1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
25 * Copyright (c) 2012 by Delphix. All rights reserved.
26 */
27
28 /*
29 * DTrace - Dynamic Tracing for Solaris
30 *
31 * This is the implementation of the Solaris Dynamic Tracing framework
32 * (DTrace). The user-visible interface to DTrace is described at length in
33 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
34 * library, the in-kernel DTrace framework, and the DTrace providers are
35 * described in the block comments in the <sys/dtrace.h> header file. The
36 * internal architecture of DTrace is described in the block comments in the
37 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
38 * implementation very much assume mastery of all of these sources; if one has
39 * an unanswered question about the implementation, one should consult them
40 * first.
41 *
42 * The functions here are ordered roughly as follows:
43 *
44 * - Probe context functions
45 * - Probe hashing functions
46 * - Non-probe context utility functions
47 * - Matching functions
48 * - Provider-to-Framework API functions
49 * - Probe management functions
50 * - DIF object functions
51 * - Format functions
52 * - Predicate functions
53 * - ECB functions
54 * - Buffer functions
55 * - Enabling functions
56 * - DOF functions
57 * - Anonymous enabling functions
58 * - Consumer state functions
59 * - Helper functions
60 * - Hook functions
61 * - Driver cookbook functions
62 *
63 * Each group of functions begins with a block comment labelled the "DTrace
64 * [Group] Functions", allowing one to find each block by searching forward
65 * on capital-f functions.
66 */
67 #include <sys/errno.h>
68 #include <sys/stat.h>
69 #include <sys/modctl.h>
70 #include <sys/conf.h>
71 #include <sys/systm.h>
72 #include <sys/ddi.h>
73 #include <sys/sunddi.h>
74 #include <sys/cpuvar.h>
75 #include <sys/kmem.h>
76 #include <sys/strsubr.h>
77 #include <sys/sysmacros.h>
78 #include <sys/dtrace_impl.h>
79 #include <sys/atomic.h>
80 #include <sys/cmn_err.h>
81 #include <sys/mutex_impl.h>
82 #include <sys/rwlock_impl.h>
83 #include <sys/ctf_api.h>
84 #include <sys/panic.h>
85 #include <sys/priv_impl.h>
86 #include <sys/policy.h>
87 #include <sys/cred_impl.h>
88 #include <sys/procfs_isa.h>
89 #include <sys/taskq.h>
90 #include <sys/mkdev.h>
91 #include <sys/kdi.h>
92 #include <sys/zone.h>
93 #include <sys/socket.h>
94 #include <netinet/in.h>
95 #include "strtolctype.h"
96
97 /*
98 * DTrace Tunable Variables
99 *
100 * The following variables may be tuned by adding a line to /etc/system that
101 * includes both the name of the DTrace module ("dtrace") and the name of the
102 * variable. For example:
103 *
104 * set dtrace:dtrace_destructive_disallow = 1
105 *
106 * In general, the only variables that one should be tuning this way are those
107 * that affect system-wide DTrace behavior, and for which the default behavior
108 * is undesirable. Most of these variables are tunable on a per-consumer
109 * basis using DTrace options, and need not be tuned on a system-wide basis.
110 * When tuning these variables, avoid pathological values; while some attempt
111 * is made to verify the integrity of these variables, they are not considered
112 * part of the supported interface to DTrace, and they are therefore not
113 * checked comprehensively. Further, these variables should not be tuned
114 * dynamically via "mdb -kw" or other means; they should only be tuned via
115 * /etc/system.
116 */
117 int dtrace_destructive_disallow = 0;
118 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
119 size_t dtrace_difo_maxsize = (256 * 1024);
120 dtrace_optval_t dtrace_dof_maxsize = (8 * 1024 * 1024);
121 size_t dtrace_global_maxsize = (16 * 1024);
122 size_t dtrace_actions_max = (16 * 1024);
123 size_t dtrace_retain_max = 1024;
124 dtrace_optval_t dtrace_helper_actions_max = 1024;
125 dtrace_optval_t dtrace_helper_providers_max = 32;
126 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
127 size_t dtrace_strsize_default = 256;
128 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
129 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
130 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
131 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
132 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
133 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
134 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
135 dtrace_optval_t dtrace_nspec_default = 1;
136 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
137 dtrace_optval_t dtrace_stackframes_default = 20;
138 dtrace_optval_t dtrace_ustackframes_default = 20;
139 dtrace_optval_t dtrace_jstackframes_default = 50;
140 dtrace_optval_t dtrace_jstackstrsize_default = 512;
141 int dtrace_msgdsize_max = 128;
142 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
143 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
144 int dtrace_devdepth_max = 32;
145 int dtrace_err_verbose;
146 hrtime_t dtrace_deadman_interval = NANOSEC;
147 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
148 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
149 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
150
151 /*
152 * DTrace External Variables
153 *
154 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
155 * available to DTrace consumers via the backtick (`) syntax. One of these,
156 * dtrace_zero, is made deliberately so: it is provided as a source of
157 * well-known, zero-filled memory. While this variable is not documented,
158 * it is used by some translators as an implementation detail.
159 */
160 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
161
162 /*
163 * DTrace Internal Variables
164 */
165 static dev_info_t *dtrace_devi; /* device info */
166 static vmem_t *dtrace_arena; /* probe ID arena */
167 static vmem_t *dtrace_minor; /* minor number arena */
168 static taskq_t *dtrace_taskq; /* task queue */
169 static dtrace_probe_t **dtrace_probes; /* array of all probes */
170 static int dtrace_nprobes; /* number of probes */
171 static dtrace_provider_t *dtrace_provider; /* provider list */
172 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
173 static int dtrace_opens; /* number of opens */
174 static int dtrace_helpers; /* number of helpers */
175 static int dtrace_getf; /* number of unpriv getf()s */
176 static void *dtrace_softstate; /* softstate pointer */
177 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
178 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
179 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
180 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
181 static int dtrace_toxranges; /* number of toxic ranges */
182 static int dtrace_toxranges_max; /* size of toxic range array */
183 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
184 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
185 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
186 static kthread_t *dtrace_panicked; /* panicking thread */
187 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
188 static dtrace_genid_t dtrace_probegen; /* current probe generation */
189 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
190 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
191 static dtrace_genid_t dtrace_retained_gen; /* current retained enab gen */
192 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
193 static int dtrace_dynvar_failclean; /* dynvars failed to clean */
194
195 /*
196 * DTrace Locking
197 * DTrace is protected by three (relatively coarse-grained) locks:
198 *
199 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
200 * including enabling state, probes, ECBs, consumer state, helper state,
201 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
202 * probe context is lock-free -- synchronization is handled via the
203 * dtrace_sync() cross call mechanism.
204 *
205 * (2) dtrace_provider_lock is required when manipulating provider state, or
206 * when provider state must be held constant.
207 *
208 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
209 * when meta provider state must be held constant.
210 *
211 * The lock ordering between these three locks is dtrace_meta_lock before
212 * dtrace_provider_lock before dtrace_lock. (In particular, there are
213 * several places where dtrace_provider_lock is held by the framework as it
214 * calls into the providers -- which then call back into the framework,
215 * grabbing dtrace_lock.)
216 *
217 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
218 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
219 * role as a coarse-grained lock; it is acquired before both of these locks.
220 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
221 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
222 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
223 * acquired _between_ dtrace_provider_lock and dtrace_lock.
224 */
225 static kmutex_t dtrace_lock; /* probe state lock */
226 static kmutex_t dtrace_provider_lock; /* provider state lock */
227 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
228
229 /*
230 * DTrace Provider Variables
231 *
232 * These are the variables relating to DTrace as a provider (that is, the
233 * provider of the BEGIN, END, and ERROR probes).
234 */
235 static dtrace_pattr_t dtrace_provider_attr = {
236 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
237 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
238 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
239 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
240 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
241 };
242
243 static void
244 dtrace_nullop(void)
245 {}
246
247 static int
248 dtrace_enable_nullop(void)
249 {
250 return (0);
251 }
252
253 static dtrace_pops_t dtrace_provider_ops = {
254 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
255 (void (*)(void *, struct modctl *))dtrace_nullop,
256 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop,
257 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
258 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
259 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
260 NULL,
261 NULL,
262 NULL,
263 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
264 };
265
266 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
267 static dtrace_id_t dtrace_probeid_end; /* special END probe */
268 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
269
270 /*
271 * DTrace Helper Tracing Variables
272 *
273 * These variables should be set dynamically to enable helper tracing. The
274 * only variables that should be set are dtrace_helptrace_enable (which should
275 * be set to a non-zero value to allocate helper tracing buffers on the next
276 * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
277 * non-zero value to deallocate helper tracing buffers on the next close of
278 * /dev/dtrace). When (and only when) helper tracing is disabled, the
279 * buffer size may also be set via dtrace_helptrace_bufsize.
280 */
281 int dtrace_helptrace_enable = 0;
282 int dtrace_helptrace_disable = 0;
283 int dtrace_helptrace_bufsize = 16 * 1024 * 1024;
284 uint32_t dtrace_helptrace_nlocals;
285 static dtrace_helptrace_t *dtrace_helptrace_buffer;
286 static uint32_t dtrace_helptrace_next = 0;
287 static int dtrace_helptrace_wrapped = 0;
288
289 /*
290 * DTrace Error Hashing
291 *
292 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
293 * table. This is very useful for checking coverage of tests that are
294 * expected to induce DIF or DOF processing errors, and may be useful for
295 * debugging problems in the DIF code generator or in DOF generation . The
296 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
297 */
298 #ifdef DEBUG
299 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
300 static const char *dtrace_errlast;
301 static kthread_t *dtrace_errthread;
302 static kmutex_t dtrace_errlock;
303 #endif
304
305 /*
306 * DTrace Macros and Constants
307 *
308 * These are various macros that are useful in various spots in the
309 * implementation, along with a few random constants that have no meaning
310 * outside of the implementation. There is no real structure to this cpp
311 * mishmash -- but is there ever?
312 */
313 #define DTRACE_HASHSTR(hash, probe) \
314 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
315
316 #define DTRACE_HASHNEXT(hash, probe) \
317 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
318
319 #define DTRACE_HASHPREV(hash, probe) \
320 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
321
322 #define DTRACE_HASHEQ(hash, lhs, rhs) \
323 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
324 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
325
326 #define DTRACE_AGGHASHSIZE_SLEW 17
327
328 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
329
330 /*
331 * The key for a thread-local variable consists of the lower 61 bits of the
332 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
333 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
334 * equal to a variable identifier. This is necessary (but not sufficient) to
335 * assure that global associative arrays never collide with thread-local
336 * variables. To guarantee that they cannot collide, we must also define the
337 * order for keying dynamic variables. That order is:
338 *
339 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
340 *
341 * Because the variable-key and the tls-key are in orthogonal spaces, there is
342 * no way for a global variable key signature to match a thread-local key
343 * signature.
344 */
345 #define DTRACE_TLS_THRKEY(where) { \
346 uint_t intr = 0; \
347 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
348 for (; actv; actv >>= 1) \
349 intr++; \
350 ASSERT(intr < (1 << 3)); \
351 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
352 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
353 }
354
355 #define DT_BSWAP_8(x) ((x) & 0xff)
356 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
357 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
358 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
359
360 #define DT_MASK_LO 0x00000000FFFFFFFFULL
361
362 #define DTRACE_STORE(type, tomax, offset, what) \
363 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
364
365 #ifndef __i386
366 #define DTRACE_ALIGNCHECK(addr, size, flags) \
367 if (addr & (size - 1)) { \
368 *flags |= CPU_DTRACE_BADALIGN; \
369 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
370 return (0); \
371 }
372 #else
373 #define DTRACE_ALIGNCHECK(addr, size, flags)
374 #endif
375
376 /*
377 * Test whether a range of memory starting at testaddr of size testsz falls
378 * within the range of memory described by addr, sz. We take care to avoid
379 * problems with overflow and underflow of the unsigned quantities, and
380 * disallow all negative sizes. Ranges of size 0 are allowed.
381 */
382 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
383 ((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
384 (testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
385 (testaddr) + (testsz) >= (testaddr))
386
387 /*
388 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
389 * alloc_sz on the righthand side of the comparison in order to avoid overflow
390 * or underflow in the comparison with it. This is simpler than the INRANGE
391 * check above, because we know that the dtms_scratch_ptr is valid in the
392 * range. Allocations of size zero are allowed.
393 */
394 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
395 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
396 (mstate)->dtms_scratch_ptr >= (alloc_sz))
397
398 #define DTRACE_LOADFUNC(bits) \
399 /*CSTYLED*/ \
400 uint##bits##_t \
401 dtrace_load##bits(uintptr_t addr) \
402 { \
403 size_t size = bits / NBBY; \
404 /*CSTYLED*/ \
405 uint##bits##_t rval; \
406 int i; \
407 volatile uint16_t *flags = (volatile uint16_t *) \
408 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags; \
409 \
410 DTRACE_ALIGNCHECK(addr, size, flags); \
411 \
412 for (i = 0; i < dtrace_toxranges; i++) { \
413 if (addr >= dtrace_toxrange[i].dtt_limit) \
414 continue; \
415 \
416 if (addr + size <= dtrace_toxrange[i].dtt_base) \
417 continue; \
418 \
419 /* \
420 * This address falls within a toxic region; return 0. \
421 */ \
422 *flags |= CPU_DTRACE_BADADDR; \
423 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr; \
424 return (0); \
425 } \
426 \
427 *flags |= CPU_DTRACE_NOFAULT; \
428 /*CSTYLED*/ \
429 rval = *((volatile uint##bits##_t *)addr); \
430 *flags &= ~CPU_DTRACE_NOFAULT; \
431 \
432 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
433 }
434
435 #ifdef _LP64
436 #define dtrace_loadptr dtrace_load64
437 #else
438 #define dtrace_loadptr dtrace_load32
439 #endif
440
441 #define DTRACE_DYNHASH_FREE 0
442 #define DTRACE_DYNHASH_SINK 1
443 #define DTRACE_DYNHASH_VALID 2
444
445 #define DTRACE_MATCH_FAIL -1
446 #define DTRACE_MATCH_NEXT 0
447 #define DTRACE_MATCH_DONE 1
448 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
449 #define DTRACE_STATE_ALIGN 64
450
451 #define DTRACE_FLAGS2FLT(flags) \
452 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
453 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
454 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
455 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
456 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
457 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
458 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
459 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
460 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
461 DTRACEFLT_UNKNOWN)
462
463 #define DTRACEACT_ISSTRING(act) \
464 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
465 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
466
467 static size_t dtrace_strlen(const char *, size_t);
468 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
469 static void dtrace_enabling_provide(dtrace_provider_t *);
470 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
471 static void dtrace_enabling_matchall(void);
472 static void dtrace_enabling_reap(void);
473 static dtrace_state_t *dtrace_anon_grab(void);
474 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
475 dtrace_state_t *, uint64_t, uint64_t);
476 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
477 static void dtrace_buffer_drop(dtrace_buffer_t *);
478 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
479 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
480 dtrace_state_t *, dtrace_mstate_t *);
481 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
482 dtrace_optval_t);
483 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
484 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
485 static int dtrace_priv_proc(dtrace_state_t *, dtrace_mstate_t *);
486 static void dtrace_getf_barrier(void);
487
488 /*
489 * DTrace Probe Context Functions
490 *
491 * These functions are called from probe context. Because probe context is
492 * any context in which C may be called, arbitrarily locks may be held,
493 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
494 * As a result, functions called from probe context may only call other DTrace
495 * support functions -- they may not interact at all with the system at large.
496 * (Note that the ASSERT macro is made probe-context safe by redefining it in
497 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
498 * loads are to be performed from probe context, they _must_ be in terms of
499 * the safe dtrace_load*() variants.
500 *
501 * Some functions in this block are not actually called from probe context;
502 * for these functions, there will be a comment above the function reading
503 * "Note: not called from probe context."
504 */
505 void
506 dtrace_panic(const char *format, ...)
507 {
508 va_list alist;
509
510 va_start(alist, format);
511 dtrace_vpanic(format, alist);
512 va_end(alist);
513 }
514
515 int
516 dtrace_assfail(const char *a, const char *f, int l)
517 {
518 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
519
520 /*
521 * We just need something here that even the most clever compiler
522 * cannot optimize away.
523 */
524 return (a[(uintptr_t)f]);
525 }
526
527 /*
528 * Atomically increment a specified error counter from probe context.
529 */
530 static void
531 dtrace_error(uint32_t *counter)
532 {
533 /*
534 * Most counters stored to in probe context are per-CPU counters.
535 * However, there are some error conditions that are sufficiently
536 * arcane that they don't merit per-CPU storage. If these counters
537 * are incremented concurrently on different CPUs, scalability will be
538 * adversely affected -- but we don't expect them to be white-hot in a
539 * correctly constructed enabling...
540 */
541 uint32_t oval, nval;
542
543 do {
544 oval = *counter;
545
546 if ((nval = oval + 1) == 0) {
547 /*
548 * If the counter would wrap, set it to 1 -- assuring
549 * that the counter is never zero when we have seen
550 * errors. (The counter must be 32-bits because we
551 * aren't guaranteed a 64-bit compare&swap operation.)
552 * To save this code both the infamy of being fingered
553 * by a priggish news story and the indignity of being
554 * the target of a neo-puritan witch trial, we're
555 * carefully avoiding any colorful description of the
556 * likelihood of this condition -- but suffice it to
557 * say that it is only slightly more likely than the
558 * overflow of predicate cache IDs, as discussed in
559 * dtrace_predicate_create().
