our input probe key.
7230 */
7231 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7232 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7233
7234 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7235 continue;
7236
7237 nmatched++;
7238
7239 if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
7240 if (rc == DTRACE_MATCH_FAIL)
7241 return (DTRACE_MATCH_FAIL);
7242 break;
7243 }
7244 }
7245
7246 return (nmatched);
7247 }
7248
7249 /*
7250 * Return the function pointer dtrace_probecmp() should use to compare the
7251 * specified pattern with a string. For NULL or empty patterns, we select
7252 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7253 * For non-empty non-glob strings, we use dtrace_match_string().
7254 */
7255 static dtrace_probekey_f *
7256 dtrace_probekey_func(const char *p)
7257 {
7258 char c;
7259
7260 if (p == NULL || *p == '\0')
7261 return (&dtrace_match_nul);
7262
7263 while ((c = *p++) != '\0') {
7264 if (c == '[' || c == '?' || c == '*' || c == '\\')
7265 return (&dtrace_match_glob);
7266 }
7267
7268 return (&dtrace_match_string);
7269 }
7270
7271 /*
7272 * Build a probe comparison key for use with dtrace_match_probe() from the
7273 * given probe description. By convention, a null key only matches anchored
7274 * probes: if each field is the empty string, reset dtpk_fmatch to
7275 * dtrace_match_nonzero().
7276 */
7277 static void
7278 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7279 {
7280 pkp->dtpk_prov = pdp->dtpd_provider;
7281 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7282
7283 pkp->dtpk_mod = pdp->dtpd_mod;
7284 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7285
7286 pkp->dtpk_func = pdp->dtpd_func;
7287 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7288
7289 pkp->dtpk_name = pdp->dtpd_name;
7290 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7291
7292 pkp->dtpk_id = pdp->dtpd_id;
7293
7294 if (pkp->dtpk_id == DTRACE_IDNONE &&
7295 pkp->dtpk_pmatch == &dtrace_match_nul &&
7296 pkp->dtpk_mmatch == &dtrace_match_nul &&
7297 pkp->dtpk_fmatch == &dtrace_match_nul &&
7298 pkp->dtpk_nmatch == &dtrace_match_nul)
7299 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7300 }
7301
7302 /*
7303 * DTrace Provider-to-Framework API Functions
7304 *
7305 * These functions implement much of the Provider-to-Framework API, as
7306 * described in <sys/dtrace.h>. The parts of the API not in this section are
7307 * the functions in the API for probe management (found below), and
7308 * dtrace_probe() itself (found above).
7309 */
7310
7311 /*
7312 * Register the calling provider with the DTrace framework. This should
7313 * generally be called by DTrace providers in their attach(9E) entry point.
7314 */
7315 int
7316 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7317 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7318 {
7319 dtrace_provider_t *provider;
7320
7321 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7322 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7323 "arguments", name ? name : "<NULL>");
7324 return (EINVAL);
7325 }
7326
7327 if (name[0] == '\0' || dtrace_badname(name)) {
7328 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7329 "provider name", name);
7330 return (EINVAL);
7331 }
7332
7333 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7334 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7335 pops->dtps_destroy == NULL ||
7336 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7337 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7338 "provider ops", name);
7339 return (EINVAL);
7340 }
7341
7342 if (dtrace_badattr(&pap->dtpa_provider) ||
7343 dtrace_badattr(&pap->dtpa_mod) ||
7344 dtrace_badattr(&pap->dtpa_func) ||
7345 dtrace_badattr(&pap->dtpa_name) ||
7346 dtrace_badattr(&pap->dtpa_args)) {
7347 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7348 "provider attributes", name);
7349 return (EINVAL);
7350 }
7351
7352 if (priv & ~DTRACE_PRIV_ALL) {
7353 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7354 "privilege attributes", name);
7355 return (EINVAL);
7356 }
7357
7358 if ((priv & DTRACE_PRIV_KERNEL) &&
7359 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7360 pops->dtps_mode == NULL) {
7361 cmn_err(CE_WARN, "failed to register provider '%s': need "
7362 "dtps_mode() op for given privilege attributes", name);
7363 return (EINVAL);
7364 }
7365
7366 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7367 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7368 (void) strcpy(provider->dtpv_name, name);
7369
7370 provider->dtpv_attr = *pap;
7371 provider->dtpv_priv.dtpp_flags = priv;
7372 if (cr != NULL) {
7373 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7374 provider->dtpv_priv.dtpp_zoneid = crgetzonedid(cr);
7375 }
7376 provider->dtpv_pops = *pops;
7377
7378 if (pops->dtps_provide == NULL) {
7379 ASSERT(pops->dtps_provide_module != NULL);
7380 provider->dtpv_pops.dtps_provide =
7381 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
7382 }
7383
7384 if (pops->dtps_provide_module == NULL) {
7385 ASSERT(pops->dtps_provide != NULL);
7386 provider->dtpv_pops.dtps_provide_module =
7387 (void (*)(void *, struct modctl *))dtrace_nullop;
7388 }
7389
7390 if (pops->dtps_suspend == NULL) {
7391 ASSERT(pops->dtps_resume == NULL);
7392 provider->dtpv_pops.dtps_suspend =
7393 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7394 provider->dtpv_pops.dtps_resume =
7395 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7396 }
7397
7398 provider->dtpv_arg = arg;
7399 *idp = (dtrace_provider_id_t)provider;
7400
7401 if (pops == &dtrace_provider_ops) {
7402 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7403 ASSERT(MUTEX_HELD(&dtrace_lock));
7404 ASSERT(dtrace_anon.dta_enabling == NULL);
7405
7406 /*
7407 * We make sure that the DTrace provider is at the head of
7408 * the provider chain.
7409 */
7410 provider->dtpv_next = dtrace_provider;
7411 dtrace_provider = provider;
7412 return (0);
7413 }
7414
7415 mutex_enter(&dtrace_provider_lock);
7416 mutex_enter(&dtrace_lock);
7417
7418 /*
7419 * If there is at least one provider registered, we'll add this
7420 * provider after the first provider.
7421 */
7422 if (dtrace_provider != NULL) {
7423 provider->dtpv_next = dtrace_provider->dtpv_next;
7424 dtrace_provider->dtpv_next = provider;
7425 } else {
7426 dtrace_provider = provider;
7427 }
7428
7429 if (dtrace_retained != NULL) {
7430 dtrace_enabling_provide(provider);
7431
7432 /*
7433 * Now we need to call dtrace_enabling_matchall() -- which
7434 * will acquire cpu_lock and dtrace_lock. We therefore need
7435 * to drop all of our locks before calling into it...
7436 */
7437 mutex_exit(&dtrace_lock);
7438 mutex_exit(&dtrace_provider_lock);
7439 dtrace_enabling_matchall();
7440
7441 return (0);
7442 }
7443
7444 mutex_exit(&dtrace_lock);
7445 mutex_exit(&dtrace_provider_lock);
7446
7447 return (0);
7448 }
7449
7450 /*
7451 * Unregister the specified provider from the DTrace framework. This should
7452 * generally be called by DTrace providers in their detach(9E) entry point.
7453 */
7454 int
7455 dtrace_unregister(dtrace_provider_id_t id)
7456 {
7457 dtrace_provider_t *old = (dtrace_provider_t *)id;
7458 dtrace_provider_t *prev = NULL;
7459 int i, self = 0, noreap = 0;
7460 dtrace_probe_t *probe, *first = NULL;
7461
7462 if (old->dtpv_pops.dtps_enable ==
7463 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
7464 /*
7465 * If DTrace itself is the provider, we're called with locks
7466 * already held.
7467 */
7468 ASSERT(old == dtrace_provider);
7469 ASSERT(dtrace_devi != NULL);
7470 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7471 ASSERT(MUTEX_HELD(&dtrace_lock));
7472 self = 1;
7473
7474 if (dtrace_provider->dtpv_next != NULL) {
7475 /*
7476 * There's another provider here; return failure.
7477 */
7478 return (EBUSY);
7479 }
7480 } else {
7481 mutex_enter(&dtrace_provider_lock);
7482 mutex_enter(&mod_lock);
7483 mutex_enter(&dtrace_lock);
7484 }
7485
7486 /*
7487 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7488 * probes, we refuse to let providers slither away, unless this
7489 * provider has already been explicitly invalidated.
7490 */
7491 if (!old->dtpv_defunct &&
7492 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7493 dtrace_anon.dta_state->dts_necbs > 0))) {
7494 if (!self) {
7495 mutex_exit(&dtrace_lock);
7496 mutex_exit(&mod_lock);
7497 mutex_exit(&dtrace_provider_lock);
7498 }
7499 return (EBUSY);
7500 }
7501
7502 /*
7503 * Attempt to destroy the probes associated with this provider.
7504 */
7505 for (i = 0; i < dtrace_nprobes; i++) {
7506 if ((probe = dtrace_probes[i]) == NULL)
7507 continue;
7508
7509 if (probe->dtpr_provider != old)
7510 continue;
7511
7512 if (probe->dtpr_ecb == NULL)
7513 continue;
7514
7515 /*
7516 * If we are trying to unregister a defunct provider, and the
7517 * provider was made defunct within the interval dictated by
7518 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7519 * attempt to reap our enablings. To denote that the provider
7520 * should reattempt to unregister itself at some point in the
7521 * future, we will return a differentiable error code (EAGAIN
7522 * instead of EBUSY) in this case.
7523 */
7524 if (dtrace_gethrtime() - old->dtpv_defunct >
7525 dtrace_unregister_defunct_reap)
7526 noreap = 1;
7527
7528 if (!self) {
7529 mutex_exit(&dtrace_lock);
7530 mutex_exit(&mod_lock);
7531 mutex_exit(&dtrace_provider_lock);
7532 }
7533
7534 if (noreap)
7535 return (EBUSY);
7536
7537 (void) taskq_dispatch(dtrace_taskq,
7538 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7539
7540 return (EAGAIN);
7541 }
7542
7543 /*
7544 * All of the probes for this provider are disabled; we can safely
7545 * remove all of them from their hash chains and from the probe array.
7546 */
7547 for (i = 0; i < dtrace_nprobes; i++) {
7548 if ((probe = dtrace_probes[i]) == NULL)
7549 continue;
7550
7551 if (probe->dtpr_provider != old)
7552 continue;
7553
7554 dtrace_probes[i] = NULL;
7555
7556 dtrace_hash_remove(dtrace_bymod, probe);
7557 dtrace_hash_remove(dtrace_byfunc, probe);
7558 dtrace_hash_remove(dtrace_byname, probe);
7559
7560 if (first == NULL) {
7561 first = probe;
7562 probe->dtpr_nextmod = NULL;
7563 } else {
7564 probe->dtpr_nextmod = first;
7565 first = probe;
7566 }
7567 }
7568
7569 /*
7570 * The provider's probes have been removed from the hash chains and
7571 * from the probe array. Now issue a dtrace_sync() to be sure that
7572 * everyone has cleared out from any probe array processing.
7573 */
7574 dtrace_sync();
7575
7576 for (probe = first; probe != NULL; probe = first) {
7577 first = probe->dtpr_nextmod;
7578
7579 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7580 probe->dtpr_arg);
7581 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7582 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7583 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7584 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7585 kmem_free(probe, sizeof (dtrace_probe_t));
7586 }
7587
7588 if ((prev = dtrace_provider) == old) {
7589 ASSERT(self || dtrace_devi == NULL);
7590 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7591 dtrace_provider = old->dtpv_next;
7592 } else {
7593 while (prev != NULL && prev->dtpv_next != old)
7594 prev = prev->dtpv_next;
7595
7596 if (prev == NULL) {
7597 panic("attempt to unregister non-existent "
7598 "dtrace provider %p\n", (void *)id);
7599 }
7600
7601 prev->dtpv_next = old->dtpv_next;
7602 }
7603
7604 if (!self) {
7605 mutex_exit(&dtrace_lock);
7606 mutex_exit(&mod_lock);
7607 mutex_exit(&dtrace_provider_lock);
7608 }
7609
7610 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7611 kmem_free(old, sizeof (dtrace_provider_t));
7612
7613 return (0);
7614 }
7615
7616 /*
7617 * Invalidate the specified provider. All subsequent probe lookups for the
7618 * specified provider will fail, but its probes will not be removed.
7619 */
7620 void
7621 dtrace_invalidate(dtrace_provider_id_t id)
7622 {
7623 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7624
7625 ASSERT(pvp->dtpv_pops.dtps_enable !=
7626 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7627
7628 mutex_enter(&dtrace_provider_lock);
7629 mutex_enter(&dtrace_lock);
7630
7631 pvp->dtpv_defunct = dtrace_gethrtime();
7632
7633 mutex_exit(&dtrace_lock);
7634 mutex_exit(&dtrace_provider_lock);
7635 }
7636
7637 /*
7638 * Indicate whether or not DTrace has attached.
7639 */
7640 int
7641 dtrace_attached(void)
7642 {
7643 /*
7644 * dtrace_provider will be non-NULL iff the DTrace driver has
7645 * attached. (It's non-NULL because DTrace is always itself a
7646 * provider.)
7647 */
7648 return (dtrace_provider != NULL);
7649 }
7650
7651 /*
7652 * Remove all the unenabled probes for the given provider. This function is
7653 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7654 * -- just as many of its associated probes as it can.
7655 */
7656 int
7657 dtrace_condense(dtrace_provider_id_t id)
7658 {
7659 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7660 int i;
7661 dtrace_probe_t *probe;
7662
7663 /*
7664 * Make sure this isn't the dtrace provider itself.
7665 */
7666 ASSERT(prov->dtpv_pops.dtps_enable !=
7667 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7668
7669 mutex_enter(&dtrace_provider_lock);
7670 mutex_enter(&dtrace_lock);
7671
7672 /*
7673 * Attempt to destroy the probes associated with this provider.
7674 */
7675 for (i = 0; i < dtrace_nprobes; i++) {
7676 if ((probe = dtrace_probes[i]) == NULL)
7677 continue;
7678
7679 if (probe->dtpr_provider != prov)
7680 continue;
7681
7682 if (probe->dtpr_ecb != NULL)
7683 continue;
7684
7685 dtrace_probes[i] = NULL;
7686
7687 dtrace_hash_remove(dtrace_bymod, probe);
7688 dtrace_hash_remove(dtrace_byfunc, probe);
7689 dtrace_hash_remove(dtrace_byname, probe);
7690
7691 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7692 probe->dtpr_arg);
7693 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7694 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7695 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7696 kmem_free(probe, sizeof (dtrace_probe_t));
7697 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7698 }
7699
7700 mutex_exit(&dtrace_lock);
7701 mutex_exit(&dtrace_provider_lock);
7702
7703 return (0);
7704 }
7705
7706 /*
7707 * DTrace Probe Management Functions
7708 *
7709 * The functions in this section perform the DTrace probe management,
7710 * including functions to create probes, look-up probes, and call into the
7711 * providers to request that probes be provided. Some of these functions are
7712 * in the Provider-to-Framework API; these functions can be identified by the
7713 * fact that they are not declared "static".
7714 */
7715
7716 /*
7717 * Create a probe with the specified module name, function name, and name.
7718 */
7719 dtrace_id_t
7720 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7721 const char *func, const char *name, int aframes, void *arg)
7722 {
7723 dtrace_probe_t *probe, **probes;
7724 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7725 dtrace_id_t id;
7726
7727 if (provider == dtrace_provider) {
7728 ASSERT(MUTEX_HELD(&dtrace_lock));
7729 } else {
7730 mutex_enter(&dtrace_lock);
7731 }
7732
7733 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7734 VM_BESTFIT | VM_SLEEP);
7735 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7736
7737 probe->dtpr_id = id;
7738 probe->dtpr_gen = dtrace_probegen++;
7739 probe->dtpr_mod = dtrace_strdup(mod);
7740 probe->dtpr_func = dtrace_strdup(func);
7741 probe->dtpr_name = dtrace_strdup(name);
7742 probe->dtpr_arg = arg;
7743 probe->dtpr_aframes = aframes;
7744 probe->dtpr_provider = provider;
7745
7746 dtrace_hash_add(dtrace_bymod, probe);
7747 dtrace_hash_add(dtrace_byfunc, probe);
7748 dtrace_hash_add(dtrace_byname, probe);
7749
7750 if (id - 1 >= dtrace_nprobes) {
7751 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7752 size_t nsize = osize << 1;
7753
7754 if (nsize == 0) {
7755 ASSERT(osize == 0);
7756 ASSERT(dtrace_probes == NULL);
7757 nsize = sizeof (dtrace_probe_t *);
7758 }
7759
7760 probes = kmem_zalloc(nsize, KM_SLEEP);
7761
7762 if (dtrace_probes == NULL) {
7763 ASSERT(osize == 0);
7764 dtrace_probes = probes;
7765 dtrace_nprobes = 1;
7766 } else {
7767 dtrace_probe_t **oprobes = dtrace_probes;
7768
7769 bcopy(oprobes, probes, osize);
7770 dtrace_membar_producer();
7771 dtrace_probes = probes;
7772
7773 dtrace_sync();
7774
7775 /*
7776 * All CPUs are now seeing the new probes array; we can
7777 * safely free the old array.
7778 */
7779 kmem_free(oprobes, osize);
7780 dtrace_nprobes <<= 1;
7781 }
7782
7783 ASSERT(id - 1 < dtrace_nprobes);
7784 }
7785
7786 ASSERT(dtrace_probes[id - 1] == NULL);
7787 dtrace_probes[id - 1] = probe;
7788
7789 if (provider != dtrace_provider)
7790 mutex_exit(&dtrace_lock);
7791
7792 return (id);
7793 }
7794
7795 static dtrace_probe_t *
7796 dtrace_probe_lookup_id(dtrace_id_t id)
7797 {
7798 ASSERT(MUTEX_HELD(&dtrace_lock));
7799
7800 if (id == 0 || id > dtrace_nprobes)
7801 return (NULL);
7802
7803 return (dtrace_probes[id - 1]);
7804 }
7805
7806 static int
7807 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7808 {
7809 *((dtrace_id_t *)arg) = probe->dtpr_id;
7810
7811 return (DTRACE_MATCH_DONE);
7812 }
7813
7814 /*
7815 * Look up a probe based on provider and one or more of module name, function
7816 * name and probe name.
7817 */
7818 dtrace_id_t
7819 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
7820 const char *func, const char *name)
7821 {
7822 dtrace_probekey_t pkey;
7823 dtrace_id_t id;
7824 int match;
7825
7826 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7827 pkey.dtpk_pmatch = &dtrace_match_string;
7828 pkey.dtpk_mod = mod;
7829 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7830 pkey.dtpk_func = func;
7831 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7832 pkey.dtpk_name = name;
7833 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7834 pkey.dtpk_id = DTRACE_IDNONE;
7835
7836 mutex_enter(&dtrace_lock);
7837 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7838 dtrace_probe_lookup_match, &id);
7839 mutex_exit(&dtrace_lock);
7840
7841 ASSERT(match == 1 || match == 0);
7842 return (match ? id : 0);
7843 }
7844
7845 /*
7846 * Returns the probe argument associated with the specified probe.
7847 */
7848 void *
7849 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7850 {
7851 dtrace_probe_t *probe;
7852 void *rval = NULL;
7853
7854 mutex_enter(&dtrace_lock);
7855
7856 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7857 probe->dtpr_provider == (dtrace_provider_t *)id)
7858 rval = probe->dtpr_arg;
7859
7860 mutex_exit(&dtrace_lock);
7861
7862 return (rval);
7863 }
7864
7865 /*
7866 * Copy a probe into a probe description.
7867 */
7868 static void
7869 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7870 {
7871 bzero(pdp, sizeof (dtrace_probedesc_t));
7872 pdp->dtpd_id = prp->dtpr_id;
7873
7874 (void) strncpy(pdp->dtpd_provider,
7875 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7876
7877 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7878 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7879 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7880 }
7881
7882 /*
7883 * Called to indicate that a probe -- or probes -- should be provided by a
7884 * specfied provider. If the specified description is NULL, the provider will
7885 * be told to provide all of its probes. (This is done whenever a new
7886 * consumer comes along, or whenever a retained enabling is to be matched.) If
7887 * the specified description is non-NULL, the provider is given the
7888 * opportunity to dynamically provide the specified probe, allowing providers
7889 * to support the creation of probes on-the-fly. (So-called _autocreated_
7890 * probes.) If the provider is NULL, the operations will be applied to all
7891 * providers; if the provider is non-NULL the operations will only be applied
7892 * to the specified provider. The dtrace_provider_lock must be held, and the
7893 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7894 * will need to grab the dtrace_lock when it reenters the framework through
7895 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7896 */
7897 static void
7898 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7899 {
7900 struct modctl *ctl;
7901 int all = 0;
7902
7903 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7904
7905 if (prv == NULL) {
7906 all = 1;
7907 prv = dtrace_provider;
7908 }
7909
7910 do {
7911 /*
7912 * First, call the blanket provide operation.
7913 */
7914 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7915
7916 /*
7917 * Now call the per-module provide operation. We will grab
7918 * mod_lock to prevent the list from being modified. Note
7919 * that this also prevents the mod_busy bits from changing.
7920 * (mod_busy can only be changed with mod_lock held.)
7921 */
7922 mutex_enter(&mod_lock);
7923
7924 ctl = &modules;
7925 do {
7926 if (ctl->mod_busy || ctl->mod_mp == NULL)
7927 continue;
7928
7929 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
7930
7931 } while ((ctl = ctl->mod_next) != &modules);
7932
7933 mutex_exit(&mod_lock);
7934 } while (all && (prv = prv->dtpv_next) != NULL);
7935 }
7936
7937 /*
7938 * Iterate over each probe, and call the Framework-to-Provider API function
7939 * denoted by offs.
7940 */
7941 static void
7942 dtrace_probe_foreach(uintptr_t offs)
7943 {
7944 dtrace_provider_t *prov;
7945 void (*func)(void *, dtrace_id_t, void *);
7946 dtrace_probe_t *probe;
7947 dtrace_icookie_t cookie;
7948 int i;
7949
7950 /*
7951 * We disable interrupts to walk through the probe array. This is
7952 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
7953 * won't see stale data.
7954 */
7955 cookie = dtrace_interrupt_disable();
7956
7957 for (i = 0; i < dtrace_nprobes; i++) {
7958 if ((probe = dtrace_probes[i]) == NULL)
7959 continue;
7960
7961 if (probe->dtpr_ecb == NULL) {
7962 /*
7963 * This probe isn't enabled -- don't call the function.
7964 */
7965 continue;
7966 }
7967
7968 prov = probe->dtpr_provider;
7969 func = *((void(**)(void *, dtrace_id_t, void *))
7970 ((uintptr_t)&prov->dtpv_pops + offs));
7971
7972 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
7973 }
7974
7975 dtrace_interrupt_enable(cookie);
7976 }
7977
7978 static int
7979 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
7980 {
7981 dtrace_probekey_t pkey;
7982 uint32_t priv;
7983 uid_t uid;
7984 zoneid_t zoneid;
7985 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
7986
7987 ASSERT(MUTEX_HELD(&dtrace_lock));
7988 dtrace_ecb_create_cache = NULL;
7989
7990 if (desc == NULL) {
7991 /*
7992 * If we're passed a NULL description, we're being asked to
7993 * create an ECB with a NULL probe.
7994 */
7995 (void) dtrace_ecb_create_enable(NULL, enab);
7996 return (0);
7997 }
7998
7999 dtrace_probekey(desc, &pkey);
8000 dtrace_cred2priv(state->dts_cred.dcr_cred, &priv, &uid, &zoneid);
8001
8002 if ((priv & DTRACE_PRIV_ZONEOWNER) &&
8003 state->dts_options[DTRACEOPT_ZONE] != DTRACEOPT_UNSET) {
8004 /*
8005 * If we have the privilege of instrumenting all zones but we
8006 * have been told to instrument but one, we will spoof this up
8007 * depriving ourselves of DTRACE_PRIV_ZONEOWNER for purposes
8008 * of dtrace_match(). (Note that DTRACEOPT_ZONE is not for
8009 * security but rather for performance: it allows the global
8010 * zone to instrument USDT probes in a local zone without
8011 * requiring all zones to be instrumented.)
8012 */
8013 priv &= ~DTRACE_PRIV_ZONEOWNER;
8014 zoneid = state->dts_options[DTRACEOPT_ZONE];
8015 }
8016
8017 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8018 enab));
8019 }
8020
8021 /*
8022 * DTrace Helper Provider Functions
8023 */
8024 static void
8025 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8026 {
8027 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8028 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8029 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8030 }
8031
8032 static void
8033 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8034 const dof_provider_t *dofprov, char *strtab)
8035 {
8036 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8037 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8038 dofprov->dofpv_provattr);
8039 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8040 dofprov->dofpv_modattr);
8041 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8042 dofprov->dofpv_funcattr);
8043 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8044 dofprov->dofpv_nameattr);
8045 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8046 dofprov->dofpv_argsattr);
8047 }
8048
8049 static void
8050 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8051 {
8052 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8053 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8054 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8055 dof_provider_t *provider;
8056 dof_probe_t *probe;
8057 uint32_t *off, *enoff;
8058 uint8_t *arg;
8059 char *strtab;
8060 uint_t i, nprobes;
8061 dtrace_helper_provdesc_t dhpv;
8062 dtrace_helper_probedesc_t dhpb;
8063 dtrace_meta_t *meta = dtrace_meta_pid;
8064 dtrace_mops_t *mops = &meta->dtm_mops;
8065 void *parg;
8066
8067 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8068 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8069 provider->dofpv_strtab * dof->dofh_secsize);
8070 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8071 provider->dofpv_probes * dof->dofh_secsize);
8072 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8073 provider->dofpv_prargs * dof->dofh_secsize);
8074 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8075 provider->dofpv_proffs * dof->dofh_secsize);
8076
8077 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8078 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8079 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8080 enoff = NULL;
8081
8082 /*
8083 * See dtrace_helper_provider_validate().
8084 */
8085 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8086 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8087 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8088 provider->dofpv_prenoffs * dof->dofh_secsize);
8089 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8090 }
8091
8092 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8093
8094 /*
8095 * Create the provider.
8096 */
8097 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8098
8099 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8100 return;
8101
8102 meta->dtm_count++;
8103
8104 /*
8105 * Create the probes.
8106 */
8107 for (i = 0; i < nprobes; i++) {
8108 probe = (dof_probe_t *)(uintptr_t)(daddr +
8109 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8110
8111 dhpb.dthpb_mod = dhp->dofhp_mod;
8112 dhpb.dthpb_func = strtab + probe->dofpr_func;
8113 dhpb.dthpb_name = strtab + probe->dofpr_name;
8114 dhpb.dthpb_base = probe->dofpr_addr;
8115 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8116 dhpb.dthpb_noffs = probe->dofpr_noffs;
8117 if (enoff != NULL) {
8118 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8119 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8120 } else {
8121 dhpb.dthpb_enoffs = NULL;
8122 dhpb.dthpb_nenoffs = 0;
8123 }
8124 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8125 dhpb.dthpb_nargc = probe->dofpr_nargc;
8126 dhpb.dthpb_xargc = probe->dofpr_xargc;
8127 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8128 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8129
8130 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8131 }
8132 }
8133
8134 static void
8135 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8136 {
8137 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8138 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8139 int i;
8140
8141 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8142
8143 for (i = 0; i < dof->dofh_secnum; i++) {
8144 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8145 dof->dofh_secoff + i * dof->dofh_secsize);
8146
8147 if (sec->dofs_type != DOF_SECT_PROVIDER)
8148 continue;
8149
8150 dtrace_helper_provide_one(dhp, sec, pid);
8151 }
8152
8153 /*
8154 * We may have just created probes, so we must now rematch against
8155 * any retained enablings. Note that this call will acquire both
8156 * cpu_lock and dtrace_lock; the fact that we are holding
8157 * dtrace_meta_lock now is what defines the ordering with respect to
8158 * these three locks.
8159 */
8160 dtrace_enabling_matchall();
8161 }
8162
8163 static void
8164 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8165 {
8166 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8167 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8168 dof_sec_t *str_sec;
8169 dof_provider_t *provider;
8170 char *strtab;
8171 dtrace_helper_provdesc_t dhpv;
8172 dtrace_meta_t *meta = dtrace_meta_pid;
8173 dtrace_mops_t *mops = &meta->dtm_mops;
8174
8175 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8176 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8177 provider->dofpv_strtab * dof->dofh_secsize);
8178
8179 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8180
8181 /*
8182 * Create the provider.
8183 */
8184 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8185
8186 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8187
8188 meta->dtm_count--;
8189 }
8190
8191 static void
8192 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8193 {
8194 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8195 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8196 int i;
8197
8198 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8199
8200 for (i = 0; i < dof->dofh_secnum; i++) {
8201 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8202 dof->dofh_secoff + i * dof->dofh_secsize);
8203
8204 if (sec->dofs_type != DOF_SECT_PROVIDER)
8205 continue;
8206
8207 dtrace_helper_provider_remove_one(dhp, sec, pid);
8208 }
8209 }
8210
8211 /*
8212 * DTrace Meta Provider-to-Framework API Functions
8213 *
8214 * These functions implement the Meta Provider-to-Framework API, as described
8215 * in <sys/dtrace.h>.
8216 */
8217 int
8218 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8219 dtrace_meta_provider_id_t *idp)
8220 {
8221 dtrace_meta_t *meta;
8222 dtrace_helpers_t *help, *next;
8223 int i;
8224
8225 *idp = DTRACE_METAPROVNONE;
8226
8227 /*
8228 * We strictly don't need the name, but we hold onto it for
8229 * debuggability. All hail error queues!
8230 */
8231 if (name == NULL) {
8232 cmn_err(CE_WARN, "failed to register meta-provider: "
8233 "invalid name");
8234 return (EINVAL);
8235 }
8236
8237 if (mops == NULL ||
8238 mops->dtms_create_probe == NULL ||
8239 mops->dtms_provide_pid == NULL ||
8240 mops->dtms_remove_pid == NULL) {
8241 cmn_err(CE_WARN, "failed to register meta-register %s: "
8242 "invalid ops", name);
8243 return (EINVAL);
8244 }
8245
8246 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8247 meta->dtm_mops = *mops;
8248 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8249 (void) strcpy(meta->dtm_name, name);
8250 meta->dtm_arg = arg;
8251
8252 mutex_enter(&dtrace_meta_lock);
8253 mutex_enter(&dtrace_lock);
8254
8255 if (dtrace_meta_pid != NULL) {
8256 mutex_exit(&dtrace_lock);
8257 mutex_exit(&dtrace_meta_lock);
8258 cmn_err(CE_WARN, "failed to register meta-register %s: "
8259 "user-land meta-provider exists", name);
8260 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8261 kmem_free(meta, sizeof (dtrace_meta_t));
8262 return (EINVAL);
8263 }
8264
8265 dtrace_meta_pid = meta;
8266 *idp = (dtrace_meta_provider_id_t)meta;
8267
8268 /*
8269 * If there are providers and probes ready to go, pass them
8270 * off to the new meta provider now.
8271 */
8272
8273 help = dtrace_deferred_pid;
8274 dtrace_deferred_pid = NULL;
8275
8276 mutex_exit(&dtrace_lock);
8277
8278 while (help != NULL) {
8279 for (i = 0; i < help->dthps_nprovs; i++) {
8280 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8281 help->dthps_pid);
8282 }
8283
8284 next = help->dthps_next;
8285 help->dthps_next = NULL;
8286 help->dthps_prev = NULL;
8287 help->dthps_deferred = 0;
8288 help = next;
8289 }
8290
8291 mutex_exit(&dtrace_meta_lock);
8292
8293 return (0);
8294 }
8295
8296 int
8297 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8298 {
8299 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8300
8301 mutex_enter(&dtrace_meta_lock);
8302 mutex_enter(&dtrace_lock);
8303
8304 if (old == dtrace_meta_pid) {
8305 pp = &dtrace_meta_pid;
8306 } else {
8307 panic("attempt to unregister non-existent "
8308 "dtrace meta-provider %p\n", (void *)old);
8309 }
8310
8311 if (old->dtm_count != 0) {
8312 mutex_exit(&dtrace_lock);
8313 mutex_exit(&dtrace_meta_lock);
8314 return (EBUSY);
8315 }
8316
8317 *pp = NULL;
8318
8319 mutex_exit(&dtrace_lock);
8320 mutex_exit(&dtrace_meta_lock);
8321
8322 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8323 kmem_free(old, sizeof (dtrace_meta_t));
8324
8325 return (0);
8326 }
8327
8328
8329 /*
8330 * DTrace DIF Object Functions
8331 */
8332 static int
8333 dtrace_difo_err(uint_t pc, const char *format, ...)