560 */
561 nval = 1;
562 }
563 } while (dtrace_cas32(counter, oval, nval) != oval);
564 }
565
566 /*
567 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
568 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
569 */
570 DTRACE_LOADFUNC(8)
571 DTRACE_LOADFUNC(16)
572 DTRACE_LOADFUNC(32)
573 DTRACE_LOADFUNC(64)
574
575 static int
576 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
577 {
578 if (dest < mstate->dtms_scratch_base)
579 return (0);
580
581 if (dest + size < dest)
582 return (0);
583
584 if (dest + size > mstate->dtms_scratch_ptr)
585 return (0);
586
587 return (1);
588 }
589
590 static int
591 dtrace_canstore_statvar(uint64_t addr, size_t sz,
592 dtrace_statvar_t **svars, int nsvars)
593 {
594 int i;
595
596 for (i = 0; i < nsvars; i++) {
597 dtrace_statvar_t *svar = svars[i];
598
599 if (svar == NULL || svar->dtsv_size == 0)
600 continue;
601
602 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
603 return (1);
604 }
605
606 return (0);
607 }
608
609 /*
610 * Check to see if the address is within a memory region to which a store may
611 * be issued. This includes the DTrace scratch areas, and any DTrace variable
612 * region. The caller of dtrace_canstore() is responsible for performing any
613 * alignment checks that are needed before stores are actually executed.
614 */
615 static int
616 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
617 dtrace_vstate_t *vstate)
618 {
619 /*
620 * First, check to see if the address is in scratch space...
621 */
622 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
623 mstate->dtms_scratch_size))
624 return (1);
625
626 /*
627 * Now check to see if it's a dynamic variable. This check will pick
628 * up both thread-local variables and any global dynamically-allocated
629 * variables.
630 */
631 if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
632 vstate->dtvs_dynvars.dtds_size)) {
633 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
634 uintptr_t base = (uintptr_t)dstate->dtds_base +
635 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
636 uintptr_t chunkoffs;
637
638 /*
639 * Before we assume that we can store here, we need to make
640 * sure that it isn't in our metadata -- storing to our
641 * dynamic variable metadata would corrupt our state. For
642 * the range to not include any dynamic variable metadata,
643 * it must:
644 *
645 * (1) Start above the hash table that is at the base of
646 * the dynamic variable space
647 *
648 * (2) Have a starting chunk offset that is beyond the
649 * dtrace_dynvar_t that is at the base of every chunk
650 *
651 * (3) Not span a chunk boundary
652 *
653 */
654 if (addr < base)
655 return (0);
656
657 chunkoffs = (addr - base) % dstate->dtds_chunksize;
658
659 if (chunkoffs < sizeof (dtrace_dynvar_t))
660 return (0);
661
662 if (chunkoffs + sz > dstate->dtds_chunksize)
663 return (0);
664
665 return (1);
666 }
667
668 /*
669 * Finally, check the static local and global variables. These checks
670 * take the longest, so we perform them last.
671 */
672 if (dtrace_canstore_statvar(addr, sz,
673 vstate->dtvs_locals, vstate->dtvs_nlocals))
674 return (1);
675
676 if (dtrace_canstore_statvar(addr, sz,
677 vstate->dtvs_globals, vstate->dtvs_nglobals))
678 return (1);
679
680 return (0);
681 }
682
683
684 /*
685 * Convenience routine to check to see if the address is within a memory
686 * region in which a load may be issued given the user's privilege level;
687 * if not, it sets the appropriate error flags and loads 'addr' into the
688 * illegal value slot.
689 *
690 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
691 * appropriate memory access protection.
692 */
693 static int
694 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
695 dtrace_vstate_t *vstate)
696 {
697 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
698 file_t *fp;
699
700 /*
701 * If we hold the privilege to read from kernel memory, then
702 * everything is readable.
703 */
704 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
705 return (1);
706
707 /*
708 * You can obviously read that which you can store.
709 */
710 if (dtrace_canstore(addr, sz, mstate, vstate))
711 return (1);
712
713 /*
714 * We're allowed to read from our own string table.
715 */
716 if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
717 mstate->dtms_difo->dtdo_strlen))
718 return (1);
719
720 if (vstate->dtvs_state != NULL &&
721 dtrace_priv_proc(vstate->dtvs_state, mstate)) {
722 proc_t *p;
723
724 /*
725 * When we have privileges to the current process, there are
726 * several context-related kernel structures that are safe to
727 * read, even absent the privilege to read from kernel memory.
728 * These reads are safe because these structures contain only
729 * state that (1) we're permitted to read, (2) is harmless or
730 * (3) contains pointers to additional kernel state that we're
731 * not permitted to read (and as such, do not present an
732 * opportunity for privilege escalation). Finally (and
733 * critically), because of the nature of their relation with
734 * the current thread context, the memory associated with these
735 * structures cannot change over the duration of probe context,
736 * and it is therefore impossible for this memory to be
737 * deallocated and reallocated as something else while it's
738 * being operated upon.
739 */
740 if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t)))
741 return (1);
742
743 if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
744 sz, curthread->t_procp, sizeof (proc_t))) {
745 return (1);
746 }
747
748 if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
749 curthread->t_cred, sizeof (cred_t))) {
750 return (1);
751 }
752
753 if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
754 &(p->p_pidp->pid_id), sizeof (pid_t))) {
755 return (1);
756 }
757
758 if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
759 curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
760 return (1);
761 }
762 }
763
764 if ((fp = mstate->dtms_getf) != NULL) {
765 uintptr_t psz = sizeof (void *);
766 vnode_t *vp;
767 vnodeops_t *op;
768
769 /*
770 * When getf() returns a file_t, the enabling is implicitly
771 * granted the (transient) right to read the returned file_t
772 * as well as the v_path and v_op->vnop_name of the underlying
773 * vnode. These accesses are allowed after a successful
774 * getf() because the members that they refer to cannot change
775 * once set -- and the barrier logic in the kernel's closef()
776 * path assures that the file_t and its referenced vode_t
777 * cannot themselves be stale (that is, it impossible for
778 * either dtms_getf itself or its f_vnode member to reference
779 * freed memory).
780 */
781 if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t)))
782 return (1);
783
784 if ((vp = fp->f_vnode) != NULL) {
785 if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz))
786 return (1);
787
788 if (vp->v_path != NULL && DTRACE_INRANGE(addr, sz,
789 vp->v_path, strlen(vp->v_path) + 1)) {
790 return (1);
791 }
792
793 if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz))
794 return (1);
795
796 if ((op = vp->v_op) != NULL &&
797 DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
798 return (1);
799 }
800
801 if (op != NULL && op->vnop_name != NULL &&
802 DTRACE_INRANGE(addr, sz, op->vnop_name,
803 strlen(op->vnop_name) + 1)) {
804 return (1);
805 }
806 }
807 }
808
809 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
810 *illval = addr;
811 return (0);
812 }
813
814 /*
815 * Convenience routine to check to see if a given string is within a memory
816 * region in which a load may be issued given the user's privilege level;
817 * this exists so that we don't need to issue unnecessary dtrace_strlen()
818 * calls in the event that the user has all privileges.
819 */
820 static int
821 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
822 dtrace_vstate_t *vstate)
823 {
824 size_t strsz;
825
826 /*
827 * If we hold the privilege to read from kernel memory, then
828 * everything is readable.
829 */
830 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
831 return (1);
832
833 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
834 if (dtrace_canload(addr, strsz, mstate, vstate))
835 return (1);
836
837 return (0);
838 }
839
840 /*
841 * Convenience routine to check to see if a given variable is within a memory
842 * region in which a load may be issued given the user's privilege level.
843 */
844 static int
845 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
846 dtrace_vstate_t *vstate)
847 {
848 size_t sz, strsize;
849 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
850
851 /*
852 * If we hold the privilege to read from kernel memory, then
853 * everything is readable.
854 */
855 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
856 return (1);
857
858 if (type->dtdt_kind == DIF_TYPE_STRING) {
859 dtrace_state_t *state = vstate->dtvs_state;
860
861 if (state != NULL) {
862 strsize = state->dts_options[DTRACEOPT_STRSIZE];
863 } else {
864 /*
865 * In helper context, we have a NULL state; fall back
866 * to using the system-wide default for the string size
867 * in this case.
868 */
869 strsize = dtrace_strsize_default;
870 }
871
872 sz = dtrace_strlen(src, strsize) + 1;
873 } else {
874 sz = type->dtdt_size;
875 }
876
877 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
878 }
879
880 /*
881 * Convert a string to a signed integer using safe loads.
882 */
883 static int64_t
884 dtrace_strtoll(char *input, int base, size_t limit)
885 {
886 uintptr_t pos = (uintptr_t)input;
887 int64_t val = 0;
888 int x;
889 boolean_t neg = B_FALSE;
890 char c, cc, ccc;
891 uintptr_t end = pos + limit;
892
893 /* eat whitespace */
894 while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
895 pos++;
896
897 /* sign? */
898 if (c == '-' || c == '+') {
899 if (c == '-')
900 neg = B_TRUE;
901 c = dtrace_load8(++pos);
902 }
903
904 /* hex prefix? */
905 if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
906 cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
907 pos += 2; /* skip over leading "0x" or "0X" */
908 c = ccc;
909 }
910
911 /* read in digits */
912 for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
913 c = dtrace_load8(++pos))
914 val = val * base + x;
915
916 return (neg ? -val : val);
917 }
918
919 /*
920 * Compare two strings using safe loads.
921 */
922 static int
923 dtrace_strncmp(char *s1, char *s2, size_t limit)
924 {
925 uint8_t c1, c2;
926 volatile uint16_t *flags;
927
928 if (s1 == s2 || limit == 0)
929 return (0);
930
931 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
932
933 do {
934 if (s1 == NULL) {
935 c1 = '\0';
936 } else {
937 c1 = dtrace_load8((uintptr_t)s1++);
938 }
939
940 if (s2 == NULL) {
941 c2 = '\0';
942 } else {
943 c2 = dtrace_load8((uintptr_t)s2++);
944 }
945
946 if (c1 != c2)
947 return (c1 - c2);
948 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
949
950 return (0);
951 }
952
953 /*
954 * Compute strlen(s) for a string using safe memory accesses. The additional
955 * len parameter is used to specify a maximum length to ensure completion.
956 */
957 static size_t
958 dtrace_strlen(const char *s, size_t lim)
959 {
960 uint_t len;
961
962 for (len = 0; len != lim; len++) {
963 if (dtrace_load8((uintptr_t)s++) == '\0')
964 break;
965 }
966
967 return (len);
968 }
969
970 /*
971 * Check if an address falls within a toxic region.
972 */
973 static int
974 dtrace_istoxic(uintptr_t kaddr, size_t size)
975 {
976 uintptr_t taddr, tsize;
977 int i;
978
979 for (i = 0; i < dtrace_toxranges; i++) {
980 taddr = dtrace_toxrange[i].dtt_base;
981 tsize = dtrace_toxrange[i].dtt_limit - taddr;
982
983 if (kaddr - taddr < tsize) {
984 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
985 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
986 return (1);
987 }
988
989 if (taddr - kaddr < size) {
990 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
991 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
992 return (1);
993 }
994 }
995
996 return (0);
997 }
998
999 /*
1000 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
1001 * memory specified by the DIF program. The dst is assumed to be safe memory
1002 * that we can store to directly because it is managed by DTrace. As with
1003 * standard bcopy, overlapping copies are handled properly.
1004 */
1005 static void
1006 dtrace_bcopy(const void *src, void *dst, size_t len)
1007 {
1008 if (len != 0) {
1009 uint8_t *s1 = dst;
1010 const uint8_t *s2 = src;
1011
1012 if (s1 <= s2) {
1013 do {
1014 *s1++ = dtrace_load8((uintptr_t)s2++);
1015 } while (--len != 0);
1016 } else {
1017 s2 += len;
1018 s1 += len;
1019
1020 do {
1021 *--s1 = dtrace_load8((uintptr_t)--s2);
1022 } while (--len != 0);
1023 }
1024 }
1025 }
1026
1027 /*
1028 * Copy src to dst using safe memory accesses, up to either the specified
1029 * length, or the point that a nul byte is encountered. The src is assumed to
1030 * be unsafe memory specified by the DIF program. The dst is assumed to be
1031 * safe memory that we can store to directly because it is managed by DTrace.
1032 * Unlike dtrace_bcopy(), overlapping regions are not handled.
1033 */
1034 static void
1035 dtrace_strcpy(const void *src, void *dst, size_t len)
1036 {
1037 if (len != 0) {
1038 uint8_t *s1 = dst, c;
1039 const uint8_t *s2 = src;
1040
1041 do {
1042 *s1++ = c = dtrace_load8((uintptr_t)s2++);
1043 } while (--len != 0 && c != '\0');
1044 }
1045 }
1046
1047 /*
1048 * Copy src to dst, deriving the size and type from the specified (BYREF)
1049 * variable type. The src is assumed to be unsafe memory specified by the DIF
1050 * program. The dst is assumed to be DTrace variable memory that is of the
1051 * specified type; we assume that we can store to directly.
1052 */
1053 static void
1054 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1055 {
1056 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1057
1058 if (type->dtdt_kind == DIF_TYPE_STRING) {
1059 dtrace_strcpy(src, dst, type->dtdt_size);
1060 } else {
1061 dtrace_bcopy(src, dst, type->dtdt_size);
1062 }
1063 }
1064
1065 /*
1066 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1067 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1068 * safe memory that we can access directly because it is managed by DTrace.
1069 */
1070 static int
1071 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1072 {
1073 volatile uint16_t *flags;
1074
1075 flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1076
1077 if (s1 == s2)
1078 return (0);
1079
1080 if (s1 == NULL || s2 == NULL)
1081 return (1);
1082
1083 if (s1 != s2 && len != 0) {
1084 const uint8_t *ps1 = s1;
1085 const uint8_t *ps2 = s2;
1086
1087 do {
1088 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1089 return (1);
1090 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1091 }
1092 return (0);
1093 }
1094
1095 /*
1096 * Zero the specified region using a simple byte-by-byte loop. Note that this
1097 * is for safe DTrace-managed memory only.
1098 */
1099 static void
1100 dtrace_bzero(void *dst, size_t len)
1101 {
1102 uchar_t *cp;
1103
1104 for (cp = dst; len != 0; len--)
1105 *cp++ = 0;
1106 }
1107
1108 static void
1109 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1110 {
1111 uint64_t result[2];
1112
1113 result[0] = addend1[0] + addend2[0];
1114 result[1] = addend1[1] + addend2[1] +
1115 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1116
1117 sum[0] = result[0];
1118 sum[1] = result[1];
1119 }
1120
1121 /*
1122 * Shift the 128-bit value in a by b. If b is positive, shift left.
1123 * If b is negative, shift right.
1124 */
1125 static void
1126 dtrace_shift_128(uint64_t *a, int b)
1127 {
1128 uint64_t mask;
1129
1130 if (b == 0)
1131 return;
1132
1133 if (b < 0) {
1134 b = -b;
1135 if (b >= 64) {
1136 a[0] = a[1] >> (b - 64);
1137 a[1] = 0;
1138 } else {
1139 a[0] >>= b;
1140 mask = 1LL << (64 - b);
1141 mask -= 1;
1142 a[0] |= ((a[1] & mask) << (64 - b));
1143 a[1] >>= b;
1144 }
1145 } else {
1146 if (b >= 64) {
1147 a[1] = a[0] << (b - 64);
1148 a[0] = 0;
1149 } else {
1150 a[1] <<= b;
1151 mask = a[0] >> (64 - b);
1152 a[1] |= mask;
1153 a[0] <<= b;
1154 }
1155 }
1156 }
1157
1158 /*
1159 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1160 * use native multiplication on those, and then re-combine into the
1161 * resulting 128-bit value.
1162 *
1163 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1164 * hi1 * hi2 << 64 +
1165 * hi1 * lo2 << 32 +
1166 * hi2 * lo1 << 32 +
1167 * lo1 * lo2
1168 */
1169 static void
1170 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1171 {
1172 uint64_t hi1, hi2, lo1, lo2;
1173 uint64_t tmp[2];
1174
1175 hi1 = factor1 >> 32;
1176 hi2 = factor2 >> 32;
1177
1178 lo1 = factor1 & DT_MASK_LO;
1179 lo2 = factor2 & DT_MASK_LO;
1180
1181 product[0] = lo1 * lo2;
1182 product[1] = hi1 * hi2;
1183
1184 tmp[0] = hi1 * lo2;
1185 tmp[1] = 0;
1186 dtrace_shift_128(tmp, 32);
1187 dtrace_add_128(product, tmp, product);
1188
1189 tmp[0] = hi2 * lo1;
1190 tmp[1] = 0;
1191 dtrace_shift_128(tmp, 32);
1192 dtrace_add_128(product, tmp, product);
1193 }
1194
1195 /*
1196 * This privilege check should be used by actions and subroutines to
1197 * verify that the user credentials of the process that enabled the
1198 * invoking ECB match the target credentials
1199 */
1200 static int
1201 dtrace_priv_proc_common_user(dtrace_state_t *state)
1202 {
1203 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1204
1205 /*
1206 * We should always have a non-NULL state cred here, since if cred
1207 * is null (anonymous tracing), we fast-path bypass this routine.