8334 {
8335 if (dtrace_err_verbose) {
8336 va_list alist;
8337
8338 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8339 va_start(alist, format);
8340 (void) vuprintf(format, alist);
8341 va_end(alist);
8342 }
8343
8344 #ifdef DTRACE_ERRDEBUG
8345 dtrace_errdebug(format);
8346 #endif
8347 return (1);
8348 }
8349
8350 /*
8351 * Validate a DTrace DIF object by checking the IR instructions. The following
8352 * rules are currently enforced by dtrace_difo_validate():
8353 *
8354 * 1. Each instruction must have a valid opcode
8355 * 2. Each register, string, variable, or subroutine reference must be valid
8356 * 3. No instruction can modify register %r0 (must be zero)
8357 * 4. All instruction reserved bits must be set to zero
8358 * 5. The last instruction must be a "ret" instruction
8359 * 6. All branch targets must reference a valid instruction _after_ the branch
8360 */
8361 static int
8362 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8363 cred_t *cr)
8364 {
8365 int err = 0, i;
8366 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8367 int kcheckload;
8368 uint_t pc;
8369
8370 kcheckload = cr == NULL ||
8371 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8372
8373 dp->dtdo_destructive = 0;
8374
8375 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8376 dif_instr_t instr = dp->dtdo_buf[pc];
8377
8378 uint_t r1 = DIF_INSTR_R1(instr);
8379 uint_t r2 = DIF_INSTR_R2(instr);
8380 uint_t rd = DIF_INSTR_RD(instr);
8381 uint_t rs = DIF_INSTR_RS(instr);
8382 uint_t label = DIF_INSTR_LABEL(instr);
8383 uint_t v = DIF_INSTR_VAR(instr);
8384 uint_t subr = DIF_INSTR_SUBR(instr);
8385 uint_t type = DIF_INSTR_TYPE(instr);
8386 uint_t op = DIF_INSTR_OP(instr);
8387
8388 switch (op) {
8389 case DIF_OP_OR:
8390 case DIF_OP_XOR:
8391 case DIF_OP_AND:
8392 case DIF_OP_SLL:
8393 case DIF_OP_SRL:
8394 case DIF_OP_SRA:
8395 case DIF_OP_SUB:
8396 case DIF_OP_ADD:
8397 case DIF_OP_MUL:
8398 case DIF_OP_SDIV:
8399 case DIF_OP_UDIV:
8400 case DIF_OP_SREM:
8401 case DIF_OP_UREM:
8402 case DIF_OP_COPYS:
8403 if (r1 >= nregs)
8404 err += efunc(pc, "invalid register %u\n", r1);
8405 if (r2 >= nregs)
8406 err += efunc(pc, "invalid register %u\n", r2);
8407 if (rd >= nregs)
8408 err += efunc(pc, "invalid register %u\n", rd);
8409 if (rd == 0)
8410 err += efunc(pc, "cannot write to %r0\n");
8411 break;
8412 case DIF_OP_NOT:
8413 case DIF_OP_MOV:
8414 case DIF_OP_ALLOCS:
8415 if (r1 >= nregs)
8416 err += efunc(pc, "invalid register %u\n", r1);
8417 if (r2 != 0)
8418 err += efunc(pc, "non-zero reserved bits\n");
8419 if (rd >= nregs)
8420 err += efunc(pc, "invalid register %u\n", rd);
8421 if (rd == 0)
8422 err += efunc(pc, "cannot write to %r0\n");
8423 break;
8424 case DIF_OP_LDSB:
8425 case DIF_OP_LDSH:
8426 case DIF_OP_LDSW:
8427 case DIF_OP_LDUB:
8428 case DIF_OP_LDUH:
8429 case DIF_OP_LDUW:
8430 case DIF_OP_LDX:
8431 if (r1 >= nregs)
8432 err += efunc(pc, "invalid register %u\n", r1);
8433 if (r2 != 0)
8434 err += efunc(pc, "non-zero reserved bits\n");
8435 if (rd >= nregs)
8436 err += efunc(pc, "invalid register %u\n", rd);
8437 if (rd == 0)
8438 err += efunc(pc, "cannot write to %r0\n");
8439 if (kcheckload)
8440 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8441 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8442 break;
8443 case DIF_OP_RLDSB:
8444 case DIF_OP_RLDSH:
8445 case DIF_OP_RLDSW:
8446 case DIF_OP_RLDUB:
8447 case DIF_OP_RLDUH:
8448 case DIF_OP_RLDUW:
8449 case DIF_OP_RLDX:
8450 if (r1 >= nregs)
8451 err += efunc(pc, "invalid register %u\n", r1);
8452 if (r2 != 0)
8453 err += efunc(pc, "non-zero reserved bits\n");
8454 if (rd >= nregs)
8455 err += efunc(pc, "invalid register %u\n", rd);
8456 if (rd == 0)
8457 err += efunc(pc, "cannot write to %r0\n");
8458 break;
8459 case DIF_OP_ULDSB:
8460 case DIF_OP_ULDSH:
8461 case DIF_OP_ULDSW:
8462 case DIF_OP_ULDUB:
8463 case DIF_OP_ULDUH:
8464 case DIF_OP_ULDUW:
8465 case DIF_OP_ULDX:
8466 if (r1 >= nregs)
8467 err += efunc(pc, "invalid register %u\n", r1);
8468 if (r2 != 0)
8469 err += efunc(pc, "non-zero reserved bits\n");
8470 if (rd >= nregs)
8471 err += efunc(pc, "invalid register %u\n", rd);
8472 if (rd == 0)
8473 err += efunc(pc, "cannot write to %r0\n");
8474 break;
8475 case DIF_OP_STB:
8476 case DIF_OP_STH:
8477 case DIF_OP_STW:
8478 case DIF_OP_STX:
8479 if (r1 >= nregs)
8480 err += efunc(pc, "invalid register %u\n", r1);
8481 if (r2 != 0)
8482 err += efunc(pc, "non-zero reserved bits\n");
8483 if (rd >= nregs)
8484 err += efunc(pc, "invalid register %u\n", rd);
8485 if (rd == 0)
8486 err += efunc(pc, "cannot write to 0 address\n");
8487 break;
8488 case DIF_OP_CMP:
8489 case DIF_OP_SCMP:
8490 if (r1 >= nregs)
8491 err += efunc(pc, "invalid register %u\n", r1);
8492 if (r2 >= nregs)
8493 err += efunc(pc, "invalid register %u\n", r2);
8494 if (rd != 0)
8495 err += efunc(pc, "non-zero reserved bits\n");
8496 break;
8497 case DIF_OP_TST:
8498 if (r1 >= nregs)
8499 err += efunc(pc, "invalid register %u\n", r1);
8500 if (r2 != 0 || rd != 0)
8501 err += efunc(pc, "non-zero reserved bits\n");
8502 break;
8503 case DIF_OP_BA:
8504 case DIF_OP_BE:
8505 case DIF_OP_BNE:
8506 case DIF_OP_BG:
8507 case DIF_OP_BGU:
8508 case DIF_OP_BGE:
8509 case DIF_OP_BGEU:
8510 case DIF_OP_BL:
8511 case DIF_OP_BLU:
8512 case DIF_OP_BLE:
8513 case DIF_OP_BLEU:
8514 if (label >= dp->dtdo_len) {
8515 err += efunc(pc, "invalid branch target %u\n",
8516 label);
8517 }
8518 if (label <= pc) {
8519 err += efunc(pc, "backward branch to %u\n",
8520 label);
8521 }
8522 break;
8523 case DIF_OP_RET:
8524 if (r1 != 0 || r2 != 0)
8525 err += efunc(pc, "non-zero reserved bits\n");
8526 if (rd >= nregs)
8527 err += efunc(pc, "invalid register %u\n", rd);
8528 break;
8529 case DIF_OP_NOP:
8530 case DIF_OP_POPTS:
8531 case DIF_OP_FLUSHTS:
8532 if (r1 != 0 || r2 != 0 || rd != 0)
8533 err += efunc(pc, "non-zero reserved bits\n");
8534 break;
8535 case DIF_OP_SETX:
8536 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8537 err += efunc(pc, "invalid integer ref %u\n",
8538 DIF_INSTR_INTEGER(instr));
8539 }
8540 if (rd >= nregs)
8541 err += efunc(pc, "invalid register %u\n", rd);
8542 if (rd == 0)
8543 err += efunc(pc, "cannot write to %r0\n");
8544 break;
8545 case DIF_OP_SETS:
8546 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8547 err += efunc(pc, "invalid string ref %u\n",
8548 DIF_INSTR_STRING(instr));
8549 }
8550 if (rd >= nregs)
8551 err += efunc(pc, "invalid register %u\n", rd);
8552 if (rd == 0)
8553 err += efunc(pc, "cannot write to %r0\n");
8554 break;
8555 case DIF_OP_LDGA:
8556 case DIF_OP_LDTA:
8557 if (r1 > DIF_VAR_ARRAY_MAX)
8558 err += efunc(pc, "invalid array %u\n", r1);
8559 if (r2 >= nregs)
8560 err += efunc(pc, "invalid register %u\n", r2);
8561 if (rd >= nregs)
8562 err += efunc(pc, "invalid register %u\n", rd);
8563 if (rd == 0)
8564 err += efunc(pc, "cannot write to %r0\n");
8565 break;
8566 case DIF_OP_LDGS:
8567 case DIF_OP_LDTS:
8568 case DIF_OP_LDLS:
8569 case DIF_OP_LDGAA:
8570 case DIF_OP_LDTAA:
8571 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8572 err += efunc(pc, "invalid variable %u\n", v);
8573 if (rd >= nregs)
8574 err += efunc(pc, "invalid register %u\n", rd);
8575 if (rd == 0)
8576 err += efunc(pc, "cannot write to %r0\n");
8577 break;
8578 case DIF_OP_STGS:
8579 case DIF_OP_STTS:
8580 case DIF_OP_STLS:
8581 case DIF_OP_STGAA:
8582 case DIF_OP_STTAA:
8583 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8584 err += efunc(pc, "invalid variable %u\n", v);
8585 if (rs >= nregs)
8586 err += efunc(pc, "invalid register %u\n", rd);
8587 break;
8588 case DIF_OP_CALL:
8589 if (subr > DIF_SUBR_MAX)
8590 err += efunc(pc, "invalid subr %u\n", subr);
8591 if (rd >= nregs)
8592 err += efunc(pc, "invalid register %u\n", rd);
8593 if (rd == 0)
8594 err += efunc(pc, "cannot write to %r0\n");
8595
8596 if (subr == DIF_SUBR_COPYOUT ||
8597 subr == DIF_SUBR_COPYOUTSTR) {
8598 dp->dtdo_destructive = 1;
8599 }
8600
8601 if (subr == DIF_SUBR_GETF) {
8602 /*
8603 * If we have a getf() we need to record that
8604 * in our state. Note that our state can be
8605 * NULL if this is a helper -- but in that
8606 * case, the call to getf() is itself illegal,
8607 * and will be caught (slightly later) when
8608 * the helper is validated.
8609 */
8610 if (vstate->dtvs_state != NULL)
8611 vstate->dtvs_state->dts_getf++;
8612 }
8613
8614 break;
8615 case DIF_OP_PUSHTR:
8616 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8617 err += efunc(pc, "invalid ref type %u\n", type);
8618 if (r2 >= nregs)
8619 err += efunc(pc, "invalid register %u\n", r2);
8620 if (rs >= nregs)
8621 err += efunc(pc, "invalid register %u\n", rs);
8622 break;
8623 case DIF_OP_PUSHTV:
8624 if (type != DIF_TYPE_CTF)
8625 err += efunc(pc, "invalid val type %u\n", type);
8626 if (r2 >= nregs)
8627 err += efunc(pc, "invalid register %u\n", r2);
8628 if (rs >= nregs)
8629 err += efunc(pc, "invalid register %u\n", rs);
8630 break;
8631 default:
8632 err += efunc(pc, "invalid opcode %u\n",
8633 DIF_INSTR_OP(instr));
8634 }
8635 }
8636
8637 if (dp->dtdo_len != 0 &&
8638 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8639 err += efunc(dp->dtdo_len - 1,
8640 "expected 'ret' as last DIF instruction\n");
8641 }
8642
8643 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8644 /*
8645 * If we're not returning by reference, the size must be either
8646 * 0 or the size of one of the base types.
8647 */
8648 switch (dp->dtdo_rtype.dtdt_size) {
8649 case 0:
8650 case sizeof (uint8_t):
8651 case sizeof (uint16_t):
8652 case sizeof (uint32_t):
8653 case sizeof (uint64_t):
8654 break;
8655
8656 default:
8657 err += efunc(dp->dtdo_len - 1, "bad return size\n");
8658 }
8659 }
8660
8661 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8662 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8663 dtrace_diftype_t *vt, *et;
8664 uint_t id, ndx;
8665
8666 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8667 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8668 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8669 err += efunc(i, "unrecognized variable scope %d\n",
8670 v->dtdv_scope);
8671 break;
8672 }
8673
8674 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8675 v->dtdv_kind != DIFV_KIND_SCALAR) {
8676 err += efunc(i, "unrecognized variable type %d\n",
8677 v->dtdv_kind);
8678 break;
8679 }
8680
8681 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8682 err += efunc(i, "%d exceeds variable id limit\n", id);
8683 break;
8684 }
8685
8686 if (id < DIF_VAR_OTHER_UBASE)
8687 continue;
8688
8689 /*
8690 * For user-defined variables, we need to check that this
8691 * definition is identical to any previous definition that we
8692 * encountered.
8693 */
8694 ndx = id - DIF_VAR_OTHER_UBASE;
8695
8696 switch (v->dtdv_scope) {
8697 case DIFV_SCOPE_GLOBAL:
8698 if (ndx < vstate->dtvs_nglobals) {
8699 dtrace_statvar_t *svar;
8700
8701 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8702 existing = &svar->dtsv_var;
8703 }
8704
8705 break;
8706
8707 case DIFV_SCOPE_THREAD:
8708 if (ndx < vstate->dtvs_ntlocals)
8709 existing = &vstate->dtvs_tlocals[ndx];
8710 break;
8711
8712 case DIFV_SCOPE_LOCAL:
8713 if (ndx < vstate->dtvs_nlocals) {
8714 dtrace_statvar_t *svar;
8715
8716 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8717 existing = &svar->dtsv_var;
8718 }
8719
8720 break;
8721 }
8722
8723 vt = &v->dtdv_type;
8724
8725 if (vt->dtdt_flags & DIF_TF_BYREF) {
8726 if (vt->dtdt_size == 0) {
8727 err += efunc(i, "zero-sized variable\n");
8728 break;
8729 }
8730
8731 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8732 vt->dtdt_size > dtrace_global_maxsize) {
8733 err += efunc(i, "oversized by-ref global\n");
8734 break;
8735 }
8736 }
8737
8738 if (existing == NULL || existing->dtdv_id == 0)
8739 continue;
8740
8741 ASSERT(existing->dtdv_id == v->dtdv_id);
8742 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8743
8744 if (existing->dtdv_kind != v->dtdv_kind)
8745 err += efunc(i, "%d changed variable kind\n", id);
8746
8747 et = &existing->dtdv_type;
8748
8749 if (vt->dtdt_flags != et->dtdt_flags) {
8750 err += efunc(i, "%d changed variable type flags\n", id);
8751 break;
8752 }
8753
8754 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8755 err += efunc(i, "%d changed variable type size\n", id);
8756 break;
8757 }
8758 }
8759
8760 return (err);
8761 }
8762
8763 /*
8764 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
8765 * are much more constrained than normal DIFOs. Specifically, they may
8766 * not:
8767 *
8768 * 1. Make calls to subroutines other than copyin(), copyinstr() or
8769 * miscellaneous string routines
8770 * 2. Access DTrace variables other than the args[] array, and the
8771 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8772 * 3. Have thread-local variables.
8773 * 4. Have dynamic variables.
8774 */
8775 static int
8776 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8777 {
8778 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8779 int err = 0;
8780 uint_t pc;
8781
8782 for (pc = 0; pc < dp->dtdo_len; pc++) {
8783 dif_instr_t instr = dp->dtdo_buf[pc];
8784
8785 uint_t v = DIF_INSTR_VAR(instr);
8786 uint_t subr = DIF_INSTR_SUBR(instr);
8787 uint_t op = DIF_INSTR_OP(instr);
8788
8789 switch (op) {
8790 case DIF_OP_OR:
8791 case DIF_OP_XOR:
8792 case DIF_OP_AND:
8793 case DIF_OP_SLL:
8794 case DIF_OP_SRL:
8795 case DIF_OP_SRA:
8796 case DIF_OP_SUB:
8797 case DIF_OP_ADD:
8798 case DIF_OP_MUL:
8799 case DIF_OP_SDIV:
8800 case DIF_OP_UDIV:
8801 case DIF_OP_SREM:
8802 case DIF_OP_UREM:
8803 case DIF_OP_COPYS:
8804 case DIF_OP_NOT:
8805 case DIF_OP_MOV:
8806 case DIF_OP_RLDSB:
8807 case DIF_OP_RLDSH:
8808 case DIF_OP_RLDSW:
8809 case DIF_OP_RLDUB:
8810 case DIF_OP_RLDUH:
8811 case DIF_OP_RLDUW:
8812 case DIF_OP_RLDX:
8813 case DIF_OP_ULDSB:
8814 case DIF_OP_ULDSH:
8815 case DIF_OP_ULDSW:
8816 case DIF_OP_ULDUB:
8817 case DIF_OP_ULDUH:
8818 case DIF_OP_ULDUW:
8819 case DIF_OP_ULDX:
8820 case DIF_OP_STB:
8821 case DIF_OP_STH:
8822 case DIF_OP_STW:
8823 case DIF_OP_STX:
8824 case DIF_OP_ALLOCS:
8825 case DIF_OP_CMP:
8826 case DIF_OP_SCMP:
8827 case DIF_OP_TST:
8828 case DIF_OP_BA:
8829 case DIF_OP_BE:
8830 case DIF_OP_BNE:
8831 case DIF_OP_BG:
8832 case DIF_OP_BGU:
8833 case DIF_OP_BGE:
8834 case DIF_OP_BGEU:
8835 case DIF_OP_BL:
8836 case DIF_OP_BLU:
8837 case DIF_OP_BLE:
8838 case DIF_OP_BLEU:
8839 case DIF_OP_RET:
8840 case DIF_OP_NOP:
8841 case DIF_OP_POPTS:
8842 case DIF_OP_FLUSHTS:
8843 case DIF_OP_SETX:
8844 case DIF_OP_SETS:
8845 case DIF_OP_LDGA:
8846 case DIF_OP_LDLS:
8847 case DIF_OP_STGS:
8848 case DIF_OP_STLS:
8849 case DIF_OP_PUSHTR:
8850 case DIF_OP_PUSHTV:
8851 break;
8852
8853 case DIF_OP_LDGS:
8854 if (v >= DIF_VAR_OTHER_UBASE)
8855 break;
8856
8857 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8858 break;
8859
8860 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8861 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8862 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8863 v == DIF_VAR_UID || v == DIF_VAR_GID)
8864 break;
8865
8866 err += efunc(pc, "illegal variable %u\n", v);
8867 break;
8868
8869 case DIF_OP_LDTA:
8870 case DIF_OP_LDTS:
8871 case DIF_OP_LDGAA:
8872 case DIF_OP_LDTAA:
8873 err += efunc(pc, "illegal dynamic variable load\n");
8874 break;
8875
8876 case DIF_OP_STTS:
8877 case DIF_OP_STGAA:
8878 case DIF_OP_STTAA:
8879 err += efunc(pc, "illegal dynamic variable store\n");
8880 break;
8881
8882 case DIF_OP_CALL:
8883 if (subr == DIF_SUBR_ALLOCA ||
8884 subr == DIF_SUBR_BCOPY ||
8885 subr == DIF_SUBR_COPYIN ||
8886 subr == DIF_SUBR_COPYINTO ||
8887 subr == DIF_SUBR_COPYINSTR ||
8888 subr == DIF_SUBR_INDEX ||
8889 subr == DIF_SUBR_INET_NTOA ||
8890 subr == DIF_SUBR_INET_NTOA6 ||
8891 subr == DIF_SUBR_INET_NTOP ||
8892 subr == DIF_SUBR_LLTOSTR ||
8893 subr == DIF_SUBR_RINDEX ||
8894 subr == DIF_SUBR_STRCHR ||
8895 subr == DIF_SUBR_STRJOIN ||
8896 subr == DIF_SUBR_STRRCHR ||
8897 subr == DIF_SUBR_STRSTR ||
8898 subr == DIF_SUBR_HTONS ||
8899 subr == DIF_SUBR_HTONL ||
8900 subr == DIF_SUBR_HTONLL ||
8901 subr == DIF_SUBR_NTOHS ||
8902 subr == DIF_SUBR_NTOHL ||
8903 subr == DIF_SUBR_NTOHLL)
8904 break;
8905
8906 err += efunc(pc, "invalid subr %u\n", subr);
8907 break;
8908
8909 default:
8910 err += efunc(pc, "invalid opcode %u\n",
8911 DIF_INSTR_OP(instr));
8912 }
8913 }
8914
8915 return (err);
8916 }
8917
8918 /*
8919 * Returns 1 if the expression in the DIF object can be cached on a per-thread
8920 * basis; 0 if not.
8921 */
8922 static int
8923 dtrace_difo_cacheable(dtrace_difo_t *dp)
8924 {
8925 int i;
8926
8927 if (dp == NULL)
8928 return (0);
8929
8930 for (i = 0; i < dp->dtdo_varlen; i++) {
8931 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8932
8933 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8934 continue;
8935
8936 switch (v->dtdv_id) {
8937 case DIF_VAR_CURTHREAD:
8938 case DIF_VAR_PID:
8939 case DIF_VAR_TID:
8940 case DIF_VAR_EXECNAME:
8941 case DIF_VAR_ZONENAME:
8942 break;
8943
8944 default:
8945 return (0);
8946 }
8947 }
8948
8949 /*
8950 * This DIF object may be cacheable. Now we need to look for any
8951 * array loading instructions, any memory loading instructions, or
8952 * any stores to thread-local variables.
8953 */
8954 for (i = 0; i < dp->dtdo_len; i++) {
8955 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
8956
8957 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
8958 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
8959 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
8960 op == DIF_OP_LDGA || op == DIF_OP_STTS)
8961 return (0);
8962 }
8963
8964 return (1);
8965 }
8966
8967 static void
8968 dtrace_difo_hold(dtrace_difo_t *dp)
8969 {
8970 int i;
8971
8972 ASSERT(MUTEX_HELD(&dtrace_lock));
8973
8974 dp->dtdo_refcnt++;
8975 ASSERT(dp->dtdo_refcnt != 0);
8976
8977 /*
8978 * We need to check this DIF object for references to the variable
8979 * DIF_VAR_VTIMESTAMP.
8980 */
8981 for (i = 0; i < dp->dtdo_varlen; i++) {
8982 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8983
8984 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8985 continue;
8986
8987 if (dtrace_vtime_references++ == 0)
8988 dtrace_vtime_enable();
8989 }
8990 }
8991
8992 /*
8993 * This routine calculates the dynamic variable chunksize for a given DIF
8994 * object. The calculation is not fool-proof, and can probably be tricked by
8995 * malicious DIF -- but it works for all compiler-generated DIF. Because this
8996 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
8997 * if a dynamic variable size exceeds the chunksize.
8998 */
8999 static void
9000 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9001 {
9002 uint64_t sval;
9003 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9004 const dif_instr_t *text = dp->dtdo_buf;
9005 uint_t pc, srd = 0;
9006 uint_t ttop = 0;
9007 size_t size, ksize;
9008 uint_t id, i;
9009
9010 for (pc = 0; pc < dp->dtdo_len; pc++) {
9011 dif_instr_t instr = text[pc];
9012 uint_t op = DIF_INSTR_OP(instr);
9013 uint_t rd = DIF_INSTR_RD(instr);
9014 uint_t r1 = DIF_INSTR_R1(instr);
9015 uint_t nkeys = 0;
9016 uchar_t scope;
9017
9018 dtrace_key_t *key = tupregs;
9019
9020 switch (op) {
9021 case DIF_OP_SETX:
9022 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9023 srd = rd;
9024 continue;
9025
9026 case DIF_OP_STTS:
9027 key = &tupregs[DIF_DTR_NREGS];
9028 key[0].dttk_size = 0;
9029 key[1].dttk_size = 0;
9030 nkeys = 2;
9031 scope = DIFV_SCOPE_THREAD;
9032 break;
9033
9034 case DIF_OP_STGAA:
9035 case DIF_OP_STTAA:
9036 nkeys = ttop;
9037
9038 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9039 key[nkeys++].dttk_size = 0;
9040
9041 key[nkeys++].dttk_size = 0;
9042
9043 if (op == DIF_OP_STTAA) {
9044 scope = DIFV_SCOPE_THREAD;
9045 } else {
9046 scope = DIFV_SCOPE_GLOBAL;
9047 }
9048
9049 break;
9050
9051 case DIF_OP_PUSHTR:
9052 if (ttop == DIF_DTR_NREGS)
9053 return;
9054
9055 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9056 /*
9057 * If the register for the size of the "pushtr"
9058 * is %r0 (or the value is 0) and the type is
9059 * a string, we'll use the system-wide default
9060 * string size.
9061 */
9062 tupregs[ttop++].dttk_size =
9063 dtrace_strsize_default;
9064 } else {
9065 if (srd == 0)
9066 return;
9067
9068 tupregs[ttop++].dttk_size = sval;
9069 }
9070
9071 break;
9072
9073 case DIF_OP_PUSHTV:
9074 if (ttop == DIF_DTR_NREGS)
9075 return;
9076
9077 tupregs[ttop++].dttk_size = 0;
9078 break;
9079
9080 case DIF_OP_FLUSHTS:
9081 ttop = 0;
9082 break;
9083
9084 case DIF_OP_POPTS:
9085 if (ttop != 0)
9086 ttop--;
9087 break;
9088 }
9089
9090 sval = 0;
9091 srd = 0;
9092
9093 if (nkeys == 0)
9094 continue;
9095
9096 /*
9097 * We have a dynamic variable allocation; calculate its size.
9098 */
9099 for (ksize = 0, i = 0; i < nkeys; i++)
9100 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9101
9102 size = sizeof (dtrace_dynvar_t);
9103 size += sizeof (dtrace_key_t) * (nkeys - 1);
9104 size += ksize;
9105
9106 /*
9107 * Now we need to determine the size of the stored data.
9108 */
9109 id = DIF_INSTR_VAR(instr);
9110
9111 for (i = 0; i < dp->dtdo_varlen; i++) {
9112 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9113
9114 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9115 size += v->dtdv_type.dtdt_size;
9116 break;
9117 }
9118 }
9119
9120 if (i == dp->dtdo_varlen)
9121 return;
9122
9123 /*
9124 * We have the size. If this is larger than the chunk size
9125 * for our dynamic variable state, reset the chunk size.
9126 */
9127 size = P2ROUNDUP(size, sizeof (uint64_t));
9128
9129 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9130 vstate->dtvs_dynvars.dtds_chunksize = size;
9131 }
9132 }
9133
9134 static void
9135 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9136 {
9137 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9138 uint_t id;
9139
9140 ASSERT(MUTEX_HELD(&dtrace_lock));
9141 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9142
9143 for (i = 0; i < dp->dtdo_varlen; i++) {
9144 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9145 dtrace_statvar_t *svar, ***svarp;
9146 size_t dsize = 0;
9147 uint8_t scope = v->dtdv_scope;
9148 int *np;
9149
9150 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9151 continue;
9152
9153 id -= DIF_VAR_OTHER_UBASE;
9154
9155 switch (scope) {
9156 case DIFV_SCOPE_THREAD:
9157 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9158 dtrace_difv_t *tlocals;
9159
9160 if ((ntlocals = (otlocals << 1)) == 0)
9161 ntlocals = 1;
9162
9163 osz = otlocals * sizeof (dtrace_difv_t);
9164 nsz = ntlocals * sizeof (dtrace_difv_t);
9165
9166 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9167
9168 if (osz != 0) {
9169 bcopy(vstate->dtvs_tlocals,
9170 tlocals, osz);
9171 kmem_free(vstate->dtvs_tlocals, osz);
9172 }
9173
9174 vstate->dtvs_tlocals = tlocals;
9175 vstate->dtvs_ntlocals = ntlocals;
9176 }
9177
9178 vstate->dtvs_tlocals[id] = *v;
9179 continue;
9180
9181 case DIFV_SCOPE_LOCAL:
9182 np = &vstate->dtvs_nlocals;
9183 svarp = &vstate->dtvs_locals;
9184
9185 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9186 dsize = NCPU * (v->dtdv_type.dtdt_size +
9187 sizeof (uint64_t));
9188 else
9189 dsize = NCPU * sizeof (uint64_t);
9190
9191 break;
9192
9193 case DIFV_SCOPE_GLOBAL:
9194 np = &vstate->dtvs_nglobals;
9195 svarp = &vstate->dtvs_globals;
9196
9197 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9198 dsize = v->dtdv_type.dtdt_size +
9199 sizeof (uint64_t);
9200
9201 break;
9202
9203 default:
9204 ASSERT(0);
9205 }
9206
9207 while (id >= (oldsvars = *np)) {
9208 dtrace_statvar_t **statics;
9209 int newsvars, oldsize, newsize;
9210
9211 if ((newsvars = (oldsvars << 1)) == 0)
9212 newsvars = 1;
9213
9214 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9215 newsize = newsvars * sizeof (dtrace_statvar_t *);
9216
9217 statics = kmem_zalloc(newsize, KM_SLEEP);
9218
9219 if (oldsize != 0) {
9220 bcopy(*svarp, statics, oldsize);
9221 kmem_free(*svarp, oldsize);
9222 }
9223
9224 *svarp = statics;
9225 *np = newsvars;
9226 }
9227
9228 if ((svar = (*svarp)[id]) == NULL) {
9229 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9230 svar->dtsv_var = *v;
9231
9232 if ((svar->dtsv_size = dsize) != 0) {
9233 svar->dtsv_data = (uint64_t)(uintptr_t)
9234 kmem_zalloc(dsize, KM_SLEEP);
9235 }
9236
9237 (*svarp)[id] = svar;
9238 }
9239
9240 svar->dtsv_refcnt++;
9241 }
9242
9243 dtrace_difo_chunksize(dp, vstate);
9244 dtrace_difo_hold(dp);
9245 }
9246
9247 static dtrace_difo_t *
9248 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9249 {
9250 dtrace_difo_t *new;
9251 size_t sz;
9252
9253 ASSERT(dp->dtdo_buf != NULL);
9254 ASSERT(dp->dtdo_refcnt != 0);
9255
9256 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9257
9258 ASSERT(dp->dtdo_buf != NULL);
9259 sz = dp->dtdo_len * sizeof (dif_instr_t);
9260 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9261 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9262 new->dtdo_len = dp->dtdo_len;
9263
9264 if (dp->dtdo_strtab != NULL) {
9265 ASSERT(dp->dtdo_strlen != 0);
9266 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9267 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9268 new->dtdo_strlen = dp->dtdo_strlen;
9269 }
9270
9271 if (dp->dtdo_inttab != NULL) {
9272 ASSERT(dp->dtdo_intlen != 0);
9273 sz = dp->dtdo_intlen * sizeof (uint64_t);
9274 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9275 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9276 new->dtdo_intlen = dp->dtdo_intlen;
9277 }
9278
9279 if (dp->dtdo_vartab != NULL) {
9280 ASSERT(dp->dtdo_varlen != 0);
9281 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9282 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9283 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9284 new->dtdo_varlen = dp->dtdo_varlen;
9285 }
9286
9287 dtrace_difo_init(new, vstate);
9288 return (new);
9289 }
9290
9291 static void
9292 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9293 {
9294 int i;
9295
9296 ASSERT(dp->dtdo_refcnt == 0);
9297
9298 for (i = 0; i < dp->dtdo_varlen; i++) {
9299 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9300 dtrace_statvar_t *svar, **svarp;
9301 uint_t id;
9302 uint8_t scope = v->dtdv_scope;
9303 int *np;
9304
9305 switch (scope) {
9306 case DIFV_SCOPE_THREAD:
9307 continue;
9308
9309 case DIFV_SCOPE_LOCAL:
9310 np = &vstate->dtvs_nlocals;
9311 svarp = vstate->dtvs_locals;
9312 break;
9313
9314 case DIFV_SCOPE_GLOBAL:
9315 np = &vstate->dtvs_nglobals;
9316 svarp = vstate->dtvs_globals;
9317 break;
9318
9319 default:
9320 ASSERT(0);
9321 }
9322
9323 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9324 continue;
9325
9326 id -= DIF_VAR_OTHER_UBASE;
9327 ASSERT(id < *np);
9328
9329 svar = svarp[id];
9330 ASSERT(svar != NULL);
9331 ASSERT(svar->dtsv_refcnt > 0);
9332
9333 if (--svar->dtsv_refcnt > 0)
9334 continue;
9335
9336 if (svar->dtsv_size != 0) {
9337 ASSERT(svar->dtsv_data != NULL);
9338 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9339 svar->dtsv_size);
9340 }
9341
9342 kmem_free(svar, sizeof (dtrace_statvar_t));
9343 svarp[id] = NULL;
9344 }
9345
9346 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9347 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9348 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9349 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9350
9351 kmem_free(dp, sizeof (dtrace_difo_t));
9352 }
9353
9354 static void
9355 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9356 {
9357 int i;
9358
9359 ASSERT(MUTEX_HELD(&dtrace_lock));
9360 ASSERT(dp->dtdo_refcnt != 0);
9361
9362 for (i = 0; i < dp->dtdo_varlen; i++) {
9363 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9364
9365 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9366 continue;
9367
9368 ASSERT(dtrace_vtime_references > 0);
9369 if (--dtrace_vtime_references == 0)
9370 dtrace_vtime_disable();
9371 }
9372
9373 if (--dp->dtdo_refcnt == 0)
9374 dtrace_difo_destroy(dp, vstate);
9375 }
9376
9377 /*
9378 * DTrace Format Functions
9379 */
9380 static uint16_t
9381 dtrace_format_add(dtrace_state_t *state, char *str)
9382 {
9383 char *fmt, **new;
9384 uint16_t ndx, len = strlen(str) + 1;
9385
9386 fmt = kmem_zalloc(len, KM_SLEEP);
9387 bcopy(str, fmt, len);
9388
9389 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9390 if (state->dts_formats[ndx] == NULL) {
9391 state->dts_formats[ndx] = fmt;
9392 return (ndx + 1);
9393 }
9394 }
9395
9396 if (state->dts_nformats == USHRT_MAX) {
9397 /*
9398 * This is only likely if a denial-of-service attack is being
9399 * attempted. As such, it's okay to fail silently here.
9400 */
9401 kmem_free(fmt, len);
9402 return (0);
9403 }
9404
9405 /*
9406 * For simplicity, we always resize the formats array to be exactly the
9407 * number of formats.
9408 */
9409 ndx = state->dts_nformats++;
9410 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9411
9412 if (state->dts_formats != NULL) {
9413 ASSERT(ndx != 0);
9414 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9415 kmem_free(state->dts_formats, ndx * sizeof (char *));
9416 }
9417
9418 state->dts_formats = new;
9419 state->dts_formats[ndx] = fmt;
9420
9421 return (ndx + 1);
9422 }
9423
9424 static void
9425 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9426 {
9427 char *fmt;
9428
9429 ASSERT(state->dts_formats != NULL);
9430 ASSERT(format <= state->dts_nformats);
9431 ASSERT(state->dts_formats[format - 1] != NULL);
9432
9433 fmt = state->dts_formats[format - 1];
9434 kmem_free(fmt, strlen(fmt) + 1);
9435 state->dts_formats[format - 1] = NULL;
9436 }
9437
9438 static void
9439 dtrace_format_destroy(dtrace_state_t *state)
9440 {
9441 int i;
9442
9443 if (state->dts_nformats == 0) {
9444 ASSERT(state->dts_formats == NULL);
9445 return;
9446 }
9447
9448 ASSERT(state->dts_formats != NULL);
9449
9450 for (i = 0; i < state->dts_nformats; i++) {
9451 char *fmt = state->dts_formats[i];
9452
9453 if (fmt == NULL)
9454 continue;
9455
9456 kmem_free(fmt, strlen(fmt) + 1);
9457 }
9458
9459 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9460 state->dts_nformats = 0;
9461 state->dts_formats = NULL;
9462 }
9463
9464 /*
9465 * DTrace Predicate Functions
9466 */
9467 static dtrace_predicate_t *
9468 dtrace_predicate_create(dtrace_difo_t *dp)
9469 {
9470 dtrace_predicate_t *pred;
9471
9472 ASSERT(MUTEX_HELD(&dtrace_lock));
9473 ASSERT(dp->dtdo_refcnt != 0);
9474
9475 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9476 pred->dtp_difo = dp;
9477 pred->dtp_refcnt = 1;
9478
9479 if (!dtrace_difo_cacheable(dp))
9480 return (pred);
9481
9482 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9483 /*
9484 * This is only theoretically possible -- we have had 2^32
9485 * cacheable predicates on this machine. We cannot allow any
9486 * more predicates to become cacheable: as unlikely as it is,
9487 * there may be a thread caching a (now stale) predicate cache
9488 * ID. (N.B.: the temptation is being successfully resisted to
9489 * have this cmn_err() "Holy shit -- we executed this code!")
9490 */
9491 return (pred);
9492 }
9493
9494 pred->dtp_cacheid = dtrace_predcache_id++;
9495
9496 return (pred);
9497 }
9498
9499 static void
9500 dtrace_predicate_hold(dtrace_predicate_t *pred)
9501 {
9502 ASSERT(MUTEX_HELD(&dtrace_lock));
9503 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9504 ASSERT(pred->dtp_refcnt > 0);
9505
9506 pred->dtp_refcnt++;
9507 }
9508
9509 static void
9510 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9511 {
9512 dtrace_difo_t *dp = pred->dtp_difo;
9513
9514 ASSERT(MUTEX_HELD(&dtrace_lock));
9515 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9516 ASSERT(pred->dtp_refcnt > 0);
9517
9518 if (--pred->dtp_refcnt == 0) {
9519 dtrace_difo_release(pred->dtp_difo, vstate);
9520 kmem_free(pred, sizeof (dtrace_predicate_t));
9521 }
9522 }
9523
9524 /*
9525 * DTrace Action Description Functions
9526 */
9527 static dtrace_actdesc_t *
9528 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9529 uint64_t uarg, uint64_t arg)
9530 {
9531 dtrace_actdesc_t *act;
9532
9533 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9534 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9535
9536 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9537 act->dtad_kind = kind;
9538 act->dtad_ntuple = ntuple;
9539 act->dtad_uarg = uarg;
9540 act->dtad_arg = arg;
9541 act->dtad_refcnt = 1;
9542
9543 return (act);
9544 }
9545
9546 static void
9547 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9548 {
9549 ASSERT(act->dtad_refcnt >= 1);
9550 act->dtad_refcnt++;
9551 }
9552
9553 static void
9554 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9555 {
9556 dtrace_actkind_t kind = act->dtad_kind;
9557 dtrace_difo_t *dp;
9558
9559 ASSERT(act->dtad_refcnt >= 1);
9560
9561 if (--act->dtad_refcnt != 0)
9562 return;
9563
9564 if ((dp = act->dtad_difo) != NULL)
9565 dtrace_difo_release(dp, vstate);
9566
9567 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9568 char *str = (char *)(uintptr_t)act->dtad_arg;
9569
9570 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9571 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9572
9573 if (str != NULL)
9574 kmem_free(str, strlen(str) + 1);
9575 }
9576
9577 kmem_free(act, sizeof (dtrace_actdesc_t));
9578 }
9579
9580 /*
9581 * DTrace ECB Functions
9582 */
9583 static dtrace_ecb_t *
9584 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9585 {
9586 dtrace_ecb_t *ecb;
9587 dtrace_epid_t epid;
9588
9589 ASSERT(MUTEX_HELD(&dtrace_lock));
9590
9591 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9592 ecb->dte_predicate = NULL;
9593 ecb->dte_probe = probe;
9594
9595 /*
9596 * The default size is the size of the default action: recording
9597 * the header.
9598 */
9599 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
9600 ecb->dte_alignment = sizeof (dtrace_epid_t);
9601
9602 epid = state->dts_epid++;
9603
9604 if (epid - 1 >= state->dts_necbs) {
9605 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9606 int necbs = state->dts_necbs << 1;
9607
9608 ASSERT(epid == state->dts_necbs + 1);
9609
9610 if (necbs == 0) {
9611 ASSERT(oecbs == NULL);
9612 necbs = 1;
9613 }
9614
9615 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9616
9617 if (oecbs != NULL)
9618 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9619
9620 dtrace_membar_producer();
9621 state->dts_ecbs = ecbs;
9622
9623 if (oecbs != NULL) {
9624 /*
9625 * If this state is active, we must dtrace_sync()
9626 * before we can free the old dts_ecbs array: we're
9627 * coming in hot, and there may be active ring
9628 * buffer processing (which indexes into the dts_ecbs
9629 * array) on another CPU.
9630 */
9631 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9632 dtrace_sync();
9633
9634 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9635 }
9636
9637 dtrace_membar_producer();
9638 state->dts_necbs = necbs;
9639 }
9640
9641 ecb->dte_state = state;
9642
9643 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9644 dtrace_membar_producer();
9645 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9646
9647 return (ecb);
9648 }
9649
9650 static int
9651 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9652 {
9653 dtrace_probe_t *probe = ecb->dte_probe;
9654
9655 ASSERT(MUTEX_HELD(&cpu_lock));
9656 ASSERT(MUTEX_HELD(&dtrace_lock));
9657 ASSERT(ecb->dte_next == NULL);
9658
9659 if (probe == NULL) {
9660 /*
9661 * This is the NULL probe -- there's nothing to do.
9662 */
9663 return (0);
9664 }
9665
9666 if (probe->dtpr_ecb == NULL) {
9667 dtrace_provider_t *prov = probe->dtpr_provider;
9668
9669 /*
9670 * We're the first ECB on this probe.
9671 */
9672 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9673
9674 if (ecb->dte_predicate != NULL)
9675 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9676
9677 return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9678 probe->dtpr_id, probe->dtpr_arg));
9679 } else {
9680 /*
9681 * This probe is already active. Swing the last pointer to
9682 * point to the new ECB, and issue a dtrace_sync() to assure
9683 * that all CPUs have seen the change.
9684 */
9685 ASSERT(probe->dtpr_ecb_last != NULL);
9686 probe->dtpr_ecb_last->dte_next = ecb;
9687 probe->dtpr_ecb_last = ecb;
9688 probe->dtpr_predcache = 0;
9689
9690 dtrace_sync();
9691 return (0);
9692 }
9693 }
9694
9695 static void
9696 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9697 {
9698 dtrace_action_t *act;
9699 uint32_t curneeded = UINT32_MAX;
9700 uint32_t aggbase = UINT32_MAX;
9701
9702 /*
9703 * If we record anything, we always record the dtrace_rechdr_t. (And
9704 * we always record it first.)
9705 */
9706 ecb->dte_size = sizeof (dtrace_rechdr_t);
9707 ecb->dte_alignment = sizeof (dtrace_epid_t);
9708
9709 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9710 dtrace_recdesc_t *rec = &act->dta_rec;
9711 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
9712
9713 ecb->dte_alignment = MAX(ecb->dte_alignment,
9714 rec->dtrd_alignment);
9715
9716 if (DTRACEACT_ISAGG(act->dta_kind)) {
9717 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9718
9719 ASSERT(rec->dtrd_size != 0);
9720 ASSERT(agg->dtag_first != NULL);
9721 ASSERT(act->dta_prev->dta_intuple);
9722 ASSERT(aggbase != UINT32_MAX);
9723 ASSERT(curneeded != UINT32_MAX);
9724
9725 agg->dtag_base = aggbase;
9726
9727 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9728 rec->dtrd_offset = curneeded;
9729 curneeded += rec->dtrd_size;
9730 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
9731
9732 aggbase = UINT32_MAX;
9733 curneeded = UINT32_MAX;
9734 } else if (act->dta_intuple) {
9735 if (curneeded == UINT32_MAX) {
9736 /*
9737 * This is the first record in a tuple. Align
9738 * curneeded to be at offset 4 in an 8-byte
9739 * aligned block.
9740 */
9741 ASSERT(act->dta_prev == NULL ||
9742 !act->dta_prev->dta_intuple);
9743 ASSERT3U(aggbase, ==, UINT32_MAX);
9744 curneeded = P2PHASEUP(ecb->dte_size,
9745 sizeof (uint64_t), sizeof (dtrace_aggid_t));
9746
9747 aggbase = curneeded - sizeof (dtrace_aggid_t);
9748 ASSERT(IS_P2ALIGNED(aggbase,
9749 sizeof (uint64_t)));
9750 }
9751 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9752 rec->dtrd_offset = curneeded;
9753 curneeded += rec->dtrd_size;
9754 } else {
9755 /* tuples must be followed by an aggregation */
9756 ASSERT(act->dta_prev == NULL ||
9757 !act->dta_prev->dta_intuple);
9758
9759 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
9760 rec->dtrd_alignment);
9761 rec->dtrd_offset = ecb->dte_size;
9762 ecb->dte_size += rec->dtrd_size;
9763 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
9764 }
9765 }
9766
9767 if ((act = ecb->dte_action) != NULL &&
9768 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9769 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
9770 /*
9771 * If the size is still sizeof (dtrace_rechdr_t), then all
9772 * actions store no data; set the size to 0.
9773 */
9774 ecb->dte_size = 0;
9775 }
9776
9777 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
9778 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
9779 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
9780 ecb->dte_needed);
9781 }
9782
9783 static dtrace_action_t *
9784 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9785 {
9786 dtrace_aggregation_t *agg;
9787 size_t size = sizeof (uint64_t);
9788 int ntuple = desc->dtad_ntuple;
9789 dtrace_action_t *act;
9790 dtrace_recdesc_t *frec;
9791 dtrace_aggid_t aggid;
9792 dtrace_state_t *state = ecb->dte_state;
9793
9794 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9795 agg->dtag_ecb = ecb;
9796
9797 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9798
9799 switch (desc->dtad_kind) {
9800 case DTRACEAGG_MIN:
9801 agg->dtag_initial = INT64_MAX;
9802 agg->dtag_aggregate = dtrace_aggregate_min;
9803 break;
9804
9805 case DTRACEAGG_MAX:
9806 agg->dtag_initial = INT64_MIN;
9807 agg->dtag_aggregate = dtrace_aggregate_max;
9808 break;
9809
9810 case DTRACEAGG_COUNT:
9811 agg->dtag_aggregate = dtrace_aggregate_count;
9812 break;
9813
9814 case DTRACEAGG_QUANTIZE:
9815 agg->dtag_aggregate = dtrace_aggregate_quantize;
9816 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9817 sizeof (uint64_t);
9818 break;
9819
9820 case DTRACEAGG_LQUANTIZE: {
9821 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9822 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9823
9824 agg->dtag_initial = desc->dtad_arg;
9825 agg->dtag_aggregate = dtrace_aggregate_lquantize;
9826
9827 if (step == 0 || levels == 0)
9828 goto err;
9829
9830 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9831 break;
9832 }
9833
9834 case DTRACEAGG_LLQUANTIZE: {
9835 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
9836 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
9837 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
9838 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
9839 int64_t v;
9840
9841 agg->dtag_initial = desc->dtad_arg;
9842 agg->dtag_aggregate = dtrace_aggregate_llquantize;
9843
9844 if (factor < 2 || low >= high || nsteps < factor)
9845 goto err;
9846
9847 /*
9848 * Now check that the number of steps evenly divides a power
9849 * of the factor. (This assures both integer bucket size and
9850 * linearity within each magnitude.)
9851 */
9852 for (v = factor; v < nsteps; v *= factor)
9853 continue;
9854
9855 if ((v % nsteps) || (nsteps % factor))
9856 goto err;
9857
9858 size = (dtrace_aggregate_llquantize_bucket(factor,
9859 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
9860 break;
9861 }
9862
9863 case DTRACEAGG_AVG:
9864 agg->dtag_aggregate = dtrace_aggregate_avg;
9865 size = sizeof (uint64_t) * 2;
9866 break;
9867
9868 case DTRACEAGG_STDDEV:
9869 agg->dtag_aggregate = dtrace_aggregate_stddev;
9870 size = sizeof (uint64_t) * 4;
9871 break;
9872
9873 case DTRACEAGG_SUM:
9874 agg->dtag_aggregate = dtrace_aggregate_sum;
9875 break;
9876
9877 default:
9878 goto err;
9879 }
9880
9881 agg->dtag_action.dta_rec.dtrd_size = size;
9882
9883 if (ntuple == 0)
9884 goto err;
9885
9886 /*
9887 * We must make sure that we have enough actions for the n-tuple.
9888 */
9889 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9890 if (DTRACEACT_ISAGG(act->dta_kind))
9891 break;
9892
9893 if (--ntuple == 0) {
9894 /*
9895 * This is the action with which our n-tuple begins.
9896 */
9897 agg->dtag_first = act;
9898 goto success;
9899 }
9900 }
9901
9902 /*
9903 * This n-tuple is short by ntuple elements. Return failure.
9904 */
9905 ASSERT(ntuple != 0);
9906 err:
9907 kmem_free(agg, sizeof (dtrace_aggregation_t));
9908 return (NULL);
9909
9910 success:
9911 /*
9912 * If the last action in the tuple has a size of zero, it's actually
9913 * an expression argument for the aggregating action.
9914 */
9915 ASSERT(ecb->dte_action_last != NULL);
9916 act = ecb->dte_action_last;
9917
9918 if (act->dta_kind == DTRACEACT_DIFEXPR) {
9919 ASSERT(act->dta_difo != NULL);
9920
9921 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
9922 agg->dtag_hasarg = 1;
9923 }
9924
9925 /*
9926 * We need to allocate an id for this aggregation.
9927 */
9928 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
9929 VM_BESTFIT | VM_SLEEP);
9930
9931 if (aggid - 1 >= state->dts_naggregations) {
9932 dtrace_aggregation_t **oaggs = state->dts_aggregations;
9933 dtrace_aggregation_t **aggs;
9934 int naggs = state->dts_naggregations << 1;
9935 int onaggs = state->dts_naggregations;
9936
9937 ASSERT(aggid == state->dts_naggregations + 1);
9938
9939 if (naggs == 0) {
9940 ASSERT(oaggs == NULL);
9941 naggs = 1;
9942 }
9943
9944 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
9945
9946 if (oaggs != NULL) {
9947 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
9948 kmem_free(oaggs, onaggs * sizeof (*aggs));
9949 }
9950
9951 state->dts_aggregations = aggs;
9952 state->dts_naggregations = naggs;
9953 }
9954
9955 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
9956 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
9957
9958 frec = &agg->dtag_first->dta_rec;
9959 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
9960 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
9961
9962 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
9963 ASSERT(!act->dta_intuple);
9964 act->dta_intuple = 1;
9965 }
9966
9967 return (&agg->dtag_action);
9968 }
9969
9970 static void
9971 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
9972 {
9973 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9974 dtrace_state_t *state = ecb->dte_state;
9975 dtrace_aggid_t aggid = agg->dtag_id;
9976
9977 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
9978 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
9979
9980 ASSERT(state->dts_aggregations[aggid - 1] == agg);
9981 state->dts_aggregations[aggid - 1] = NULL;
9982
9983 kmem_free(agg, sizeof (dtrace_aggregation_t));
9984 }
9985
9986 static int
9987 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9988 {
9989 dtrace_action_t *action, *last;
9990 dtrace_difo_t *dp = desc->dtad_difo;
9991 uint32_t size = 0, align = sizeof (uint8_t), mask;
9992 uint16_t format = 0;
9993 dtrace_recdesc_t *rec;
9994 dtrace_state_t *state = ecb->dte_state;
9995 dtrace_optval_t *opt = state->dts_options, nframes, strsize;
9996 uint64_t arg = desc->dtad_arg;
9997
9998 ASSERT(MUTEX_HELD(&dtrace_lock));
9999 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10000
10001 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10002 /*
10003 * If this is an aggregating action, there must be neither
10004 * a speculate nor a commit on the action chain.
10005 */
10006 dtrace_action_t *act;
10007
10008 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10009 if (act->dta_kind == DTRACEACT_COMMIT)
10010 return (EINVAL);
10011
10012 if (act->dta_kind == DTRACEACT_SPECULATE)
10013 return (EINVAL);
10014 }
10015
10016 action = dtrace_ecb_aggregation_create(ecb, desc);
10017
10018 if (action == NULL)
10019 return (EINVAL);
10020 } else {
10021 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10022 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10023 dp != NULL && dp->dtdo_destructive)) {
10024 state->dts_destructive = 1;
10025 }
10026
10027 switch (desc->dtad_kind) {
10028 case DTRACEACT_PRINTF:
10029 case DTRACEACT_PRINTA:
10030 case DTRACEACT_SYSTEM:
10031 case DTRACEACT_FREOPEN:
10032 case DTRACEACT_DIFEXPR:
10033 /*
10034 * We know that our arg is a string -- turn it into a
10035 * format.
10036 */
10037 if (arg == NULL) {
10038 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10039 desc->dtad_kind == DTRACEACT_DIFEXPR);
10040 format = 0;
10041 } else {
10042 ASSERT(arg != NULL);
10043 ASSERT(arg > KERNELBASE);
10044 format = dtrace_format_add(state,
10045 (char *)(uintptr_t)arg);
10046 }
10047
10048 /*FALLTHROUGH*/
10049 case DTRACEACT_LIBACT:
10050 case DTRACEACT_TRACEMEM:
10051 case DTRACEACT_TRACEMEM_DYNSIZE:
10052 if (dp == NULL)
10053 return (EINVAL);
10054
10055 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10056 break;
10057
10058 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10059 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10060 return (EINVAL);
10061
10062 size = opt[DTRACEOPT_STRSIZE];
10063 }
10064
10065 break;
10066
10067 case DTRACEACT_STACK:
10068 if ((nframes = arg) == 0) {
10069 nframes = opt[DTRACEOPT_STACKFRAMES];
10070 ASSERT(nframes > 0);
10071 arg = nframes;
10072 }
10073
10074 size = nframes * sizeof (pc_t);
10075 break;
10076
10077 case DTRACEACT_JSTACK:
10078 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10079 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10080
10081 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10082 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10083
10084 arg = DTRACE_USTACK_ARG(nframes, strsize);
10085
10086 /*FALLTHROUGH*/
10087 case DTRACEACT_USTACK:
10088 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10089 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10090 strsize = DTRACE_USTACK_STRSIZE(arg);
10091 nframes = opt[DTRACEOPT_USTACKFRAMES];
10092 ASSERT(nframes > 0);
10093 arg = DTRACE_USTACK_ARG(nframes, strsize);
10094 }
10095
10096 /*
10097 * Save a slot for the pid.
10098 */
10099 size = (nframes + 1) * sizeof (uint64_t);
10100 size += DTRACE_USTACK_STRSIZE(arg);
10101 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10102
10103 break;
10104
10105 case DTRACEACT_SYM:
10106 case DTRACEACT_MOD:
10107 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10108 sizeof (uint64_t)) ||
10109 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10110 return (EINVAL);
10111 break;
10112
10113 case DTRACEACT_USYM:
10114 case DTRACEACT_UMOD:
10115 case DTRACEACT_UADDR:
10116 if (dp == NULL ||
10117 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10118 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10119 return (EINVAL);
10120
10121 /*
10122 * We have a slot for the pid, plus a slot for the
10123 * argument. To keep things simple (aligned with
10124 * bitness-neutral sizing), we store each as a 64-bit
10125 * quantity.
10126 */
10127 size = 2 * sizeof (uint64_t);
10128 break;
10129
10130 case DTRACEACT_STOP:
10131 case DTRACEACT_BREAKPOINT:
10132 case DTRACEACT_PANIC:
10133 break;
10134
10135 case DTRACEACT_CHILL:
10136 case DTRACEACT_DISCARD:
10137 case DTRACEACT_RAISE:
10138 if (dp == NULL)
10139 return (EINVAL);
10140 break;
10141
10142 case DTRACEACT_EXIT:
10143 if (dp == NULL ||
10144 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10145 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10146 return (EINVAL);
10147 break;
10148
10149 case DTRACEACT_SPECULATE:
10150 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10151 return (EINVAL);
10152
10153 if (dp == NULL)
10154 return (EINVAL);
10155
10156 state->dts_speculates = 1;
10157 break;
10158
10159 case DTRACEACT_COMMIT: {
10160 dtrace_action_t *act = ecb->dte_action;
10161
10162 for (; act != NULL; act = act->dta_next) {
10163 if (act->dta_kind == DTRACEACT_COMMIT)
10164 return (EINVAL);
10165 }
10166
10167 if (dp == NULL)
10168 return (EINVAL);
10169 break;
10170 }
10171
10172 default:
10173 return (EINVAL);
10174 }
10175
10176 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10177 /*
10178 * If this is a data-storing action or a speculate,
10179 * we must be sure that there isn't a commit on the
10180 * action chain.
10181 */
10182 dtrace_action_t *act = ecb->dte_action;
10183
10184 for (; act != NULL; act = act->dta_next) {
10185 if (act->dta_kind == DTRACEACT_COMMIT)
10186 return (EINVAL);
10187 }
10188 }
10189
10190 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10191 action->dta_rec.dtrd_size = size;
10192 }
10193
10194 action->dta_refcnt = 1;
10195 rec = &action->dta_rec;
10196 size = rec->dtrd_size;
10197
10198 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10199 if (!(size & mask)) {
10200 align = mask + 1;
10201 break;
10202 }
10203 }
10204
10205 action->dta_kind = desc->dtad_kind;
10206
10207 if ((action->dta_difo = dp) != NULL)
10208 dtrace_difo_hold(dp);
10209
10210 rec->dtrd_action = action->dta_kind;
10211 rec->dtrd_arg = arg;
10212 rec->dtrd_uarg = desc->dtad_uarg;
10213 rec->dtrd_alignment = (uint16_t)align;
10214 rec->dtrd_format = format;
10215
10216 if ((last = ecb->dte_action_last) != NULL) {
10217 ASSERT(ecb->dte_action != NULL);
10218 action->dta_prev = last;
10219 last->dta_next = action;
10220 } else {
10221 ASSERT(ecb->dte_action == NULL);
10222 ecb->dte_action = action;
10223 }
10224
10225 ecb->dte_action_last = action;
10226
10227 return (0);
10228 }
10229
10230 static void
10231 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10232 {
10233 dtrace_action_t *act = ecb->dte_action, *next;
10234 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10235 dtrace_difo_t *dp;
10236 uint16_t format;
10237
10238 if (act != NULL && act->dta_refcnt > 1) {
10239 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10240 act->dta_refcnt--;
10241 } else {
10242 for (; act != NULL; act = next) {
10243 next = act->dta_next;
10244 ASSERT(next != NULL || act == ecb->dte_action_last);
10245 ASSERT(act->dta_refcnt == 1);
10246
10247 if ((format = act->dta_rec.dtrd_format) != 0)
10248 dtrace_format_remove(ecb->dte_state, format);
10249
10250 if ((dp = act->dta_difo) != NULL)
10251 dtrace_difo_release(dp, vstate);
10252
10253 if (DTRACEACT_ISAGG(act->dta_kind)) {
10254 dtrace_ecb_aggregation_destroy(ecb, act);
10255 } else {
10256 kmem_free(act, sizeof (dtrace_action_t));
10257 }
10258 }
10259 }
10260
10261 ecb->dte_action = NULL;
10262 ecb->dte_action_last = NULL;
10263 ecb->dte_size = 0;
10264 }
10265
10266 static void
10267 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10268 {
10269 /*
10270 * We disable the ECB by removing it from its probe.
10271 */
10272 dtrace_ecb_t *pecb, *prev = NULL;
10273 dtrace_probe_t *probe = ecb->dte_probe;
10274
10275 ASSERT(MUTEX_HELD(&dtrace_lock));
10276
10277 if (probe == NULL) {
10278 /*
10279 * This is the NULL probe; there is nothing to disable.
10280 */
10281 return;
10282 }
10283
10284 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10285 if (pecb == ecb)
10286 break;
10287 prev = pecb;
10288 }
10289
10290 ASSERT(pecb != NULL);
10291
10292 if (prev == NULL) {
10293 probe->dtpr_ecb = ecb->dte_next;
10294 } else {
10295 prev->dte_next = ecb->dte_next;
10296 }
10297
10298 if (ecb == probe->dtpr_ecb_last) {
10299 ASSERT(ecb->dte_next == NULL);
10300 probe->dtpr_ecb_last = prev;
10301 }
10302
10303 /*
10304 * The ECB has been disconnected from the probe; now sync to assure
10305 * that all CPUs have seen the change before returning.
10306 */
10307 dtrace_sync();
10308
10309 if (probe->dtpr_ecb == NULL) {
10310 /*
10311 * That was the last ECB on the probe; clear the predicate
10312 * cache ID for the probe, disable it and sync one more time
10313 * to assure that we'll never hit it again.
10314 */
10315 dtrace_provider_t *prov = probe->dtpr_provider;
10316
10317 ASSERT(ecb->dte_next == NULL);
10318 ASSERT(probe->dtpr_ecb_last == NULL);
10319 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10320 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10321 probe->dtpr_id, probe->dtpr_arg);
10322 dtrace_sync();
10323 } else {
10324 /*
10325 * There is at least one ECB remaining on the probe. If there
10326 * is _exactly_ one, set the probe's predicate cache ID to be
10327 * the predicate cache ID of the remaining ECB.
10328 */
10329 ASSERT(probe->dtpr_ecb_last != NULL);
10330 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10331
10332 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10333 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10334
10335 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10336
10337 if (p != NULL)
10338 probe->dtpr_predcache = p->dtp_cacheid;
10339 }
10340
10341 ecb->dte_next = NULL;
10342 }
10343 }
10344
10345 static void
10346 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10347 {
10348 dtrace_state_t *state = ecb->dte_state;
10349 dtrace_vstate_t *vstate = &state->dts_vstate;
10350 dtrace_predicate_t *pred;
10351 dtrace_epid_t epid = ecb->dte_epid;
10352
10353 ASSERT(MUTEX_HELD(&dtrace_lock));
10354 ASSERT(ecb->dte_next == NULL);
10355 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10356
10357 if ((pred = ecb->dte_predicate) != NULL)
10358 dtrace_predicate_release(pred, vstate);
10359
10360 dtrace_ecb_action_remove(ecb);
10361
10362 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10363 state->dts_ecbs[epid - 1] = NULL;
10364
10365 kmem_free(ecb, sizeof (dtrace_ecb_t));
10366 }
10367
10368 static dtrace_ecb_t *
10369 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10370 dtrace_enabling_t *enab)
10371 {
10372 dtrace_ecb_t *ecb;
10373 dtrace_predicate_t *pred;
10374 dtrace_actdesc_t *act;
10375 dtrace_provider_t *prov;
10376 dtrace_ecbdesc_t *desc = enab->dten_current;
10377
10378 ASSERT(MUTEX_HELD(&dtrace_lock));
10379 ASSERT(state != NULL);
10380
10381 ecb = dtrace_ecb_add(state, probe);
10382 ecb->dte_uarg = desc->dted_uarg;
10383
10384 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10385 dtrace_predicate_hold(pred);
10386 ecb->dte_predicate = pred;
10387 }
10388
10389 if (probe != NULL) {
10390 /*
10391 * If the provider shows more leg than the consumer is old
10392 * enough to see, we need to enable the appropriate implicit
10393 * predicate bits to prevent the ecb from activating at
10394 * revealing times.
10395 *
10396 * Providers specifying DTRACE_PRIV_USER at register time
10397 * are stating that they need the /proc-style privilege
10398 * model to be enforced, and this is what DTRACE_COND_OWNER
10399 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10400 */
10401 prov = probe->dtpr_provider;
10402 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10403 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10404 ecb->dte_cond |= DTRACE_COND_OWNER;
10405
10406 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10407 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10408 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10409
10410 /*
10411 * If the provider shows us kernel innards and the user
10412 * is lacking sufficient privilege, enable the
10413 * DTRACE_COND_USERMODE implicit predicate.
10414 */
10415 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10416 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10417 ecb->dte_cond |= DTRACE_COND_USERMODE;
10418 }
10419
10420 if (dtrace_ecb_create_cache != NULL) {
10421 /*
10422 * If we have a cached ecb, we'll use its action list instead
10423 * of creating our own (saving both time and space).
10424 */
10425 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10426 dtrace_action_t *act = cached->dte_action;
10427
10428 if (act != NULL) {
10429 ASSERT(act->dta_refcnt > 0);
10430 act->dta_refcnt++;
10431 ecb->dte_action = act;
10432 ecb->dte_action_last = cached->dte_action_last;
10433 ecb->dte_needed = cached->dte_needed;
10434 ecb->dte_size = cached->dte_size;
10435 ecb->dte_alignment = cached->dte_alignment;
10436 }
10437
10438 return (ecb);
10439 }
10440
10441 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10442 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10443 dtrace_ecb_destroy(ecb);
10444 return (NULL);
10445 }
10446 }
10447
10448 dtrace_ecb_resize(ecb);
10449
10450 return (dtrace_ecb_create_cache = ecb);
10451 }
10452
10453 static int
10454 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10455 {
10456 dtrace_ecb_t *ecb;
10457 dtrace_enabling_t *enab = arg;
10458 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10459
10460 ASSERT(state != NULL);
10461
10462 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10463 /*
10464 * This probe was created in a generation for which this
10465 * enabling has previously created ECBs; we don't want to
10466 * enable it again, so just kick out.
10467 */
10468 return (DTRACE_MATCH_NEXT);
10469 }
10470
10471 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10472 return (DTRACE_MATCH_DONE);
10473
10474 if (dtrace_ecb_enable(ecb) < 0)
10475 return (DTRACE_MATCH_FAIL);
10476
10477 return (DTRACE_MATCH_NEXT);
10478 }
10479
10480 static dtrace_ecb_t *
10481 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10482 {
10483 dtrace_ecb_t *ecb;
10484
10485 ASSERT(MUTEX_HELD(&dtrace_lock));
10486
10487 if (id == 0 || id > state->dts_necbs)
10488 return (NULL);
10489
10490 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10491 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10492
10493 return (state->dts_ecbs[id - 1]);
10494 }
10495
10496 static dtrace_aggregation_t *
10497 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10498 {
10499 dtrace_aggregation_t *agg;
10500
10501 ASSERT(MUTEX_HELD(&dtrace_lock));
10502
10503 if (id == 0 || id > state->dts_naggregations)
10504 return (NULL);
10505
10506 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10507 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10508 agg->dtag_id == id);
10509
10510 return (state->dts_aggregations[id - 1]);
10511 }
10512
10513 /*
10514 * DTrace Buffer Functions
10515 *
10516 * The following functions manipulate DTrace buffers. Most of these functions
10517 * are called in the context of establishing or processing consumer state;
10518 * exceptions are explicitly noted.
10519 */
10520
10521 /*
10522 * Note: called from cross call context. This function switches the two
10523 * buffers on a given CPU. The atomicity of this operation is assured by
10524 * disabling interrupts while the actual switch takes place; the disabling of
10525 * interrupts serializes the execution with any execution of dtrace_probe() on
10526 * the same CPU.