1208 */
1209 ASSERT(s_cr != NULL);
1210
1211 if ((cr = CRED()) != NULL &&
1212 s_cr->cr_uid == cr->cr_uid &&
1213 s_cr->cr_uid == cr->cr_ruid &&
1214 s_cr->cr_uid == cr->cr_suid &&
1215 s_cr->cr_gid == cr->cr_gid &&
1216 s_cr->cr_gid == cr->cr_rgid &&
1217 s_cr->cr_gid == cr->cr_sgid)
1218 return (1);
1219
1220 return (0);
1221 }
1222
1223 /*
1224 * This privilege check should be used by actions and subroutines to
1225 * verify that the zone of the process that enabled the invoking ECB
1226 * matches the target credentials
1227 */
1228 static int
1229 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1230 {
1231 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1232
1233 /*
1234 * We should always have a non-NULL state cred here, since if cred
1235 * is null (anonymous tracing), we fast-path bypass this routine.
1236 */
1237 ASSERT(s_cr != NULL);
1238
1239 if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1240 return (1);
1241
1242 return (0);
1243 }
1244
1245 /*
1246 * This privilege check should be used by actions and subroutines to
1247 * verify that the process has not setuid or changed credentials.
1248 */
1249 static int
1250 dtrace_priv_proc_common_nocd()
1251 {
1252 proc_t *proc;
1253
1254 if ((proc = ttoproc(curthread)) != NULL &&
1255 !(proc->p_flag & SNOCD))
1256 return (1);
1257
1258 return (0);
1259 }
1260
1261 static int
1262 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1263 {
1264 int action = state->dts_cred.dcr_action;
1265
1266 if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1267 goto bad;
1268
1269 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1270 dtrace_priv_proc_common_zone(state) == 0)
1271 goto bad;
1272
1273 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1274 dtrace_priv_proc_common_user(state) == 0)
1275 goto bad;
1276
1277 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1278 dtrace_priv_proc_common_nocd() == 0)
1279 goto bad;
1280
1281 return (1);
1282
1283 bad:
1284 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1285
1286 return (0);
1287 }
1288
1289 static int
1290 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1291 {
1292 if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1293 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1294 return (1);
1295
1296 if (dtrace_priv_proc_common_zone(state) &&
1297 dtrace_priv_proc_common_user(state) &&
1298 dtrace_priv_proc_common_nocd())
1299 return (1);
1300 }
1301
1302 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1303
1304 return (0);
1305 }
1306
1307 static int
1308 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1309 {
1310 if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1311 (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1312 return (1);
1313
1314 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1315
1316 return (0);
1317 }
1318
1319 static int
1320 dtrace_priv_kernel(dtrace_state_t *state)
1321 {
1322 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1323 return (1);
1324
1325 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1326
1327 return (0);
1328 }
1329
1330 static int
1331 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1332 {
1333 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1334 return (1);
1335
1336 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1337
1338 return (0);
1339 }
1340
1341 /*
1342 * Determine if the dte_cond of the specified ECB allows for processing of
1343 * the current probe to continue. Note that this routine may allow continued
1344 * processing, but with access(es) stripped from the mstate's dtms_access
1345 * field.
1346 */
1347 static int
1348 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1349 dtrace_ecb_t *ecb)
1350 {
1351 dtrace_probe_t *probe = ecb->dte_probe;
1352 dtrace_provider_t *prov = probe->dtpr_provider;
1353 dtrace_pops_t *pops = &prov->dtpv_pops;
1354 int mode = DTRACE_MODE_NOPRIV_DROP;
1355
1356 ASSERT(ecb->dte_cond);
1357
1358 if (pops->dtps_mode != NULL) {
1359 mode = pops->dtps_mode(prov->dtpv_arg,
1360 probe->dtpr_id, probe->dtpr_arg);
1361
1362 ASSERT(mode & (DTRACE_MODE_USER | DTRACE_MODE_KERNEL));
1363 ASSERT(mode & (DTRACE_MODE_NOPRIV_RESTRICT |
1364 DTRACE_MODE_NOPRIV_DROP));
1365 }
1366
1367 /*
1368 * If the dte_cond bits indicate that this consumer is only allowed to
1369 * see user-mode firings of this probe, check that the probe was fired
1370 * while in a user context. If that's not the case, use the policy
1371 * specified by the provider to determine if we drop the probe or
1372 * merely restrict operation.
1373 */
1374 if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1375 ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1376
1377 if (!(mode & DTRACE_MODE_USER)) {
1378 if (mode & DTRACE_MODE_NOPRIV_DROP)
1379 return (0);
1380
1381 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1382 }
1383 }
1384
1385 /*
1386 * This is more subtle than it looks. We have to be absolutely certain
1387 * that CRED() isn't going to change out from under us so it's only
1388 * legit to examine that structure if we're in constrained situations.
1389 * Currently, the only times we'll this check is if a non-super-user
1390 * has enabled the profile or syscall providers -- providers that
1391 * allow visibility of all processes. For the profile case, the check
1392 * above will ensure that we're examining a user context.
1393 */
1394 if (ecb->dte_cond & DTRACE_COND_OWNER) {
1395 cred_t *cr;
1396 cred_t *s_cr = state->dts_cred.dcr_cred;
1397 proc_t *proc;
1398
1399 ASSERT(s_cr != NULL);
1400
1401 if ((cr = CRED()) == NULL ||
1402 s_cr->cr_uid != cr->cr_uid ||
1403 s_cr->cr_uid != cr->cr_ruid ||
1404 s_cr->cr_uid != cr->cr_suid ||
1405 s_cr->cr_gid != cr->cr_gid ||
1406 s_cr->cr_gid != cr->cr_rgid ||
1407 s_cr->cr_gid != cr->cr_sgid ||
1408 (proc = ttoproc(curthread)) == NULL ||
1409 (proc->p_flag & SNOCD)) {
1410 if (mode & DTRACE_MODE_NOPRIV_DROP)
1411 return (0);
1412
1413 mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1414 }
1415 }
1416
1417 /*
1418 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1419 * in our zone, check to see if our mode policy is to restrict rather
1420 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1421 * and DTRACE_ACCESS_ARGS
1422 */
1423 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1424 cred_t *cr;
1425 cred_t *s_cr = state->dts_cred.dcr_cred;
1426
1427 ASSERT(s_cr != NULL);
1428
1429 if ((cr = CRED()) == NULL ||
1430 s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1431 if (mode & DTRACE_MODE_NOPRIV_DROP)
1432 return (0);
1433
1434 mstate->dtms_access &=
1435 ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1436 }
1437 }
1438
1439 /*
1440 * By merits of being in this code path at all, we have limited
1441 * privileges. If the provider has indicated that limited privileges
1442 * are to denote restricted operation, strip off the ability to access
1443 * arguments.
1444 */
1445 if (mode & DTRACE_MODE_LIMITEDPRIV_RESTRICT)
1446 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1447
1448 return (1);
1449 }
1450
1451 /*
1452 * Note: not called from probe context. This function is called
1453 * asynchronously (and at a regular interval) from outside of probe context to
1454 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1455 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1456 */
1457 void
1458 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1459 {
1460 dtrace_dynvar_t *dirty;
1461 dtrace_dstate_percpu_t *dcpu;
1462 dtrace_dynvar_t **rinsep;
1463 int i, j, work = 0;
1464
1465 for (i = 0; i < NCPU; i++) {
1466 dcpu = &dstate->dtds_percpu[i];
1467 rinsep = &dcpu->dtdsc_rinsing;
1468
1469 /*
1470 * If the dirty list is NULL, there is no dirty work to do.
1471 */
1472 if (dcpu->dtdsc_dirty == NULL)
1473 continue;
1474
1475 if (dcpu->dtdsc_rinsing != NULL) {
1476 /*
1477 * If the rinsing list is non-NULL, then it is because
1478 * this CPU was selected to accept another CPU's
1479 * dirty list -- and since that time, dirty buffers
1480 * have accumulated. This is a highly unlikely
1481 * condition, but we choose to ignore the dirty
1482 * buffers -- they'll be picked up a future cleanse.
1483 */
1484 continue;
1485 }
1486
1487 if (dcpu->dtdsc_clean != NULL) {
1488 /*
1489 * If the clean list is non-NULL, then we're in a
1490 * situation where a CPU has done deallocations (we
1491 * have a non-NULL dirty list) but no allocations (we
1492 * also have a non-NULL clean list). We can't simply
1493 * move the dirty list into the clean list on this
1494 * CPU, yet we also don't want to allow this condition
1495 * to persist, lest a short clean list prevent a
1496 * massive dirty list from being cleaned (which in
1497 * turn could lead to otherwise avoidable dynamic
1498 * drops). To deal with this, we look for some CPU
1499 * with a NULL clean list, NULL dirty list, and NULL
1500 * rinsing list -- and then we borrow this CPU to
1501 * rinse our dirty list.
1502 */
1503 for (j = 0; j < NCPU; j++) {
1504 dtrace_dstate_percpu_t *rinser;
1505
1506 rinser = &dstate->dtds_percpu[j];
1507
1508 if (rinser->dtdsc_rinsing != NULL)
1509 continue;
1510
1511 if (rinser->dtdsc_dirty != NULL)
1512 continue;
1513
1514 if (rinser->dtdsc_clean != NULL)
1515 continue;
1516
1517 rinsep = &rinser->dtdsc_rinsing;
1518 break;
1519 }
1520
1521 if (j == NCPU) {
1522 /*
1523 * We were unable to find another CPU that
1524 * could accept this dirty list -- we are
1525 * therefore unable to clean it now.
1526 */
1527 dtrace_dynvar_failclean++;
1528 continue;
1529 }
1530 }
1531
1532 work = 1;
1533
1534 /*
1535 * Atomically move the dirty list aside.
1536 */
1537 do {
1538 dirty = dcpu->dtdsc_dirty;
1539
1540 /*
1541 * Before we zap the dirty list, set the rinsing list.
1542 * (This allows for a potential assertion in
1543 * dtrace_dynvar(): if a free dynamic variable appears
1544 * on a hash chain, either the dirty list or the
1545 * rinsing list for some CPU must be non-NULL.)
1546 */
1547 *rinsep = dirty;
1548 dtrace_membar_producer();
1549 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1550 dirty, NULL) != dirty);
1551 }
1552
1553 if (!work) {
1554 /*
1555 * We have no work to do; we can simply return.
1556 */
1557 return;
1558 }
1559
1560 dtrace_sync();
1561
1562 for (i = 0; i < NCPU; i++) {
1563 dcpu = &dstate->dtds_percpu[i];
1564
1565 if (dcpu->dtdsc_rinsing == NULL)
1566 continue;
1567
1568 /*
1569 * We are now guaranteed that no hash chain contains a pointer
1570 * into this dirty list; we can make it clean.
1571 */
1572 ASSERT(dcpu->dtdsc_clean == NULL);
1573 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1574 dcpu->dtdsc_rinsing = NULL;
1575 }
1576
1577 /*
1578 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1579 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1580 * This prevents a race whereby a CPU incorrectly decides that
1581 * the state should be something other than DTRACE_DSTATE_CLEAN
1582 * after dtrace_dynvar_clean() has completed.
1583 */
1584 dtrace_sync();
1585
1586 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1587 }
1588
1589 /*
1590 * Depending on the value of the op parameter, this function looks-up,
1591 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1592 * allocation is requested, this function will return a pointer to a
1593 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1594 * variable can be allocated. If NULL is returned, the appropriate counter
1595 * will be incremented.
1596 */
1597 dtrace_dynvar_t *
1598 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1599 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1600 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1601 {
1602 uint64_t hashval = DTRACE_DYNHASH_VALID;
1603 dtrace_dynhash_t *hash = dstate->dtds_hash;
1604 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1605 processorid_t me = CPU->cpu_id, cpu = me;
1606 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1607 size_t bucket, ksize;
1608 size_t chunksize = dstate->dtds_chunksize;
1609 uintptr_t kdata, lock, nstate;
1610 uint_t i;
1611
1612 ASSERT(nkeys != 0);
1613
1614 /*
1615 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1616 * algorithm. For the by-value portions, we perform the algorithm in
1617 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1618 * bit, and seems to have only a minute effect on distribution. For
1619 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1620 * over each referenced byte. It's painful to do this, but it's much
1621 * better than pathological hash distribution. The efficacy of the
1622 * hashing algorithm (and a comparison with other algorithms) may be
1623 * found by running the ::dtrace_dynstat MDB dcmd.
1624 */
1625 for (i = 0; i < nkeys; i++) {
1626 if (key[i].dttk_size == 0) {
1627 uint64_t val = key[i].dttk_value;
1628
1629 hashval += (val >> 48) & 0xffff;
1630 hashval += (hashval << 10);
1631 hashval ^= (hashval >> 6);
1632
1633 hashval += (val >> 32) & 0xffff;
1634 hashval += (hashval << 10);
1635 hashval ^= (hashval >> 6);
1636
1637 hashval += (val >> 16) & 0xffff;
1638 hashval += (hashval << 10);
1639 hashval ^= (hashval >> 6);
1640
1641 hashval += val & 0xffff;
1642 hashval += (hashval << 10);
1643 hashval ^= (hashval >> 6);
1644 } else {
1645 /*
1646 * This is incredibly painful, but it beats the hell
1647 * out of the alternative.
1648 */
1649 uint64_t j, size = key[i].dttk_size;
1650 uintptr_t base = (uintptr_t)key[i].dttk_value;
1651
1652 if (!dtrace_canload(base, size, mstate, vstate))
1653 break;
1654
1655 for (j = 0; j < size; j++) {
1656 hashval += dtrace_load8(base + j);
1657 hashval += (hashval << 10);
1658 hashval ^= (hashval >> 6);
1659 }
1660 }
1661 }
1662
1663 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1664 return (NULL);
1665
1666 hashval += (hashval << 3);
1667 hashval ^= (hashval >> 11);
1668 hashval += (hashval << 15);
1669
1670 /*
1671 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1672 * comes out to be one of our two sentinel hash values. If this
1673 * actually happens, we set the hashval to be a value known to be a
1674 * non-sentinel value.
1675 */
1676 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1677 hashval = DTRACE_DYNHASH_VALID;
1678
1679 /*
1680 * Yes, it's painful to do a divide here. If the cycle count becomes
1681 * important here, tricks can be pulled to reduce it. (However, it's
1682 * critical that hash collisions be kept to an absolute minimum;
1683 * they're much more painful than a divide.) It's better to have a
1684 * solution that generates few collisions and still keeps things
1685 * relatively simple.
1686 */
1687 bucket = hashval % dstate->dtds_hashsize;
1688
1689 if (op == DTRACE_DYNVAR_DEALLOC) {
1690 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1691
1692 for (;;) {
1693 while ((lock = *lockp) & 1)
1694 continue;
1695
1696 if (dtrace_casptr((void *)lockp,
1697 (void *)lock, (void *)(lock + 1)) == (void *)lock)
1698 break;
1699 }
1700
1701 dtrace_membar_producer();
1702 }
1703
1704 top:
1705 prev = NULL;
1706 lock = hash[bucket].dtdh_lock;
1707
1708 dtrace_membar_consumer();
1709
1710 start = hash[bucket].dtdh_chain;
1711 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1712 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1713 op != DTRACE_DYNVAR_DEALLOC));
1714
1715 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1716 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1717 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1718
1719 if (dvar->dtdv_hashval != hashval) {
1720 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1721 /*
1722 * We've reached the sink, and therefore the
1723 * end of the hash chain; we can kick out of
1724 * the loop knowing that we have seen a valid
1725 * snapshot of state.
1726 */
1727 ASSERT(dvar->dtdv_next == NULL);
1728 ASSERT(dvar == &dtrace_dynhash_sink);
1729 break;
1730 }
1731
1732 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1733 /*
1734 * We've gone off the rails: somewhere along
1735 * the line, one of the members of this hash
1736 * chain was deleted. Note that we could also
1737 * detect this by simply letting this loop run
1738 * to completion, as we would eventually hit
1739 * the end of the dirty list. However, we
1740 * want to avoid running the length of the
1741 * dirty list unnecessarily (it might be quite
1742 * long), so we catch this as early as
1743 * possible by detecting the hash marker. In
1744 * this case, we simply set dvar to NULL and
1745 * break; the conditional after the loop will
1746 * send us back to top.
1747 */
1748 dvar = NULL;
1749 break;
1750 }
1751
1752 goto next;
1753 }
1754
1755 if (dtuple->dtt_nkeys != nkeys)
1756 goto next;
1757
1758 for (i = 0; i < nkeys; i++, dkey++) {
1759 if (dkey->dttk_size != key[i].dttk_size)
1760 goto next; /* size or type mismatch */
1761
1762 if (dkey->dttk_size != 0) {
1763 if (dtrace_bcmp(
1764 (void *)(uintptr_t)key[i].dttk_value,
1765 (void *)(uintptr_t)dkey->dttk_value,
1766 dkey->dttk_size))
1767 goto next;
1768 } else {
1769 if (dkey->dttk_value != key[i].dttk_value)
1770 goto next;
1771 }
1772 }
1773
1774 if (op != DTRACE_DYNVAR_DEALLOC)
1775 return (dvar);
1776
1777 ASSERT(dvar->dtdv_next == NULL ||
1778 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1779
1780 if (prev != NULL) {
1781 ASSERT(hash[bucket].dtdh_chain != dvar);
1782 ASSERT(start != dvar);
1783 ASSERT(prev->dtdv_next == dvar);
1784 prev->dtdv_next = dvar->dtdv_next;
1785 } else {
1786 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1787 start, dvar->dtdv_next) != start) {
1788 /*
1789 * We have failed to atomically swing the
1790 * hash table head pointer, presumably because
1791 * of a conflicting allocation on another CPU.
1792 * We need to reread the hash chain and try
1793 * again.