10527 */
10528 static void
10529 dtrace_buffer_switch(dtrace_buffer_t *buf)
10530 {
10531 caddr_t tomax = buf->dtb_tomax;
10532 caddr_t xamot = buf->dtb_xamot;
10533 dtrace_icookie_t cookie;
10534 hrtime_t now;
10535
10536 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10537 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10538
10539 cookie = dtrace_interrupt_disable();
10540 now = dtrace_gethrtime();
10541 buf->dtb_tomax = xamot;
10542 buf->dtb_xamot = tomax;
10543 buf->dtb_xamot_drops = buf->dtb_drops;
10544 buf->dtb_xamot_offset = buf->dtb_offset;
10545 buf->dtb_xamot_errors = buf->dtb_errors;
10546 buf->dtb_xamot_flags = buf->dtb_flags;
10547 buf->dtb_offset = 0;
10548 buf->dtb_drops = 0;
10549 buf->dtb_errors = 0;
10550 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10551 buf->dtb_interval = now - buf->dtb_switched;
10552 buf->dtb_switched = now;
10553 dtrace_interrupt_enable(cookie);
10554 }
10555
10556 /*
10557 * Note: called from cross call context. This function activates a buffer
10558 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10559 * is guaranteed by the disabling of interrupts.
10560 */
10561 static void
10562 dtrace_buffer_activate(dtrace_state_t *state)
10563 {
10564 dtrace_buffer_t *buf;
10565 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10566
10567 buf = &state->dts_buffer[CPU->cpu_id];
10568
10569 if (buf->dtb_tomax != NULL) {
10570 /*
10571 * We might like to assert that the buffer is marked inactive,
10572 * but this isn't necessarily true: the buffer for the CPU
10573 * that processes the BEGIN probe has its buffer activated
10574 * manually. In this case, we take the (harmless) action
10575 * re-clearing the bit INACTIVE bit.
10576 */
10577 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10578 }
10579
10580 dtrace_interrupt_enable(cookie);
10581 }
10582
10583 static int
10584 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10585 processorid_t cpu, int *factor)
10586 {
10587 cpu_t *cp;
10588 dtrace_buffer_t *buf;
10589 int allocated = 0, desired = 0;
10590
10591 ASSERT(MUTEX_HELD(&cpu_lock));
10592 ASSERT(MUTEX_HELD(&dtrace_lock));
10593
10594 *factor = 1;
10595
10596 if (size > dtrace_nonroot_maxsize &&
10597 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10598 return (EFBIG);
10599
10600 cp = cpu_list;
10601
10602 do {
10603 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10604 continue;
10605
10606 buf = &bufs[cp->cpu_id];
10607
10608 /*
10609 * If there is already a buffer allocated for this CPU, it
10610 * is only possible that this is a DR event. In this case,
10611 * the buffer size must match our specified size.
10612 */
10613 if (buf->dtb_tomax != NULL) {
10614 ASSERT(buf->dtb_size == size);
10615 continue;
10616 }
10617
10618 ASSERT(buf->dtb_xamot == NULL);
10619
10620 if ((buf->dtb_tomax = kmem_zalloc(size,
10621 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10622 goto err;
10623
10624 buf->dtb_size = size;
10625 buf->dtb_flags = flags;
10626 buf->dtb_offset = 0;
10627 buf->dtb_drops = 0;
10628
10629 if (flags & DTRACEBUF_NOSWITCH)
10630 continue;
10631
10632 if ((buf->dtb_xamot = kmem_zalloc(size,
10633 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10634 goto err;
10635 } while ((cp = cp->cpu_next) != cpu_list);
10636
10637 return (0);
10638
10639 err:
10640 cp = cpu_list;
10641
10642 do {
10643 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10644 continue;
10645
10646 buf = &bufs[cp->cpu_id];
10647 desired += 2;
10648
10649 if (buf->dtb_xamot != NULL) {
10650 ASSERT(buf->dtb_tomax != NULL);
10651 ASSERT(buf->dtb_size == size);
10652 kmem_free(buf->dtb_xamot, size);
10653 allocated++;
10654 }
10655
10656 if (buf->dtb_tomax != NULL) {
10657 ASSERT(buf->dtb_size == size);
10658 kmem_free(buf->dtb_tomax, size);
10659 allocated++;
10660 }
10661
10662 buf->dtb_tomax = NULL;
10663 buf->dtb_xamot = NULL;
10664 buf->dtb_size = 0;
10665 } while ((cp = cp->cpu_next) != cpu_list);
10666
10667 *factor = desired / (allocated > 0 ? allocated : 1);
10668
10669 return (ENOMEM);
10670 }
10671
10672 /*
10673 * Note: called from probe context. This function just increments the drop
10674 * count on a buffer. It has been made a function to allow for the
10675 * possibility of understanding the source of mysterious drop counts. (A
10676 * problem for which one may be particularly disappointed that DTrace cannot
10677 * be used to understand DTrace.)
10678 */
10679 static void
10680 dtrace_buffer_drop(dtrace_buffer_t *buf)
10681 {
10682 buf->dtb_drops++;
10683 }
10684
10685 /*
10686 * Note: called from probe context. This function is called to reserve space
10687 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
10688 * mstate. Returns the new offset in the buffer, or a negative value if an
10689 * error has occurred.
10690 */
10691 static intptr_t
10692 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10693 dtrace_state_t *state, dtrace_mstate_t *mstate)
10694 {
10695 intptr_t offs = buf->dtb_offset, soffs;
10696 intptr_t woffs;
10697 caddr_t tomax;
10698 size_t total;
10699
10700 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10701 return (-1);
10702
10703 if ((tomax = buf->dtb_tomax) == NULL) {
10704 dtrace_buffer_drop(buf);
10705 return (-1);
10706 }
10707
10708 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10709 while (offs & (align - 1)) {
10710 /*
10711 * Assert that our alignment is off by a number which
10712 * is itself sizeof (uint32_t) aligned.
10713 */
10714 ASSERT(!((align - (offs & (align - 1))) &
10715 (sizeof (uint32_t) - 1)));
10716 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10717 offs += sizeof (uint32_t);
10718 }
10719
10720 if ((soffs = offs + needed) > buf->dtb_size) {
10721 dtrace_buffer_drop(buf);
10722 return (-1);
10723 }
10724
10725 if (mstate == NULL)
10726 return (offs);
10727
10728 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10729 mstate->dtms_scratch_size = buf->dtb_size - soffs;
10730 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10731
10732 return (offs);
10733 }
10734
10735 if (buf->dtb_flags & DTRACEBUF_FILL) {
10736 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10737 (buf->dtb_flags & DTRACEBUF_FULL))
10738 return (-1);
10739 goto out;
10740 }
10741
10742 total = needed + (offs & (align - 1));
10743
10744 /*
10745 * For a ring buffer, life is quite a bit more complicated. Before
10746 * we can store any padding, we need to adjust our wrapping offset.
10747 * (If we've never before wrapped or we're not about to, no adjustment
10748 * is required.)
10749 */
10750 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10751 offs + total > buf->dtb_size) {
10752 woffs = buf->dtb_xamot_offset;
10753
10754 if (offs + total > buf->dtb_size) {
10755 /*
10756 * We can't fit in the end of the buffer. First, a
10757 * sanity check that we can fit in the buffer at all.
10758 */
10759 if (total > buf->dtb_size) {
10760 dtrace_buffer_drop(buf);
10761 return (-1);
10762 }
10763
10764 /*
10765 * We're going to be storing at the top of the buffer,
10766 * so now we need to deal with the wrapped offset. We
10767 * only reset our wrapped offset to 0 if it is
10768 * currently greater than the current offset. If it
10769 * is less than the current offset, it is because a
10770 * previous allocation induced a wrap -- but the
10771 * allocation didn't subsequently take the space due
10772 * to an error or false predicate evaluation. In this
10773 * case, we'll just leave the wrapped offset alone: if
10774 * the wrapped offset hasn't been advanced far enough
10775 * for this allocation, it will be adjusted in the
10776 * lower loop.
10777 */
10778 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10779 if (woffs >= offs)
10780 woffs = 0;
10781 } else {
10782 woffs = 0;
10783 }
10784
10785 /*
10786 * Now we know that we're going to be storing to the
10787 * top of the buffer and that there is room for us
10788 * there. We need to clear the buffer from the current
10789 * offset to the end (there may be old gunk there).
10790 */
10791 while (offs < buf->dtb_size)
10792 tomax[offs++] = 0;
10793
10794 /*
10795 * We need to set our offset to zero. And because we
10796 * are wrapping, we need to set the bit indicating as
10797 * much. We can also adjust our needed space back
10798 * down to the space required by the ECB -- we know
10799 * that the top of the buffer is aligned.
10800 */
10801 offs = 0;
10802 total = needed;
10803 buf->dtb_flags |= DTRACEBUF_WRAPPED;
10804 } else {
10805 /*
10806 * There is room for us in the buffer, so we simply
10807 * need to check the wrapped offset.
10808 */
10809 if (woffs < offs) {
10810 /*
10811 * The wrapped offset is less than the offset.
10812 * This can happen if we allocated buffer space
10813 * that induced a wrap, but then we didn't
10814 * subsequently take the space due to an error
10815 * or false predicate evaluation. This is
10816 * okay; we know that _this_ allocation isn't
10817 * going to induce a wrap. We still can't
10818 * reset the wrapped offset to be zero,
10819 * however: the space may have been trashed in
10820 * the previous failed probe attempt. But at
10821 * least the wrapped offset doesn't need to
10822 * be adjusted at all...
10823 */
10824 goto out;
10825 }
10826 }
10827
10828 while (offs + total > woffs) {
10829 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
10830 size_t size;
10831
10832 if (epid == DTRACE_EPIDNONE) {
10833 size = sizeof (uint32_t);
10834 } else {
10835 ASSERT3U(epid, <=, state->dts_necbs);
10836 ASSERT(state->dts_ecbs[epid - 1] != NULL);
10837
10838 size = state->dts_ecbs[epid - 1]->dte_size;
10839 }
10840
10841 ASSERT(woffs + size <= buf->dtb_size);
10842 ASSERT(size != 0);
10843
10844 if (woffs + size == buf->dtb_size) {
10845 /*
10846 * We've reached the end of the buffer; we want
10847 * to set the wrapped offset to 0 and break
10848 * out. However, if the offs is 0, then we're
10849 * in a strange edge-condition: the amount of
10850 * space that we want to reserve plus the size
10851 * of the record that we're overwriting is
10852 * greater than the size of the buffer. This
10853 * is problematic because if we reserve the
10854 * space but subsequently don't consume it (due
10855 * to a failed predicate or error) the wrapped
10856 * offset will be 0 -- yet the EPID at offset 0
10857 * will not be committed. This situation is
10858 * relatively easy to deal with: if we're in
10859 * this case, the buffer is indistinguishable
10860 * from one that hasn't wrapped; we need only
10861 * finish the job by clearing the wrapped bit,
10862 * explicitly setting the offset to be 0, and
10863 * zero'ing out the old data in the buffer.
10864 */
10865 if (offs == 0) {
10866 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
10867 buf->dtb_offset = 0;
10868 woffs = total;
10869
10870 while (woffs < buf->dtb_size)
10871 tomax[woffs++] = 0;
10872 }
10873
10874 woffs = 0;
10875 break;
10876 }
10877
10878 woffs += size;
10879 }
10880
10881 /*
10882 * We have a wrapped offset. It may be that the wrapped offset
10883 * has become zero -- that's okay.
10884 */
10885 buf->dtb_xamot_offset = woffs;
10886 }
10887
10888 out:
10889 /*
10890 * Now we can plow the buffer with any necessary padding.
10891 */
10892 while (offs & (align - 1)) {
10893 /*
10894 * Assert that our alignment is off by a number which
10895 * is itself sizeof (uint32_t) aligned.
10896 */
10897 ASSERT(!((align - (offs & (align - 1))) &
10898 (sizeof (uint32_t) - 1)));
10899 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10900 offs += sizeof (uint32_t);
10901 }
10902
10903 if (buf->dtb_flags & DTRACEBUF_FILL) {
10904 if (offs + needed > buf->dtb_size - state->dts_reserve) {
10905 buf->dtb_flags |= DTRACEBUF_FULL;
10906 return (-1);
10907 }
10908 }
10909
10910 if (mstate == NULL)
10911 return (offs);
10912
10913 /*
10914 * For ring buffers and fill buffers, the scratch space is always
10915 * the inactive buffer.
10916 */
10917 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
10918 mstate->dtms_scratch_size = buf->dtb_size;
10919 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10920
10921 return (offs);
10922 }
10923
10924 static void
10925 dtrace_buffer_polish(dtrace_buffer_t *buf)
10926 {
10927 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
10928 ASSERT(MUTEX_HELD(&dtrace_lock));
10929
10930 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
10931 return;
10932
10933 /*
10934 * We need to polish the ring buffer. There are three cases:
10935 *
10936 * - The first (and presumably most common) is that there is no gap
10937 * between the buffer offset and the wrapped offset. In this case,
10938 * there is nothing in the buffer that isn't valid data; we can
10939 * mark the buffer as polished and return.
10940 *
10941 * - The second (less common than the first but still more common
10942 * than the third) is that there is a gap between the buffer offset
10943 * and the wrapped offset, and the wrapped offset is larger than the
10944 * buffer offset. This can happen because of an alignment issue, or
10945 * can happen because of a call to dtrace_buffer_reserve() that
10946 * didn't subsequently consume the buffer space. In this case,
10947 * we need to zero the data from the buffer offset to the wrapped
10948 * offset.
10949 *
10950 * - The third (and least common) is that there is a gap between the
10951 * buffer offset and the wrapped offset, but the wrapped offset is
10952 * _less_ than the buffer offset. This can only happen because a
10953 * call to dtrace_buffer_reserve() induced a wrap, but the space
10954 * was not subsequently consumed. In this case, we need to zero the
10955 * space from the offset to the end of the buffer _and_ from the
10956 * top of the buffer to the wrapped offset.
10957 */
10958 if (buf->dtb_offset < buf->dtb_xamot_offset) {
10959 bzero(buf->dtb_tomax + buf->dtb_offset,
10960 buf->dtb_xamot_offset - buf->dtb_offset);
10961 }
10962
10963 if (buf->dtb_offset > buf->dtb_xamot_offset) {
10964 bzero(buf->dtb_tomax + buf->dtb_offset,
10965 buf->dtb_size - buf->dtb_offset);
10966 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
10967 }
10968 }
10969
10970 /*
10971 * This routine determines if data generated at the specified time has likely
10972 * been entirely consumed at user-level. This routine is called to determine
10973 * if an ECB on a defunct probe (but for an active enabling) can be safely
10974 * disabled and destroyed.
10975 */
10976 static int
10977 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
10978 {
10979 int i;
10980
10981 for (i = 0; i < NCPU; i++) {
10982 dtrace_buffer_t *buf = &bufs[i];
10983
10984 if (buf->dtb_size == 0)
10985 continue;
10986
10987 if (buf->dtb_flags & DTRACEBUF_RING)
10988 return (0);
10989
10990 if (!buf->dtb_switched && buf->dtb_offset != 0)
10991 return (0);
10992
10993 if (buf->dtb_switched - buf->dtb_interval < when)
10994 return (0);
10995 }
10996
10997 return (1);
10998 }
10999
11000 static void
11001 dtrace_buffer_free(dtrace_buffer_t *bufs)
11002 {
11003 int i;
11004
11005 for (i = 0; i < NCPU; i++) {
11006 dtrace_buffer_t *buf = &bufs[i];
11007
11008 if (buf->dtb_tomax == NULL) {
11009 ASSERT(buf->dtb_xamot == NULL);
11010 ASSERT(buf->dtb_size == 0);
11011 continue;
11012 }
11013
11014 if (buf->dtb_xamot != NULL) {
11015 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11016 kmem_free(buf->dtb_xamot, buf->dtb_size);
11017 }
11018
11019 kmem_free(buf->dtb_tomax, buf->dtb_size);
11020 buf->dtb_size = 0;
11021 buf->dtb_tomax = NULL;
11022 buf->dtb_xamot = NULL;
11023 }
11024 }
11025
11026 /*
11027 * DTrace Enabling Functions
11028 */
11029 static dtrace_enabling_t *
11030 dtrace_enabling_create(dtrace_vstate_t *vstate)
11031 {
11032 dtrace_enabling_t *enab;
11033
11034 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11035 enab->dten_vstate = vstate;
11036
11037 return (enab);
11038 }
11039
11040 static void
11041 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11042 {
11043 dtrace_ecbdesc_t **ndesc;
11044 size_t osize, nsize;
11045
11046 /*
11047 * We can't add to enablings after we've enabled them, or after we've
11048 * retained them.
11049 */
11050 ASSERT(enab->dten_probegen == 0);
11051 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11052
11053 if (enab->dten_ndesc < enab->dten_maxdesc) {
11054 enab->dten_desc[enab->dten_ndesc++] = ecb;
11055 return;
11056 }
11057
11058 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11059
11060 if (enab->dten_maxdesc == 0) {
11061 enab->dten_maxdesc = 1;
11062 } else {
11063 enab->dten_maxdesc <<= 1;
11064 }
11065
11066 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11067
11068 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11069 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11070 bcopy(enab->dten_desc, ndesc, osize);
11071 kmem_free(enab->dten_desc, osize);
11072
11073 enab->dten_desc = ndesc;
11074 enab->dten_desc[enab->dten_ndesc++] = ecb;
11075 }
11076
11077 static void
11078 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11079 dtrace_probedesc_t *pd)
11080 {
11081 dtrace_ecbdesc_t *new;
11082 dtrace_predicate_t *pred;
11083 dtrace_actdesc_t *act;
11084
11085 /*
11086 * We're going to create a new ECB description that matches the
11087 * specified ECB in every way, but has the specified probe description.
11088 */
11089 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11090
11091 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11092 dtrace_predicate_hold(pred);
11093
11094 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11095 dtrace_actdesc_hold(act);
11096
11097 new->dted_action = ecb->dted_action;
11098 new->dted_pred = ecb->dted_pred;
11099 new->dted_probe = *pd;
11100 new->dted_uarg = ecb->dted_uarg;
11101
11102 dtrace_enabling_add(enab, new);
11103 }
11104
11105 static void
11106 dtrace_enabling_dump(dtrace_enabling_t *enab)
11107 {
11108 int i;
11109
11110 for (i = 0; i < enab->dten_ndesc; i++) {
11111 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11112
11113 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11114 desc->dtpd_provider, desc->dtpd_mod,
11115 desc->dtpd_func, desc->dtpd_name);
11116 }
11117 }
11118
11119 static void
11120 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11121 {
11122 int i;
11123 dtrace_ecbdesc_t *ep;
11124 dtrace_vstate_t *vstate = enab->dten_vstate;
11125
11126 ASSERT(MUTEX_HELD(&dtrace_lock));
11127
11128 for (i = 0; i < enab->dten_ndesc; i++) {
11129 dtrace_actdesc_t *act, *next;
11130 dtrace_predicate_t *pred;
11131
11132 ep = enab->dten_desc[i];
11133
11134 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11135 dtrace_predicate_release(pred, vstate);
11136
11137 for (act = ep->dted_action; act != NULL; act = next) {
11138 next = act->dtad_next;
11139 dtrace_actdesc_release(act, vstate);
11140 }
11141
11142 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11143 }
11144
11145 kmem_free(enab->dten_desc,
11146 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11147
11148 /*
11149 * If this was a retained enabling, decrement the dts_nretained count
11150 * and take it off of the dtrace_retained list.
11151 */
11152 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11153 dtrace_retained == enab) {
11154 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11155 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11156 enab->dten_vstate->dtvs_state->dts_nretained--;
11157 dtrace_retained_gen++;
11158 }
11159
11160 if (enab->dten_prev == NULL) {
11161 if (dtrace_retained == enab) {
11162 dtrace_retained = enab->dten_next;
11163
11164 if (dtrace_retained != NULL)
11165 dtrace_retained->dten_prev = NULL;
11166 }
11167 } else {
11168 ASSERT(enab != dtrace_retained);
11169 ASSERT(dtrace_retained != NULL);
11170 enab->dten_prev->dten_next = enab->dten_next;
11171 }
11172
11173 if (enab->dten_next != NULL) {
11174 ASSERT(dtrace_retained != NULL);
11175 enab->dten_next->dten_prev = enab->dten_prev;
11176 }
11177
11178 kmem_free(enab, sizeof (dtrace_enabling_t));
11179 }
11180
11181 static int
11182 dtrace_enabling_retain(dtrace_enabling_t *enab)
11183 {
11184 dtrace_state_t *state;
11185
11186 ASSERT(MUTEX_HELD(&dtrace_lock));
11187 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11188 ASSERT(enab->dten_vstate != NULL);
11189
11190 state = enab->dten_vstate->dtvs_state;
11191 ASSERT(state != NULL);
11192
11193 /*
11194 * We only allow each state to retain dtrace_retain_max enablings.
11195 */
11196 if (state->dts_nretained >= dtrace_retain_max)
11197 return (ENOSPC);
11198
11199 state->dts_nretained++;
11200 dtrace_retained_gen++;
11201
11202 if (dtrace_retained == NULL) {
11203 dtrace_retained = enab;
11204 return (0);
11205 }
11206
11207 enab->dten_next = dtrace_retained;
11208 dtrace_retained->dten_prev = enab;
11209 dtrace_retained = enab;
11210
11211 return (0);
11212 }
11213
11214 static int
11215 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11216 dtrace_probedesc_t *create)
11217 {
11218 dtrace_enabling_t *new, *enab;
11219 int found = 0, err = ENOENT;
11220
11221 ASSERT(MUTEX_HELD(&dtrace_lock));
11222 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11223 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11224 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11225 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11226
11227 new = dtrace_enabling_create(&state->dts_vstate);
11228
11229 /*
11230 * Iterate over all retained enablings, looking for enablings that
11231 * match the specified state.
11232 */
11233 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11234 int i;
11235
11236 /*
11237 * dtvs_state can only be NULL for helper enablings -- and
11238 * helper enablings can't be retained.
11239 */
11240 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11241
11242 if (enab->dten_vstate->dtvs_state != state)
11243 continue;
11244
11245 /*
11246 * Now iterate over each probe description; we're looking for
11247 * an exact match to the specified probe description.
11248 */
11249 for (i = 0; i < enab->dten_ndesc; i++) {
11250 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11251 dtrace_probedesc_t *pd = &ep->dted_probe;
11252
11253 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11254 continue;
11255
11256 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11257 continue;
11258
11259 if (strcmp(pd->dtpd_func, match->dtpd_func))
11260 continue;
11261
11262 if (strcmp(pd->dtpd_name, match->dtpd_name))
11263 continue;
11264
11265 /*
11266 * We have a winning probe! Add it to our growing
11267 * enabling.
11268 */
11269 found = 1;
11270 dtrace_enabling_addlike(new, ep, create);
11271 }
11272 }
11273
11274 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11275 dtrace_enabling_destroy(new);
11276 return (err);
11277 }
11278
11279 return (0);
11280 }
11281
11282 static void
11283 dtrace_enabling_retract(dtrace_state_t *state)
11284 {
11285 dtrace_enabling_t *enab, *next;
11286
11287 ASSERT(MUTEX_HELD(&dtrace_lock));
11288
11289 /*
11290 * Iterate over all retained enablings, destroy the enablings retained
11291 * for the specified state.
11292 */
11293 for (enab = dtrace_retained; enab != NULL; enab = next) {
11294 next = enab->dten_next;
11295
11296 /*
11297 * dtvs_state can only be NULL for helper enablings -- and
11298 * helper enablings can't be retained.
11299 */
11300 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11301
11302 if (enab->dten_vstate->dtvs_state == state) {
11303 ASSERT(state->dts_nretained > 0);
11304 dtrace_enabling_destroy(enab);
11305 }
11306 }
11307
11308 ASSERT(state->dts_nretained == 0);
11309 }
11310
11311 static int
11312 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11313 {
11314 int i = 0;
11315 int total_matched = 0, matched = 0;
11316
11317 ASSERT(MUTEX_HELD(&cpu_lock));
11318 ASSERT(MUTEX_HELD(&dtrace_lock));
11319
11320 for (i = 0; i < enab->dten_ndesc; i++) {
11321 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11322
11323 enab->dten_current = ep;
11324 enab->dten_error = 0;
11325
11326 /*
11327 * If a provider failed to enable a probe then get out and
11328 * let the consumer know we failed.
11329 */
11330 if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
11331 return (EBUSY);
11332
11333 total_matched += matched;
11334
11335 if (enab->dten_error != 0) {
11336 /*
11337 * If we get an error half-way through enabling the
11338 * probes, we kick out -- perhaps with some number of
11339 * them enabled. Leaving enabled probes enabled may
11340 * be slightly confusing for user-level, but we expect
11341 * that no one will attempt to actually drive on in
11342 * the face of such errors. If this is an anonymous
11343 * enabling (indicated with a NULL nmatched pointer),
11344 * we cmn_err() a message. We aren't expecting to
11345 * get such an error -- such as it can exist at all,
11346 * it would be a result of corrupted DOF in the driver
11347 * properties.
11348 */
11349 if (nmatched == NULL) {
11350 cmn_err(CE_WARN, "dtrace_enabling_match() "
11351 "error on %p: %d", (void *)ep,
11352 enab->dten_error);
11353 }
11354
11355 return (enab->dten_error);
11356 }
11357 }
11358
11359 enab->dten_probegen = dtrace_probegen;
11360 if (nmatched != NULL)
11361 *nmatched = total_matched;
11362
11363 return (0);
11364 }
11365
11366 static void
11367 dtrace_enabling_matchall(void)
11368 {
11369 dtrace_enabling_t *enab;
11370
11371 mutex_enter(&cpu_lock);
11372 mutex_enter(&dtrace_lock);
11373
11374 /*
11375 * Iterate over all retained enablings to see if any probes match
11376 * against them. We only perform this operation on enablings for which
11377 * we have sufficient permissions by virtue of being in the global zone
11378 * or in the same zone as the DTrace client. Because we can be called
11379 * after dtrace_detach() has been called, we cannot assert that there
11380 * are retained enablings. We can safely load from dtrace_retained,
11381 * however: the taskq_destroy() at the end of dtrace_detach() will
11382 * block pending our completion.
11383 */
11384 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11385 dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
11386 cred_t *cr = dcr->dcr_cred;
11387 zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
11388
11389 if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
11390 (zone == GLOBAL_ZONEID || getzoneid() == zone)))
11391 (void) dtrace_enabling_match(enab, NULL);
11392 }
11393
11394 mutex_exit(&dtrace_lock);
11395 mutex_exit(&cpu_lock);
11396 }
11397
11398 /*
11399 * If an enabling is to be enabled without having matched probes (that is, if
11400 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11401 * enabling must be _primed_ by creating an ECB for every ECB description.
11402 * This must be done to assure that we know the number of speculations, the
11403 * number of aggregations, the minimum buffer size needed, etc. before we
11404 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11405 * enabling any probes, we create ECBs for every ECB decription, but with a
11406 * NULL probe -- which is exactly what this function does.
11407 */
11408 static void
11409 dtrace_enabling_prime(dtrace_state_t *state)
11410 {
11411 dtrace_enabling_t *enab;
11412 int i;
11413
11414 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11415 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11416
11417 if (enab->dten_vstate->dtvs_state != state)
11418 continue;
11419
11420 /*
11421 * We don't want to prime an enabling more than once, lest
11422 * we allow a malicious user to induce resource exhaustion.
11423 * (The ECBs that result from priming an enabling aren't
11424 * leaked -- but they also aren't deallocated until the
11425 * consumer state is destroyed.)
11426 */
11427 if (enab->dten_primed)
11428 continue;
11429
11430 for (i = 0; i < enab->dten_ndesc; i++) {
11431 enab->dten_current = enab->dten_desc[i];
11432 (void) dtrace_probe_enable(NULL, enab);
11433 }
11434
11435 enab->dten_primed = 1;
11436 }
11437 }
11438
11439 /*
11440 * Called to indicate that probes should be provided due to retained
11441 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11442 * must take an initial lap through the enabling calling the dtps_provide()
11443 * entry point explicitly to allow for autocreated probes.
11444 */
11445 static void
11446 dtrace_enabling_provide(dtrace_provider_t *prv)
11447 {
11448 int i, all = 0;
11449 dtrace_probedesc_t desc;
11450 dtrace_genid_t gen;
11451
11452 ASSERT(MUTEX_HELD(&dtrace_lock));
11453 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11454
11455 if (prv == NULL) {
11456 all = 1;
11457 prv = dtrace_provider;
11458 }
11459
11460 do {
11461 dtrace_enabling_t *enab;
11462 void *parg = prv->dtpv_arg;
11463
11464 retry:
11465 gen = dtrace_retained_gen;
11466 for (enab = dtrace_retained; enab != NULL;
11467 enab = enab->dten_next) {
11468 for (i = 0; i < enab->dten_ndesc; i++) {
11469 desc = enab->dten_desc[i]->dted_probe;
11470 mutex_exit(&dtrace_lock);
11471 prv->dtpv_pops.dtps_provide(parg, &desc);
11472 mutex_enter(&dtrace_lock);
11473 /*
11474 * Process the retained enablings again if
11475 * they have changed while we weren't holding
11476 * dtrace_lock.
11477 */
11478 if (gen != dtrace_retained_gen)
11479 goto retry;
11480 }
11481 }
11482 } while (all && (prv = prv->dtpv_next) != NULL);
11483
11484 mutex_exit(&dtrace_lock);
11485 dtrace_probe_provide(NULL, all ? NULL : prv);
11486 mutex_enter(&dtrace_lock);
11487 }
11488
11489 /*
11490 * Called to reap ECBs that are attached to probes from defunct providers.
11491 */
11492 static void
11493 dtrace_enabling_reap(void)
11494 {
11495 dtrace_provider_t *prov;
11496 dtrace_probe_t *probe;
11497 dtrace_ecb_t *ecb;
11498 hrtime_t when;
11499 int i;
11500
11501 mutex_enter(&cpu_lock);
11502 mutex_enter(&dtrace_lock);
11503
11504 for (i = 0; i < dtrace_nprobes; i++) {
11505 if ((probe = dtrace_probes[i]) == NULL)
11506 continue;
11507
11508 if (probe->dtpr_ecb == NULL)
11509 continue;
11510
11511 prov = probe->dtpr_provider;
11512
11513 if ((when = prov->dtpv_defunct) == 0)
11514 continue;
11515
11516 /*
11517 * We have ECBs on a defunct provider: we want to reap these
11518 * ECBs to allow the provider to unregister. The destruction
11519 * of these ECBs must be done carefully: if we destroy the ECB
11520 * and the consumer later wishes to consume an EPID that
11521 * corresponds to the destroyed ECB (and if the EPID metadata
11522 * has not been previously consumed), the consumer will abort
11523 * processing on the unknown EPID. To reduce (but not, sadly,
11524 * eliminate) the possibility of this, we will only destroy an
11525 * ECB for a defunct provider if, for the state that
11526 * corresponds to the ECB:
11527 *
11528 * (a) There is no speculative tracing (which can effectively
11529 * cache an EPID for an arbitrary amount of time).
11530 *
11531 * (b) The principal buffers have been switched twice since the
11532 * provider became defunct.
11533 *
11534 * (c) The aggregation buffers are of zero size or have been
11535 * switched twice since the provider became defunct.
11536 *
11537 * We use dts_speculates to determine (a) and call a function
11538 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
11539 * that as soon as we've been unable to destroy one of the ECBs
11540 * associated with the probe, we quit trying -- reaping is only
11541 * fruitful in as much as we can destroy all ECBs associated
11542 * with the defunct provider's probes.
11543 */
11544 while ((ecb = probe->dtpr_ecb) != NULL) {
11545 dtrace_state_t *state = ecb->dte_state;
11546 dtrace_buffer_t *buf = state->dts_buffer;
11547 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11548
11549 if (state->dts_speculates)
11550 break;
11551
11552 if (!dtrace_buffer_consumed(buf, when))
11553 break;
11554
11555 if (!dtrace_buffer_consumed(aggbuf, when))
11556 break;
11557
11558 dtrace_ecb_disable(ecb);
11559 ASSERT(probe->dtpr_ecb != ecb);
11560 dtrace_ecb_destroy(ecb);
11561 }
11562 }
11563
11564 mutex_exit(&dtrace_lock);
11565 mutex_exit(&cpu_lock);
11566 }
11567
11568 /*
11569 * DTrace DOF Functions
11570 */
11571 /*ARGSUSED*/
11572 static void
11573 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11574 {
11575 if (dtrace_err_verbose)
11576 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11577
11578 #ifdef DTRACE_ERRDEBUG
11579 dtrace_errdebug(str);
11580 #endif
11581 }
11582
11583 /*
11584 * Create DOF out of a currently enabled state. Right now, we only create
11585 * DOF containing the run-time options -- but this could be expanded to create
11586 * complete DOF representing the enabled state.
11587 */
11588 static dof_hdr_t *
11589 dtrace_dof_create(dtrace_state_t *state)
11590 {
11591 dof_hdr_t *dof;
11592 dof_sec_t *sec;
11593 dof_optdesc_t *opt;
11594 int i, len = sizeof (dof_hdr_t) +
11595 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11596 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11597
11598 ASSERT(MUTEX_HELD(&dtrace_lock));
11599
11600 dof = kmem_zalloc(len, KM_SLEEP);
11601 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11602 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11603 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11604 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11605
11606 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11607 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11608 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11609 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11610 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11611 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11612
11613 dof->dofh_flags = 0;
11614 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11615 dof->dofh_secsize = sizeof (dof_sec_t);
11616 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11617 dof->dofh_secoff = sizeof (dof_hdr_t);
11618 dof->dofh_loadsz = len;
11619 dof->dofh_filesz = len;
11620 dof->dofh_pad = 0;
11621
11622 /*
11623 * Fill in the option section header...