1794 */
1795 goto top;
1796 }
1797 }
1798
1799 dtrace_membar_producer();
1800
1801 /*
1802 * Now set the hash value to indicate that it's free.
1803 */
1804 ASSERT(hash[bucket].dtdh_chain != dvar);
1805 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1806
1807 dtrace_membar_producer();
1808
1809 /*
1810 * Set the next pointer to point at the dirty list, and
1811 * atomically swing the dirty pointer to the newly freed dvar.
1812 */
1813 do {
1814 next = dcpu->dtdsc_dirty;
1815 dvar->dtdv_next = next;
1816 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1817
1818 /*
1819 * Finally, unlock this hash bucket.
1820 */
1821 ASSERT(hash[bucket].dtdh_lock == lock);
1822 ASSERT(lock & 1);
1823 hash[bucket].dtdh_lock++;
1824
1825 return (NULL);
1826 next:
1827 prev = dvar;
1828 continue;
1829 }
1830
1831 if (dvar == NULL) {
1832 /*
1833 * If dvar is NULL, it is because we went off the rails:
1834 * one of the elements that we traversed in the hash chain
1835 * was deleted while we were traversing it. In this case,
1836 * we assert that we aren't doing a dealloc (deallocs lock
1837 * the hash bucket to prevent themselves from racing with
1838 * one another), and retry the hash chain traversal.
1839 */
1840 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1841 goto top;
1842 }
1843
1844 if (op != DTRACE_DYNVAR_ALLOC) {
1845 /*
1846 * If we are not to allocate a new variable, we want to
1847 * return NULL now. Before we return, check that the value
1848 * of the lock word hasn't changed. If it has, we may have
1849 * seen an inconsistent snapshot.
1850 */
1851 if (op == DTRACE_DYNVAR_NOALLOC) {
1852 if (hash[bucket].dtdh_lock != lock)
1853 goto top;
1854 } else {
1855 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1856 ASSERT(hash[bucket].dtdh_lock == lock);
1857 ASSERT(lock & 1);
1858 hash[bucket].dtdh_lock++;
1859 }
1860
1861 return (NULL);
1862 }
1863
1864 /*
1865 * We need to allocate a new dynamic variable. The size we need is the
1866 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1867 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1868 * the size of any referred-to data (dsize). We then round the final
1869 * size up to the chunksize for allocation.
1870 */
1871 for (ksize = 0, i = 0; i < nkeys; i++)
1872 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1873
1874 /*
1875 * This should be pretty much impossible, but could happen if, say,
1876 * strange DIF specified the tuple. Ideally, this should be an
1877 * assertion and not an error condition -- but that requires that the
1878 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1879 * bullet-proof. (That is, it must not be able to be fooled by
1880 * malicious DIF.) Given the lack of backwards branches in DIF,
1881 * solving this would presumably not amount to solving the Halting
1882 * Problem -- but it still seems awfully hard.
1883 */
1884 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1885 ksize + dsize > chunksize) {
1886 dcpu->dtdsc_drops++;
1887 return (NULL);
1888 }
1889
1890 nstate = DTRACE_DSTATE_EMPTY;
1891
1892 do {
1893 retry:
1894 free = dcpu->dtdsc_free;
1895
1896 if (free == NULL) {
1897 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1898 void *rval;
1899
1900 if (clean == NULL) {
1901 /*
1902 * We're out of dynamic variable space on
1903 * this CPU. Unless we have tried all CPUs,
1904 * we'll try to allocate from a different
1905 * CPU.
1906 */
1907 switch (dstate->dtds_state) {
1908 case DTRACE_DSTATE_CLEAN: {
1909 void *sp = &dstate->dtds_state;
1910
1911 if (++cpu >= NCPU)
1912 cpu = 0;
1913
1914 if (dcpu->dtdsc_dirty != NULL &&
1915 nstate == DTRACE_DSTATE_EMPTY)
1916 nstate = DTRACE_DSTATE_DIRTY;
1917
1918 if (dcpu->dtdsc_rinsing != NULL)
1919 nstate = DTRACE_DSTATE_RINSING;
1920
1921 dcpu = &dstate->dtds_percpu[cpu];
1922
1923 if (cpu != me)
1924 goto retry;
1925
1926 (void) dtrace_cas32(sp,
1927 DTRACE_DSTATE_CLEAN, nstate);
1928
1929 /*
1930 * To increment the correct bean
1931 * counter, take another lap.
1932 */
1933 goto retry;
1934 }
1935
1936 case DTRACE_DSTATE_DIRTY:
1937 dcpu->dtdsc_dirty_drops++;
1938 break;
1939
1940 case DTRACE_DSTATE_RINSING:
1941 dcpu->dtdsc_rinsing_drops++;
1942 break;
1943
1944 case DTRACE_DSTATE_EMPTY:
1945 dcpu->dtdsc_drops++;
1946 break;
1947 }
1948
1949 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1950 return (NULL);
1951 }
1952
1953 /*
1954 * The clean list appears to be non-empty. We want to
1955 * move the clean list to the free list; we start by
1956 * moving the clean pointer aside.
1957 */
1958 if (dtrace_casptr(&dcpu->dtdsc_clean,
1959 clean, NULL) != clean) {
1960 /*
1961 * We are in one of two situations:
1962 *
1963 * (a) The clean list was switched to the
1964 * free list by another CPU.
1965 *
1966 * (b) The clean list was added to by the
1967 * cleansing cyclic.
1968 *
1969 * In either of these situations, we can
1970 * just reattempt the free list allocation.
1971 */
1972 goto retry;
1973 }
1974
1975 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1976
1977 /*
1978 * Now we'll move the clean list to our free list.
1979 * It's impossible for this to fail: the only way
1980 * the free list can be updated is through this
1981 * code path, and only one CPU can own the clean list.
1982 * Thus, it would only be possible for this to fail if
1983 * this code were racing with dtrace_dynvar_clean().
1984 * (That is, if dtrace_dynvar_clean() updated the clean
1985 * list, and we ended up racing to update the free
1986 * list.) This race is prevented by the dtrace_sync()
1987 * in dtrace_dynvar_clean() -- which flushes the
1988 * owners of the clean lists out before resetting
1989 * the clean lists.
1990 */
1991 dcpu = &dstate->dtds_percpu[me];
1992 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1993 ASSERT(rval == NULL);
1994 goto retry;
1995 }
1996
1997 dvar = free;
1998 new_free = dvar->dtdv_next;
1999 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2000
2001 /*
2002 * We have now allocated a new chunk. We copy the tuple keys into the
2003 * tuple array and copy any referenced key data into the data space
2004 * following the tuple array. As we do this, we relocate dttk_value
2005 * in the final tuple to point to the key data address in the chunk.
2006 */
2007 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2008 dvar->dtdv_data = (void *)(kdata + ksize);
2009 dvar->dtdv_tuple.dtt_nkeys = nkeys;
2010
2011 for (i = 0; i < nkeys; i++) {
2012 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2013 size_t kesize = key[i].dttk_size;
2014
2015 if (kesize != 0) {
2016 dtrace_bcopy(
2017 (const void *)(uintptr_t)key[i].dttk_value,
2018 (void *)kdata, kesize);
2019 dkey->dttk_value = kdata;
2020 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2021 } else {
2022 dkey->dttk_value = key[i].dttk_value;
2023 }
2024
2025 dkey->dttk_size = kesize;
2026 }
2027
2028 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2029 dvar->dtdv_hashval = hashval;
2030 dvar->dtdv_next = start;
2031
2032 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2033 return (dvar);
2034
2035 /*
2036 * The cas has failed. Either another CPU is adding an element to
2037 * this hash chain, or another CPU is deleting an element from this
2038 * hash chain. The simplest way to deal with both of these cases
2039 * (though not necessarily the most efficient) is to free our
2040 * allocated block and tail-call ourselves. Note that the free is
2041 * to the dirty list and _not_ to the free list. This is to prevent
2042 * races with allocators, above.
2043 */
2044 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2045
2046 dtrace_membar_producer();
2047
2048 do {
2049 free = dcpu->dtdsc_dirty;
2050 dvar->dtdv_next = free;
2051 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2052
2053 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
2054 }
2055
2056 /*ARGSUSED*/
2057 static void
2058 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2059 {
2060 if ((int64_t)nval < (int64_t)*oval)
2061 *oval = nval;
2062 }
2063
2064 /*ARGSUSED*/
2065 static void
2066 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2067 {
2068 if ((int64_t)nval > (int64_t)*oval)
2069 *oval = nval;
2070 }
2071
2072 static void
2073 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2074 {
2075 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2076 int64_t val = (int64_t)nval;
2077
2078 if (val < 0) {
2079 for (i = 0; i < zero; i++) {
2080 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2081 quanta[i] += incr;
2082 return;
2083 }
2084 }
2085 } else {
2086 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2087 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2088 quanta[i - 1] += incr;
2089 return;
2090 }
2091 }
2092
2093 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2094 return;
2095 }
2096
2097 ASSERT(0);
2098 }
2099
2100 static void
2101 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2102 {
2103 uint64_t arg = *lquanta++;
2104 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2105 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2106 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2107 int32_t val = (int32_t)nval, level;
2108
2109 ASSERT(step != 0);
2110 ASSERT(levels != 0);
2111
2112 if (val < base) {
2113 /*
2114 * This is an underflow.
2115 */
2116 lquanta[0] += incr;
2117 return;
2118 }
2119
2120 level = (val - base) / step;
2121
2122 if (level < levels) {
2123 lquanta[level + 1] += incr;
2124 return;
2125 }
2126
2127 /*
2128 * This is an overflow.
2129 */
2130 lquanta[levels + 1] += incr;
2131 }
2132
2133 static int
2134 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2135 uint16_t high, uint16_t nsteps, int64_t value)
2136 {
2137 int64_t this = 1, last, next;
2138 int base = 1, order;
2139
2140 ASSERT(factor <= nsteps);
2141 ASSERT(nsteps % factor == 0);
2142
2143 for (order = 0; order < low; order++)
2144 this *= factor;
2145
2146 /*
2147 * If our value is less than our factor taken to the power of the
2148 * low order of magnitude, it goes into the zeroth bucket.
2149 */
2150 if (value < (last = this))
2151 return (0);
2152
2153 for (this *= factor; order <= high; order++) {
2154 int nbuckets = this > nsteps ? nsteps : this;
2155
2156 if ((next = this * factor) < this) {
2157 /*
2158 * We should not generally get log/linear quantizations
2159 * with a high magnitude that allows 64-bits to
2160 * overflow, but we nonetheless protect against this
2161 * by explicitly checking for overflow, and clamping
2162 * our value accordingly.
2163 */
2164 value = this - 1;
2165 }
2166
2167 if (value < this) {
2168 /*
2169 * If our value lies within this order of magnitude,
2170 * determine its position by taking the offset within
2171 * the order of magnitude, dividing by the bucket
2172 * width, and adding to our (accumulated) base.
2173 */
2174 return (base + (value - last) / (this / nbuckets));
2175 }
2176
2177 base += nbuckets - (nbuckets / factor);
2178 last = this;
2179 this = next;
2180 }
2181
2182 /*
2183 * Our value is greater than or equal to our factor taken to the
2184 * power of one plus the high magnitude -- return the top bucket.
2185 */
2186 return (base);
2187 }
2188
2189 static void
2190 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2191 {
2192 uint64_t arg = *llquanta++;
2193 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2194 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2195 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2196 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2197
2198 llquanta[dtrace_aggregate_llquantize_bucket(factor,
2199 low, high, nsteps, nval)] += incr;
2200 }
2201
2202 /*ARGSUSED*/
2203 static void
2204 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2205 {
2206 data[0]++;
2207 data[1] += nval;
2208 }
2209
2210 /*ARGSUSED*/
2211 static void
2212 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2213 {
2214 int64_t snval = (int64_t)nval;
2215 uint64_t tmp[2];
2216
2217 data[0]++;
2218 data[1] += nval;
2219
2220 /*
2221 * What we want to say here is:
2222 *
2223 * data[2] += nval * nval;
2224 *
2225 * But given that nval is 64-bit, we could easily overflow, so
2226 * we do this as 128-bit arithmetic.
2227 */
2228 if (snval < 0)
2229 snval = -snval;
2230
2231 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2232 dtrace_add_128(data + 2, tmp, data + 2);
2233 }
2234
2235 /*ARGSUSED*/
2236 static void
2237 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2238 {
2239 *oval = *oval + 1;
2240 }
2241
2242 /*ARGSUSED*/
2243 static void
2244 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2245 {
2246 *oval += nval;
2247 }
2248
2249 /*
2250 * Aggregate given the tuple in the principal data buffer, and the aggregating
2251 * action denoted by the specified dtrace_aggregation_t. The aggregation
2252 * buffer is specified as the buf parameter. This routine does not return
2253 * failure; if there is no space in the aggregation buffer, the data will be
2254 * dropped, and a corresponding counter incremented.
2255 */
2256 static void
2257 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2258 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2259 {
2260 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2261 uint32_t i, ndx, size, fsize;
2262 uint32_t align = sizeof (uint64_t) - 1;
2263 dtrace_aggbuffer_t *agb;
2264 dtrace_aggkey_t *key;
2265 uint32_t hashval = 0, limit, isstr;
2266 caddr_t tomax, data, kdata;
2267 dtrace_actkind_t action;
2268 dtrace_action_t *act;
2269 uintptr_t offs;
2270
2271 if (buf == NULL)
2272 return;
2273
2274 if (!agg->dtag_hasarg) {
2275 /*
2276 * Currently, only quantize() and lquantize() take additional
2277 * arguments, and they have the same semantics: an increment
2278 * value that defaults to 1 when not present. If additional
2279 * aggregating actions take arguments, the setting of the
2280 * default argument value will presumably have to become more
2281 * sophisticated...
2282 */
2283 arg = 1;
2284 }
2285
2286 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2287 size = rec->dtrd_offset - agg->dtag_base;
2288 fsize = size + rec->dtrd_size;
2289
2290 ASSERT(dbuf->dtb_tomax != NULL);
2291 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2292
2293 if ((tomax = buf->dtb_tomax) == NULL) {
2294 dtrace_buffer_drop(buf);
2295 return;
2296 }
2297
2298 /*
2299 * The metastructure is always at the bottom of the buffer.
2300 */
2301 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2302 sizeof (dtrace_aggbuffer_t));
2303
2304 if (buf->dtb_offset == 0) {
2305 /*
2306 * We just kludge up approximately 1/8th of the size to be
2307 * buckets. If this guess ends up being routinely
2308 * off-the-mark, we may need to dynamically readjust this
2309 * based on past performance.
2310 */
2311 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2312
2313 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2314 (uintptr_t)tomax || hashsize == 0) {
2315 /*
2316 * We've been given a ludicrously small buffer;
2317 * increment our drop count and leave.
2318 */
2319 dtrace_buffer_drop(buf);
2320 return;
2321 }
2322
2323 /*
2324 * And now, a pathetic attempt to try to get a an odd (or
2325 * perchance, a prime) hash size for better hash distribution.
2326 */
2327 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2328 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2329
2330 agb->dtagb_hashsize = hashsize;
2331 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2332 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2333 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2334
2335 for (i = 0; i < agb->dtagb_hashsize; i++)
2336 agb->dtagb_hash[i] = NULL;
2337 }
2338
2339 ASSERT(agg->dtag_first != NULL);
2340 ASSERT(agg->dtag_first->dta_intuple);
2341
2342 /*
2343 * Calculate the hash value based on the key. Note that we _don't_
2344 * include the aggid in the hashing (but we will store it as part of
2345 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2346 * algorithm: a simple, quick algorithm that has no known funnels, and
2347 * gets good distribution in practice. The efficacy of the hashing
2348 * algorithm (and a comparison with other algorithms) may be found by
2349 * running the ::dtrace_aggstat MDB dcmd.
2350 */
2351 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2352 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2353 limit = i + act->dta_rec.dtrd_size;
2354 ASSERT(limit <= size);
2355 isstr = DTRACEACT_ISSTRING(act);
2356
2357 for (; i < limit; i++) {
2358 hashval += data[i];
2359 hashval += (hashval << 10);
2360 hashval ^= (hashval >> 6);
2361
2362 if (isstr && data[i] == '\0')
2363 break;
2364 }
2365 }
2366
2367 hashval += (hashval << 3);
2368 hashval ^= (hashval >> 11);
2369 hashval += (hashval << 15);
2370
2371 /*
2372 * Yes, the divide here is expensive -- but it's generally the least
2373 * of the performance issues given the amount of data that we iterate
2374 * over to compute hash values, compare data, etc.
2375 */
2376 ndx = hashval % agb->dtagb_hashsize;
2377
2378 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2379 ASSERT((caddr_t)key >= tomax);
2380 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2381
2382 if (hashval != key->dtak_hashval || key->dtak_size != size)
2383 continue;
2384
2385 kdata = key->dtak_data;
2386 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2387
2388 for (act = agg->dtag_first; act->dta_intuple;
2389 act = act->dta_next) {
2390 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2391 limit = i + act->dta_rec.dtrd_size;
2392 ASSERT(limit <= size);
2393 isstr = DTRACEACT_ISSTRING(act);
2394
2395 for (; i < limit; i++) {
2396 if (kdata[i] != data[i])
2397 goto next;
2398
2399 if (isstr && data[i] == '\0')
2400 break;
2401 }
2402 }
2403
2404 if (action != key->dtak_action) {
2405 /*
2406 * We are aggregating on the same value in the same
2407 * aggregation with two different aggregating actions.
2408 * (This should have been picked up in the compiler,
2409 * so we may be dealing with errant or devious DIF.)
2410 * This is an error condition; we indicate as much,
2411 * and return.
2412 */
2413 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2414 return;
2415 }
2416
2417 /*
2418 * This is a hit: we need to apply the aggregator to
2419 * the value at this key.
2420 */
2421 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2422 return;
2423 next:
2424 continue;
2425 }
2426
2427 /*
2428 * We didn't find it. We need to allocate some zero-filled space,
2429 * link it into the hash table appropriately, and apply the aggregator
2430 * to the (zero-filled) value.
2431 */
2432 offs = buf->dtb_offset;
2433 while (offs & (align - 1))
2434 offs += sizeof (uint32_t);
2435
2436 /*
2437 * If we don't have enough room to both allocate a new key _and_
2438 * its associated data, increment the drop count and return.
2439 */
2440 if ((uintptr_t)tomax + offs + fsize >
2441 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2442 dtrace_buffer_drop(buf);
2443 return;
2444 }
2445
2446 /*CONSTCOND*/
2447 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2448 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2449 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2450
2451 key->dtak_data = kdata = tomax + offs;
2452 buf->dtb_offset = offs + fsize;
2453
2454 /*
2455 * Now copy the data across.
2456 */
2457 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2458
2459 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2460 kdata[i] = data[i];
2461
2462 /*
2463 * Because strings are not zeroed out by default, we need to iterate
2464 * looking for actions that store strings, and we need to explicitly
2465 * pad these strings out with zeroes.
2466 */
2467 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2468 int nul;
2469
2470 if (!DTRACEACT_ISSTRING(act))
2471 continue;
2472
2473 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2474 limit = i + act->dta_rec.dtrd_size;
2475 ASSERT(limit <= size);
2476
2477 for (nul = 0; i < limit; i++) {
2478 if (nul) {
2479 kdata[i] = '\0';
2480 continue;
2481 }
2482
2483 if (data[i] != '\0')
2484 continue;
2485
2486 nul = 1;
2487 }
2488 }
2489
2490 for (i = size; i < fsize; i++)
2491 kdata[i] = 0;
2492
2493 key->dtak_hashval = hashval;
2494 key->dtak_size = size;
2495 key->dtak_action = action;
2496 key->dtak_next = agb->dtagb_hash[ndx];
2497 agb->dtagb_hash[ndx] = key;
2498
2499 /*
2500 * Finally, apply the aggregator.
2501 */
2502 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2503 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2504 }
2505
2506 /*
2507 * Given consumer state, this routine finds a speculation in the INACTIVE
2508 * state and transitions it into the ACTIVE state. If there is no speculation
2509 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2510 * incremented -- it is up to the caller to take appropriate action.
2511 */
2512 static int
2513 dtrace_speculation(dtrace_state_t *state)
2514 {
2515 int i = 0;
2516 dtrace_speculation_state_t current;
2517 uint32_t *stat = &state->dts_speculations_unavail, count;
2518
2519 while (i < state->dts_nspeculations) {
2520 dtrace_speculation_t *spec = &state->dts_speculations[i];
2521
2522 current = spec->dtsp_state;
2523
2524 if (current != DTRACESPEC_INACTIVE) {
2525 if (current == DTRACESPEC_COMMITTINGMANY ||
2526 current == DTRACESPEC_COMMITTING ||
2527 current == DTRACESPEC_DISCARDING)
2528 stat = &state->dts_speculations_busy;
2529 i++;
2530 continue;
2531 }
2532
2533 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2534 current, DTRACESPEC_ACTIVE) == current)
2535 return (i + 1);
2536 }
2537
2538 /*
2539 * We couldn't find a speculation. If we found as much as a single
2540 * busy speculation buffer, we'll attribute this failure as "busy"
2541 * instead of "unavail".
2542 */
2543 do {
2544 count = *stat;
2545 } while (dtrace_cas32(stat, count, count + 1) != count);
2546
2547 return (0);
2548 }
2549
2550 /*
2551 * This routine commits an active speculation. If the specified speculation
2552 * is not in a valid state to perform a commit(), this routine will silently do
2553 * nothing. The state of the specified speculation is transitioned according
2554 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2555 */
2556 static void
2557 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2558 dtrace_specid_t which)
2559 {
2560 dtrace_speculation_t *spec;
2561 dtrace_buffer_t *src, *dest;
2562 uintptr_t daddr, saddr, dlimit, slimit;
2563 dtrace_speculation_state_t current, new;
2564 intptr_t offs;
2565 uint64_t timestamp;
2566
2567 if (which == 0)
2568 return;
2569
2570 if (which > state->dts_nspeculations) {
2571 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2572 return;
2573 }
2574
2575 spec = &state->dts_speculations[which - 1];
2576 src = &spec->dtsp_buffer[cpu];
2577 dest = &state->dts_buffer[cpu];
2578
2579 do {
2580 current = spec->dtsp_state;
2581
2582 if (current == DTRACESPEC_COMMITTINGMANY)
2583 break;
2584
2585 switch (current) {
2586 case DTRACESPEC_INACTIVE:
2587 case DTRACESPEC_DISCARDING:
2588 return;
2589
2590 case DTRACESPEC_COMMITTING:
2591 /*
2592 * This is only possible if we are (a) commit()'ing
2593 * without having done a prior speculate() on this CPU
2594 * and (b) racing with another commit() on a different
2595 * CPU. There's nothing to do -- we just assert that
2596 * our offset is 0.
2597 */
2598 ASSERT(src->dtb_offset == 0);
2599 return;
2600
2601 case DTRACESPEC_ACTIVE:
2602 new = DTRACESPEC_COMMITTING;
2603 break;
2604
2605 case DTRACESPEC_ACTIVEONE:
2606 /*
2607 * This speculation is active on one CPU. If our
2608 * buffer offset is non-zero, we know that the one CPU
2609 * must be us. Otherwise, we are committing on a
2610 * different CPU from the speculate(), and we must
2611 * rely on being asynchronously cleaned.
2612 */
2613 if (src->dtb_offset != 0) {
2614 new = DTRACESPEC_COMMITTING;
2615 break;
2616 }
2617 /*FALLTHROUGH*/
2618
2619 case DTRACESPEC_ACTIVEMANY:
2620 new = DTRACESPEC_COMMITTINGMANY;
2621 break;
2622
2623 default:
2624 ASSERT(0);
2625 }
2626 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2627 current, new) != current);
2628
2629 /*
2630 * We have set the state to indicate that we are committing this
2631 * speculation. Now reserve the necessary space in the destination
2632 * buffer.
2633 */
2634 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2635 sizeof (uint64_t), state, NULL)) < 0) {
2636 dtrace_buffer_drop(dest);
2637 goto out;
2638 }
2639
2640 /*
2641 * We have sufficient space to copy the speculative buffer into the
2642 * primary buffer. First, modify the speculative buffer, filling
2643 * in the timestamp of all entries with the current time. The data
2644 * must have the commit() time rather than the time it was traced,
2645 * so that all entries in the primary buffer are in timestamp order.
2646 */
2647 timestamp = dtrace_gethrtime();
2648 saddr = (uintptr_t)src->dtb_tomax;
2649 slimit = saddr + src->dtb_offset;
2650 while (saddr < slimit) {
2651 size_t size;
2652 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2653
2654 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2655 saddr += sizeof (dtrace_epid_t);
2656 continue;
2657 }
2658 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2659 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2660
2661 ASSERT3U(saddr + size, <=, slimit);
2662 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2663 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2664
2665 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2666
2667 saddr += size;
2668 }
2669
2670 /*
2671 * Copy the buffer across. (Note that this is a
2672 * highly subobtimal bcopy(); in the unlikely event that this becomes
2673 * a serious performance issue, a high-performance DTrace-specific
2674 * bcopy() should obviously be invented.)
2675 */
2676 daddr = (uintptr_t)dest->dtb_tomax + offs;
2677 dlimit = daddr + src->dtb_offset;
2678 saddr = (uintptr_t)src->dtb_tomax;
2679
2680 /*
2681 * First, the aligned portion.
2682 */
2683 while (dlimit - daddr >= sizeof (uint64_t)) {
2684 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2685
2686 daddr += sizeof (uint64_t);
2687 saddr += sizeof (uint64_t);
2688 }
2689
2690 /*
2691 * Now any left-over bit...
2692 */
2693 while (dlimit - daddr)
2694 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2695
2696 /*
2697 * Finally, commit the reserved space in the destination buffer.
2698 */
2699 dest->dtb_offset = offs + src->dtb_offset;
2700
2701 out:
2702 /*
2703 * If we're lucky enough to be the only active CPU on this speculation
2704 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2705 */
2706 if (current == DTRACESPEC_ACTIVE ||
2707 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2708 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2709 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2710
2711 ASSERT(rval == DTRACESPEC_COMMITTING);
2712 }
2713
2714 src->dtb_offset = 0;
2715 src->dtb_xamot_drops += src->dtb_drops;
2716 src->dtb_drops = 0;
2717 }
2718
2719 /*
2720 * This routine discards an active speculation. If the specified speculation
2721 * is not in a valid state to perform a discard(), this routine will silently
2722 * do nothing. The state of the specified speculation is transitioned
2723 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2724 */
2725 static void
2726 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2727 dtrace_specid_t which)
2728 {
2729 dtrace_speculation_t *spec;
2730 dtrace_speculation_state_t current, new;
2731 dtrace_buffer_t *buf;
2732
2733 if (which == 0)
2734 return;
2735
2736 if (which > state->dts_nspeculations) {
2737 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2738 return;
2739 }
2740
2741 spec = &state->dts_speculations[which - 1];
2742 buf = &spec->dtsp_buffer[cpu];
2743
2744 do {
2745 current = spec->dtsp_state;
2746
2747 switch (current) {
2748 case DTRACESPEC_INACTIVE:
2749 case DTRACESPEC_COMMITTINGMANY:
2750 case DTRACESPEC_COMMITTING:
2751 case DTRACESPEC_DISCARDING:
2752 return;
2753
2754 case DTRACESPEC_ACTIVE:
2755 case DTRACESPEC_ACTIVEMANY:
2756 new = DTRACESPEC_DISCARDING;
2757 break;
2758
2759 case DTRACESPEC_ACTIVEONE:
2760 if (buf->dtb_offset != 0) {
2761 new = DTRACESPEC_INACTIVE;
2762 } else {
2763 new = DTRACESPEC_DISCARDING;
2764 }
2765 break;
2766
2767 default:
2768 ASSERT(0);
2769 }
2770 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2771 current, new) != current);
2772
2773 buf->dtb_offset = 0;
2774 buf->dtb_drops = 0;
2775 }
2776
2777 /*
2778 * Note: not called from probe context. This function is called
2779 * asynchronously from cross call context to clean any speculations that are
2780 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2781 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2782 * speculation.
2783 */
2784 static void
2785 dtrace_speculation_clean_here(dtrace_state_t *state)
2786 {
2787 dtrace_icookie_t cookie;
2788 processorid_t cpu = CPU->cpu_id;
2789 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2790 dtrace_specid_t i;
2791
2792 cookie = dtrace_interrupt_disable();
2793
2794 if (dest->dtb_tomax == NULL) {
2795 dtrace_interrupt_enable(cookie);
2796 return;
2797 }
2798
2799 for (i = 0; i < state->dts_nspeculations; i++) {
2800 dtrace_speculation_t *spec = &state->dts_speculations[i];
2801 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2802
2803 if (src->dtb_tomax == NULL)
2804 continue;
2805
2806 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2807 src->dtb_offset = 0;
2808 continue;
2809 }
2810
2811 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2812 continue;
2813
2814 if (src->dtb_offset == 0)
2815 continue;
2816
2817 dtrace_speculation_commit(state, cpu, i + 1);
2818 }
2819
2820 dtrace_interrupt_enable(cookie);
2821 }
2822
2823 /*
2824 * Note: not called from probe context. This function is called
2825 * asynchronously (and at a regular interval) to clean any speculations that
2826 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2827 * is work to be done, it cross calls all CPUs to perform that work;
2828 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2829 * INACTIVE state until they have been cleaned by all CPUs.
2830 */
2831 static void
2832 dtrace_speculation_clean(dtrace_state_t *state)
2833 {
2834 int work = 0, rv;
2835 dtrace_specid_t i;
2836
2837 for (i = 0; i < state->dts_nspeculations; i++) {
2838 dtrace_speculation_t *spec = &state->dts_speculations[i];
2839
2840 ASSERT(!spec->dtsp_cleaning);
2841
2842 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2843 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2844 continue;
2845
2846 work++;
2847 spec->dtsp_cleaning = 1;
2848 }
2849
2850 if (!work)
2851 return;
2852
2853 dtrace_xcall(DTRACE_CPUALL,
2854 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2855
2856 /*
2857 * We now know that all CPUs have committed or discarded their
2858 * speculation buffers, as appropriate. We can now set the state
2859 * to inactive.
2860 */
2861 for (i = 0; i < state->dts_nspeculations; i++) {
2862 dtrace_speculation_t *spec = &state->dts_speculations[i];
2863 dtrace_speculation_state_t current, new;
2864
2865 if (!spec->dtsp_cleaning)
2866 continue;
2867
2868 current = spec->dtsp_state;
2869 ASSERT(current == DTRACESPEC_DISCARDING ||
2870 current == DTRACESPEC_COMMITTINGMANY);
2871
2872 new = DTRACESPEC_INACTIVE;
2873
2874 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2875 ASSERT(rv == current);
2876 spec->dtsp_cleaning = 0;
2877 }
2878 }
2879
2880 /*
2881 * Called as part of a speculate() to get the speculative buffer associated
2882 * with a given speculation. Returns NULL if the specified speculation is not
2883 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2884 * the active CPU is not the specified CPU -- the speculation will be
2885 * atomically transitioned into the ACTIVEMANY state.
2886 */
2887 static dtrace_buffer_t *
2888 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2889 dtrace_specid_t which)
2890 {
2891 dtrace_speculation_t *spec;
2892 dtrace_speculation_state_t current, new;
2893 dtrace_buffer_t *buf;
2894
2895 if (which == 0)
2896 return (NULL);
2897
2898 if (which > state->dts_nspeculations) {
2899 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2900 return (NULL);
2901 }
2902
2903 spec = &state->dts_speculations[which - 1];
2904 buf = &spec->dtsp_buffer[cpuid];
2905
2906 do {
2907 current = spec->dtsp_state;
2908
2909 switch (current) {
2910 case DTRACESPEC_INACTIVE:
2911 case DTRACESPEC_COMMITTINGMANY:
2912 case DTRACESPEC_DISCARDING:
2913 return (NULL);
2914
2915 case DTRACESPEC_COMMITTING:
2916 ASSERT(buf->dtb_offset == 0);
2917 return (NULL);
2918
2919 case DTRACESPEC_ACTIVEONE:
2920 /*
2921 * This speculation is currently active on one CPU.
2922 * Check the offset in the buffer; if it's non-zero,
2923 * that CPU must be us (and we leave the state alone).
2924 * If it's zero, assume that we're starting on a new
2925 * CPU -- and change the state to indicate that the
2926 * speculation is active on more than one CPU.
2927 */
2928 if (buf->dtb_offset != 0)
2929 return (buf);
2930
2931 new = DTRACESPEC_ACTIVEMANY;
2932 break;
2933
2934 case DTRACESPEC_ACTIVEMANY:
2935 return (buf);
2936
2937 case DTRACESPEC_ACTIVE:
2938 new = DTRACESPEC_ACTIVEONE;
2939 break;
2940
2941 default:
2942 ASSERT(0);
2943 }
2944 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2945 current, new) != current);
2946
2947 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2948 return (buf);
2949 }
2950
2951 /*
2952 * Return a string. In the event that the user lacks the privilege to access
2953 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2954 * don't fail access checking.
2955 *
2956 * dtrace_dif_variable() uses this routine as a helper for various
2957 * builtin values such as 'execname' and 'probefunc.'
2958 */
2959 uintptr_t
2960 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2961 dtrace_mstate_t *mstate)
2962 {
2963 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2964 uintptr_t ret;
2965 size_t strsz;
2966
2967 /*
2968 * The easy case: this probe is allowed to read all of memory, so
2969 * we can just return this as a vanilla pointer.
2970 */
2971 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2972 return (addr);
2973
2974 /*
2975 * This is the tougher case: we copy the string in question from
2976 * kernel memory into scratch memory and return it that way: this
2977 * ensures that we won't trip up when access checking tests the
2978 * BYREF return value.
2979 */
2980 strsz = dtrace_strlen((char *)addr, size) + 1;
2981
2982 if (mstate->dtms_scratch_ptr + strsz >
2983 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2984 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2985 return (NULL);
2986 }
2987
2988 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2989 strsz);
2990 ret = mstate->dtms_scratch_ptr;
2991 mstate->dtms_scratch_ptr += strsz;
2992 return (ret);
2993 }
2994
2995 /*
2996 * This function implements the DIF emulator's variable lookups. The emulator
2997 * passes a reserved variable identifier and optional built-in array index.
2998 */
2999 static uint64_t
3000 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3001 uint64_t ndx)
3002 {
3003 /*
3004 * If we're accessing one of the uncached arguments, we'll turn this
3005 * into a reference in the args array.