11624 */
11625 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11626 sec->dofs_type = DOF_SECT_OPTDESC;
11627 sec->dofs_align = sizeof (uint64_t);
11628 sec->dofs_flags = DOF_SECF_LOAD;
11629 sec->dofs_entsize = sizeof (dof_optdesc_t);
11630
11631 opt = (dof_optdesc_t *)((uintptr_t)sec +
11632 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11633
11634 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11635 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11636
11637 for (i = 0; i < DTRACEOPT_MAX; i++) {
11638 opt[i].dofo_option = i;
11639 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11640 opt[i].dofo_value = state->dts_options[i];
11641 }
11642
11643 return (dof);
11644 }
11645
11646 static dof_hdr_t *
11647 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11648 {
11649 dof_hdr_t hdr, *dof;
11650
11651 ASSERT(!MUTEX_HELD(&dtrace_lock));
11652
11653 /*
11654 * First, we're going to copyin() the sizeof (dof_hdr_t).
11655 */
11656 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11657 dtrace_dof_error(NULL, "failed to copyin DOF header");
11658 *errp = EFAULT;
11659 return (NULL);
11660 }
11661
11662 /*
11663 * Now we'll allocate the entire DOF and copy it in -- provided
11664 * that the length isn't outrageous.
11665 */
11666 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11667 dtrace_dof_error(&hdr, "load size exceeds maximum");
11668 *errp = E2BIG;
11669 return (NULL);
11670 }
11671
11672 if (hdr.dofh_loadsz < sizeof (hdr)) {
11673 dtrace_dof_error(&hdr, "invalid load size");
11674 *errp = EINVAL;
11675 return (NULL);
11676 }
11677
11678 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11679
11680 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
11681 dof->dofh_loadsz != hdr.dofh_loadsz) {
11682 kmem_free(dof, hdr.dofh_loadsz);
11683 *errp = EFAULT;
11684 return (NULL);
11685 }
11686
11687 return (dof);
11688 }
11689
11690 static dof_hdr_t *
11691 dtrace_dof_property(const char *name)
11692 {
11693 uchar_t *buf;
11694 uint64_t loadsz;
11695 unsigned int len, i;
11696 dof_hdr_t *dof;
11697
11698 /*
11699 * Unfortunately, array of values in .conf files are always (and
11700 * only) interpreted to be integer arrays. We must read our DOF
11701 * as an integer array, and then squeeze it into a byte array.
11702 */
11703 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11704 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11705 return (NULL);
11706
11707 for (i = 0; i < len; i++)
11708 buf[i] = (uchar_t)(((int *)buf)[i]);
11709
11710 if (len < sizeof (dof_hdr_t)) {
11711 ddi_prop_free(buf);
11712 dtrace_dof_error(NULL, "truncated header");
11713 return (NULL);
11714 }
11715
11716 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11717 ddi_prop_free(buf);
11718 dtrace_dof_error(NULL, "truncated DOF");
11719 return (NULL);
11720 }
11721
11722 if (loadsz >= dtrace_dof_maxsize) {
11723 ddi_prop_free(buf);
11724 dtrace_dof_error(NULL, "oversized DOF");
11725 return (NULL);
11726 }
11727
11728 dof = kmem_alloc(loadsz, KM_SLEEP);
11729 bcopy(buf, dof, loadsz);
11730 ddi_prop_free(buf);
11731
11732 return (dof);
11733 }
11734
11735 static void
11736 dtrace_dof_destroy(dof_hdr_t *dof)
11737 {
11738 kmem_free(dof, dof->dofh_loadsz);
11739 }
11740
11741 /*
11742 * Return the dof_sec_t pointer corresponding to a given section index. If the
11743 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
11744 * a type other than DOF_SECT_NONE is specified, the header is checked against
11745 * this type and NULL is returned if the types do not match.
11746 */
11747 static dof_sec_t *
11748 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11749 {
11750 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11751 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11752
11753 if (i >= dof->dofh_secnum) {
11754 dtrace_dof_error(dof, "referenced section index is invalid");
11755 return (NULL);
11756 }
11757
11758 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11759 dtrace_dof_error(dof, "referenced section is not loadable");
11760 return (NULL);
11761 }
11762
11763 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11764 dtrace_dof_error(dof, "referenced section is the wrong type");
11765 return (NULL);
11766 }
11767
11768 return (sec);
11769 }
11770
11771 static dtrace_probedesc_t *
11772 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11773 {
11774 dof_probedesc_t *probe;
11775 dof_sec_t *strtab;
11776 uintptr_t daddr = (uintptr_t)dof;
11777 uintptr_t str;
11778 size_t size;
11779
11780 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11781 dtrace_dof_error(dof, "invalid probe section");
11782 return (NULL);
11783 }
11784
11785 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11786 dtrace_dof_error(dof, "bad alignment in probe description");
11787 return (NULL);
11788 }
11789
11790 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11791 dtrace_dof_error(dof, "truncated probe description");
11792 return (NULL);
11793 }
11794
11795 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11796 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11797
11798 if (strtab == NULL)
11799 return (NULL);
11800
11801 str = daddr + strtab->dofs_offset;
11802 size = strtab->dofs_size;
11803
11804 if (probe->dofp_provider >= strtab->dofs_size) {
11805 dtrace_dof_error(dof, "corrupt probe provider");
11806 return (NULL);
11807 }
11808
11809 (void) strncpy(desc->dtpd_provider,
11810 (char *)(str + probe->dofp_provider),
11811 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11812
11813 if (probe->dofp_mod >= strtab->dofs_size) {
11814 dtrace_dof_error(dof, "corrupt probe module");
11815 return (NULL);
11816 }
11817
11818 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11819 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11820
11821 if (probe->dofp_func >= strtab->dofs_size) {
11822 dtrace_dof_error(dof, "corrupt probe function");
11823 return (NULL);
11824 }
11825
11826 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11827 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11828
11829 if (probe->dofp_name >= strtab->dofs_size) {
11830 dtrace_dof_error(dof, "corrupt probe name");
11831 return (NULL);
11832 }
11833
11834 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11835 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11836
11837 return (desc);
11838 }
11839
11840 static dtrace_difo_t *
11841 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11842 cred_t *cr)
11843 {
11844 dtrace_difo_t *dp;
11845 size_t ttl = 0;
11846 dof_difohdr_t *dofd;
11847 uintptr_t daddr = (uintptr_t)dof;
11848 size_t max = dtrace_difo_maxsize;
11849 int i, l, n;
11850
11851 static const struct {
11852 int section;
11853 int bufoffs;
11854 int lenoffs;
11855 int entsize;
11856 int align;
11857 const char *msg;
11858 } difo[] = {
11859 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11860 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11861 sizeof (dif_instr_t), "multiple DIF sections" },
11862
11863 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11864 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11865 sizeof (uint64_t), "multiple integer tables" },
11866
11867 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11868 offsetof(dtrace_difo_t, dtdo_strlen), 0,
11869 sizeof (char), "multiple string tables" },
11870
11871 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11872 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11873 sizeof (uint_t), "multiple variable tables" },
11874
11875 { DOF_SECT_NONE, 0, 0, 0, NULL }
11876 };
11877
11878 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11879 dtrace_dof_error(dof, "invalid DIFO header section");
11880 return (NULL);
11881 }
11882
11883 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11884 dtrace_dof_error(dof, "bad alignment in DIFO header");
11885 return (NULL);
11886 }
11887
11888 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
11889 sec->dofs_size % sizeof (dof_secidx_t)) {
11890 dtrace_dof_error(dof, "bad size in DIFO header");
11891 return (NULL);
11892 }
11893
11894 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11895 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
11896
11897 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
11898 dp->dtdo_rtype = dofd->dofd_rtype;
11899
11900 for (l = 0; l < n; l++) {
11901 dof_sec_t *subsec;
11902 void **bufp;
11903 uint32_t *lenp;
11904
11905 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
11906 dofd->dofd_links[l])) == NULL)
11907 goto err; /* invalid section link */
11908
11909 if (ttl + subsec->dofs_size > max) {
11910 dtrace_dof_error(dof, "exceeds maximum size");
11911 goto err;
11912 }
11913
11914 ttl += subsec->dofs_size;
11915
11916 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
11917 if (subsec->dofs_type != difo[i].section)
11918 continue;
11919
11920 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
11921 dtrace_dof_error(dof, "section not loaded");
11922 goto err;
11923 }
11924
11925 if (subsec->dofs_align != difo[i].align) {
11926 dtrace_dof_error(dof, "bad alignment");
11927 goto err;
11928 }
11929
11930 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
11931 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
11932
11933 if (*bufp != NULL) {
11934 dtrace_dof_error(dof, difo[i].msg);
11935 goto err;
11936 }
11937
11938 if (difo[i].entsize != subsec->dofs_entsize) {
11939 dtrace_dof_error(dof, "entry size mismatch");
11940 goto err;
11941 }
11942
11943 if (subsec->dofs_entsize != 0 &&
11944 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
11945 dtrace_dof_error(dof, "corrupt entry size");
11946 goto err;
11947 }
11948
11949 *lenp = subsec->dofs_size;
11950 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
11951 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
11952 *bufp, subsec->dofs_size);
11953
11954 if (subsec->dofs_entsize != 0)
11955 *lenp /= subsec->dofs_entsize;
11956
11957 break;
11958 }
11959
11960 /*
11961 * If we encounter a loadable DIFO sub-section that is not
11962 * known to us, assume this is a broken program and fail.
11963 */
11964 if (difo[i].section == DOF_SECT_NONE &&
11965 (subsec->dofs_flags & DOF_SECF_LOAD)) {
11966 dtrace_dof_error(dof, "unrecognized DIFO subsection");
11967 goto err;
11968 }
11969 }
11970
11971 if (dp->dtdo_buf == NULL) {
11972 /*
11973 * We can't have a DIF object without DIF text.
11974 */
11975 dtrace_dof_error(dof, "missing DIF text");
11976 goto err;
11977 }
11978
11979 /*
11980 * Before we validate the DIF object, run through the variable table
11981 * looking for the strings -- if any of their size are under, we'll set
11982 * their size to be the system-wide default string size. Note that
11983 * this should _not_ happen if the "strsize" option has been set --
11984 * in this case, the compiler should have set the size to reflect the
11985 * setting of the option.
11986 */
11987 for (i = 0; i < dp->dtdo_varlen; i++) {
11988 dtrace_difv_t *v = &dp->dtdo_vartab[i];
11989 dtrace_diftype_t *t = &v->dtdv_type;
11990
11991 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
11992 continue;
11993
11994 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
11995 t->dtdt_size = dtrace_strsize_default;
11996 }
11997
11998 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
11999 goto err;
12000
12001 dtrace_difo_init(dp, vstate);
12002 return (dp);
12003
12004 err:
12005 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12006 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12007 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12008 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12009
12010 kmem_free(dp, sizeof (dtrace_difo_t));
12011 return (NULL);
12012 }
12013
12014 static dtrace_predicate_t *
12015 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12016 cred_t *cr)
12017 {
12018 dtrace_difo_t *dp;
12019
12020 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12021 return (NULL);
12022
12023 return (dtrace_predicate_create(dp));
12024 }
12025
12026 static dtrace_actdesc_t *
12027 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12028 cred_t *cr)
12029 {
12030 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12031 dof_actdesc_t *desc;
12032 dof_sec_t *difosec;
12033 size_t offs;
12034 uintptr_t daddr = (uintptr_t)dof;
12035 uint64_t arg;
12036 dtrace_actkind_t kind;
12037
12038 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12039 dtrace_dof_error(dof, "invalid action section");
12040 return (NULL);
12041 }
12042
12043 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12044 dtrace_dof_error(dof, "truncated action description");
12045 return (NULL);
12046 }
12047
12048 if (sec->dofs_align != sizeof (uint64_t)) {
12049 dtrace_dof_error(dof, "bad alignment in action description");
12050 return (NULL);
12051 }
12052
12053 if (sec->dofs_size < sec->dofs_entsize) {
12054 dtrace_dof_error(dof, "section entry size exceeds total size");
12055 return (NULL);
12056 }
12057
12058 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12059 dtrace_dof_error(dof, "bad entry size in action description");
12060 return (NULL);
12061 }
12062
12063 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12064 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12065 return (NULL);
12066 }
12067
12068 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12069 desc = (dof_actdesc_t *)(daddr +
12070 (uintptr_t)sec->dofs_offset + offs);
12071 kind = (dtrace_actkind_t)desc->dofa_kind;
12072
12073 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12074 (kind != DTRACEACT_PRINTA ||
12075 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12076 (kind == DTRACEACT_DIFEXPR &&
12077 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12078 dof_sec_t *strtab;
12079 char *str, *fmt;
12080 uint64_t i;
12081
12082 /*
12083 * The argument to these actions is an index into the
12084 * DOF string table. For printf()-like actions, this
12085 * is the format string. For print(), this is the
12086 * CTF type of the expression result.
12087 */
12088 if ((strtab = dtrace_dof_sect(dof,
12089 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12090 goto err;
12091
12092 str = (char *)((uintptr_t)dof +
12093 (uintptr_t)strtab->dofs_offset);
12094
12095 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12096 if (str[i] == '\0')
12097 break;
12098 }
12099
12100 if (i >= strtab->dofs_size) {
12101 dtrace_dof_error(dof, "bogus format string");
12102 goto err;
12103 }
12104
12105 if (i == desc->dofa_arg) {
12106 dtrace_dof_error(dof, "empty format string");
12107 goto err;
12108 }
12109
12110 i -= desc->dofa_arg;
12111 fmt = kmem_alloc(i + 1, KM_SLEEP);
12112 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12113 arg = (uint64_t)(uintptr_t)fmt;
12114 } else {
12115 if (kind == DTRACEACT_PRINTA) {
12116 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12117 arg = 0;
12118 } else {
12119 arg = desc->dofa_arg;
12120 }
12121 }
12122
12123 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12124 desc->dofa_uarg, arg);
12125
12126 if (last != NULL) {
12127 last->dtad_next = act;
12128 } else {
12129 first = act;
12130 }
12131
12132 last = act;
12133
12134 if (desc->dofa_difo == DOF_SECIDX_NONE)
12135 continue;
12136
12137 if ((difosec = dtrace_dof_sect(dof,
12138 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12139 goto err;
12140
12141 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12142
12143 if (act->dtad_difo == NULL)
12144 goto err;
12145 }
12146
12147 ASSERT(first != NULL);
12148 return (first);
12149
12150 err:
12151 for (act = first; act != NULL; act = next) {
12152 next = act->dtad_next;
12153 dtrace_actdesc_release(act, vstate);
12154 }
12155
12156 return (NULL);
12157 }
12158
12159 static dtrace_ecbdesc_t *
12160 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12161 cred_t *cr)
12162 {
12163 dtrace_ecbdesc_t *ep;
12164 dof_ecbdesc_t *ecb;
12165 dtrace_probedesc_t *desc;
12166 dtrace_predicate_t *pred = NULL;
12167
12168 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12169 dtrace_dof_error(dof, "truncated ECB description");
12170 return (NULL);
12171 }
12172
12173 if (sec->dofs_align != sizeof (uint64_t)) {
12174 dtrace_dof_error(dof, "bad alignment in ECB description");
12175 return (NULL);
12176 }
12177
12178 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12179 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12180
12181 if (sec == NULL)
12182 return (NULL);
12183
12184 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12185 ep->dted_uarg = ecb->dofe_uarg;
12186 desc = &ep->dted_probe;
12187
12188 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12189 goto err;
12190
12191 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12192 if ((sec = dtrace_dof_sect(dof,
12193 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12194 goto err;
12195
12196 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12197 goto err;
12198
12199 ep->dted_pred.dtpdd_predicate = pred;
12200 }
12201
12202 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12203 if ((sec = dtrace_dof_sect(dof,
12204 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12205 goto err;
12206
12207 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12208
12209 if (ep->dted_action == NULL)
12210 goto err;
12211 }
12212
12213 return (ep);
12214
12215 err:
12216 if (pred != NULL)
12217 dtrace_predicate_release(pred, vstate);
12218 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12219 return (NULL);
12220 }
12221
12222 /*
12223 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12224 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12225 * site of any user SETX relocations to account for load object base address.
12226 * In the future, if we need other relocations, this function can be extended.
12227 */
12228 static int
12229 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12230 {
12231 uintptr_t daddr = (uintptr_t)dof;
12232 dof_relohdr_t *dofr =
12233 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12234 dof_sec_t *ss, *rs, *ts;
12235 dof_relodesc_t *r;
12236 uint_t i, n;
12237
12238 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12239 sec->dofs_align != sizeof (dof_secidx_t)) {
12240 dtrace_dof_error(dof, "invalid relocation header");
12241 return (-1);
12242 }
12243
12244 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12245 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12246 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12247
12248 if (ss == NULL || rs == NULL || ts == NULL)
12249 return (-1); /* dtrace_dof_error() has been called already */
12250
12251 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12252 rs->dofs_align != sizeof (uint64_t)) {
12253 dtrace_dof_error(dof, "invalid relocation section");
12254 return (-1);
12255 }
12256
12257 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12258 n = rs->dofs_size / rs->dofs_entsize;
12259
12260 for (i = 0; i < n; i++) {
12261 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12262
12263 switch (r->dofr_type) {
12264 case DOF_RELO_NONE:
12265 break;
12266 case DOF_RELO_SETX:
12267 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12268 sizeof (uint64_t) > ts->dofs_size) {
12269 dtrace_dof_error(dof, "bad relocation offset");
12270 return (-1);
12271 }
12272
12273 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12274 dtrace_dof_error(dof, "misaligned setx relo");
12275 return (-1);
12276 }
12277
12278 *(uint64_t *)taddr += ubase;
12279 break;
12280 default:
12281 dtrace_dof_error(dof, "invalid relocation type");
12282 return (-1);
12283 }
12284
12285 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12286 }
12287
12288 return (0);
12289 }
12290
12291 /*
12292 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12293 * header: it should be at the front of a memory region that is at least
12294 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12295 * size. It need not be validated in any other way.
12296 */
12297 static int
12298 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12299 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12300 {
12301 uint64_t len = dof->dofh_loadsz, seclen;
12302 uintptr_t daddr = (uintptr_t)dof;
12303 dtrace_ecbdesc_t *ep;
12304 dtrace_enabling_t *enab;
12305 uint_t i;
12306
12307 ASSERT(MUTEX_HELD(&dtrace_lock));
12308 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12309
12310 /*
12311 * Check the DOF header identification bytes. In addition to checking
12312 * valid settings, we also verify that unused bits/bytes are zeroed so
12313 * we can use them later without fear of regressing existing binaries.
12314 */
12315 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12316 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12317 dtrace_dof_error(dof, "DOF magic string mismatch");
12318 return (-1);
12319 }
12320
12321 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12322 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12323 dtrace_dof_error(dof, "DOF has invalid data model");
12324 return (-1);
12325 }
12326
12327 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12328 dtrace_dof_error(dof, "DOF encoding mismatch");
12329 return (-1);
12330 }
12331
12332 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12333 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12334 dtrace_dof_error(dof, "DOF version mismatch");
12335 return (-1);
12336 }
12337
12338 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12339 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12340 return (-1);
12341 }
12342
12343 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12344 dtrace_dof_error(dof, "DOF uses too many integer registers");
12345 return (-1);
12346 }
12347
12348 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12349 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12350 return (-1);
12351 }
12352
12353 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12354 if (dof->dofh_ident[i] != 0) {
12355 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12356 return (-1);
12357 }
12358 }
12359
12360 if (dof->dofh_flags & ~DOF_FL_VALID) {
12361 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12362 return (-1);
12363 }
12364
12365 if (dof->dofh_secsize == 0) {
12366 dtrace_dof_error(dof, "zero section header size");
12367 return (-1);
12368 }
12369
12370 /*
12371 * Check that the section headers don't exceed the amount of DOF
12372 * data. Note that we cast the section size and number of sections
12373 * to uint64_t's to prevent possible overflow in the multiplication.
12374 */
12375 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12376
12377 if (dof->dofh_secoff > len || seclen > len ||
12378 dof->dofh_secoff + seclen > len) {
12379 dtrace_dof_error(dof, "truncated section headers");
12380 return (-1);
12381 }
12382
12383 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12384 dtrace_dof_error(dof, "misaligned section headers");
12385 return (-1);
12386 }
12387
12388 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12389 dtrace_dof_error(dof, "misaligned section size");
12390 return (-1);
12391 }
12392
12393 /*
12394 * Take an initial pass through the section headers to be sure that
12395 * the headers don't have stray offsets. If the 'noprobes' flag is
12396 * set, do not permit sections relating to providers, probes, or args.
12397 */
12398 for (i = 0; i < dof->dofh_secnum; i++) {
12399 dof_sec_t *sec = (dof_sec_t *)(daddr +
12400 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12401
12402 if (noprobes) {
12403 switch (sec->dofs_type) {
12404 case DOF_SECT_PROVIDER:
12405 case DOF_SECT_PROBES:
12406 case DOF_SECT_PRARGS:
12407 case DOF_SECT_PROFFS:
12408 dtrace_dof_error(dof, "illegal sections "
12409 "for enabling");
12410 return (-1);
12411 }
12412 }
12413
12414 if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
12415 !(sec->dofs_flags & DOF_SECF_LOAD)) {
12416 dtrace_dof_error(dof, "loadable section with load "
12417 "flag unset");
12418 return (-1);
12419 }
12420
12421 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12422 continue; /* just ignore non-loadable sections */
12423
12424 if (sec->dofs_align & (sec->dofs_align - 1)) {
12425 dtrace_dof_error(dof, "bad section alignment");
12426 return (-1);
12427 }
12428
12429 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12430 dtrace_dof_error(dof, "misaligned section");
12431 return (-1);
12432 }
12433
12434 if (sec->dofs_offset > len || sec->dofs_size > len ||
12435 sec->dofs_offset + sec->dofs_size > len) {
12436 dtrace_dof_error(dof, "corrupt section header");
12437 return (-1);
12438 }
12439
12440 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12441 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12442 dtrace_dof_error(dof, "non-terminating string table");
12443 return (-1);
12444 }
12445 }
12446
12447 /*
12448 * Take a second pass through the sections and locate and perform any
12449 * relocations that are present. We do this after the first pass to
12450 * be sure that all sections have had their headers validated.
12451 */
12452 for (i = 0; i < dof->dofh_secnum; i++) {
12453 dof_sec_t *sec = (dof_sec_t *)(daddr +
12454 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12455
12456 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12457 continue; /* skip sections that are not loadable */
12458
12459 switch (sec->dofs_type) {
12460 case DOF_SECT_URELHDR:
12461 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12462 return (-1);
12463 break;
12464 }
12465 }
12466
12467 if ((enab = *enabp) == NULL)
12468 enab = *enabp = dtrace_enabling_create(vstate);
12469
12470 for (i = 0; i < dof->dofh_secnum; i++) {
12471 dof_sec_t *sec = (dof_sec_t *)(daddr +
12472 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12473
12474 if (sec->dofs_type != DOF_SECT_ECBDESC)
12475 continue;
12476
12477 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12478 dtrace_enabling_destroy(enab);
12479 *enabp = NULL;
12480 return (-1);
12481 }
12482
12483 dtrace_enabling_add(enab, ep);
12484 }
12485
12486 return (0);
12487 }
12488
12489 /*
12490 * Process DOF for any options. This routine assumes that the DOF has been
12491 * at least processed by dtrace_dof_slurp().
12492 */
12493 static int
12494 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12495 {
12496 int i, rval;
12497 uint32_t entsize;
12498 size_t offs;
12499 dof_optdesc_t *desc;
12500
12501 for (i = 0; i < dof->dofh_secnum; i++) {
12502 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12503 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12504
12505 if (sec->dofs_type != DOF_SECT_OPTDESC)
12506 continue;
12507
12508 if (sec->dofs_align != sizeof (uint64_t)) {
12509 dtrace_dof_error(dof, "bad alignment in "
12510 "option description");
12511 return (EINVAL);
12512 }
12513
12514 if ((entsize = sec->dofs_entsize) == 0) {
12515 dtrace_dof_error(dof, "zeroed option entry size");
12516 return (EINVAL);
12517 }
12518
12519 if (entsize < sizeof (dof_optdesc_t)) {
12520 dtrace_dof_error(dof, "bad option entry size");
12521 return (EINVAL);
12522 }
12523
12524 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12525 desc = (dof_optdesc_t *)((uintptr_t)dof +
12526 (uintptr_t)sec->dofs_offset + offs);
12527
12528 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12529 dtrace_dof_error(dof, "non-zero option string");
12530 return (EINVAL);
12531 }
12532
12533 if (desc->dofo_value == DTRACEOPT_UNSET) {
12534 dtrace_dof_error(dof, "unset option");
12535 return (EINVAL);
12536 }
12537
12538 if ((rval = dtrace_state_option(state,
12539 desc->dofo_option, desc->dofo_value)) != 0) {
12540 dtrace_dof_error(dof, "rejected option");
12541 return (rval);
12542 }
12543 }
12544 }
12545
12546 return (0);
12547 }
12548
12549 /*
12550 * DTrace Consumer State Functions
12551 */
12552 int
12553 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12554 {
12555 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12556 void *base;
12557 uintptr_t limit;
12558 dtrace_dynvar_t *dvar, *next, *start;
12559 int i;
12560
12561 ASSERT(MUTEX_HELD(&dtrace_lock));
12562 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12563
12564 bzero(dstate, sizeof (dtrace_dstate_t));
12565
12566 if ((dstate->dtds_chunksize = chunksize) == 0)
12567 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12568
12569 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12570 size = min;
12571
12572 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12573 return (ENOMEM);
12574
12575 dstate->dtds_size = size;
12576 dstate->dtds_base = base;
12577 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12578 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12579
12580 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12581
12582 if (hashsize != 1 && (hashsize & 1))
12583 hashsize--;
12584
12585 dstate->dtds_hashsize = hashsize;
12586 dstate->dtds_hash = dstate->dtds_base;
12587
12588 /*
12589 * Set all of our hash buckets to point to the single sink, and (if
12590 * it hasn't already been set), set the sink's hash value to be the
12591 * sink sentinel value. The sink is needed for dynamic variable
12592 * lookups to know that they have iterated over an entire, valid hash
12593 * chain.
12594 */
12595 for (i = 0; i < hashsize; i++)
12596 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12597
12598 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12599 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12600
12601 /*
12602 * Determine number of active CPUs. Divide free list evenly among
12603 * active CPUs.
12604 */
12605 start = (dtrace_dynvar_t *)
12606 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12607 limit = (uintptr_t)base + size;
12608
12609 maxper = (limit - (uintptr_t)start) / NCPU;
12610 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12611
12612 for (i = 0; i < NCPU; i++) {
12613 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12614
12615 /*
12616 * If we don't even have enough chunks to make it once through
12617 * NCPUs, we're just going to allocate everything to the first
12618 * CPU. And if we're on the last CPU, we're going to allocate
12619 * whatever is left over. In either case, we set the limit to
12620 * be the limit of the dynamic variable space.
12621 */
12622 if (maxper == 0 || i == NCPU - 1) {
12623 limit = (uintptr_t)base + size;
12624 start = NULL;
12625 } else {
12626 limit = (uintptr_t)start + maxper;
12627 start = (dtrace_dynvar_t *)limit;
12628 }
12629
12630 ASSERT(limit <= (uintptr_t)base + size);
12631
12632 for (;;) {
12633 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12634 dstate->dtds_chunksize);
12635
12636 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12637 break;
12638
12639 dvar->dtdv_next = next;
12640 dvar = next;
12641 }
12642
12643 if (maxper == 0)
12644 break;
12645 }
12646
12647 return (0);
12648 }
12649
12650 void
12651 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12652 {
12653 ASSERT(MUTEX_HELD(&cpu_lock));
12654
12655 if (dstate->dtds_base == NULL)
12656 return;
12657
12658 kmem_free(dstate->dtds_base, dstate->dtds_size);
12659 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12660 }
12661
12662 static void
12663 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12664 {
12665 /*
12666 * Logical XOR, where are you?
12667 */
12668 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12669
12670 if (vstate->dtvs_nglobals > 0) {
12671 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12672 sizeof (dtrace_statvar_t *));
12673 }
12674
12675 if (vstate->dtvs_ntlocals > 0) {
12676 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12677 sizeof (dtrace_difv_t));
12678 }
12679
12680 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
12681
12682 if (vstate->dtvs_nlocals > 0) {
12683 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
12684 sizeof (dtrace_statvar_t *));
12685 }
12686 }
12687
12688 static void
12689 dtrace_state_clean(dtrace_state_t *state)
12690 {
12691 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12692 return;
12693
12694 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12695 dtrace_speculation_clean(state);
12696 }
12697
12698 static void
12699 dtrace_state_deadman(dtrace_state_t *state)
12700 {
12701 hrtime_t now;
12702
12703 dtrace_sync();
12704
12705 now = dtrace_gethrtime();
12706
12707 if (state != dtrace_anon.dta_state &&
12708 now - state->dts_laststatus >= dtrace_deadman_user)
12709 return;
12710
12711 /*
12712 * We must be sure that dts_alive never appears to be less than the
12713 * value upon entry to dtrace_state_deadman(), and because we lack a
12714 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12715 * store INT64_MAX to it, followed by a memory barrier, followed by
12716 * the new value. This assures that dts_alive never appears to be
12717 * less than its true value, regardless of the order in which the
12718 * stores to the underlying storage are issued.
12719 */
12720 state->dts_alive = INT64_MAX;
12721 dtrace_membar_producer();
12722 state->dts_alive = now;
12723 }
12724
12725 dtrace_state_t *
12726 dtrace_state_create(dev_t *devp, cred_t *cr)
12727 {
12728 minor_t minor;
12729 major_t major;
12730 char c[30];
12731 dtrace_state_t *state;
12732 dtrace_optval_t *opt;
12733 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12734
12735 ASSERT(MUTEX_HELD(&dtrace_lock));
12736 ASSERT(MUTEX_HELD(&cpu_lock));
12737
12738 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12739 VM_BESTFIT | VM_SLEEP);
12740
12741 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12742 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12743 return (NULL);
12744 }
12745
12746 state = ddi_get_soft_state(dtrace_softstate, minor);
12747 state->dts_epid = DTRACE_EPIDNONE + 1;
12748
12749 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
12750 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12751 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12752
12753 if (devp != NULL) {
12754 major = getemajor(*devp);
12755 } else {
12756 major = ddi_driver_major(dtrace_devi);
12757 }
12758
12759 state->dts_dev = makedevice(major, minor);
12760
12761 if (devp != NULL)
12762 *devp = state->dts_dev;
12763
12764 /*
12765 * We allocate NCPU buffers. On the one hand, this can be quite
12766 * a bit of memory per instance (nearly 36K on a Starcat). On the
12767 * other hand, it saves an additional memory reference in the probe
12768 * path.
12769 */
12770 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12771 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12772 state->dts_cleaner = CYCLIC_NONE;
12773 state->dts_deadman = CYCLIC_NONE;
12774 state->dts_vstate.dtvs_state = state;
12775
12776 for (i = 0; i < DTRACEOPT_MAX; i++)
12777 state->dts_options[i] = DTRACEOPT_UNSET;
12778
12779 /*
12780 * Set the default options.
12781 */
12782 opt = state->dts_options;
12783 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12784 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12785 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12786 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12787 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12788 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12789 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12790 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12791 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12792 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12793 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12794 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12795 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12796 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12797
12798 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12799
12800 /*
12801 * Depending on the user credentials, we set flag bits which alter probe
12802 * visibility or the amount of destructiveness allowed. In the case of
12803 * actual anonymous tracing, or the possession of all privileges, all of
12804 * the normal checks are bypassed.
12805 */
12806 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
12807 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
12808 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
12809 } else {
12810 /*
12811 * Set up the credentials for this instantiation. We take a
12812 * hold on the credential to prevent it from disappearing on
12813 * us; this in turn prevents the zone_t referenced by this
12814 * credential from disappearing. This means that we can
12815 * examine the credential and the zone from probe context.
12816 */
12817 crhold(cr);
12818 state->dts_cred.dcr_cred = cr;
12819
12820 /*
12821 * CRA_PROC means "we have *some* privilege for dtrace" and
12822 * unlocks the use of variables like pid, zonename, etc.
12823 */
12824 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
12825 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12826 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
12827 }
12828
12829 /*
12830 * dtrace_user allows use of syscall and profile providers.
12831 * If the user also has proc_owner and/or proc_zone, we
12832 * extend the scope to include additional visibility and
12833 * destructive power.
12834 */
12835 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
12836 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
12837 state->dts_cred.dcr_visible |=
12838 DTRACE_CRV_ALLPROC;
12839
12840 state->dts_cred.dcr_action |=
12841 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12842 }
12843
12844 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
12845 state->dts_cred.dcr_visible |=
12846 DTRACE_CRV_ALLZONE;
12847
12848 state->dts_cred.dcr_action |=
12849 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12850 }
12851
12852 /*
12853 * If we have all privs in whatever zone this is,
12854 * we can do destructive things to processes which
12855 * have altered credentials.