3006 */
3007 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3008 ndx = v - DIF_VAR_ARG0;
3009 v = DIF_VAR_ARGS;
3010 }
3011
3012 switch (v) {
3013 case DIF_VAR_ARGS:
3014 if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
3015 cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
3016 CPU_DTRACE_KPRIV;
3017 return (0);
3018 }
3019
3020 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3021 if (ndx >= sizeof (mstate->dtms_arg) /
3022 sizeof (mstate->dtms_arg[0])) {
3023 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3024 dtrace_provider_t *pv;
3025 uint64_t val;
3026
3027 pv = mstate->dtms_probe->dtpr_provider;
3028 if (pv->dtpv_pops.dtps_getargval != NULL)
3029 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3030 mstate->dtms_probe->dtpr_id,
3031 mstate->dtms_probe->dtpr_arg, ndx, aframes);
3032 else
3033 val = dtrace_getarg(ndx, aframes);
3034
3035 /*
3036 * This is regrettably required to keep the compiler
3037 * from tail-optimizing the call to dtrace_getarg().
3038 * The condition always evaluates to true, but the
3039 * compiler has no way of figuring that out a priori.
3040 * (None of this would be necessary if the compiler
3041 * could be relied upon to _always_ tail-optimize
3042 * the call to dtrace_getarg() -- but it can't.)
3043 */
3044 if (mstate->dtms_probe != NULL)
3045 return (val);
3046
3047 ASSERT(0);
3048 }
3049
3050 return (mstate->dtms_arg[ndx]);
3051
3052 case DIF_VAR_UREGS: {
3053 klwp_t *lwp;
3054
3055 if (!dtrace_priv_proc(state, mstate))
3056 return (0);
3057
3058 if ((lwp = curthread->t_lwp) == NULL) {
3059 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3060 cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
3061 return (0);
3062 }
3063
3064 return (dtrace_getreg(lwp->lwp_regs, ndx));
3065 }
3066
3067 case DIF_VAR_VMREGS: {
3068 uint64_t rval;
3069
3070 if (!dtrace_priv_kernel(state))
3071 return (0);
3072
3073 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3074
3075 rval = dtrace_getvmreg(ndx,
3076 &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
3077
3078 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3079
3080 return (rval);
3081 }
3082
3083 case DIF_VAR_CURTHREAD:
3084 if (!dtrace_priv_proc(state, mstate))
3085 return (0);
3086 return ((uint64_t)(uintptr_t)curthread);
3087
3088 case DIF_VAR_TIMESTAMP:
3089 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3090 mstate->dtms_timestamp = dtrace_gethrtime();
3091 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3092 }
3093 return (mstate->dtms_timestamp);
3094
3095 case DIF_VAR_VTIMESTAMP:
3096 ASSERT(dtrace_vtime_references != 0);
3097 return (curthread->t_dtrace_vtime);
3098
3099 case DIF_VAR_WALLTIMESTAMP:
3100 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3101 mstate->dtms_walltimestamp = dtrace_gethrestime();
3102 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3103 }
3104 return (mstate->dtms_walltimestamp);
3105
3106 case DIF_VAR_IPL:
3107 if (!dtrace_priv_kernel(state))
3108 return (0);
3109 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3110 mstate->dtms_ipl = dtrace_getipl();
3111 mstate->dtms_present |= DTRACE_MSTATE_IPL;
3112 }
3113 return (mstate->dtms_ipl);
3114
3115 case DIF_VAR_EPID:
3116 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3117 return (mstate->dtms_epid);
3118
3119 case DIF_VAR_ID:
3120 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3121 return (mstate->dtms_probe->dtpr_id);
3122
3123 case DIF_VAR_STACKDEPTH:
3124 if (!dtrace_priv_kernel(state))
3125 return (0);
3126 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3127 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3128
3129 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3130 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3131 }
3132 return (mstate->dtms_stackdepth);
3133
3134 case DIF_VAR_USTACKDEPTH:
3135 if (!dtrace_priv_proc(state, mstate))
3136 return (0);
3137 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3138 /*
3139 * See comment in DIF_VAR_PID.
3140 */
3141 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3142 CPU_ON_INTR(CPU)) {
3143 mstate->dtms_ustackdepth = 0;
3144 } else {
3145 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3146 mstate->dtms_ustackdepth =
3147 dtrace_getustackdepth();
3148 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3149 }
3150 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3151 }
3152 return (mstate->dtms_ustackdepth);
3153
3154 case DIF_VAR_CALLER:
3155 if (!dtrace_priv_kernel(state))
3156 return (0);
3157 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3158 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3159
3160 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3161 /*
3162 * If this is an unanchored probe, we are
3163 * required to go through the slow path:
3164 * dtrace_caller() only guarantees correct
3165 * results for anchored probes.
3166 */
3167 pc_t caller[2];
3168
3169 dtrace_getpcstack(caller, 2, aframes,
3170 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3171 mstate->dtms_caller = caller[1];
3172 } else if ((mstate->dtms_caller =
3173 dtrace_caller(aframes)) == -1) {
3174 /*
3175 * We have failed to do this the quick way;
3176 * we must resort to the slower approach of
3177 * calling dtrace_getpcstack().
3178 */
3179 pc_t caller;
3180
3181 dtrace_getpcstack(&caller, 1, aframes, NULL);
3182 mstate->dtms_caller = caller;
3183 }
3184
3185 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3186 }
3187 return (mstate->dtms_caller);
3188
3189 case DIF_VAR_UCALLER:
3190 if (!dtrace_priv_proc(state, mstate))
3191 return (0);
3192
3193 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3194 uint64_t ustack[3];
3195
3196 /*
3197 * dtrace_getupcstack() fills in the first uint64_t
3198 * with the current PID. The second uint64_t will
3199 * be the program counter at user-level. The third
3200 * uint64_t will contain the caller, which is what
3201 * we're after.
3202 */
3203 ustack[2] = NULL;
3204 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3205 dtrace_getupcstack(ustack, 3);
3206 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3207 mstate->dtms_ucaller = ustack[2];
3208 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3209 }
3210
3211 return (mstate->dtms_ucaller);
3212
3213 case DIF_VAR_PROBEPROV:
3214 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3215 return (dtrace_dif_varstr(
3216 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3217 state, mstate));
3218
3219 case DIF_VAR_PROBEMOD:
3220 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3221 return (dtrace_dif_varstr(
3222 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3223 state, mstate));
3224
3225 case DIF_VAR_PROBEFUNC:
3226 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3227 return (dtrace_dif_varstr(
3228 (uintptr_t)mstate->dtms_probe->dtpr_func,
3229 state, mstate));
3230
3231 case DIF_VAR_PROBENAME:
3232 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3233 return (dtrace_dif_varstr(
3234 (uintptr_t)mstate->dtms_probe->dtpr_name,
3235 state, mstate));
3236
3237 case DIF_VAR_PID:
3238 if (!dtrace_priv_proc(state, mstate))
3239 return (0);
3240
3241 /*
3242 * Note that we are assuming that an unanchored probe is
3243 * always due to a high-level interrupt. (And we're assuming
3244 * that there is only a single high level interrupt.)
3245 */
3246 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3247 return (pid0.pid_id);
3248
3249 /*
3250 * It is always safe to dereference one's own t_procp pointer:
3251 * it always points to a valid, allocated proc structure.
3252 * Further, it is always safe to dereference the p_pidp member
3253 * of one's own proc structure. (These are truisms becuase
3254 * threads and processes don't clean up their own state --
3255 * they leave that task to whomever reaps them.)
3256 */
3257 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3258
3259 case DIF_VAR_PPID:
3260 if (!dtrace_priv_proc(state, mstate))
3261 return (0);
3262
3263 /*
3264 * See comment in DIF_VAR_PID.
3265 */
3266 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3267 return (pid0.pid_id);
3268
3269 /*
3270 * It is always safe to dereference one's own t_procp pointer:
3271 * it always points to a valid, allocated proc structure.
3272 * (This is true because threads don't clean up their own
3273 * state -- they leave that task to whomever reaps them.)
3274 */
3275 return ((uint64_t)curthread->t_procp->p_ppid);
3276
3277 case DIF_VAR_TID:
3278 /*
3279 * See comment in DIF_VAR_PID.
3280 */
3281 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3282 return (0);
3283
3284 return ((uint64_t)curthread->t_tid);
3285
3286 case DIF_VAR_EXECNAME:
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)(uintptr_t)p0.p_user.u_comm);
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 return (dtrace_dif_varstr(
3303 (uintptr_t)curthread->t_procp->p_user.u_comm,
3304 state, mstate));
3305
3306 case DIF_VAR_ZONENAME:
3307 if (!dtrace_priv_proc(state, mstate))
3308 return (0);
3309
3310 /*
3311 * See comment in DIF_VAR_PID.
3312 */
3313 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3314 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3315
3316 /*
3317 * It is always safe to dereference one's own t_procp pointer:
3318 * it always points to a valid, allocated proc structure.
3319 * (This is true because threads don't clean up their own
3320 * state -- they leave that task to whomever reaps them.)
3321 */
3322 return (dtrace_dif_varstr(
3323 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3324 state, mstate));
3325
3326 case DIF_VAR_UID:
3327 if (!dtrace_priv_proc(state, mstate))
3328 return (0);
3329
3330 /*
3331 * See comment in DIF_VAR_PID.
3332 */
3333 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3334 return ((uint64_t)p0.p_cred->cr_uid);
3335
3336 /*
3337 * It is always safe to dereference one's own t_procp pointer:
3338 * it always points to a valid, allocated proc structure.
3339 * (This is true because threads don't clean up their own
3340 * state -- they leave that task to whomever reaps them.)
3341 *
3342 * Additionally, it is safe to dereference one's own process
3343 * credential, since this is never NULL after process birth.
3344 */
3345 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3346
3347 case DIF_VAR_GID:
3348 if (!dtrace_priv_proc(state, mstate))
3349 return (0);
3350
3351 /*
3352 * See comment in DIF_VAR_PID.
3353 */
3354 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3355 return ((uint64_t)p0.p_cred->cr_gid);
3356
3357 /*
3358 * It is always safe to dereference one's own t_procp pointer:
3359 * it always points to a valid, allocated proc structure.
3360 * (This is true because threads don't clean up their own
3361 * state -- they leave that task to whomever reaps them.)
3362 *
3363 * Additionally, it is safe to dereference one's own process
3364 * credential, since this is never NULL after process birth.
3365 */
3366 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3367
3368 case DIF_VAR_ERRNO: {
3369 klwp_t *lwp;
3370 if (!dtrace_priv_proc(state, mstate))
3371 return (0);
3372
3373 /*
3374 * See comment in DIF_VAR_PID.
3375 */
3376 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3377 return (0);
3378
3379 /*
3380 * It is always safe to dereference one's own t_lwp pointer in
3381 * the event that this pointer is non-NULL. (This is true
3382 * because threads and lwps don't clean up their own state --
3383 * they leave that task to whomever reaps them.)
3384 */
3385 if ((lwp = curthread->t_lwp) == NULL)
3386 return (0);
3387
3388 return ((uint64_t)lwp->lwp_errno);
3389 }
3390 default:
3391 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3392 return (0);
3393 }
3394 }
3395
3396
3397 typedef enum json_state {
3398 JSON_REST = 1,
3399 JSON_OBJECT,
3400 JSON_STRING,
3401 JSON_STRING_ESCAPE,
3402 JSON_STRING_ESCAPE_UNICODE,
3403 JSON_COLON,
3404 JSON_COMMA,
3405 JSON_VALUE,
3406 JSON_IDENTIFIER,
3407 JSON_NUMBER,
3408 JSON_NUMBER_FRAC,
3409 JSON_NUMBER_EXP,
3410 JSON_COLLECT_OBJECT
3411 } json_state_t;
3412
3413 /*
3414 * This function possesses just enough knowledge about JSON to extract a single
3415 * value from a JSON string and store it in the scratch buffer. It is able
3416 * to extract nested object values, and members of arrays by index.
3417 *
3418 * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3419 * be looked up as we descend into the object tree. e.g.
3420 *
3421 * foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3422 * with nelems = 5.
3423 */
3424 static char *
3425 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3426 char *dest)
3427 {
3428 json_state_t state = JSON_REST;
3429 uint64_t i;
3430 int64_t array_elem = INT64_MIN;
3431 int64_t array_pos = 0;
3432 uint8_t escape_unicount = 0;
3433 boolean_t string_is_key = B_FALSE;
3434 boolean_t collect_object = B_FALSE;
3435 boolean_t found_key = B_FALSE;
3436 boolean_t in_array = B_FALSE;
3437 uint8_t braces = 0, brackets = 0;
3438 char *elem = elemlist;
3439 char *dd = dest;
3440 uintptr_t cur;
3441
3442 for (cur = json; cur < json + size; cur++) {
3443 char cc = dtrace_load8(cur);
3444 if (cc == '\0' || braces > 250)
3445 return (NULL);
3446
3447 switch (state) {
3448 case JSON_REST:
3449 if (cc == ' ' || cc == '\t' || cc == '\n' || cc == '\r')
3450 break; /* eat whitespace */
3451
3452 if (cc == '{') {
3453 state = JSON_OBJECT;
3454 break;
3455 }
3456
3457 if (cc == '[') {
3458 in_array = B_TRUE;
3459 array_pos = 0;
3460 array_elem = dtrace_strtoll(elem, 10, size);
3461 found_key = !!(array_elem == 0);
3462 state = JSON_VALUE;
3463 break;
3464 }
3465
3466 /* ERROR: expected object or array */
3467 return (NULL);
3468 case JSON_OBJECT:
3469 if (cc == ' ' || cc == '\t' || cc == '\n' || cc == '\r')
3470 break; /* eat whitespace */
3471
3472 if (cc == '"') {
3473 state = JSON_STRING;
3474 string_is_key = B_TRUE;
3475 break;
3476 }
3477
3478 /* ERROR: key not found! */
3479 return (NULL);
3480 case JSON_STRING:
3481 if (cc == '\\') {
3482 *dd++ = '\\';
3483 state = JSON_STRING_ESCAPE;
3484 break;
3485 }
3486
3487 if (cc == '"') {
3488 if (collect_object) {
3489 /*
3490 * We don't reset the dest here, as
3491 * the string is part of a larger
3492 * object being collected.
3493 */
3494 *dd++ = cc;
3495 collect_object = B_FALSE;
3496 state = JSON_COLLECT_OBJECT;
3497 break;
3498 }
3499 *dd = '\0';
3500 dd = dest; /* reset string buffer */
3501 if (string_is_key) {
3502 if (dtrace_strncmp(dest, elem,
3503 size) == 0)
3504 found_key = B_TRUE;
3505 } else if (found_key) {
3506 if (nelems > 1) {
3507 /*
3508 * We expected an object, not
3509 * this string.
3510 */
3511 return (NULL);
3512 }
3513 return (dest);
3514 }
3515 state = string_is_key ? JSON_COLON :
3516 JSON_COMMA;
3517 string_is_key = B_FALSE;
3518 break;
3519 }
3520
3521 *dd++ = cc;
3522 break;
3523 case JSON_STRING_ESCAPE:
3524 *dd++ = cc;
3525 if (cc == 'u') {
3526 escape_unicount = 0;
3527 state = JSON_STRING_ESCAPE_UNICODE;
3528 } else {
3529 state = JSON_STRING;
3530 }
3531 break;
3532 case JSON_STRING_ESCAPE_UNICODE:
3533 if (!isxdigit(cc))
3534 /* ERROR: unvalid unicode escape */
3535 return (NULL);
3536
3537 *dd++ = cc;
3538 if (++escape_unicount == 4)
3539 state = JSON_STRING;
3540 break;
3541 case JSON_COLON:
3542 if (cc == ' ' || cc == '\t' || cc == '\n' || cc == '\r')
3543 break; /* eat whitespace */
3544
3545 if (cc == ':') {
3546 state = JSON_VALUE;
3547 break;
3548 }
3549
3550 /* ERROR: expected colon */
3551 return (NULL);
3552 case JSON_COMMA:
3553 if (cc == ' ' || cc == '\t' || cc == '\n' || cc == '\r')
3554 break; /* eat whitespace */
3555
3556 if (cc == ',') {
3557 if (in_array) {
3558 state = JSON_VALUE;
3559 if (++array_pos == array_elem)
3560 found_key = B_TRUE;
3561 } else {
3562 state = JSON_OBJECT;
3563 }
3564 break;
3565 }
3566
3567 /* ERROR: key not found or expected comma */
3568 return (NULL);
3569 case JSON_IDENTIFIER:
3570 if (cc >= 'a' && cc <= 'z') {
3571 *dd++ = cc;
3572 break;
3573 }
3574
3575 *dd = '\0';
3576 dd = dest; /* reset string buffer */
3577
3578 if (dtrace_strncmp(dest, "true", 5) == 0 ||
3579 dtrace_strncmp(dest, "false", 6) == 0 ||
3580 dtrace_strncmp(dest, "null", 5) == 0) {
3581 if (found_key) {
3582 if (nelems > 1) {
3583 /*
3584 * We expected an object, not
3585 * this identifier.
3586 */
3587 return (NULL);
3588 }
3589 return (dest);
3590 } else {
3591 cur--;
3592 state = JSON_COMMA;
3593 break;
3594 }
3595 }
3596
3597 /* ERROR: unexpected identifier */
3598 return (NULL);
3599 case JSON_NUMBER:
3600 if (cc == '.') {
3601 *dd++ = cc;
3602 state = JSON_NUMBER_FRAC;
3603 break;
3604 }
3605
3606 if (cc == 'x' || cc == 'X')
3607 /* ERROR: spec explicitly excludes hex */
3608 return (NULL);
3609
3610 /* FALLTHRU */
3611 case JSON_NUMBER_FRAC:
3612 if (cc == 'e' || cc == 'E') {
3613 *dd++ = cc;
3614 state = JSON_NUMBER_EXP;
3615 break;
3616 }
3617
3618 if (cc == '+' || cc == '-') {
3619 /*
3620 * ERROR: expect sign as part of exponent only
3621 */
3622 return (NULL);
3623 }
3624 /* FALLTHRU */
3625 case JSON_NUMBER_EXP:
3626 if ((cc >= '0' && cc <= '9') || cc == '+' ||
3627 cc == '-') {
3628 *dd++ = cc;
3629 break;
3630 }
3631
3632 *dd = '\0';
3633 dd = dest; /* reset string buffer */
3634 if (found_key) {
3635 if (nelems > 1) {
3636 /*
3637 * We expected an object, not this
3638 * number.