12856 */
12857 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12858 cr->cr_zone->zone_privset)) {
12859 state->dts_cred.dcr_action |=
12860 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12861 }
12862 }
12863
12864 /*
12865 * Holding the dtrace_kernel privilege also implies that
12866 * the user has the dtrace_user privilege from a visibility
12867 * perspective. But without further privileges, some
12868 * destructive actions are not available.
12869 */
12870 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
12871 /*
12872 * Make all probes in all zones visible. However,
12873 * this doesn't mean that all actions become available
12874 * to all zones.
12875 */
12876 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
12877 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
12878
12879 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
12880 DTRACE_CRA_PROC;
12881 /*
12882 * Holding proc_owner means that destructive actions
12883 * for *this* zone are allowed.
12884 */
12885 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12886 state->dts_cred.dcr_action |=
12887 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12888
12889 /*
12890 * Holding proc_zone means that destructive actions
12891 * for this user/group ID in all zones is allowed.
12892 */
12893 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12894 state->dts_cred.dcr_action |=
12895 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12896
12897 /*
12898 * If we have all privs in whatever zone this is,
12899 * we can do destructive things to processes which
12900 * have altered credentials.
12901 */
12902 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12903 cr->cr_zone->zone_privset)) {
12904 state->dts_cred.dcr_action |=
12905 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12906 }
12907 }
12908
12909 /*
12910 * Holding the dtrace_proc privilege gives control over fasttrap
12911 * and pid providers. We need to grant wider destructive
12912 * privileges in the event that the user has proc_owner and/or
12913 * proc_zone.
12914 */
12915 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12916 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12917 state->dts_cred.dcr_action |=
12918 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12919
12920 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12921 state->dts_cred.dcr_action |=
12922 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12923 }
12924 }
12925
12926 return (state);
12927 }
12928
12929 static int
12930 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
12931 {
12932 dtrace_optval_t *opt = state->dts_options, size;
12933 processorid_t cpu;
12934 int flags = 0, rval, factor, divisor = 1;
12935
12936 ASSERT(MUTEX_HELD(&dtrace_lock));
12937 ASSERT(MUTEX_HELD(&cpu_lock));
12938 ASSERT(which < DTRACEOPT_MAX);
12939 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
12940 (state == dtrace_anon.dta_state &&
12941 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
12942
12943 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
12944 return (0);
12945
12946 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
12947 cpu = opt[DTRACEOPT_CPU];
12948
12949 if (which == DTRACEOPT_SPECSIZE)
12950 flags |= DTRACEBUF_NOSWITCH;
12951
12952 if (which == DTRACEOPT_BUFSIZE) {
12953 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
12954 flags |= DTRACEBUF_RING;
12955
12956 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
12957 flags |= DTRACEBUF_FILL;
12958
12959 if (state != dtrace_anon.dta_state ||
12960 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
12961 flags |= DTRACEBUF_INACTIVE;
12962 }
12963
12964 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
12965 /*
12966 * The size must be 8-byte aligned. If the size is not 8-byte
12967 * aligned, drop it down by the difference.
12968 */
12969 if (size & (sizeof (uint64_t) - 1))
12970 size -= size & (sizeof (uint64_t) - 1);
12971
12972 if (size < state->dts_reserve) {
12973 /*
12974 * Buffers always must be large enough to accommodate
12975 * their prereserved space. We return E2BIG instead
12976 * of ENOMEM in this case to allow for user-level
12977 * software to differentiate the cases.
12978 */
12979 return (E2BIG);
12980 }
12981
12982 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
12983
12984 if (rval != ENOMEM) {
12985 opt[which] = size;
12986 return (rval);
12987 }
12988
12989 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
12990 return (rval);
12991
12992 for (divisor = 2; divisor < factor; divisor <<= 1)
12993 continue;
12994 }
12995
12996 return (ENOMEM);
12997 }
12998
12999 static int
13000 dtrace_state_buffers(dtrace_state_t *state)
13001 {
13002 dtrace_speculation_t *spec = state->dts_speculations;
13003 int rval, i;
13004
13005 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13006 DTRACEOPT_BUFSIZE)) != 0)
13007 return (rval);
13008
13009 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13010 DTRACEOPT_AGGSIZE)) != 0)
13011 return (rval);
13012
13013 for (i = 0; i < state->dts_nspeculations; i++) {
13014 if ((rval = dtrace_state_buffer(state,
13015 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13016 return (rval);
13017 }
13018
13019 return (0);
13020 }
13021
13022 static void
13023 dtrace_state_prereserve(dtrace_state_t *state)
13024 {
13025 dtrace_ecb_t *ecb;
13026 dtrace_probe_t *probe;
13027
13028 state->dts_reserve = 0;
13029
13030 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13031 return;
13032
13033 /*
13034 * If our buffer policy is a "fill" buffer policy, we need to set the
13035 * prereserved space to be the space required by the END probes.
13036 */
13037 probe = dtrace_probes[dtrace_probeid_end - 1];
13038 ASSERT(probe != NULL);
13039
13040 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13041 if (ecb->dte_state != state)
13042 continue;
13043
13044 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13045 }
13046 }
13047
13048 static int
13049 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13050 {
13051 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13052 dtrace_speculation_t *spec;
13053 dtrace_buffer_t *buf;
13054 cyc_handler_t hdlr;
13055 cyc_time_t when;
13056 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13057 dtrace_icookie_t cookie;
13058
13059 mutex_enter(&cpu_lock);
13060 mutex_enter(&dtrace_lock);
13061
13062 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13063 rval = EBUSY;
13064 goto out;
13065 }
13066
13067 /*
13068 * Before we can perform any checks, we must prime all of the
13069 * retained enablings that correspond to this state.
13070 */
13071 dtrace_enabling_prime(state);
13072
13073 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13074 rval = EACCES;
13075 goto out;
13076 }
13077
13078 dtrace_state_prereserve(state);
13079
13080 /*
13081 * Now we want to do is try to allocate our speculations.
13082 * We do not automatically resize the number of speculations; if
13083 * this fails, we will fail the operation.
13084 */
13085 nspec = opt[DTRACEOPT_NSPEC];
13086 ASSERT(nspec != DTRACEOPT_UNSET);
13087
13088 if (nspec > INT_MAX) {
13089 rval = ENOMEM;
13090 goto out;
13091 }
13092
13093 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
13094 KM_NOSLEEP | KM_NORMALPRI);
13095
13096 if (spec == NULL) {
13097 rval = ENOMEM;
13098 goto out;
13099 }
13100
13101 state->dts_speculations = spec;
13102 state->dts_nspeculations = (int)nspec;
13103
13104 for (i = 0; i < nspec; i++) {
13105 if ((buf = kmem_zalloc(bufsize,
13106 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
13107 rval = ENOMEM;
13108 goto err;
13109 }
13110
13111 spec[i].dtsp_buffer = buf;
13112 }
13113
13114 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13115 if (dtrace_anon.dta_state == NULL) {
13116 rval = ENOENT;
13117 goto out;
13118 }
13119
13120 if (state->dts_necbs != 0) {
13121 rval = EALREADY;
13122 goto out;
13123 }
13124
13125 state->dts_anon = dtrace_anon_grab();
13126 ASSERT(state->dts_anon != NULL);
13127 state = state->dts_anon;
13128
13129 /*
13130 * We want "grabanon" to be set in the grabbed state, so we'll
13131 * copy that option value from the grabbing state into the
13132 * grabbed state.
13133 */
13134 state->dts_options[DTRACEOPT_GRABANON] =
13135 opt[DTRACEOPT_GRABANON];
13136
13137 *cpu = dtrace_anon.dta_beganon;
13138
13139 /*
13140 * If the anonymous state is active (as it almost certainly
13141 * is if the anonymous enabling ultimately matched anything),
13142 * we don't allow any further option processing -- but we
13143 * don't return failure.
13144 */
13145 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13146 goto out;
13147 }
13148
13149 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13150 opt[DTRACEOPT_AGGSIZE] != 0) {
13151 if (state->dts_aggregations == NULL) {
13152 /*
13153 * We're not going to create an aggregation buffer
13154 * because we don't have any ECBs that contain
13155 * aggregations -- set this option to 0.
13156 */
13157 opt[DTRACEOPT_AGGSIZE] = 0;
13158 } else {
13159 /*
13160 * If we have an aggregation buffer, we must also have
13161 * a buffer to use as scratch.
13162 */
13163 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13164 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13165 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13166 }
13167 }
13168 }
13169
13170 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13171 opt[DTRACEOPT_SPECSIZE] != 0) {
13172 if (!state->dts_speculates) {
13173 /*
13174 * We're not going to create speculation buffers
13175 * because we don't have any ECBs that actually
13176 * speculate -- set the speculation size to 0.
13177 */
13178 opt[DTRACEOPT_SPECSIZE] = 0;
13179 }
13180 }
13181
13182 /*
13183 * The bare minimum size for any buffer that we're actually going to
13184 * do anything to is sizeof (uint64_t).
13185 */
13186 sz = sizeof (uint64_t);
13187
13188 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13189 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13190 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13191 /*
13192 * A buffer size has been explicitly set to 0 (or to a size
13193 * that will be adjusted to 0) and we need the space -- we
13194 * need to return failure. We return ENOSPC to differentiate
13195 * it from failing to allocate a buffer due to failure to meet
13196 * the reserve (for which we return E2BIG).
13197 */
13198 rval = ENOSPC;
13199 goto out;
13200 }
13201
13202 if ((rval = dtrace_state_buffers(state)) != 0)
13203 goto err;
13204
13205 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13206 sz = dtrace_dstate_defsize;
13207
13208 do {
13209 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13210
13211 if (rval == 0)
13212 break;
13213
13214 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13215 goto err;
13216 } while (sz >>= 1);
13217
13218 opt[DTRACEOPT_DYNVARSIZE] = sz;
13219
13220 if (rval != 0)
13221 goto err;
13222
13223 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13224 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13225
13226 if (opt[DTRACEOPT_CLEANRATE] == 0)
13227 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13228
13229 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13230 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13231
13232 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13233 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13234
13235 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13236 hdlr.cyh_arg = state;
13237 hdlr.cyh_level = CY_LOW_LEVEL;
13238
13239 when.cyt_when = 0;
13240 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13241
13242 state->dts_cleaner = cyclic_add(&hdlr, &when);
13243
13244 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13245 hdlr.cyh_arg = state;
13246 hdlr.cyh_level = CY_LOW_LEVEL;
13247
13248 when.cyt_when = 0;
13249 when.cyt_interval = dtrace_deadman_interval;
13250
13251 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13252 state->dts_deadman = cyclic_add(&hdlr, &when);
13253
13254 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13255
13256 if (state->dts_getf != 0 &&
13257 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13258 /*
13259 * We don't have kernel privs but we have at least one call
13260 * to getf(); we need to bump our zone's count, and (if
13261 * this is the first enabling to have an unprivileged call
13262 * to getf()) we need to hook into closef().
13263 */
13264 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
13265
13266 if (dtrace_getf++ == 0) {
13267 ASSERT(dtrace_closef == NULL);
13268 dtrace_closef = dtrace_getf_barrier;
13269 }
13270 }
13271
13272 /*
13273 * Now it's time to actually fire the BEGIN probe. We need to disable
13274 * interrupts here both to record the CPU on which we fired the BEGIN
13275 * probe (the data from this CPU will be processed first at user
13276 * level) and to manually activate the buffer for this CPU.
13277 */
13278 cookie = dtrace_interrupt_disable();
13279 *cpu = CPU->cpu_id;
13280 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13281 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13282
13283 dtrace_probe(dtrace_probeid_begin,
13284 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13285 dtrace_interrupt_enable(cookie);
13286 /*
13287 * We may have had an exit action from a BEGIN probe; only change our
13288 * state to ACTIVE if we're still in WARMUP.
13289 */
13290 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13291 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13292
13293 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13294 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13295
13296 /*
13297 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13298 * want each CPU to transition its principal buffer out of the
13299 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13300 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13301 * atomically transition from processing none of a state's ECBs to
13302 * processing all of them.
13303 */
13304 dtrace_xcall(DTRACE_CPUALL,
13305 (dtrace_xcall_t)dtrace_buffer_activate, state);
13306 goto out;
13307
13308 err:
13309 dtrace_buffer_free(state->dts_buffer);
13310 dtrace_buffer_free(state->dts_aggbuffer);
13311
13312 if ((nspec = state->dts_nspeculations) == 0) {
13313 ASSERT(state->dts_speculations == NULL);
13314 goto out;
13315 }
13316
13317 spec = state->dts_speculations;
13318 ASSERT(spec != NULL);
13319
13320 for (i = 0; i < state->dts_nspeculations; i++) {
13321 if ((buf = spec[i].dtsp_buffer) == NULL)
13322 break;
13323
13324 dtrace_buffer_free(buf);
13325 kmem_free(buf, bufsize);
13326 }
13327
13328 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13329 state->dts_nspeculations = 0;
13330 state->dts_speculations = NULL;
13331
13332 out:
13333 mutex_exit(&dtrace_lock);
13334 mutex_exit(&cpu_lock);
13335
13336 return (rval);
13337 }
13338
13339 static int
13340 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13341 {
13342 dtrace_icookie_t cookie;
13343
13344 ASSERT(MUTEX_HELD(&dtrace_lock));
13345
13346 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13347 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13348 return (EINVAL);
13349
13350 /*
13351 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13352 * to be sure that every CPU has seen it. See below for the details
13353 * on why this is done.
13354 */
13355 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13356 dtrace_sync();
13357
13358 /*
13359 * By this point, it is impossible for any CPU to be still processing
13360 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13361 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13362 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13363 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13364 * iff we're in the END probe.
13365 */
13366 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13367 dtrace_sync();
13368 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13369
13370 /*
13371 * Finally, we can release the reserve and call the END probe. We
13372 * disable interrupts across calling the END probe to allow us to
13373 * return the CPU on which we actually called the END probe. This
13374 * allows user-land to be sure that this CPU's principal buffer is
13375 * processed last.
13376 */
13377 state->dts_reserve = 0;
13378
13379 cookie = dtrace_interrupt_disable();
13380 *cpu = CPU->cpu_id;
13381 dtrace_probe(dtrace_probeid_end,
13382 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13383 dtrace_interrupt_enable(cookie);
13384
13385 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13386 dtrace_sync();
13387
13388 if (state->dts_getf != 0 &&
13389 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13390 /*
13391 * We don't have kernel privs but we have at least one call
13392 * to getf(); we need to lower our zone's count, and (if
13393 * this is the last enabling to have an unprivileged call
13394 * to getf()) we need to clear the closef() hook.
13395 */
13396 ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
13397 ASSERT(dtrace_closef == dtrace_getf_barrier);
13398 ASSERT(dtrace_getf > 0);
13399
13400 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
13401
13402 if (--dtrace_getf == 0)
13403 dtrace_closef = NULL;
13404 }
13405
13406 return (0);
13407 }
13408
13409 static int
13410 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13411 dtrace_optval_t val)
13412 {
13413 ASSERT(MUTEX_HELD(&dtrace_lock));
13414
13415 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13416 return (EBUSY);
13417
13418 if (option >= DTRACEOPT_MAX)
13419 return (EINVAL);
13420
13421 if (option != DTRACEOPT_CPU && val < 0)
13422 return (EINVAL);
13423
13424 switch (option) {
13425 case DTRACEOPT_DESTRUCTIVE:
13426 if (dtrace_destructive_disallow)
13427 return (EACCES);
13428
13429 state->dts_cred.dcr_destructive = 1;
13430 break;
13431
13432 case DTRACEOPT_BUFSIZE:
13433 case DTRACEOPT_DYNVARSIZE:
13434 case DTRACEOPT_AGGSIZE:
13435 case DTRACEOPT_SPECSIZE:
13436 case DTRACEOPT_STRSIZE:
13437 if (val < 0)
13438 return (EINVAL);
13439
13440 if (val >= LONG_MAX) {
13441 /*
13442 * If this is an otherwise negative value, set it to
13443 * the highest multiple of 128m less than LONG_MAX.
13444 * Technically, we're adjusting the size without
13445 * regard to the buffer resizing policy, but in fact,
13446 * this has no effect -- if we set the buffer size to
13447 * ~LONG_MAX and the buffer policy is ultimately set to
13448 * be "manual", the buffer allocation is guaranteed to
13449 * fail, if only because the allocation requires two
13450 * buffers. (We set the the size to the highest
13451 * multiple of 128m because it ensures that the size
13452 * will remain a multiple of a megabyte when
13453 * repeatedly halved -- all the way down to 15m.)
13454 */
13455 val = LONG_MAX - (1 << 27) + 1;
13456 }
13457 }
13458
13459 state->dts_options[option] = val;
13460
13461 return (0);
13462 }
13463
13464 static void
13465 dtrace_state_destroy(dtrace_state_t *state)
13466 {
13467 dtrace_ecb_t *ecb;
13468 dtrace_vstate_t *vstate = &state->dts_vstate;
13469 minor_t minor = getminor(state->dts_dev);
13470 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13471 dtrace_speculation_t *spec = state->dts_speculations;
13472 int nspec = state->dts_nspeculations;
13473 uint32_t match;
13474
13475 ASSERT(MUTEX_HELD(&dtrace_lock));
13476 ASSERT(MUTEX_HELD(&cpu_lock));
13477
13478 /*
13479 * First, retract any retained enablings for this state.
13480 */
13481 dtrace_enabling_retract(state);
13482 ASSERT(state->dts_nretained == 0);
13483
13484 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13485 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13486 /*
13487 * We have managed to come into dtrace_state_destroy() on a
13488 * hot enabling -- almost certainly because of a disorderly
13489 * shutdown of a consumer. (That is, a consumer that is
13490 * exiting without having called dtrace_stop().) In this case,
13491 * we're going to set our activity to be KILLED, and then
13492 * issue a sync to be sure that everyone is out of probe
13493 * context before we start blowing away ECBs.
13494 */
13495 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13496 dtrace_sync();
13497 }
13498
13499 /*
13500 * Release the credential hold we took in dtrace_state_create().
13501 */
13502 if (state->dts_cred.dcr_cred != NULL)
13503 crfree(state->dts_cred.dcr_cred);
13504
13505 /*
13506 * Now we can safely disable and destroy any enabled probes. Because
13507 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13508 * (especially if they're all enabled), we take two passes through the
13509 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13510 * in the second we disable whatever is left over.
13511 */
13512 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13513 for (i = 0; i < state->dts_necbs; i++) {
13514 if ((ecb = state->dts_ecbs[i]) == NULL)
13515 continue;
13516
13517 if (match && ecb->dte_probe != NULL) {
13518 dtrace_probe_t *probe = ecb->dte_probe;
13519 dtrace_provider_t *prov = probe->dtpr_provider;
13520
13521 if (!(prov->dtpv_priv.dtpp_flags & match))
13522 continue;
13523 }
13524
13525 dtrace_ecb_disable(ecb);
13526 dtrace_ecb_destroy(ecb);
13527 }
13528
13529 if (!match)
13530 break;
13531 }
13532
13533 /*
13534 * Before we free the buffers, perform one more sync to assure that
13535 * every CPU is out of probe context.
13536 */
13537 dtrace_sync();
13538
13539 dtrace_buffer_free(state->dts_buffer);
13540 dtrace_buffer_free(state->dts_aggbuffer);
13541
13542 for (i = 0; i < nspec; i++)
13543 dtrace_buffer_free(spec[i].dtsp_buffer);
13544
13545 if (state->dts_cleaner != CYCLIC_NONE)
13546 cyclic_remove(state->dts_cleaner);
13547
13548 if (state->dts_deadman != CYCLIC_NONE)
13549 cyclic_remove(state->dts_deadman);
13550
13551 dtrace_dstate_fini(&vstate->dtvs_dynvars);
13552 dtrace_vstate_fini(vstate);
13553 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13554
13555 if (state->dts_aggregations != NULL) {
13556 #ifdef DEBUG
13557 for (i = 0; i < state->dts_naggregations; i++)
13558 ASSERT(state->dts_aggregations[i] == NULL);
13559 #endif
13560 ASSERT(state->dts_naggregations > 0);
13561 kmem_free(state->dts_aggregations,
13562 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13563 }
13564
13565 kmem_free(state->dts_buffer, bufsize);
13566 kmem_free(state->dts_aggbuffer, bufsize);
13567
13568 for (i = 0; i < nspec; i++)
13569 kmem_free(spec[i].dtsp_buffer, bufsize);
13570
13571 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13572
13573 dtrace_format_destroy(state);
13574
13575 vmem_destroy(state->dts_aggid_arena);
13576 ddi_soft_state_free(dtrace_softstate, minor);
13577 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13578 }
13579
13580 /*
13581 * DTrace Anonymous Enabling Functions
13582 */
13583 static dtrace_state_t *
13584 dtrace_anon_grab(void)
13585 {
13586 dtrace_state_t *state;
13587
13588 ASSERT(MUTEX_HELD(&dtrace_lock));
13589
13590 if ((state = dtrace_anon.dta_state) == NULL) {
13591 ASSERT(dtrace_anon.dta_enabling == NULL);
13592 return (NULL);
13593 }
13594
13595 ASSERT(dtrace_anon.dta_enabling != NULL);
13596 ASSERT(dtrace_retained != NULL);
13597
13598 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13599 dtrace_anon.dta_enabling = NULL;
13600 dtrace_anon.dta_state = NULL;
13601
13602 return (state);
13603 }
13604
13605 static void
13606 dtrace_anon_property(void)
13607 {
13608 int i, rv;
13609 dtrace_state_t *state;
13610 dof_hdr_t *dof;
13611 char c[32]; /* enough for "dof-data-" + digits */
13612
13613 ASSERT(MUTEX_HELD(&dtrace_lock));
13614 ASSERT(MUTEX_HELD(&cpu_lock));
13615
13616 for (i = 0; ; i++) {
13617 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
13618
13619 dtrace_err_verbose = 1;
13620
13621 if ((dof = dtrace_dof_property(c)) == NULL) {
13622 dtrace_err_verbose = 0;
13623 break;
13624 }
13625
13626 /*
13627 * We want to create anonymous state, so we need to transition
13628 * the kernel debugger to indicate that DTrace is active. If
13629 * this fails (e.g. because the debugger has modified text in
13630 * some way), we won't continue with the processing.
13631 */
13632 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13633 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
13634 "enabling ignored.");
13635 dtrace_dof_destroy(dof);
13636 break;
13637 }
13638
13639 /*
13640 * If we haven't allocated an anonymous state, we'll do so now.
13641 */
13642 if ((state = dtrace_anon.dta_state) == NULL) {
13643 state = dtrace_state_create(NULL, NULL);
13644 dtrace_anon.dta_state = state;
13645
13646 if (state == NULL) {
13647 /*
13648 * This basically shouldn't happen: the only
13649 * failure mode from dtrace_state_create() is a
13650 * failure of ddi_soft_state_zalloc() that
13651 * itself should never happen. Still, the
13652 * interface allows for a failure mode, and
13653 * we want to fail as gracefully as possible:
13654 * we'll emit an error message and cease
13655 * processing anonymous state in this case.
13656 */
13657 cmn_err(CE_WARN, "failed to create "
13658 "anonymous state");
13659 dtrace_dof_destroy(dof);
13660 break;
13661 }
13662 }
13663
13664 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
13665 &dtrace_anon.dta_enabling, 0, B_TRUE);
13666
13667 if (rv == 0)
13668 rv = dtrace_dof_options(dof, state);
13669
13670 dtrace_err_verbose = 0;
13671 dtrace_dof_destroy(dof);
13672
13673 if (rv != 0) {
13674 /*
13675 * This is malformed DOF; chuck any anonymous state
13676 * that we created.
13677 */
13678 ASSERT(dtrace_anon.dta_enabling == NULL);
13679 dtrace_state_destroy(state);
13680 dtrace_anon.dta_state = NULL;
13681 break;
13682 }
13683
13684 ASSERT(dtrace_anon.dta_enabling != NULL);
13685 }
13686
13687 if (dtrace_anon.dta_enabling != NULL) {
13688 int rval;
13689
13690 /*
13691 * dtrace_enabling_retain() can only fail because we are
13692 * trying to retain more enablings than are allowed -- but
13693 * we only have one anonymous enabling, and we are guaranteed
13694 * to be allowed at least one retained enabling; we assert
13695 * that dtrace_enabling_retain() returns success.
13696 */
13697 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
13698 ASSERT(rval == 0);
13699
13700 dtrace_enabling_dump(dtrace_anon.dta_enabling);
13701 }
13702 }
13703
13704 /*
13705 * DTrace Helper Functions
13706 */
13707 static void
13708 dtrace_helper_trace(dtrace_helper_action_t *helper,
13709 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
13710 {
13711 uint32_t size, next, nnext, i;
13712 dtrace_helptrace_t *ent, *buffer;
13713 uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13714
13715 if ((buffer = dtrace_helptrace_buffer) == NULL)
13716 return;
13717
13718 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
13719
13720 /*
13721 * What would a tracing framework be without its own tracing
13722 * framework? (Well, a hell of a lot simpler, for starters...)
13723 */
13724 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
13725 sizeof (uint64_t) - sizeof (uint64_t);
13726
13727 /*
13728 * Iterate until we can allocate a slot in the trace buffer.
13729 */
13730 do {
13731 next = dtrace_helptrace_next;
13732
13733 if (next + size < dtrace_helptrace_bufsize) {
13734 nnext = next + size;
13735 } else {
13736 nnext = size;
13737 }
13738 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
13739
13740 /*
13741 * We have our slot; fill it in.
13742 */
13743 if (nnext == size) {
13744 dtrace_helptrace_wrapped++;
13745 next = 0;
13746 }
13747
13748 ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
13749 ent->dtht_helper = helper;
13750 ent->dtht_where = where;
13751 ent->dtht_nlocals = vstate->dtvs_nlocals;
13752
13753 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13754 mstate->dtms_fltoffs : -1;
13755 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13756 ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
13757
13758 for (i = 0; i < vstate->dtvs_nlocals; i++) {
13759 dtrace_statvar_t *svar;
13760
13761 if ((svar = vstate->dtvs_locals[i]) == NULL)
13762 continue;
13763
13764 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13765 ent->dtht_locals[i] =
13766 ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
13767 }
13768 }
13769
13770 static uint64_t
13771 dtrace_helper(int which, dtrace_mstate_t *mstate,
13772 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13773 {
13774 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13775 uint64_t sarg0 = mstate->dtms_arg[0];
13776 uint64_t sarg1 = mstate->dtms_arg[1];
13777 uint64_t rval;
13778 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13779 dtrace_helper_action_t *helper;
13780 dtrace_vstate_t *vstate;
13781 dtrace_difo_t *pred;
13782 int i, trace = dtrace_helptrace_buffer != NULL;
13783
13784 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13785
13786 if (helpers == NULL)
13787 return (0);
13788
13789 if ((helper = helpers->dthps_actions[which]) == NULL)
13790 return (0);
13791
13792 vstate = &helpers->dthps_vstate;
13793 mstate->dtms_arg[0] = arg0;
13794 mstate->dtms_arg[1] = arg1;
13795
13796 /*
13797 * Now iterate over each helper. If its predicate evaluates to 'true',
13798 * we'll call the corresponding actions. Note that the below calls
13799 * to dtrace_dif_emulate() may set faults in machine state. This is
13800 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
13801 * the stored DIF offset with its own (which is the desired behavior).
13802 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13803 * from machine state; this is okay, too.
13804 */
13805 for (; helper != NULL; helper = helper->dtha_next) {
13806 if ((pred = helper->dtha_predicate) != NULL) {
13807 if (trace)
13808 dtrace_helper_trace(helper, mstate, vstate, 0);
13809
13810 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
13811 goto next;
13812
13813 if (*flags & CPU_DTRACE_FAULT)
13814 goto err;
13815 }
13816
13817 for (i = 0; i < helper->dtha_nactions; i++) {
13818 if (trace)
13819 dtrace_helper_trace(helper,
13820 mstate, vstate, i + 1);
13821
13822 rval = dtrace_dif_emulate(helper->dtha_actions[i],
13823 mstate, vstate, state);
13824
13825 if (*flags & CPU_DTRACE_FAULT)
13826 goto err;
13827 }
13828
13829 next:
13830 if (trace)
13831 dtrace_helper_trace(helper, mstate, vstate,
13832 DTRACE_HELPTRACE_NEXT);
13833 }
13834
13835 if (trace)
13836 dtrace_helper_trace(helper, mstate, vstate,
13837 DTRACE_HELPTRACE_DONE);
13838
13839 /*
13840 * Restore the arg0 that we saved upon entry.
13841 */
13842 mstate->dtms_arg[0] = sarg0;
13843 mstate->dtms_arg[1] = sarg1;
13844
13845 return (rval);
13846
13847 err:
13848 if (trace)
13849 dtrace_helper_trace(helper, mstate, vstate,
13850 DTRACE_HELPTRACE_ERR);
13851
13852 /*
13853 * Restore the arg0 that we saved upon entry.
13854 */
13855 mstate->dtms_arg[0] = sarg0;
13856 mstate->dtms_arg[1] = sarg1;
13857
13858 return (NULL);
13859 }
13860
13861 static void
13862 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
13863 dtrace_vstate_t *vstate)
13864 {
13865 int i;
13866
13867 if (helper->dtha_predicate != NULL)
13868 dtrace_difo_release(helper->dtha_predicate, vstate);
13869
13870 for (i = 0; i < helper->dtha_nactions; i++) {
13871 ASSERT(helper->dtha_actions[i] != NULL);
13872 dtrace_difo_release(helper->dtha_actions[i], vstate);
13873 }
13874
13875 kmem_free(helper->dtha_actions,
13876 helper->dtha_nactions * sizeof (dtrace_difo_t *));
13877 kmem_free(helper, sizeof (dtrace_helper_action_t));
13878 }
13879
13880 static int
13881 dtrace_helper_destroygen(int gen)
13882 {
13883 proc_t *p = curproc;
13884 dtrace_helpers_t *help = p->p_dtrace_helpers;
13885 dtrace_vstate_t *vstate;
13886 int i;
13887
13888 ASSERT(MUTEX_HELD(&dtrace_lock));
13889
13890 if (help == NULL || gen > help->dthps_generation)
13891 return (EINVAL);
13892
13893 vstate = &help->dthps_vstate;
13894
13895 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13896 dtrace_helper_action_t *last = NULL, *h, *next;
13897
13898 for (h = help->dthps_actions[i]; h != NULL; h = next) {
13899 next = h->dtha_next;
13900
13901 if (h->dtha_generation == gen) {
13902 if (last != NULL) {
13903 last->dtha_next = next;
13904 } else {
13905 help->dthps_actions[i] = next;
13906 }
13907
13908 dtrace_helper_action_destroy(h, vstate);
13909 } else {
13910 last = h;
13911 }
13912 }
13913 }
13914
13915 /*
13916 * Interate until we've cleared out all helper providers with the
13917 * given generation number.
13918 */
13919 for (;;) {
13920 dtrace_helper_provider_t *prov;
13921
13922 /*
13923 * Look for a helper provider with the right generation. We
13924 * have to start back at the beginning of the list each time
13925 * because we drop dtrace_lock. It's unlikely that we'll make
13926 * more than two passes.
13927 */
13928 for (i = 0; i < help->dthps_nprovs; i++) {
13929 prov = help->dthps_provs[i];
13930
13931 if (prov->dthp_generation == gen)
13932 break;
13933 }
13934
13935 /*
13936 * If there were no matches, we're done.
13937 */
13938 if (i == help->dthps_nprovs)
13939 break;
13940
13941 /*
13942 * Move the last helper provider into this slot.
13943 */
13944 help->dthps_nprovs--;
13945 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
13946 help->dthps_provs[help->dthps_nprovs] = NULL;
13947
13948 mutex_exit(&dtrace_lock);
13949
13950 /*
13951 * If we have a meta provider, remove this helper provider.
13952 */
13953 mutex_enter(&dtrace_meta_lock);
13954 if (dtrace_meta_pid != NULL) {
13955 ASSERT(dtrace_deferred_pid == NULL);
13956 dtrace_helper_provider_remove(&prov->dthp_prov,
13957 p->p_pid);
13958 }
13959 mutex_exit(&dtrace_meta_lock);
13960
13961 dtrace_helper_provider_destroy(prov);
13962
13963 mutex_enter(&dtrace_lock);
13964 }
13965
13966 return (0);
13967 }
13968
13969 static int
13970 dtrace_helper_validate(dtrace_helper_action_t *helper)
13971 {
13972 int err = 0, i;
13973 dtrace_difo_t *dp;
13974
13975 if ((dp = helper->dtha_predicate) != NULL)
13976 err += dtrace_difo_validate_helper(dp);
13977
13978 for (i = 0; i < helper->dtha_nactions; i++)
13979 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
13980
13981 return (err == 0);
13982 }
13983
13984 static int
13985 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
13986 {
13987 dtrace_helpers_t *help;
13988 dtrace_helper_action_t *helper, *last;
13989 dtrace_actdesc_t *act;
13990 dtrace_vstate_t *vstate;
13991 dtrace_predicate_t *pred;
13992 int count = 0, nactions = 0, i;
13993
13994 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
13995 return (EINVAL);
13996
13997 help = curproc->p_dtrace_helpers;
13998 last = help->dthps_actions[which];
13999 vstate = &help->dthps_vstate;
14000
14001 for (count = 0; last != NULL; last = last->dtha_next) {
14002 count++;
14003 if (last->dtha_next == NULL)
14004 break;
14005 }
14006
14007 /*
14008 * If we already have dtrace_helper_actions_max helper actions for this
14009 * helper action type, we'll refuse to add a new one.
14010 */
14011 if (count >= dtrace_helper_actions_max)
14012 return (ENOSPC);
14013
14014 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14015 helper->dtha_generation = help->dthps_generation;
14016
14017 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14018 ASSERT(pred->dtp_difo != NULL);
14019 dtrace_difo_hold(pred->dtp_difo);
14020 helper->dtha_predicate = pred->dtp_difo;
14021 }
14022
14023 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14024 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14025 goto err;
14026
14027 if (act->dtad_difo == NULL)
14028 goto err;
14029
14030 nactions++;
14031 }
14032
14033 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14034 (helper->dtha_nactions = nactions), KM_SLEEP);
14035
14036 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14037 dtrace_difo_hold(act->dtad_difo);
14038 helper->dtha_actions[i++] = act->dtad_difo;
14039 }
14040
14041 if (!dtrace_helper_validate(helper))
14042 goto err;
14043
14044 if (last == NULL) {
14045 help->dthps_actions[which] = helper;
14046 } else {
14047 last->dtha_next = helper;
14048 }
14049
14050 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14051 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14052 dtrace_helptrace_next = 0;
14053 }
14054
14055 return (0);
14056 err:
14057 dtrace_helper_action_destroy(helper, vstate);
14058 return (EINVAL);
14059 }
14060
14061 static void
14062 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14063 dof_helper_t *dofhp)
14064 {
14065 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14066
14067 mutex_enter(&dtrace_meta_lock);
14068 mutex_enter(&dtrace_lock);
14069
14070 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14071 /*
14072 * If the dtrace module is loaded but not attached, or if
14073 * there aren't isn't a meta provider registered to deal with
14074 * these provider descriptions, we need to postpone creating
14075 * the actual providers until later.
14076 */
14077
14078 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14079 dtrace_deferred_pid != help) {
14080 help->dthps_deferred = 1;
14081 help->dthps_pid = p->p_pid;
14082 help->dthps_next = dtrace_deferred_pid;
14083 help->dthps_prev = NULL;
14084 if (dtrace_deferred_pid != NULL)
14085 dtrace_deferred_pid->dthps_prev = help;
14086 dtrace_deferred_pid = help;
14087 }
14088
14089 mutex_exit(&dtrace_lock);
14090
14091 } else if (dofhp != NULL) {
14092 /*
14093 * If the dtrace module is loaded and we have a particular
14094 * helper provider description, pass that off to the
14095 * meta provider.
14096 */
14097
14098 mutex_exit(&dtrace_lock);
14099
14100 dtrace_helper_provide(dofhp, p->p_pid);
14101
14102 } else {
14103 /*
14104 * Otherwise, just pass all the helper provider descriptions
14105 * off to the meta provider.
14106 */
14107
14108 int i;
14109 mutex_exit(&dtrace_lock);
14110
14111 for (i = 0; i < help->dthps_nprovs; i++) {
14112 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14113 p->p_pid);
14114 }
14115 }
14116
14117 mutex_exit(&dtrace_meta_lock);
14118 }
14119
14120 static int
14121 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14122 {
14123 dtrace_helpers_t *help;
14124 dtrace_helper_provider_t *hprov, **tmp_provs;
14125 uint_t tmp_maxprovs, i;
14126
14127 ASSERT(MUTEX_HELD(&dtrace_lock));
14128
14129 help = curproc->p_dtrace_helpers;
14130 ASSERT(help != NULL);
14131
14132 /*
14133 * If we already have dtrace_helper_providers_max helper providers,
14134 * we're refuse to add a new one.
14135 */
14136 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14137 return (ENOSPC);
14138
14139 /*
14140 * Check to make sure this isn't a duplicate.
14141 */
14142 for (i = 0; i < help->dthps_nprovs; i++) {
14143 if (dofhp->dofhp_addr ==
14144 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14145 return (EALREADY);
14146 }
14147
14148 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14149 hprov->dthp_prov = *dofhp;
14150 hprov->dthp_ref = 1;
14151 hprov->dthp_generation = gen;
14152
14153 /*
14154 * Allocate a bigger table for helper providers if it's already full.
14155 */
14156 if (help->dthps_maxprovs == help->dthps_nprovs) {
14157 tmp_maxprovs = help->dthps_maxprovs;
14158 tmp_provs = help->dthps_provs;
14159
14160 if (help->dthps_maxprovs == 0)
14161 help->dthps_maxprovs = 2;
14162 else
14163 help->dthps_maxprovs *= 2;
14164 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14165 help->dthps_maxprovs = dtrace_helper_providers_max;
14166
14167 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14168
14169 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14170 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14171
14172 if (tmp_provs != NULL) {
14173 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14174 sizeof (dtrace_helper_provider_t *));
14175 kmem_free(tmp_provs, tmp_maxprovs *
14176 sizeof (dtrace_helper_provider_t *));
14177 }
14178 }
14179
14180 help->dthps_provs[help->dthps_nprovs] = hprov;
14181 help->dthps_nprovs++;
14182
14183 return (0);
14184 }
14185
14186 static void
14187 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14188 {
14189 mutex_enter(&dtrace_lock);
14190
14191 if (--hprov->dthp_ref == 0) {
14192 dof_hdr_t *dof;
14193 mutex_exit(&dtrace_lock);
14194 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14195 dtrace_dof_destroy(dof);
14196 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14197 } else {
14198 mutex_exit(&dtrace_lock);
14199 }
14200 }
14201
14202 static int
14203 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14204 {
14205 uintptr_t daddr = (uintptr_t)dof;
14206 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14207 dof_provider_t *provider;
14208 dof_probe_t *probe;
14209 uint8_t *arg;
14210 char *strtab, *typestr;
14211 dof_stridx_t typeidx;
14212 size_t typesz;
14213 uint_t nprobes, j, k;
14214
14215 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14216
14217 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14218 dtrace_dof_error(dof, "misaligned section offset");
14219 return (-1);
14220 }
14221
14222 /*
14223 * The section needs to be large enough to contain the DOF provider
14224 * structure appropriate for the given version.
14225 */
14226 if (sec->dofs_size <
14227 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14228 offsetof(dof_provider_t, dofpv_prenoffs) :
14229 sizeof (dof_provider_t))) {
14230 dtrace_dof_error(dof, "provider section too small");
14231 return (-1);
14232 }
14233
14234 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14235 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14236 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14237 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14238 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14239
14240 if (str_sec == NULL || prb_sec == NULL ||
14241 arg_sec == NULL || off_sec == NULL)
14242 return (-1);
14243
14244 enoff_sec = NULL;
14245
14246 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14247 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14248 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14249 provider->dofpv_prenoffs)) == NULL)
14250 return (-1);
14251
14252 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14253
14254 if (provider->dofpv_name >= str_sec->dofs_size ||
14255 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14256 dtrace_dof_error(dof, "invalid provider name");
14257 return (-1);
14258 }
14259
14260 if (prb_sec->dofs_entsize == 0 ||
14261 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14262 dtrace_dof_error(dof, "invalid entry size");
14263 return (-1);
14264 }
14265
14266 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14267 dtrace_dof_error(dof, "misaligned entry size");
14268 return (-1);
14269 }
14270
14271 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14272 dtrace_dof_error(dof, "invalid entry size");
14273 return (-1);
14274 }
14275
14276 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14277 dtrace_dof_error(dof, "misaligned section offset");
14278 return (-1);
14279 }
14280
14281 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14282 dtrace_dof_error(dof, "invalid entry size");
14283 return (-1);
14284 }
14285
14286 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14287
14288 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14289
14290 /*
14291 * Take a pass through the probes to check for errors.
14292 */
14293 for (j = 0; j < nprobes; j++) {
14294 probe = (dof_probe_t *)(uintptr_t)(daddr +
14295 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14296
14297 if (probe->dofpr_func >= str_sec->dofs_size) {
14298 dtrace_dof_error(dof, "invalid function name");
14299 return (-1);
14300 }
14301
14302 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14303 dtrace_dof_error(dof, "function name too long");
14304 return (-1);
14305 }
14306
14307 if (probe->dofpr_name >= str_sec->dofs_size ||
14308 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14309 dtrace_dof_error(dof, "invalid probe name");
14310 return (-1);
14311 }
14312
14313 /*
14314 * The offset count must not wrap the index, and the offsets
14315 * must also not overflow the section's data.
14316 */
14317 if (probe->dofpr_offidx + probe->dofpr_noffs <
14318 probe->dofpr_offidx ||
14319 (probe->dofpr_offidx + probe->dofpr_noffs) *
14320 off_sec->dofs_entsize > off_sec->dofs_size) {
14321 dtrace_dof_error(dof, "invalid probe offset");
14322 return (-1);
14323 }
14324
14325 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14326 /*
14327 * If there's no is-enabled offset section, make sure
14328 * there aren't any is-enabled offsets. Otherwise
14329 * perform the same checks as for probe offsets
14330 * (immediately above).
14331 */
14332 if (enoff_sec == NULL) {
14333 if (probe->dofpr_enoffidx != 0 ||
14334 probe->dofpr_nenoffs != 0) {
14335 dtrace_dof_error(dof, "is-enabled "
14336 "offsets with null section");
14337 return (-1);
14338 }
14339 } else if (probe->dofpr_enoffidx +
14340 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14341 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14342 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14343 dtrace_dof_error(dof, "invalid is-enabled "
14344 "offset");
14345 return (-1);
14346 }
14347
14348 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14349 dtrace_dof_error(dof, "zero probe and "
14350 "is-enabled offsets");
14351 return (-1);
14352 }
14353 } else if (probe->dofpr_noffs == 0) {
14354 dtrace_dof_error(dof, "zero probe offsets");
14355 return (-1);
14356 }
14357
14358 if (probe->dofpr_argidx + probe->dofpr_xargc <
14359 probe->dofpr_argidx ||
14360 (probe->dofpr_argidx + probe->dofpr_xargc) *
14361 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14362 dtrace_dof_error(dof, "invalid args");
14363 return (-1);
14364 }
14365
14366 typeidx = probe->dofpr_nargv;
14367 typestr = strtab + probe->dofpr_nargv;
14368 for (k = 0; k < probe->dofpr_nargc; k++) {
14369 if (typeidx >= str_sec->dofs_size) {
14370 dtrace_dof_error(dof, "bad "
14371 "native argument type");
14372 return (-1);
14373 }
14374
14375 typesz = strlen(typestr) + 1;
14376 if (typesz > DTRACE_ARGTYPELEN) {
14377 dtrace_dof_error(dof, "native "
14378 "argument type too long");
14379 return (-1);
14380 }
14381 typeidx += typesz;
14382 typestr += typesz;
14383 }
14384
14385 typeidx = probe->dofpr_xargv;
14386 typestr = strtab + probe->dofpr_xargv;
14387 for (k = 0; k < probe->dofpr_xargc; k++) {
14388 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14389 dtrace_dof_error(dof, "bad "
14390 "native argument index");
14391 return (-1);
14392 }
14393
14394 if (typeidx >= str_sec->dofs_size) {
14395 dtrace_dof_error(dof, "bad "
14396 "translated argument type");
14397 return (-1);
14398 }
14399
14400 typesz = strlen(typestr) + 1;
14401 if (typesz > DTRACE_ARGTYPELEN) {
14402 dtrace_dof_error(dof, "translated argument "
14403 "type too long");
14404 return (-1);
14405 }
14406
14407 typeidx += typesz;
14408 typestr += typesz;
14409 }
14410 }
14411
14412 return (0);
14413 }
14414
14415 static int
14416 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14417 {
14418 dtrace_helpers_t *help;
14419 dtrace_vstate_t *vstate;
14420 dtrace_enabling_t *enab = NULL;
14421 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14422 uintptr_t daddr = (uintptr_t)dof;
14423
14424 ASSERT(MUTEX_HELD(&dtrace_lock));
14425
14426 if ((help = curproc->p_dtrace_helpers) == NULL)
14427 help = dtrace_helpers_create(curproc);
14428
14429 vstate = &help->dthps_vstate;
14430
14431 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14432 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14433 dtrace_dof_destroy(dof);
14434 return (rv);
14435 }
14436
14437 /*
14438 * Look for helper providers and validate their descriptions.
14439 */
14440 if (dhp != NULL) {
14441 for (i = 0; i < dof->dofh_secnum; i++) {
14442 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14443 dof->dofh_secoff + i * dof->dofh_secsize);
14444
14445 if (sec->dofs_type != DOF_SECT_PROVIDER)
14446 continue;
14447
14448 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14449 dtrace_enabling_destroy(enab);
14450 dtrace_dof_destroy(dof);
14451 return (-1);
14452 }
14453
14454 nprovs++;
14455 }
14456 }
14457
14458 /*
14459 * Now we need to walk through the ECB descriptions in the enabling.
14460 */
14461 for (i = 0; i < enab->dten_ndesc; i++) {
14462 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14463 dtrace_probedesc_t *desc = &ep->dted_probe;
14464
14465 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14466 continue;
14467
14468 if (strcmp(desc->dtpd_mod, "helper") != 0)
14469 continue;
14470
14471 if (strcmp(desc->dtpd_func, "ustack") != 0)
14472 continue;
14473
14474 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14475 ep)) != 0) {
14476 /*
14477 * Adding this helper action failed -- we are now going
14478 * to rip out the entire generation and return failure.
14479 */
14480 (void) dtrace_helper_destroygen(help->dthps_generation);
14481 dtrace_enabling_destroy(enab);
14482 dtrace_dof_destroy(dof);
14483 return (-1);
14484 }
14485
14486 nhelpers++;
14487 }
14488
14489 if (nhelpers < enab->dten_ndesc)
14490 dtrace_dof_error(dof, "unmatched helpers");
14491
14492 gen = help->dthps_generation++;
14493 dtrace_enabling_destroy(enab);
14494
14495 if (dhp != NULL && nprovs > 0) {
14496 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14497 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14498 mutex_exit(&dtrace_lock);
14499 dtrace_helper_provider_register(curproc, help, dhp);
14500 mutex_enter(&dtrace_lock);
14501
14502 destroy = 0;
14503 }
14504 }
14505
14506 if (destroy)
14507 dtrace_dof_destroy(dof);
14508
14509 return (gen);
14510 }
14511
14512 static dtrace_helpers_t *
14513 dtrace_helpers_create(proc_t *p)
14514 {
14515 dtrace_helpers_t *help;
14516
14517 ASSERT(MUTEX_HELD(&dtrace_lock));
14518 ASSERT(p->p_dtrace_helpers == NULL);
14519
14520 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14521 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14522 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14523
14524 p->p_dtrace_helpers = help;
14525 dtrace_helpers++;
14526
14527 return (help);
14528 }
14529
14530 static void
14531 dtrace_helpers_destroy(void)
14532 {
14533 dtrace_helpers_t *help;
14534 dtrace_vstate_t *vstate;
14535 proc_t *p = curproc;
14536 int i;
14537
14538 mutex_enter(&dtrace_lock);
14539
14540 ASSERT(p->p_dtrace_helpers != NULL);
14541 ASSERT(dtrace_helpers > 0);
14542
14543 help = p->p_dtrace_helpers;
14544 vstate = &help->dthps_vstate;
14545
14546 /*
14547 * We're now going to lose the help from this process.
14548 */
14549 p->p_dtrace_helpers = NULL;
14550 dtrace_sync();
14551
14552 /*
14553 * Destory the helper actions.
14554 */
14555 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14556 dtrace_helper_action_t *h, *next;
14557
14558 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14559 next = h->dtha_next;
14560 dtrace_helper_action_destroy(h, vstate);
14561 h = next;
14562 }
14563 }
14564
14565 mutex_exit(&dtrace_lock);
14566
14567 /*
14568 * Destroy the helper providers.
14569 */
14570 if (help->dthps_maxprovs > 0) {
14571 mutex_enter(&dtrace_meta_lock);
14572 if (dtrace_meta_pid != NULL) {
14573 ASSERT(dtrace_deferred_pid == NULL);
14574
14575 for (i = 0; i < help->dthps_nprovs; i++) {
14576 dtrace_helper_provider_remove(
14577 &help->dthps_provs[i]->dthp_prov, p->p_pid);
14578 }
14579 } else {
14580 mutex_enter(&dtrace_lock);
14581 ASSERT(help->dthps_deferred == 0 ||
14582 help->dthps_next != NULL ||
14583 help->dthps_prev != NULL ||
14584 help == dtrace_deferred_pid);
14585
14586 /*
14587 * Remove the helper from the deferred list.
14588 */
14589 if (help->dthps_next != NULL)
14590 help->dthps_next->dthps_prev = help->dthps_prev;
14591 if (help->dthps_prev != NULL)
14592 help->dthps_prev->dthps_next = help->dthps_next;
14593 if (dtrace_deferred_pid == help) {
14594 dtrace_deferred_pid = help->dthps_next;
14595 ASSERT(help->dthps_prev == NULL);
14596 }
14597
14598 mutex_exit(&dtrace_lock);
14599 }
14600
14601 mutex_exit(&dtrace_meta_lock);
14602
14603 for (i = 0; i < help->dthps_nprovs; i++) {
14604 dtrace_helper_provider_destroy(help->dthps_provs[i]);
14605 }
14606
14607 kmem_free(help->dthps_provs, help->dthps_maxprovs *
14608 sizeof (dtrace_helper_provider_t *));
14609 }
14610
14611 mutex_enter(&dtrace_lock);
14612
14613 dtrace_vstate_fini(&help->dthps_vstate);
14614 kmem_free(help->dthps_actions,
14615 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
14616 kmem_free(help, sizeof (dtrace_helpers_t));
14617
14618 --dtrace_helpers;
14619 mutex_exit(&dtrace_lock);
14620 }
14621
14622 static void
14623 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
14624 {
14625 dtrace_helpers_t *help, *newhelp;
14626 dtrace_helper_action_t *helper, *new, *last;
14627 dtrace_difo_t *dp;
14628 dtrace_vstate_t *vstate;
14629 int i, j, sz, hasprovs = 0;
14630
14631 mutex_enter(&dtrace_lock);
14632 ASSERT(from->p_dtrace_helpers != NULL);
14633 ASSERT(dtrace_helpers > 0);
14634
14635 help = from->p_dtrace_helpers;
14636 newhelp = dtrace_helpers_create(to);
14637 ASSERT(to->p_dtrace_helpers != NULL);
14638
14639 newhelp->dthps_generation = help->dthps_generation;
14640 vstate = &newhelp->dthps_vstate;
14641
14642 /*
14643 * Duplicate the helper actions.
14644 */
14645 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14646 if ((helper = help->dthps_actions[i]) == NULL)
14647 continue;
14648
14649 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
14650 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
14651 KM_SLEEP);
14652 new->dtha_generation = helper->dtha_generation;
14653
14654 if ((dp = helper->dtha_predicate) != NULL) {
14655 dp = dtrace_difo_duplicate(dp, vstate);
14656 new->dtha_predicate = dp;
14657 }
14658
14659 new->dtha_nactions = helper->dtha_nactions;
14660 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
14661 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
14662
14663 for (j = 0; j < new->dtha_nactions; j++) {
14664 dtrace_difo_t *dp = helper->dtha_actions[j];
14665
14666 ASSERT(dp != NULL);
14667 dp = dtrace_difo_duplicate(dp, vstate);
14668 new->dtha_actions[j] = dp;
14669 }
14670
14671 if (last != NULL) {
14672 last->dtha_next = new;
14673 } else {
14674 newhelp->dthps_actions[i] = new;
14675 }
14676
14677 last = new;
14678 }
14679 }
14680
14681 /*
14682 * Duplicate the helper providers and register them with the
14683 * DTrace framework.
14684 */
14685 if (help->dthps_nprovs > 0) {
14686 newhelp->dthps_nprovs = help->dthps_nprovs;
14687 newhelp->dthps_maxprovs = help->dthps_nprovs;
14688 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
14689 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14690 for (i = 0; i < newhelp->dthps_nprovs; i++) {
14691 newhelp->dthps_provs[i] = help->dthps_provs[i];
14692 newhelp->dthps_provs[i]->dthp_ref++;
14693 }
14694
14695 hasprovs = 1;
14696 }
14697
14698 mutex_exit(&dtrace_lock);
14699
14700 if (hasprovs)
14701 dtrace_helper_provider_register(to, newhelp, NULL);
14702 }
14703
14704 /*
14705 * DTrace Hook Functions
14706 */
14707 static void
14708 dtrace_module_loaded(struct modctl *ctl)
14709 {
14710 dtrace_provider_t *prv;
14711
14712 mutex_enter(&dtrace_provider_lock);
14713 mutex_enter(&mod_lock);
14714
14715 ASSERT(ctl->mod_busy);
14716
14717 /*
14718 * We're going to call each providers per-module provide operation
14719 * specifying only this module.
14720 */
14721 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
14722 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
14723
14724 mutex_exit(&mod_lock);
14725 mutex_exit(&dtrace_provider_lock);
14726
14727 /*
14728 * If we have any retained enablings, we need to match against them.
14729 * Enabling probes requires that cpu_lock be held, and we cannot hold
14730 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
14731 * module. (In particular, this happens when loading scheduling
14732 * classes.) So if we have any retained enablings, we need to dispatch
14733 * our task queue to do the match for us.
14734 */
14735 mutex_enter(&dtrace_lock);
14736
14737 if (dtrace_retained == NULL) {
14738 mutex_exit(&dtrace_lock);
14739 return;
14740 }
14741
14742 (void) taskq_dispatch(dtrace_taskq,
14743 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
14744
14745 mutex_exit(&dtrace_lock);
14746
14747 /*
14748 * And now, for a little heuristic sleaze: in general, we want to
14749 * match modules as soon as they load. However, we cannot guarantee
14750 * this, because it would lead us to the lock ordering violation
14751 * outlined above. The common case, of course, is that cpu_lock is
14752 * _not_ held -- so we delay here for a clock tick, hoping that that's
14753 * long enough for the task queue to do its work. If it's not, it's
14754 * not a serious problem -- it just means that the module that we
14755 * just loaded may not be immediately instrumentable.
14756 */
14757 delay(1);
14758 }
14759
14760 static void
14761 dtrace_module_unloaded(struct modctl *ctl)
14762 {
14763 dtrace_probe_t template, *probe, *first, *next;
14764 dtrace_provider_t *prov;
14765
14766 template.dtpr_mod = ctl->mod_modname;
14767
14768 mutex_enter(&dtrace_provider_lock);
14769 mutex_enter(&mod_lock);
14770 mutex_enter(&dtrace_lock);
14771
14772 if (dtrace_bymod == NULL) {
14773 /*
14774 * The DTrace module is loaded (obviously) but not attached;
14775 * we don't have any work to do.
14776 */
14777 mutex_exit(&dtrace_provider_lock);
14778 mutex_exit(&mod_lock);
14779 mutex_exit(&dtrace_lock);
14780 return;
14781 }
14782
14783 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14784 probe != NULL; probe = probe->dtpr_nextmod) {
14785 if (probe->dtpr_ecb != NULL) {
14786 mutex_exit(&dtrace_provider_lock);
14787 mutex_exit(&mod_lock);
14788 mutex_exit(&dtrace_lock);
14789
14790 /*
14791 * This shouldn't _actually_ be possible -- we're
14792 * unloading a module that has an enabled probe in it.
14793 * (It's normally up to the provider to make sure that
14794 * this can't happen.) However, because dtps_enable()
14795 * doesn't have a failure mode, there can be an
14796 * enable/unload race. Upshot: we don't want to
14797 * assert, but we're not going to disable the
14798 * probe, either.
14799 */
14800 if (dtrace_err_verbose) {
14801 cmn_err(CE_WARN, "unloaded module '%s' had "
14802 "enabled probes", ctl->mod_modname);
14803 }
14804
14805 return;
14806 }
14807 }
14808
14809 probe = first;
14810
14811 for (first = NULL; probe != NULL; probe = next) {
14812 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
14813
14814 dtrace_probes[probe->dtpr_id - 1] = NULL;
14815
14816 next = probe->dtpr_nextmod;
14817 dtrace_hash_remove(dtrace_bymod, probe);
14818 dtrace_hash_remove(dtrace_byfunc, probe);
14819 dtrace_hash_remove(dtrace_byname, probe);
14820
14821 if (first == NULL) {
14822 first = probe;
14823 probe->dtpr_nextmod = NULL;
14824 } else {
14825 probe->dtpr_nextmod = first;
14826 first = probe;
14827 }
14828 }
14829
14830 /*
14831 * We've removed all of the module's probes from the hash chains and
14832 * from the probe array. Now issue a dtrace_sync() to be sure that
14833 * everyone has cleared out from any probe array processing.
14834 */
14835 dtrace_sync();
14836
14837 for (probe = first; probe != NULL; probe = first) {
14838 first = probe->dtpr_nextmod;
14839 prov = probe->dtpr_provider;
14840 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
14841 probe->dtpr_arg);
14842 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
14843 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
14844 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
14845 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
14846 kmem_free(probe, sizeof (dtrace_probe_t));
14847 }
14848
14849 mutex_exit(&dtrace_lock);
14850 mutex_exit(&mod_lock);
14851 mutex_exit(&dtrace_provider_lock);
14852 }
14853
14854 void
14855 dtrace_suspend(void)
14856 {
14857 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
14858 }
14859
14860 void
14861 dtrace_resume(void)
14862 {
14863 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
14864 }
14865
14866 static int
14867 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
14868 {
14869 ASSERT(MUTEX_HELD(&cpu_lock));
14870 mutex_enter(&dtrace_lock);
14871
14872 switch (what) {
14873 case CPU_CONFIG: {
14874 dtrace_state_t *state;
14875 dtrace_optval_t *opt, rs, c;
14876
14877 /*
14878 * For now, we only allocate a new buffer for anonymous state.
14879 */
14880 if ((state = dtrace_anon.dta_state) == NULL)
14881 break;
14882
14883 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14884 break;
14885
14886 opt = state->dts_options;
14887 c = opt[DTRACEOPT_CPU];
14888
14889 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
14890 break;
14891
14892 /*
14893 * Regardless of what the actual policy is, we're going to
14894 * temporarily set our resize policy to be manual. We're
14895 * also going to temporarily set our CPU option to denote
14896 * the newly configured CPU.
14897 */
14898 rs = opt[DTRACEOPT_BUFRESIZE];
14899 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
14900 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
14901
14902 (void) dtrace_state_buffers(state);
14903
14904 opt[DTRACEOPT_BUFRESIZE] = rs;
14905 opt[DTRACEOPT_CPU] = c;
14906
14907 break;
14908 }
14909
14910 case CPU_UNCONFIG:
14911 /*
14912 * We don't free the buffer in the CPU_UNCONFIG case. (The
14913 * buffer will be freed when the consumer exits.)
14914 */
14915 break;
14916
14917 default:
14918 break;
14919 }
14920
14921 mutex_exit(&dtrace_lock);
14922 return (0);
14923 }
14924
14925 static void
14926 dtrace_cpu_setup_initial(processorid_t cpu)
14927 {
14928 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
14929 }
14930
14931 static void
14932 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
14933 {
14934 if (dtrace_toxranges >= dtrace_toxranges_max) {
14935 int osize, nsize;
14936 dtrace_toxrange_t *range;
14937
14938 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14939
14940 if (osize == 0) {
14941 ASSERT(dtrace_toxrange == NULL);
14942 ASSERT(dtrace_toxranges_max == 0);
14943 dtrace_toxranges_max = 1;
14944 } else {
14945 dtrace_toxranges_max <<= 1;
14946 }
14947
14948 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14949 range = kmem_zalloc(nsize, KM_SLEEP);
14950
14951 if (dtrace_toxrange != NULL) {
14952 ASSERT(osize != 0);
14953 bcopy(dtrace_toxrange, range, osize);
14954 kmem_free(dtrace_toxrange, osize);
14955 }
14956
14957 dtrace_toxrange = range;
14958 }
14959
14960 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
14961 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
14962
14963 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
14964 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
14965 dtrace_toxranges++;
14966 }
14967
14968 static void
14969 dtrace_getf_barrier()
14970 {
14971 /*
14972 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
14973 * that contain calls to getf(), this routine will be called on every
14974 * closef() before either the underlying vnode is released or the
14975 * file_t itself is freed. By the time we are here, it is essential
14976 * that the file_t can no longer be accessed from a call to getf()
14977 * in probe context -- that assures that a dtrace_sync() can be used
14978 * to clear out any enablings referring to the old structures.
14979 */
14980 if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
14981 kcred->cr_zone->zone_dtrace_getf != 0)
14982 dtrace_sync();
14983 }
14984
14985 /*
14986 * DTrace Driver Cookbook Functions
14987 */
14988 /*ARGSUSED*/
14989 static int
14990 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
14991 {
14992 dtrace_provider_id_t id;
14993 dtrace_state_t *state = NULL;
14994 dtrace_enabling_t *enab;
14995
14996 mutex_enter(&cpu_lock);
14997 mutex_enter(&dtrace_provider_lock);
14998 mutex_enter(&dtrace_lock);
14999
15000 if (ddi_soft_state_init(&dtrace_softstate,
15001 sizeof (dtrace_state_t), 0) != 0) {
15002 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15003 mutex_exit(&cpu_lock);
15004 mutex_exit(&dtrace_provider_lock);
15005 mutex_exit(&dtrace_lock);
15006 return (DDI_FAILURE);
15007 }
15008
15009 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15010 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15011 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15012 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15013 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15014 ddi_remove_minor_node(devi, NULL);
15015 ddi_soft_state_fini(&dtrace_softstate);
15016 mutex_exit(&cpu_lock);
15017 mutex_exit(&dtrace_provider_lock);
15018 mutex_exit(&dtrace_lock);
15019 return (DDI_FAILURE);
15020 }
15021
15022 ddi_report_dev(devi);
15023 dtrace_devi = devi;
15024
15025 dtrace_modload = dtrace_module_loaded;
15026 dtrace_modunload = dtrace_module_unloaded;
15027 dtrace_cpu_init = dtrace_cpu_setup_initial;
15028 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15029 dtrace_helpers_fork = dtrace_helpers_duplicate;
15030 dtrace_cpustart_init = dtrace_suspend;
15031 dtrace_cpustart_fini = dtrace_resume;
15032 dtrace_debugger_init = dtrace_suspend;
15033 dtrace_debugger_fini = dtrace_resume;
15034
15035 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15036
15037 ASSERT(MUTEX_HELD(&cpu_lock));
15038
15039 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15040 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15041 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15042 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15043 VM_SLEEP | VMC_IDENTIFIER);
15044 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15045 1, INT_MAX, 0);
15046
15047 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15048 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15049 NULL, NULL, NULL, NULL, NULL, 0);
15050
15051 ASSERT(MUTEX_HELD(&cpu_lock));
15052 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15053 offsetof(dtrace_probe_t, dtpr_nextmod),
15054 offsetof(dtrace_probe_t, dtpr_prevmod));
15055
15056 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15057 offsetof(dtrace_probe_t, dtpr_nextfunc),
15058 offsetof(dtrace_probe_t, dtpr_prevfunc));
15059
15060 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15061 offsetof(dtrace_probe_t, dtpr_nextname),
15062 offsetof(dtrace_probe_t, dtpr_prevname));
15063
15064 if (dtrace_retain_max < 1) {
15065 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15066 "setting to 1", dtrace_retain_max);
15067 dtrace_retain_max = 1;
15068 }
15069
15070 /*
15071 * Now discover our toxic ranges.
15072 */
15073 dtrace_toxic_ranges(dtrace_toxrange_add);
15074
15075 /*
15076 * Before we register ourselves as a provider to our own framework,
15077 * we would like to assert that dtrace_provider is NULL -- but that's
15078 * not true if we were loaded as a dependency of a DTrace provider.
15079 * Once we've registered, we can assert that dtrace_provider is our
15080 * pseudo provider.
15081 */
15082 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15083 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15084
15085 ASSERT(dtrace_provider != NULL);
15086 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15087
15088 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15089 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15090 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15091 dtrace_provider, NULL, NULL, "END", 0, NULL);
15092 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15093 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15094
15095 dtrace_anon_property();
15096 mutex_exit(&cpu_lock);
15097
15098 /*
15099 * If there are already providers, we must ask them to provide their
15100 * probes, and then match any anonymous enabling against them. Note
15101 * that there should be no other retained enablings at this time:
15102 * the only retained enablings at this time should be the anonymous
15103 * enabling.
15104 */
15105 if (dtrace_anon.dta_enabling != NULL) {
15106 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15107
15108 dtrace_enabling_provide(NULL);
15109 state = dtrace_anon.dta_state;
15110
15111 /*
15112 * We couldn't hold cpu_lock across the above call to
15113 * dtrace_enabling_provide(), but we must hold it to actually
15114 * enable the probes. We have to drop all of our locks, pick
15115 * up cpu_lock, and regain our locks before matching the
15116 * retained anonymous enabling.
15117 */
15118 mutex_exit(&dtrace_lock);
15119 mutex_exit(&dtrace_provider_lock);
15120
15121 mutex_enter(&cpu_lock);
15122 mutex_enter(&dtrace_provider_lock);
15123 mutex_enter(&dtrace_lock);
15124
15125 if ((enab = dtrace_anon.dta_enabling) != NULL)
15126 (void) dtrace_enabling_match(enab, NULL);
15127
15128 mutex_exit(&cpu_lock);
15129 }
15130
15131 mutex_exit(&dtrace_lock);
15132 mutex_exit(&dtrace_provider_lock);
15133
15134 if (state != NULL) {
15135 /*
15136 * If we created any anonymous state, set it going now.