3639 */
3640 return (NULL);
3641 }
3642 return (dest);
3643 }
3644
3645 cur--;
3646 state = JSON_COMMA;
3647 break;
3648 case JSON_VALUE:
3649 if (cc == ' ' || cc == '\t' || cc == '\n' || cc == '\r')
3650 break; /* eat whitespace */
3651
3652 if (cc == '{' || cc == '[') {
3653 if (nelems > 1 && found_key) {
3654 in_array = !!(cc == '[');
3655 /*
3656 * If our element selector directs us
3657 * to descend into this nested object,
3658 * then move to the next selector
3659 * element in the list and restart the
3660 * state machine.
3661 */
3662 while (*elem != '\0')
3663 elem++;
3664 elem++; /* skip the inter-element NUL */
3665 nelems--;
3666 dd = dest;
3667 if (in_array) {
3668 state = JSON_VALUE;
3669 array_pos = 0;
3670 array_elem = dtrace_strtoll(
3671 elem, 10, size);
3672 found_key = !!(array_elem == 0);
3673 } else {
3674 found_key = B_FALSE;
3675 state = JSON_OBJECT;
3676 }
3677 break;
3678 }
3679
3680 /*
3681 * Otherwise, we wish to either skip this
3682 * nested object or return it in full.
3683 */
3684 if (cc == '[')
3685 brackets = 1;
3686 else
3687 braces = 1;
3688 *dd++ = cc;
3689 state = JSON_COLLECT_OBJECT;
3690 break;
3691 }
3692
3693 if (cc == '"') {
3694 state = JSON_STRING;
3695 break;
3696 }
3697
3698 if (cc >= 'a' && cc <= 'z') {
3699 /* Here we deal with true, false and null */
3700 *dd++ = cc;
3701 state = JSON_IDENTIFIER;
3702 break;
3703 }
3704
3705 if (cc == '-' || (cc >= '0' && cc <= '9')) {
3706 *dd++ = cc;
3707 state = JSON_NUMBER;
3708 break;
3709 }
3710
3711 /* ERROR: unexpected character */
3712 return (NULL);
3713 case JSON_COLLECT_OBJECT:
3714 if (cc == '\0')
3715 /* ERROR: unexpected end of input */
3716 return (NULL);
3717
3718 *dd++ = cc;
3719 if (cc == '"') {
3720 collect_object = B_TRUE;
3721 state = JSON_STRING;
3722 break;
3723 }
3724
3725 if (cc == ']') {
3726 if (brackets-- == 0) {
3727 /* ERROR: unbalanced brackets */
3728 return (NULL);
3729 }
3730 } else if (cc == '}') {
3731 if (braces-- == 0) {
3732 /* ERROR: unbalanced braces */
3733 return (NULL);
3734 }
3735 } else if (cc == '{') {
3736 braces++;
3737 } else if (cc == '[') {
3738 brackets++;
3739 }
3740
3741 if (brackets == 0 && braces == 0) {
3742 if (found_key) {
3743 *dd = '\0';
3744 return (dest);
3745 }
3746 dd = dest; /* reset string buffer */
3747 state = JSON_COMMA;
3748 }
3749 break;
3750 }
3751 }
3752 return (NULL);
3753 }
3754
3755 /*
3756 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3757 * Notice that we don't bother validating the proper number of arguments or
3758 * their types in the tuple stack. This isn't needed because all argument
3759 * interpretation is safe because of our load safety -- the worst that can
3760 * happen is that a bogus program can obtain bogus results.
3761 */
3762 static void
3763 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3764 dtrace_key_t *tupregs, int nargs,
3765 dtrace_mstate_t *mstate, dtrace_state_t *state)
3766 {
3767 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
3768 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
3769 dtrace_vstate_t *vstate = &state->dts_vstate;
3770
3771 union {
3772 mutex_impl_t mi;
3773 uint64_t mx;
3774 } m;
3775
3776 union {
3777 krwlock_t ri;
3778 uintptr_t rw;
3779 } r;
3780
3781 switch (subr) {
3782 case DIF_SUBR_RAND:
3783 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3784 break;
3785
3786 case DIF_SUBR_MUTEX_OWNED:
3787 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3788 mstate, vstate)) {
3789 regs[rd] = NULL;
3790 break;
3791 }
3792
3793 m.mx = dtrace_load64(tupregs[0].dttk_value);
3794 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3795 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3796 else
3797 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3798 break;
3799
3800 case DIF_SUBR_MUTEX_OWNER:
3801 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3802 mstate, vstate)) {
3803 regs[rd] = NULL;
3804 break;
3805 }
3806
3807 m.mx = dtrace_load64(tupregs[0].dttk_value);
3808 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3809 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3810 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3811 else
3812 regs[rd] = 0;
3813 break;
3814
3815 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3816 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3817 mstate, vstate)) {
3818 regs[rd] = NULL;
3819 break;
3820 }
3821
3822 m.mx = dtrace_load64(tupregs[0].dttk_value);
3823 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3824 break;
3825
3826 case DIF_SUBR_MUTEX_TYPE_SPIN:
3827 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3828 mstate, vstate)) {
3829 regs[rd] = NULL;
3830 break;
3831 }
3832
3833 m.mx = dtrace_load64(tupregs[0].dttk_value);
3834 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3835 break;
3836
3837 case DIF_SUBR_RW_READ_HELD: {
3838 uintptr_t tmp;
3839
3840 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3841 mstate, vstate)) {
3842 regs[rd] = NULL;
3843 break;
3844 }
3845
3846 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3847 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3848 break;
3849 }
3850
3851 case DIF_SUBR_RW_WRITE_HELD:
3852 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3853 mstate, vstate)) {
3854 regs[rd] = NULL;
3855 break;
3856 }
3857
3858 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3859 regs[rd] = _RW_WRITE_HELD(&r.ri);
3860 break;
3861
3862 case DIF_SUBR_RW_ISWRITER:
3863 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3864 mstate, vstate)) {
3865 regs[rd] = NULL;
3866 break;
3867 }
3868
3869 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3870 regs[rd] = _RW_ISWRITER(&r.ri);
3871 break;
3872
3873 case DIF_SUBR_BCOPY: {
3874 /*
3875 * We need to be sure that the destination is in the scratch
3876 * region -- no other region is allowed.
3877 */
3878 uintptr_t src = tupregs[0].dttk_value;
3879 uintptr_t dest = tupregs[1].dttk_value;
3880 size_t size = tupregs[2].dttk_value;
3881
3882 if (!dtrace_inscratch(dest, size, mstate)) {
3883 *flags |= CPU_DTRACE_BADADDR;
3884 *illval = regs[rd];
3885 break;
3886 }
3887
3888 if (!dtrace_canload(src, size, mstate, vstate)) {
3889 regs[rd] = NULL;
3890 break;
3891 }
3892
3893 dtrace_bcopy((void *)src, (void *)dest, size);
3894 break;
3895 }
3896
3897 case DIF_SUBR_ALLOCA:
3898 case DIF_SUBR_COPYIN: {
3899 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3900 uint64_t size =
3901 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3902 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3903
3904 /*
3905 * This action doesn't require any credential checks since
3906 * probes will not activate in user contexts to which the
3907 * enabling user does not have permissions.
3908 */
3909
3910 /*
3911 * Rounding up the user allocation size could have overflowed
3912 * a large, bogus allocation (like -1ULL) to 0.
3913 */
3914 if (scratch_size < size ||
3915 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3916 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3917 regs[rd] = NULL;
3918 break;
3919 }
3920
3921 if (subr == DIF_SUBR_COPYIN) {
3922 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3923 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3924 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3925 }
3926
3927 mstate->dtms_scratch_ptr += scratch_size;
3928 regs[rd] = dest;
3929 break;
3930 }
3931
3932 case DIF_SUBR_COPYINTO: {
3933 uint64_t size = tupregs[1].dttk_value;
3934 uintptr_t dest = tupregs[2].dttk_value;
3935
3936 /*
3937 * This action doesn't require any credential checks since
3938 * probes will not activate in user contexts to which the
3939 * enabling user does not have permissions.
3940 */
3941 if (!dtrace_inscratch(dest, size, mstate)) {
3942 *flags |= CPU_DTRACE_BADADDR;
3943 *illval = regs[rd];
3944 break;
3945 }
3946
3947 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3948 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3949 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3950 break;
3951 }
3952
3953 case DIF_SUBR_COPYINSTR: {
3954 uintptr_t dest = mstate->dtms_scratch_ptr;
3955 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3956
3957 if (nargs > 1 && tupregs[1].dttk_value < size)
3958 size = tupregs[1].dttk_value + 1;
3959
3960 /*
3961 * This action doesn't require any credential checks since
3962 * probes will not activate in user contexts to which the
3963 * enabling user does not have permissions.
3964 */
3965 if (!DTRACE_INSCRATCH(mstate, size)) {
3966 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3967 regs[rd] = NULL;
3968 break;
3969 }
3970
3971 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3972 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3973 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3974
3975 ((char *)dest)[size - 1] = '\0';
3976 mstate->dtms_scratch_ptr += size;
3977 regs[rd] = dest;
3978 break;
3979 }
3980
3981 case DIF_SUBR_MSGSIZE:
3982 case DIF_SUBR_MSGDSIZE: {
3983 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3984 uintptr_t wptr, rptr;
3985 size_t count = 0;
3986 int cont = 0;
3987
3988 while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3989
3990 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3991 vstate)) {
3992 regs[rd] = NULL;
3993 break;
3994 }
3995
3996 wptr = dtrace_loadptr(baddr +
3997 offsetof(mblk_t, b_wptr));
3998
3999 rptr = dtrace_loadptr(baddr +
4000 offsetof(mblk_t, b_rptr));
4001
4002 if (wptr < rptr) {
4003 *flags |= CPU_DTRACE_BADADDR;
4004 *illval = tupregs[0].dttk_value;
4005 break;
4006 }
4007
4008 daddr = dtrace_loadptr(baddr +
4009 offsetof(mblk_t, b_datap));
4010
4011 baddr = dtrace_loadptr(baddr +
4012 offsetof(mblk_t, b_cont));
4013
4014 /*
4015 * We want to prevent against denial-of-service here,
4016 * so we're only going to search the list for
4017 * dtrace_msgdsize_max mblks.
4018 */
4019 if (cont++ > dtrace_msgdsize_max) {
4020 *flags |= CPU_DTRACE_ILLOP;
4021 break;
4022 }
4023
4024 if (subr == DIF_SUBR_MSGDSIZE) {
4025 if (dtrace_load8(daddr +
4026 offsetof(dblk_t, db_type)) != M_DATA)
4027 continue;
4028 }
4029
4030 count += wptr - rptr;
4031 }
4032
4033 if (!(*flags & CPU_DTRACE_FAULT))
4034 regs[rd] = count;
4035
4036 break;
4037 }
4038
4039 case DIF_SUBR_PROGENYOF: {
4040 pid_t pid = tupregs[0].dttk_value;
4041 proc_t *p;
4042 int rval = 0;
4043
4044 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4045
4046 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4047 if (p->p_pidp->pid_id == pid) {
4048 rval = 1;
4049 break;
4050 }
4051 }
4052
4053 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4054
4055 regs[rd] = rval;
4056 break;
4057 }
4058
4059 case DIF_SUBR_SPECULATION:
4060 regs[rd] = dtrace_speculation(state);
4061 break;
4062
4063 case DIF_SUBR_COPYOUT: {
4064 uintptr_t kaddr = tupregs[0].dttk_value;
4065 uintptr_t uaddr = tupregs[1].dttk_value;
4066 uint64_t size = tupregs[2].dttk_value;
4067
4068 if (!dtrace_destructive_disallow &&
4069 dtrace_priv_proc_control(state, mstate) &&
4070 !dtrace_istoxic(kaddr, size)) {
4071 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4072 dtrace_copyout(kaddr, uaddr, size, flags);
4073 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4074 }
4075 break;
4076 }
4077
4078 case DIF_SUBR_COPYOUTSTR: {
4079 uintptr_t kaddr = tupregs[0].dttk_value;
4080 uintptr_t uaddr = tupregs[1].dttk_value;
4081 uint64_t size = tupregs[2].dttk_value;
4082
4083 if (!dtrace_destructive_disallow &&
4084 dtrace_priv_proc_control(state, mstate) &&
4085 !dtrace_istoxic(kaddr, size)) {
4086 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4087 dtrace_copyoutstr(kaddr, uaddr, size, flags);
4088 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4089 }
4090 break;
4091 }
4092
4093 case DIF_SUBR_STRLEN: {
4094 size_t sz;
4095 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4096 sz = dtrace_strlen((char *)addr,
4097 state->dts_options[DTRACEOPT_STRSIZE]);
4098
4099 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
4100 regs[rd] = NULL;
4101 break;
4102 }
4103
4104 regs[rd] = sz;
4105
4106 break;
4107 }
4108
4109 case DIF_SUBR_STRCHR:
4110 case DIF_SUBR_STRRCHR: {
4111 /*
4112 * We're going to iterate over the string looking for the
4113 * specified character. We will iterate until we have reached
4114 * the string length or we have found the character. If this
4115 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4116 * of the specified character instead of the first.
4117 */
4118 uintptr_t saddr = tupregs[0].dttk_value;
4119 uintptr_t addr = tupregs[0].dttk_value;
4120 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
4121 char c, target = (char)tupregs[1].dttk_value;
4122
4123 for (regs[rd] = NULL; addr < limit; addr++) {
4124 if ((c = dtrace_load8(addr)) == target) {
4125 regs[rd] = addr;
4126
4127 if (subr == DIF_SUBR_STRCHR)
4128 break;
4129 }
4130
4131 if (c == '\0')
4132 break;
4133 }
4134
4135 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
4136 regs[rd] = NULL;
4137 break;
4138 }
4139
4140 break;
4141 }
4142
4143 case DIF_SUBR_STRSTR:
4144 case DIF_SUBR_INDEX:
4145 case DIF_SUBR_RINDEX: {
4146 /*
4147 * We're going to iterate over the string looking for the
4148 * specified string. We will iterate until we have reached
4149 * the string length or we have found the string. (Yes, this
4150 * is done in the most naive way possible -- but considering
4151 * that the string we're searching for is likely to be
4152 * relatively short, the complexity of Rabin-Karp or similar
4153 * hardly seems merited.)
4154 */
4155 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4156 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4157 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4158 size_t len = dtrace_strlen(addr, size);
4159 size_t sublen = dtrace_strlen(substr, size);
4160 char *limit = addr + len, *orig = addr;
4161 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4162 int inc = 1;
4163
4164 regs[rd] = notfound;
4165
4166 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4167 regs[rd] = NULL;
4168 break;
4169 }
4170
4171 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4172 vstate)) {
4173 regs[rd] = NULL;
4174 break;
4175 }
4176
4177 /*
4178 * strstr() and index()/rindex() have similar semantics if
4179 * both strings are the empty string: strstr() returns a
4180 * pointer to the (empty) string, and index() and rindex()
4181 * both return index 0 (regardless of any position argument).
4182 */
4183 if (sublen == 0 && len == 0) {
4184 if (subr == DIF_SUBR_STRSTR)
4185 regs[rd] = (uintptr_t)addr;
4186 else
4187 regs[rd] = 0;
4188 break;
4189 }
4190
4191 if (subr != DIF_SUBR_STRSTR) {
4192 if (subr == DIF_SUBR_RINDEX) {
4193 limit = orig - 1;
4194 addr += len;
4195 inc = -1;
4196 }
4197
4198 /*
4199 * Both index() and rindex() take an optional position
4200 * argument that denotes the starting position.
4201 */
4202 if (nargs == 3) {
4203 int64_t pos = (int64_t)tupregs[2].dttk_value;
4204
4205 /*
4206 * If the position argument to index() is
4207 * negative, Perl implicitly clamps it at
4208 * zero. This semantic is a little surprising
4209 * given the special meaning of negative
4210 * positions to similar Perl functions like
4211 * substr(), but it appears to reflect a
4212 * notion that index() can start from a
4213 * negative index and increment its way up to
4214 * the string. Given this notion, Perl's
4215 * rindex() is at least self-consistent in
4216 * that it implicitly clamps positions greater
4217 * than the string length to be the string
4218 * length. Where Perl completely loses
4219 * coherence, however, is when the specified
4220 * substring is the empty string (""). In
4221 * this case, even if the position is
4222 * negative, rindex() returns 0 -- and even if
4223 * the position is greater than the length,
4224 * index() returns the string length. These
4225 * semantics violate the notion that index()
4226 * should never return a value less than the
4227 * specified position and that rindex() should
4228 * never return a value greater than the
4229 * specified position. (One assumes that
4230 * these semantics are artifacts of Perl's
4231 * implementation and not the results of
4232 * deliberate design -- it beggars belief that
4233 * even Larry Wall could desire such oddness.)