15137 */
15138 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15139 }
15140
15141 return (DDI_SUCCESS);
15142 }
15143
15144 /*ARGSUSED*/
15145 static int
15146 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15147 {
15148 dtrace_state_t *state;
15149 uint32_t priv;
15150 uid_t uid;
15151 zoneid_t zoneid;
15152
15153 if (getminor(*devp) == DTRACEMNRN_HELPER)
15154 return (0);
15155
15156 /*
15157 * If this wasn't an open with the "helper" minor, then it must be
15158 * the "dtrace" minor.
15159 */
15160 if (getminor(*devp) != DTRACEMNRN_DTRACE)
15161 return (ENXIO);
15162
15163 /*
15164 * If no DTRACE_PRIV_* bits are set in the credential, then the
15165 * caller lacks sufficient permission to do anything with DTrace.
15166 */
15167 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15168 if (priv == DTRACE_PRIV_NONE)
15169 return (EACCES);
15170
15171 /*
15172 * Ask all providers to provide all their probes.
15173 */
15174 mutex_enter(&dtrace_provider_lock);
15175 dtrace_probe_provide(NULL, NULL);
15176 mutex_exit(&dtrace_provider_lock);
15177
15178 mutex_enter(&cpu_lock);
15179 mutex_enter(&dtrace_lock);
15180 dtrace_opens++;
15181 dtrace_membar_producer();
15182
15183 /*
15184 * If the kernel debugger is active (that is, if the kernel debugger
15185 * modified text in some way), we won't allow the open.
15186 */
15187 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15188 dtrace_opens--;
15189 mutex_exit(&cpu_lock);
15190 mutex_exit(&dtrace_lock);
15191 return (EBUSY);
15192 }
15193
15194 if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
15195 /*
15196 * If DTrace helper tracing is enabled, we need to allocate the
15197 * trace buffer and initialize the values.
15198 */
15199 dtrace_helptrace_buffer =
15200 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15201 dtrace_helptrace_next = 0;
15202 dtrace_helptrace_wrapped = 0;
15203 dtrace_helptrace_enable = 0;
15204 }
15205
15206 state = dtrace_state_create(devp, cred_p);
15207 mutex_exit(&cpu_lock);
15208
15209 if (state == NULL) {
15210 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15211 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15212 mutex_exit(&dtrace_lock);
15213 return (EAGAIN);
15214 }
15215
15216 mutex_exit(&dtrace_lock);
15217
15218 return (0);
15219 }
15220
15221 /*ARGSUSED*/
15222 static int
15223 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15224 {
15225 minor_t minor = getminor(dev);
15226 dtrace_state_t *state;
15227 dtrace_helptrace_t *buf = NULL;
15228
15229 if (minor == DTRACEMNRN_HELPER)
15230 return (0);
15231
15232 state = ddi_get_soft_state(dtrace_softstate, minor);
15233
15234 mutex_enter(&cpu_lock);
15235 mutex_enter(&dtrace_lock);
15236
15237 if (state->dts_anon) {
15238 /*
15239 * There is anonymous state. Destroy that first.
15240 */
15241 ASSERT(dtrace_anon.dta_state == NULL);
15242 dtrace_state_destroy(state->dts_anon);
15243 }
15244
15245 if (dtrace_helptrace_disable) {
15246 /*
15247 * If we have been told to disable helper tracing, set the
15248 * buffer to NULL before calling into dtrace_state_destroy();
15249 * we take advantage of its dtrace_sync() to know that no
15250 * CPU is in probe context with enabled helper tracing
15251 * after it returns.
15252 */
15253 buf = dtrace_helptrace_buffer;
15254 dtrace_helptrace_buffer = NULL;
15255 }
15256
15257 dtrace_state_destroy(state);
15258 ASSERT(dtrace_opens > 0);
15259
15260 /*
15261 * Only relinquish control of the kernel debugger interface when there
15262 * are no consumers and no anonymous enablings.
15263 */
15264 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15265 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15266
15267 if (buf != NULL) {
15268 kmem_free(buf, dtrace_helptrace_bufsize);
15269 dtrace_helptrace_disable = 0;
15270 }
15271
15272 mutex_exit(&dtrace_lock);
15273 mutex_exit(&cpu_lock);
15274
15275 return (0);
15276 }
15277
15278 /*ARGSUSED*/
15279 static int
15280 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15281 {
15282 int rval;
15283 dof_helper_t help, *dhp = NULL;
15284
15285 switch (cmd) {
15286 case DTRACEHIOC_ADDDOF:
15287 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15288 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15289 return (EFAULT);
15290 }
15291
15292 dhp = &help;
15293 arg = (intptr_t)help.dofhp_dof;
15294 /*FALLTHROUGH*/
15295
15296 case DTRACEHIOC_ADD: {
15297 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15298
15299 if (dof == NULL)
15300 return (rval);
15301
15302 mutex_enter(&dtrace_lock);
15303
15304 /*
15305 * dtrace_helper_slurp() takes responsibility for the dof --
15306 * it may free it now or it may save it and free it later.
15307 */
15308 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15309 *rv = rval;
15310 rval = 0;
15311 } else {
15312 rval = EINVAL;
15313 }
15314
15315 mutex_exit(&dtrace_lock);
15316 return (rval);
15317 }
15318
15319 case DTRACEHIOC_REMOVE: {
15320 mutex_enter(&dtrace_lock);
15321 rval = dtrace_helper_destroygen(arg);
15322 mutex_exit(&dtrace_lock);
15323
15324 return (rval);
15325 }
15326
15327 default:
15328 break;
15329 }
15330
15331 return (ENOTTY);
15332 }
15333
15334 /*ARGSUSED*/
15335 static int
15336 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15337 {
15338 minor_t minor = getminor(dev);
15339 dtrace_state_t *state;
15340 int rval;
15341
15342 if (minor == DTRACEMNRN_HELPER)
15343 return (dtrace_ioctl_helper(cmd, arg, rv));
15344
15345 state = ddi_get_soft_state(dtrace_softstate, minor);
15346
15347 if (state->dts_anon) {
15348 ASSERT(dtrace_anon.dta_state == NULL);
15349 state = state->dts_anon;
15350 }
15351
15352 switch (cmd) {
15353 case DTRACEIOC_PROVIDER: {
15354 dtrace_providerdesc_t pvd;
15355 dtrace_provider_t *pvp;
15356
15357 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15358 return (EFAULT);
15359
15360 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15361 mutex_enter(&dtrace_provider_lock);
15362
15363 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15364 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15365 break;
15366 }
15367
15368 mutex_exit(&dtrace_provider_lock);
15369
15370 if (pvp == NULL)
15371 return (ESRCH);
15372
15373 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15374 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15375 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15376 return (EFAULT);
15377
15378 return (0);
15379 }
15380
15381 case DTRACEIOC_EPROBE: {
15382 dtrace_eprobedesc_t epdesc;
15383 dtrace_ecb_t *ecb;
15384 dtrace_action_t *act;
15385 void *buf;
15386 size_t size;
15387 uintptr_t dest;
15388 int nrecs;
15389
15390 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15391 return (EFAULT);
15392
15393 mutex_enter(&dtrace_lock);
15394
15395 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15396 mutex_exit(&dtrace_lock);
15397 return (EINVAL);
15398 }
15399
15400 if (ecb->dte_probe == NULL) {
15401 mutex_exit(&dtrace_lock);
15402 return (EINVAL);
15403 }
15404
15405 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15406 epdesc.dtepd_uarg = ecb->dte_uarg;
15407 epdesc.dtepd_size = ecb->dte_size;
15408
15409 nrecs = epdesc.dtepd_nrecs;
15410 epdesc.dtepd_nrecs = 0;
15411 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15412 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15413 continue;
15414
15415 epdesc.dtepd_nrecs++;
15416 }
15417
15418 /*
15419 * Now that we have the size, we need to allocate a temporary
15420 * buffer in which to store the complete description. We need
15421 * the temporary buffer to be able to drop dtrace_lock()
15422 * across the copyout(), below.
15423 */
15424 size = sizeof (dtrace_eprobedesc_t) +
15425 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15426
15427 buf = kmem_alloc(size, KM_SLEEP);
15428 dest = (uintptr_t)buf;
15429
15430 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15431 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15432
15433 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15434 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15435 continue;
15436
15437 if (nrecs-- == 0)
15438 break;
15439
15440 bcopy(&act->dta_rec, (void *)dest,
15441 sizeof (dtrace_recdesc_t));
15442 dest += sizeof (dtrace_recdesc_t);
15443 }
15444
15445 mutex_exit(&dtrace_lock);
15446
15447 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15448 kmem_free(buf, size);
15449 return (EFAULT);
15450 }
15451
15452 kmem_free(buf, size);
15453 return (0);
15454 }
15455
15456 case DTRACEIOC_AGGDESC: {
15457 dtrace_aggdesc_t aggdesc;
15458 dtrace_action_t *act;
15459 dtrace_aggregation_t *agg;
15460 int nrecs;
15461 uint32_t offs;
15462 dtrace_recdesc_t *lrec;
15463 void *buf;
15464 size_t size;
15465 uintptr_t dest;
15466
15467 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15468 return (EFAULT);
15469
15470 mutex_enter(&dtrace_lock);
15471
15472 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15473 mutex_exit(&dtrace_lock);
15474 return (EINVAL);
15475 }
15476
15477 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15478
15479 nrecs = aggdesc.dtagd_nrecs;
15480 aggdesc.dtagd_nrecs = 0;
15481
15482 offs = agg->dtag_base;
15483 lrec = &agg->dtag_action.dta_rec;
15484 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15485
15486 for (act = agg->dtag_first; ; act = act->dta_next) {
15487 ASSERT(act->dta_intuple ||
15488 DTRACEACT_ISAGG(act->dta_kind));
15489
15490 /*
15491 * If this action has a record size of zero, it
15492 * denotes an argument to the aggregating action.
15493 * Because the presence of this record doesn't (or
15494 * shouldn't) affect the way the data is interpreted,
15495 * we don't copy it out to save user-level the
15496 * confusion of dealing with a zero-length record.
15497 */
15498 if (act->dta_rec.dtrd_size == 0) {
15499 ASSERT(agg->dtag_hasarg);
15500 continue;
15501 }
15502
15503 aggdesc.dtagd_nrecs++;
15504
15505 if (act == &agg->dtag_action)
15506 break;
15507 }
15508
15509 /*
15510 * Now that we have the size, we need to allocate a temporary
15511 * buffer in which to store the complete description. We need
15512 * the temporary buffer to be able to drop dtrace_lock()
15513 * across the copyout(), below.
15514 */
15515 size = sizeof (dtrace_aggdesc_t) +
15516 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15517
15518 buf = kmem_alloc(size, KM_SLEEP);
15519 dest = (uintptr_t)buf;
15520
15521 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15522 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15523
15524 for (act = agg->dtag_first; ; act = act->dta_next) {
15525 dtrace_recdesc_t rec = act->dta_rec;
15526
15527 /*
15528 * See the comment in the above loop for why we pass
15529 * over zero-length records.
15530 */
15531 if (rec.dtrd_size == 0) {
15532 ASSERT(agg->dtag_hasarg);
15533 continue;
15534 }
15535
15536 if (nrecs-- == 0)
15537 break;
15538
15539 rec.dtrd_offset -= offs;
15540 bcopy(&rec, (void *)dest, sizeof (rec));
15541 dest += sizeof (dtrace_recdesc_t);
15542
15543 if (act == &agg->dtag_action)
15544 break;
15545 }
15546
15547 mutex_exit(&dtrace_lock);
15548
15549 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15550 kmem_free(buf, size);
15551 return (EFAULT);
15552 }
15553
15554 kmem_free(buf, size);
15555 return (0);
15556 }
15557
15558 case DTRACEIOC_ENABLE: {
15559 dof_hdr_t *dof;
15560 dtrace_enabling_t *enab = NULL;
15561 dtrace_vstate_t *vstate;
15562 int err = 0;
15563
15564 *rv = 0;
15565
15566 /*
15567 * If a NULL argument has been passed, we take this as our
15568 * cue to reevaluate our enablings.
15569 */
15570 if (arg == NULL) {
15571 dtrace_enabling_matchall();
15572
15573 return (0);
15574 }
15575
15576 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
15577 return (rval);
15578
15579 mutex_enter(&cpu_lock);
15580 mutex_enter(&dtrace_lock);
15581 vstate = &state->dts_vstate;
15582
15583 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
15584 mutex_exit(&dtrace_lock);
15585 mutex_exit(&cpu_lock);
15586 dtrace_dof_destroy(dof);
15587 return (EBUSY);
15588 }
15589
15590 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
15591 mutex_exit(&dtrace_lock);
15592 mutex_exit(&cpu_lock);
15593 dtrace_dof_destroy(dof);
15594 return (EINVAL);
15595 }
15596
15597 if ((rval = dtrace_dof_options(dof, state)) != 0) {
15598 dtrace_enabling_destroy(enab);
15599 mutex_exit(&dtrace_lock);
15600 mutex_exit(&cpu_lock);
15601 dtrace_dof_destroy(dof);
15602 return (rval);
15603 }
15604
15605 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
15606 err = dtrace_enabling_retain(enab);
15607 } else {
15608 dtrace_enabling_destroy(enab);
15609 }
15610
15611 mutex_exit(&cpu_lock);
15612 mutex_exit(&dtrace_lock);
15613 dtrace_dof_destroy(dof);
15614
15615 return (err);
15616 }
15617
15618 case DTRACEIOC_REPLICATE: {
15619 dtrace_repldesc_t desc;
15620 dtrace_probedesc_t *match = &desc.dtrpd_match;
15621 dtrace_probedesc_t *create = &desc.dtrpd_create;
15622 int err;
15623
15624 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15625 return (EFAULT);
15626
15627 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15628 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15629 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15630 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15631
15632 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15633 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15634 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15635 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15636
15637 mutex_enter(&dtrace_lock);
15638 err = dtrace_enabling_replicate(state, match, create);
15639 mutex_exit(&dtrace_lock);
15640
15641 return (err);
15642 }
15643
15644 case DTRACEIOC_PROBEMATCH:
15645 case DTRACEIOC_PROBES: {
15646 dtrace_probe_t *probe = NULL;
15647 dtrace_probedesc_t desc;
15648 dtrace_probekey_t pkey;
15649 dtrace_id_t i;
15650 int m = 0;
15651 uint32_t priv;
15652 uid_t uid;
15653 zoneid_t zoneid;
15654
15655 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15656 return (EFAULT);
15657
15658 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15659 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15660 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15661 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15662
15663 /*
15664 * Before we attempt to match this probe, we want to give
15665 * all providers the opportunity to provide it.
15666 */
15667 if (desc.dtpd_id == DTRACE_IDNONE) {
15668 mutex_enter(&dtrace_provider_lock);
15669 dtrace_probe_provide(&desc, NULL);
15670 mutex_exit(&dtrace_provider_lock);
15671 desc.dtpd_id++;
15672 }
15673
15674 if (cmd == DTRACEIOC_PROBEMATCH) {
15675 dtrace_probekey(&desc, &pkey);
15676 pkey.dtpk_id = DTRACE_IDNONE;
15677 }
15678
15679 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
15680
15681 mutex_enter(&dtrace_lock);
15682
15683 if (cmd == DTRACEIOC_PROBEMATCH) {
15684 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15685 if ((probe = dtrace_probes[i - 1]) != NULL &&
15686 (m = dtrace_match_probe(probe, &pkey,
15687 priv, uid, zoneid)) != 0)
15688 break;
15689 }
15690
15691 if (m < 0) {
15692 mutex_exit(&dtrace_lock);
15693 return (EINVAL);
15694 }
15695
15696 } else {
15697 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15698 if ((probe = dtrace_probes[i - 1]) != NULL &&
15699 dtrace_match_priv(probe, priv, uid, zoneid))
15700 break;
15701 }
15702 }
15703
15704 if (probe == NULL) {
15705 mutex_exit(&dtrace_lock);
15706 return (ESRCH);
15707 }
15708
15709 dtrace_probe_description(probe, &desc);
15710 mutex_exit(&dtrace_lock);
15711
15712 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15713 return (EFAULT);
15714
15715 return (0);
15716 }
15717
15718 case DTRACEIOC_PROBEARG: {
15719 dtrace_argdesc_t desc;
15720 dtrace_probe_t *probe;
15721 dtrace_provider_t *prov;
15722
15723 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15724 return (EFAULT);
15725
15726 if (desc.dtargd_id == DTRACE_IDNONE)
15727 return (EINVAL);
15728
15729 if (desc.dtargd_ndx == DTRACE_ARGNONE)
15730 return (EINVAL);
15731
15732 mutex_enter(&dtrace_provider_lock);
15733 mutex_enter(&mod_lock);
15734 mutex_enter(&dtrace_lock);
15735
15736 if (desc.dtargd_id > dtrace_nprobes) {
15737 mutex_exit(&dtrace_lock);
15738 mutex_exit(&mod_lock);
15739 mutex_exit(&dtrace_provider_lock);
15740 return (EINVAL);
15741 }
15742
15743 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
15744 mutex_exit(&dtrace_lock);
15745 mutex_exit(&mod_lock);
15746 mutex_exit(&dtrace_provider_lock);
15747 return (EINVAL);
15748 }
15749
15750 mutex_exit(&dtrace_lock);
15751
15752 prov = probe->dtpr_provider;
15753
15754 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
15755 /*
15756 * There isn't any typed information for this probe.
15757 * Set the argument number to DTRACE_ARGNONE.
15758 */
15759 desc.dtargd_ndx = DTRACE_ARGNONE;
15760 } else {
15761 desc.dtargd_native[0] = '\0';
15762 desc.dtargd_xlate[0] = '\0';
15763 desc.dtargd_mapping = desc.dtargd_ndx;
15764
15765 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
15766 probe->dtpr_id, probe->dtpr_arg, &desc);
15767 }
15768
15769 mutex_exit(&mod_lock);
15770 mutex_exit(&dtrace_provider_lock);
15771
15772 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15773 return (EFAULT);
15774
15775 return (0);
15776 }
15777
15778 case DTRACEIOC_GO: {
15779 processorid_t cpuid;
15780 rval = dtrace_state_go(state, &cpuid);
15781
15782 if (rval != 0)
15783 return (rval);
15784
15785 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15786 return (EFAULT);
15787
15788 return (0);
15789 }
15790
15791 case DTRACEIOC_STOP: {
15792 processorid_t cpuid;
15793
15794 mutex_enter(&dtrace_lock);
15795 rval = dtrace_state_stop(state, &cpuid);
15796 mutex_exit(&dtrace_lock);
15797
15798 if (rval != 0)
15799 return (rval);
15800
15801 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15802 return (EFAULT);
15803
15804 return (0);
15805 }
15806
15807 case DTRACEIOC_DOFGET: {
15808 dof_hdr_t hdr, *dof;
15809 uint64_t len;
15810
15811 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
15812 return (EFAULT);
15813
15814 mutex_enter(&dtrace_lock);
15815 dof = dtrace_dof_create(state);
15816 mutex_exit(&dtrace_lock);
15817
15818 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
15819 rval = copyout(dof, (void *)arg, len);
15820 dtrace_dof_destroy(dof);
15821
15822 return (rval == 0 ? 0 : EFAULT);
15823 }
15824
15825 case DTRACEIOC_AGGSNAP:
15826 case DTRACEIOC_BUFSNAP: {
15827 dtrace_bufdesc_t desc;
15828 caddr_t cached;
15829 dtrace_buffer_t *buf;
15830
15831 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15832 return (EFAULT);
15833
15834 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
15835 return (EINVAL);
15836
15837 mutex_enter(&dtrace_lock);
15838
15839 if (cmd == DTRACEIOC_BUFSNAP) {
15840 buf = &state->dts_buffer[desc.dtbd_cpu];
15841 } else {
15842 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
15843 }
15844
15845 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
15846 size_t sz = buf->dtb_offset;
15847
15848 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
15849 mutex_exit(&dtrace_lock);
15850 return (EBUSY);
15851 }
15852
15853 /*
15854 * If this buffer has already been consumed, we're
15855 * going to indicate that there's nothing left here
15856 * to consume.
15857 */
15858 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
15859 mutex_exit(&dtrace_lock);
15860
15861 desc.dtbd_size = 0;
15862 desc.dtbd_drops = 0;
15863 desc.dtbd_errors = 0;
15864 desc.dtbd_oldest = 0;
15865 sz = sizeof (desc);
15866
15867 if (copyout(&desc, (void *)arg, sz) != 0)
15868 return (EFAULT);
15869
15870 return (0);
15871 }
15872
15873 /*
15874 * If this is a ring buffer that has wrapped, we want
15875 * to copy the whole thing out.
15876 */
15877 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
15878 dtrace_buffer_polish(buf);
15879 sz = buf->dtb_size;
15880 }
15881
15882 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
15883 mutex_exit(&dtrace_lock);
15884 return (EFAULT);
15885 }
15886
15887 desc.dtbd_size = sz;
15888 desc.dtbd_drops = buf->dtb_drops;
15889 desc.dtbd_errors = buf->dtb_errors;
15890 desc.dtbd_oldest = buf->dtb_xamot_offset;
15891 desc.dtbd_timestamp = dtrace_gethrtime();
15892
15893 mutex_exit(&dtrace_lock);
15894
15895 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15896 return (EFAULT);
15897
15898 buf->dtb_flags |= DTRACEBUF_CONSUMED;
15899
15900 return (0);
15901 }
15902
15903 if (buf->dtb_tomax == NULL) {
15904 ASSERT(buf->dtb_xamot == NULL);
15905 mutex_exit(&dtrace_lock);
15906 return (ENOENT);
15907 }
15908
15909 cached = buf->dtb_tomax;
15910 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
15911
15912 dtrace_xcall(desc.dtbd_cpu,
15913 (dtrace_xcall_t)dtrace_buffer_switch, buf);
15914
15915 state->dts_errors += buf->dtb_xamot_errors;
15916
15917 /*
15918 * If the buffers did not actually switch, then the cross call
15919 * did not take place -- presumably because the given CPU is
15920 * not in the ready set. If this is the case, we'll return
15921 * ENOENT.
15922 */
15923 if (buf->dtb_tomax == cached) {
15924 ASSERT(buf->dtb_xamot != cached);
15925 mutex_exit(&dtrace_lock);
15926 return (ENOENT);
15927 }
15928
15929 ASSERT(cached == buf->dtb_xamot);
15930
15931 /*
15932 * We have our snapshot; now copy it out.
15933 */
15934 if (copyout(buf->dtb_xamot, desc.dtbd_data,
15935 buf->dtb_xamot_offset) != 0) {
15936 mutex_exit(&dtrace_lock);
15937 return (EFAULT);
15938 }
15939
15940 desc.dtbd_size = buf->dtb_xamot_offset;
15941 desc.dtbd_drops = buf->dtb_xamot_drops;
15942 desc.dtbd_errors = buf->dtb_xamot_errors;
15943 desc.dtbd_oldest = 0;
15944 desc.dtbd_timestamp = buf->dtb_switched;
15945
15946 mutex_exit(&dtrace_lock);
15947
15948 /*
15949 * Finally, copy out the buffer description.
15950 */
15951 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15952 return (EFAULT);
15953
15954 return (0);
15955 }
15956
15957 case DTRACEIOC_CONF: {
15958 dtrace_conf_t conf;
15959
15960 bzero(&conf, sizeof (conf));
15961 conf.dtc_difversion = DIF_VERSION;
15962 conf.dtc_difintregs = DIF_DIR_NREGS;
15963 conf.dtc_diftupregs = DIF_DTR_NREGS;
15964 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
15965
15966 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
15967 return (EFAULT);
15968
15969 return (0);
15970 }
15971
15972 case DTRACEIOC_STATUS: {
15973 dtrace_status_t stat;
15974 dtrace_dstate_t *dstate;
15975 int i, j;
15976 uint64_t nerrs;
15977
15978 /*
15979 * See the comment in dtrace_state_deadman() for the reason
15980 * for setting dts_laststatus to INT64_MAX before setting
15981 * it to the correct value.
15982 */
15983 state->dts_laststatus = INT64_MAX;
15984 dtrace_membar_producer();
15985 state->dts_laststatus = dtrace_gethrtime();
15986
15987 bzero(&stat, sizeof (stat));
15988
15989 mutex_enter(&dtrace_lock);
15990
15991 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
15992 mutex_exit(&dtrace_lock);
15993 return (ENOENT);
15994 }
15995
15996 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
15997 stat.dtst_exiting = 1;
15998
15999 nerrs = state->dts_errors;
16000 dstate = &state->dts_vstate.dtvs_dynvars;
16001
16002 for (i = 0; i < NCPU; i++) {
16003 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16004
16005 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16006 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16007 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16008
16009 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16010 stat.dtst_filled++;
16011
16012 nerrs += state->dts_buffer[i].dtb_errors;
16013
16014 for (j = 0; j < state->dts_nspeculations; j++) {
16015 dtrace_speculation_t *spec;
16016 dtrace_buffer_t *buf;
16017
16018 spec = &state->dts_speculations[j];
16019 buf = &spec->dtsp_buffer[i];
16020 stat.dtst_specdrops += buf->dtb_xamot_drops;
16021 }
16022 }
16023
16024 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16025 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16026 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16027 stat.dtst_dblerrors = state->dts_dblerrors;
16028 stat.dtst_killed =
16029 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16030 stat.dtst_errors = nerrs;
16031
16032 mutex_exit(&dtrace_lock);
16033
16034 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16035 return (EFAULT);
16036
16037 return (0);
16038 }
16039
16040 case DTRACEIOC_FORMAT: {
16041 dtrace_fmtdesc_t fmt;
16042 char *str;
16043 int len;
16044
16045 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16046 return (EFAULT);
16047
16048 mutex_enter(&dtrace_lock);
16049
16050 if (fmt.dtfd_format == 0 ||
16051 fmt.dtfd_format > state->dts_nformats) {
16052 mutex_exit(&dtrace_lock);
16053 return (EINVAL);
16054 }
16055
16056 /*
16057 * Format strings are allocated contiguously and they are
16058 * never freed; if a format index is less than the number
16059 * of formats, we can assert that the format map is non-NULL
16060 * and that the format for the specified index is non-NULL.
16061 */
16062 ASSERT(state->dts_formats != NULL);
16063 str = state->dts_formats[fmt.dtfd_format - 1];
16064 ASSERT(str != NULL);
16065
16066 len = strlen(str) + 1;
16067
16068 if (len > fmt.dtfd_length) {
16069 fmt.dtfd_length = len;
16070
16071 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16072 mutex_exit(&dtrace_lock);
16073 return (EINVAL);
16074 }
16075 } else {
16076 if (copyout(str, fmt.dtfd_string, len) != 0) {
16077 mutex_exit(&dtrace_lock);
16078 return (EINVAL);
16079 }
16080 }
16081
16082 mutex_exit(&dtrace_lock);
16083 return (0);
16084 }
16085
16086 default:
16087 break;
16088 }
16089
16090 return (ENOTTY);
16091 }
16092
16093 /*ARGSUSED*/
16094 static int
16095 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16096 {
16097 dtrace_state_t *state;
16098
16099 switch (cmd) {
16100 case DDI_DETACH:
16101 break;
16102
16103 case DDI_SUSPEND:
16104 return (DDI_SUCCESS);
16105
16106 default:
16107 return (DDI_FAILURE);
16108 }
16109
16110 mutex_enter(&cpu_lock);
16111 mutex_enter(&dtrace_provider_lock);
16112 mutex_enter(&dtrace_lock);
16113
16114 ASSERT(dtrace_opens == 0);
16115
16116 if (dtrace_helpers > 0) {
16117 mutex_exit(&dtrace_provider_lock);
16118 mutex_exit(&dtrace_lock);
16119 mutex_exit(&cpu_lock);
16120 return (DDI_FAILURE);
16121 }
16122
16123 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16124 mutex_exit(&dtrace_provider_lock);
16125 mutex_exit(&dtrace_lock);
16126 mutex_exit(&cpu_lock);
16127 return (DDI_FAILURE);
16128 }
16129
16130 dtrace_provider = NULL;
16131
16132 if ((state = dtrace_anon_grab()) != NULL) {
16133 /*
16134 * If there were ECBs on this state, the provider should
16135 * have not been allowed to detach; assert that there is
16136 * none.
16137 */
16138 ASSERT(state->dts_necbs == 0);
16139 dtrace_state_destroy(state);
16140
16141 /*
16142 * If we're being detached with anonymous state, we need to
16143 * indicate to the kernel debugger that DTrace is now inactive.
16144 */
16145 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16146 }
16147
16148 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16149 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16150 dtrace_cpu_init = NULL;
16151 dtrace_helpers_cleanup = NULL;
16152 dtrace_helpers_fork = NULL;
16153 dtrace_cpustart_init = NULL;
16154 dtrace_cpustart_fini = NULL;
16155 dtrace_debugger_init = NULL;
16156 dtrace_debugger_fini = NULL;
16157 dtrace_modload = NULL;
16158 dtrace_modunload = NULL;
16159
16160 ASSERT(dtrace_getf == 0);
16161 ASSERT(dtrace_closef == NULL);
16162
16163 mutex_exit(&cpu_lock);
16164
16165 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16166 dtrace_probes = NULL;
16167 dtrace_nprobes = 0;
16168
16169 dtrace_hash_destroy(dtrace_bymod);
16170 dtrace_hash_destroy(dtrace_byfunc);
16171 dtrace_hash_destroy(dtrace_byname);
16172 dtrace_bymod = NULL;
16173 dtrace_byfunc = NULL;
16174 dtrace_byname = NULL;
16175
16176 kmem_cache_destroy(dtrace_state_cache);
16177 vmem_destroy(dtrace_minor);
16178 vmem_destroy(dtrace_arena);
16179
16180 if (dtrace_toxrange != NULL) {
16181 kmem_free(dtrace_toxrange,
16182 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16183 dtrace_toxrange = NULL;
16184 dtrace_toxranges = 0;
16185 dtrace_toxranges_max = 0;
16186 }
16187
16188 ddi_remove_minor_node(dtrace_devi, NULL);
16189 dtrace_devi = NULL;
16190
16191 ddi_soft_state_fini(&dtrace_softstate);
16192
16193 ASSERT(dtrace_vtime_references == 0);
16194 ASSERT(dtrace_opens == 0);
16195 ASSERT(dtrace_retained == NULL);
16196
16197 mutex_exit(&dtrace_lock);
16198 mutex_exit(&dtrace_provider_lock);
16199
16200 /*
16201 * We don't destroy the task queue until after we have dropped our
16202 * locks (taskq_destroy() may block on running tasks). To prevent
16203 * attempting to do work after we have effectively detached but before
16204 * the task queue has been destroyed, all tasks dispatched via the
16205 * task queue must check that DTrace is still attached before
16206 * performing any operation.
16207 */
16208 taskq_destroy(dtrace_taskq);
16209 dtrace_taskq = NULL;
16210
16211 return (DDI_SUCCESS);
16212 }
16213
16214 /*ARGSUSED*/
16215 static int
16216 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16217 {
16218 int error;
16219
16220 switch (infocmd) {
16221 case DDI_INFO_DEVT2DEVINFO:
16222 *result = (void *)dtrace_devi;
16223 error = DDI_SUCCESS;
16224 break;
16225 case DDI_INFO_DEVT2INSTANCE:
16226 *result = (void *)0;
16227 error = DDI_SUCCESS;
16228 break;
16229 default:
16230 error = DDI_FAILURE;
16231 }
16232 return (error);
16233 }
16234
16235 static struct cb_ops dtrace_cb_ops = {
16236 dtrace_open, /* open */
16237 dtrace_close, /* close */
16238 nulldev, /* strategy */
16239 nulldev, /* print */
16240 nodev, /* dump */
16241 nodev, /* read */
16242 nodev, /* write */
16243 dtrace_ioctl, /* ioctl */
16244 nodev, /* devmap */
16245 nodev, /* mmap */
16246 nodev, /* segmap */
16247 nochpoll, /* poll */
16248 ddi_prop_op, /* cb_prop_op */
16249 0, /* streamtab */
16250 D_NEW | D_MP /* Driver compatibility flag */
16251 };
16252
16253 static struct dev_ops dtrace_ops = {
16254 DEVO_REV, /* devo_rev */
16255 0, /* refcnt */
16256 dtrace_info, /* get_dev_info */
16257 nulldev, /* identify */
16258 nulldev, /* probe */
16259 dtrace_attach, /* attach */
16260 dtrace_detach, /* detach */
16261 nodev, /* reset */
16262 &dtrace_cb_ops, /* driver operations */
16263 NULL, /* bus operations */
16264 nodev, /* dev power */
16265 ddi_quiesce_not_needed, /* quiesce */
16266 };
16267
16268 static struct modldrv modldrv = {
16269 &mod_driverops, /* module type (this is a pseudo driver) */
16270 "Dynamic Tracing", /* name of module */
16271 &dtrace_ops, /* driver ops */
16272 };
16273
16274 static struct modlinkage modlinkage = {
16275 MODREV_1,
16276 (void *)&modldrv,
16277 NULL
16278 };
16279
16280 int
16281 _init(void)
16282 {
16283 return (mod_install(&modlinkage));
16284 }
16285
16286 int
16287 _info(struct modinfo *modinfop)
16288 {
16289 return (mod_info(&modlinkage, modinfop));
16290 }
16291
16292 int
16293 _fini(void)
16294 {
16295 return (mod_remove(&modlinkage));
16296 }