4234 * While in the abstract one would wish for
4235 * consistent position semantics across
4236 * substr(), index() and rindex() -- or at the
4237 * very least self-consistent position
4238 * semantics for index() and rindex() -- we
4239 * instead opt to keep with the extant Perl
4240 * semantics, in all their broken glory. (Do
4241 * we have more desire to maintain Perl's
4242 * semantics than Perl does? Probably.)
4243 */
4244 if (subr == DIF_SUBR_RINDEX) {
4245 if (pos < 0) {
4246 if (sublen == 0)
4247 regs[rd] = 0;
4248 break;
4249 }
4250
4251 if (pos > len)
4252 pos = len;
4253 } else {
4254 if (pos < 0)
4255 pos = 0;
4256
4257 if (pos >= len) {
4258 if (sublen == 0)
4259 regs[rd] = len;
4260 break;
4261 }
4262 }
4263
4264 addr = orig + pos;
4265 }
4266 }
4267
4268 for (regs[rd] = notfound; addr != limit; addr += inc) {
4269 if (dtrace_strncmp(addr, substr, sublen) == 0) {
4270 if (subr != DIF_SUBR_STRSTR) {
4271 /*
4272 * As D index() and rindex() are
4273 * modeled on Perl (and not on awk),
4274 * we return a zero-based (and not a
4275 * one-based) index. (For you Perl
4276 * weenies: no, we're not going to add
4277 * $[ -- and shouldn't you be at a con
4278 * or something?)
4279 */
4280 regs[rd] = (uintptr_t)(addr - orig);
4281 break;
4282 }
4283
4284 ASSERT(subr == DIF_SUBR_STRSTR);
4285 regs[rd] = (uintptr_t)addr;
4286 break;
4287 }
4288 }
4289
4290 break;
4291 }
4292
4293 case DIF_SUBR_STRTOK: {
4294 uintptr_t addr = tupregs[0].dttk_value;
4295 uintptr_t tokaddr = tupregs[1].dttk_value;
4296 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4297 uintptr_t limit, toklimit = tokaddr + size;
4298 uint8_t c, tokmap[32]; /* 256 / 8 */
4299 char *dest = (char *)mstate->dtms_scratch_ptr;
4300 int i;
4301
4302 /*
4303 * Check both the token buffer and (later) the input buffer,
4304 * since both could be non-scratch addresses.
4305 */
4306 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
4307 regs[rd] = NULL;
4308 break;
4309 }
4310
4311 if (!DTRACE_INSCRATCH(mstate, size)) {
4312 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4313 regs[rd] = NULL;
4314 break;
4315 }
4316
4317 if (addr == NULL) {
4318 /*
4319 * If the address specified is NULL, we use our saved
4320 * strtok pointer from the mstate. Note that this
4321 * means that the saved strtok pointer is _only_
4322 * valid within multiple enablings of the same probe --
4323 * it behaves like an implicit clause-local variable.
4324 */
4325 addr = mstate->dtms_strtok;
4326 } else {
4327 /*
4328 * If the user-specified address is non-NULL we must
4329 * access check it. This is the only time we have
4330 * a chance to do so, since this address may reside
4331 * in the string table of this clause-- future calls
4332 * (when we fetch addr from mstate->dtms_strtok)
4333 * would fail this access check.
4334 */
4335 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
4336 regs[rd] = NULL;
4337 break;
4338 }
4339 }
4340
4341 /*
4342 * First, zero the token map, and then process the token
4343 * string -- setting a bit in the map for every character
4344 * found in the token string.
4345 */
4346 for (i = 0; i < sizeof (tokmap); i++)
4347 tokmap[i] = 0;
4348
4349 for (; tokaddr < toklimit; tokaddr++) {
4350 if ((c = dtrace_load8(tokaddr)) == '\0')
4351 break;
4352
4353 ASSERT((c >> 3) < sizeof (tokmap));
4354 tokmap[c >> 3] |= (1 << (c & 0x7));
4355 }
4356
4357 for (limit = addr + size; addr < limit; addr++) {
4358 /*
4359 * We're looking for a character that is _not_ contained
4360 * in the token string.
4361 */
4362 if ((c = dtrace_load8(addr)) == '\0')
4363 break;
4364
4365 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4366 break;
4367 }
4368
4369 if (c == '\0') {
4370 /*
4371 * We reached the end of the string without finding
4372 * any character that was not in the token string.
4373 * We return NULL in this case, and we set the saved
4374 * address to NULL as well.
4375 */
4376 regs[rd] = NULL;
4377 mstate->dtms_strtok = NULL;
4378 break;
4379 }
4380
4381 /*
4382 * From here on, we're copying into the destination string.
4383 */
4384 for (i = 0; addr < limit && i < size - 1; addr++) {
4385 if ((c = dtrace_load8(addr)) == '\0')
4386 break;
4387
4388 if (tokmap[c >> 3] & (1 << (c & 0x7)))
4389 break;
4390
4391 ASSERT(i < size);
4392 dest[i++] = c;
4393 }
4394
4395 ASSERT(i < size);
4396 dest[i] = '\0';
4397 regs[rd] = (uintptr_t)dest;
4398 mstate->dtms_scratch_ptr += size;
4399 mstate->dtms_strtok = addr;
4400 break;
4401 }
4402
4403 case DIF_SUBR_SUBSTR: {
4404 uintptr_t s = tupregs[0].dttk_value;
4405 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4406 char *d = (char *)mstate->dtms_scratch_ptr;
4407 int64_t index = (int64_t)tupregs[1].dttk_value;
4408 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4409 size_t len = dtrace_strlen((char *)s, size);
4410 int64_t i;
4411
4412 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4413 regs[rd] = NULL;
4414 break;
4415 }
4416
4417 if (!DTRACE_INSCRATCH(mstate, size)) {
4418 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4419 regs[rd] = NULL;
4420 break;
4421 }
4422
4423 if (nargs <= 2)
4424 remaining = (int64_t)size;
4425
4426 if (index < 0) {
4427 index += len;
4428
4429 if (index < 0 && index + remaining > 0) {
4430 remaining += index;
4431 index = 0;
4432 }
4433 }
4434
4435 if (index >= len || index < 0) {
4436 remaining = 0;
4437 } else if (remaining < 0) {
4438 remaining += len - index;
4439 } else if (index + remaining > size) {
4440 remaining = size - index;
4441 }
4442
4443 for (i = 0; i < remaining; i++) {
4444 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4445 break;
4446 }
4447
4448 d[i] = '\0';
4449
4450 mstate->dtms_scratch_ptr += size;
4451 regs[rd] = (uintptr_t)d;
4452 break;
4453 }
4454
4455 case DIF_SUBR_JSON: {
4456 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4457 uintptr_t json = tupregs[0].dttk_value;
4458 size_t jsonlen = dtrace_strlen((char *)json, size);
4459 uintptr_t elem = tupregs[1].dttk_value;
4460 size_t elemlen = dtrace_strlen((char *)elem, size);
4461
4462 char *dest = (char *)mstate->dtms_scratch_ptr;
4463 char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
4464 char *ee = elemlist;
4465 int nelems = 1;
4466 uintptr_t cur;
4467
4468 if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
4469 !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
4470 regs[rd] = NULL;
4471 break;
4472 }
4473
4474 if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
4475 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4476 regs[rd] = NULL;
4477 break;
4478 }
4479
4480 /*
4481 * Read the element selector and split it up into a packed list
4482 * of strings.
4483 */
4484 for (cur = elem; cur < elem + elemlen; cur++) {
4485 char cc = dtrace_load8(cur);
4486
4487 if (cur == elem && cc == '[')
4488 /* first element selector may be an array */
4489 continue;
4490
4491 if (cc == ']')
4492 continue;
4493
4494 if (cc == '.' || cc == '[') {
4495 nelems++;
4496 cc = '\0';
4497 }
4498
4499 *ee++ = cc;
4500 }
4501 *ee++ = '\0';
4502
4503 if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
4504 nelems, dest)) != NULL)
4505 mstate->dtms_scratch_ptr += jsonlen + 1;
4506 break;
4507 }
4508
4509 case DIF_SUBR_TOUPPER:
4510 case DIF_SUBR_TOLOWER: {
4511 uintptr_t s = tupregs[0].dttk_value;
4512 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4513 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4514 size_t len = dtrace_strlen((char *)s, size);
4515 char lower, upper, convert;
4516 int64_t i;
4517
4518 if (subr == DIF_SUBR_TOUPPER) {
4519 lower = 'a';
4520 upper = 'z';
4521 convert = 'A';
4522 } else {
4523 lower = 'A';
4524 upper = 'Z';
4525 convert = 'a';
4526 }
4527
4528 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4529 regs[rd] = NULL;
4530 break;
4531 }
4532
4533 if (!DTRACE_INSCRATCH(mstate, size)) {
4534 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4535 regs[rd] = NULL;
4536 break;
4537 }
4538
4539 for (i = 0; i < size - 1; i++) {
4540 if ((c = dtrace_load8(s + i)) == '\0')
4541 break;
4542
4543 if (c >= lower && c <= upper)
4544 c = convert + (c - lower);
4545
4546 dest[i] = c;
4547 }
4548
4549 ASSERT(i < size);
4550 dest[i] = '\0';
4551 regs[rd] = (uintptr_t)dest;
4552 mstate->dtms_scratch_ptr += size;
4553 break;
4554 }
4555
4556 case DIF_SUBR_GETMAJOR:
4557 #ifdef _LP64
4558 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4559 #else
4560 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4561 #endif
4562 break;
4563
4564 case DIF_SUBR_GETMINOR:
4565 #ifdef _LP64
4566 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4567 #else
4568 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4569 #endif
4570 break;
4571
4572 case DIF_SUBR_DDI_PATHNAME: {
4573 /*
4574 * This one is a galactic mess. We are going to roughly
4575 * emulate ddi_pathname(), but it's made more complicated
4576 * by the fact that we (a) want to include the minor name and
4577 * (b) must proceed iteratively instead of recursively.
4578 */
4579 uintptr_t dest = mstate->dtms_scratch_ptr;
4580 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4581 char *start = (char *)dest, *end = start + size - 1;
4582 uintptr_t daddr = tupregs[0].dttk_value;
4583 int64_t minor = (int64_t)tupregs[1].dttk_value;
4584 char *s;
4585 int i, len, depth = 0;
4586
4587 /*
4588 * Due to all the pointer jumping we do and context we must
4589 * rely upon, we just mandate that the user must have kernel
4590 * read privileges to use this routine.
4591 */
4592 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4593 *flags |= CPU_DTRACE_KPRIV;
4594 *illval = daddr;
4595 regs[rd] = NULL;
4596 }
4597
4598 if (!DTRACE_INSCRATCH(mstate, size)) {
4599 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4600 regs[rd] = NULL;
4601 break;
4602 }
4603
4604 *end = '\0';
4605
4606 /*
4607 * We want to have a name for the minor. In order to do this,
4608 * we need to walk the minor list from the devinfo. We want
4609 * to be sure that we don't infinitely walk a circular list,
4610 * so we check for circularity by sending a scout pointer
4611 * ahead two elements for every element that we iterate over;
4612 * if the list is circular, these will ultimately point to the
4613 * same element. You may recognize this little trick as the
4614 * answer to a stupid interview question -- one that always
4615 * seems to be asked by those who had to have it laboriously
4616 * explained to them, and who can't even concisely describe
4617 * the conditions under which one would be forced to resort to
4618 * this technique. Needless to say, those conditions are
4619 * found here -- and probably only here. Is this the only use
4620 * of this infamous trick in shipping, production code? If it
4621 * isn't, it probably should be...
4622 */
4623 if (minor != -1) {
4624 uintptr_t maddr = dtrace_loadptr(daddr +
4625 offsetof(struct dev_info, devi_minor));
4626
4627 uintptr_t next = offsetof(struct ddi_minor_data, next);
4628 uintptr_t name = offsetof(struct ddi_minor_data,
4629 d_minor) + offsetof(struct ddi_minor, name);
4630 uintptr_t dev = offsetof(struct ddi_minor_data,
4631 d_minor) + offsetof(struct ddi_minor, dev);
4632 uintptr_t scout;
4633
4634 if (maddr != NULL)
4635 scout = dtrace_loadptr(maddr + next);
4636
4637 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4638 uint64_t m;
4639 #ifdef _LP64
4640 m = dtrace_load64(maddr + dev) & MAXMIN64;
4641 #else
4642 m = dtrace_load32(maddr + dev) & MAXMIN;
4643 #endif
4644 if (m != minor) {
4645 maddr = dtrace_loadptr(maddr + next);
4646
4647 if (scout == NULL)
4648 continue;
4649
4650 scout = dtrace_loadptr(scout + next);
4651
4652 if (scout == NULL)
4653 continue;
4654
4655 scout = dtrace_loadptr(scout + next);
4656
4657 if (scout == NULL)
4658 continue;
4659
4660 if (scout == maddr) {
4661 *flags |= CPU_DTRACE_ILLOP;
4662 break;
4663 }
4664
4665 continue;
4666 }
4667
4668 /*
4669 * We have the minor data. Now we need to
4670 * copy the minor's name into the end of the
4671 * pathname.
4672 */
4673 s = (char *)dtrace_loadptr(maddr + name);
4674 len = dtrace_strlen(s, size);
4675
4676 if (*flags & CPU_DTRACE_FAULT)
4677 break;
4678
4679 if (len != 0) {
4680 if ((end -= (len + 1)) < start)
4681 break;
4682
4683 *end = ':';
4684 }
4685
4686 for (i = 1; i <= len; i++)
4687 end[i] = dtrace_load8((uintptr_t)s++);
4688 break;
4689 }
4690 }
4691
4692 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4693 ddi_node_state_t devi_state;
4694
4695 devi_state = dtrace_load32(daddr +
4696 offsetof(struct dev_info, devi_node_state));
4697
4698 if (*flags & CPU_DTRACE_FAULT)
4699 break;
4700
4701 if (devi_state >= DS_INITIALIZED) {
4702 s = (char *)dtrace_loadptr(daddr +
4703 offsetof(struct dev_info, devi_addr));
4704 len = dtrace_strlen(s, size);
4705
4706 if (*flags & CPU_DTRACE_FAULT)
4707 break;
4708
4709 if (len != 0) {
4710 if ((end -= (len + 1)) < start)
4711 break;
4712
4713 *end = '@';
4714 }
4715
4716 for (i = 1; i <= len; i++)
4717 end[i] = dtrace_load8((uintptr_t)s++);
4718 }
4719
4720 /*
4721 * Now for the node name...
4722 */
4723 s = (char *)dtrace_loadptr(daddr +
4724 offsetof(struct dev_info, devi_node_name));
4725
4726 daddr = dtrace_loadptr(daddr +
4727 offsetof(struct dev_info, devi_parent));
4728
4729 /*
4730 * If our parent is NULL (that is, if we're the root
4731 * node), we're going to use the special path
4732 * "devices".
4733 */
4734 if (daddr == NULL)
4735 s = "devices";
4736
4737 len = dtrace_strlen(s, size);
4738 if (*flags & CPU_DTRACE_FAULT)
4739 break;
4740
4741 if ((end -= (len + 1)) < start)
4742 break;
4743
4744 for (i = 1; i <= len; i++)
4745 end[i] = dtrace_load8((uintptr_t)s++);
4746 *end = '/';
4747
4748 if (depth++ > dtrace_devdepth_max) {
4749 *flags |= CPU_DTRACE_ILLOP;
4750 break;
4751 }
4752 }
4753
4754 if (end < start)
4755 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4756
4757 if (daddr == NULL) {
4758 regs[rd] = (uintptr_t)end;
4759 mstate->dtms_scratch_ptr += size;
4760 }
4761
4762 break;
4763 }
4764
4765 case DIF_SUBR_STRJOIN: {
4766 char *d = (char *)mstate->dtms_scratch_ptr;
4767 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4768 uintptr_t s1 = tupregs[0].dttk_value;
4769 uintptr_t s2 = tupregs[1].dttk_value;
4770 int i = 0;
4771
4772 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4773 !dtrace_strcanload(s2, size, mstate, vstate)) {
4774 regs[rd] = NULL;
4775 break;
4776 }
4777
4778 if (!DTRACE_INSCRATCH(mstate, size)) {
4779 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4780 regs[rd] = NULL;
4781 break;
4782 }
4783
4784 for (;;) {
4785 if (i >= size) {
4786 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4787 regs[rd] = NULL;
4788 break;
4789 }
4790
4791 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4792 i--;
4793 break;
4794 }
4795 }
4796
4797 for (;;) {
4798 if (i >= size) {
4799 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4800 regs[rd] = NULL;
4801 break;
4802 }
4803
4804 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4805 break;
4806 }
4807
4808 if (i < size) {
4809 mstate->dtms_scratch_ptr += i;
4810 regs[rd] = (uintptr_t)d;
4811 }
4812
4813 break;
4814 }
4815
4816 case DIF_SUBR_STRTOLL: {
4817 uintptr_t s = tupregs[0].dttk_value;
4818 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4819 int base = 10;
4820
4821 if (nargs > 1) {
4822 if ((base = tupregs[1].dttk_value) <= 1 ||
4823 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4824 *flags |= CPU_DTRACE_ILLOP;
4825 break;
4826 }
4827 }
4828
4829 if (!dtrace_strcanload(s, size, mstate, vstate)) {
4830 regs[rd] = INT64_MIN;
4831 break;
4832 }
4833
4834 regs[rd] = dtrace_strtoll((char *)s, base, size);
4835 break;
4836 }
4837
4838 case DIF_SUBR_LLTOSTR: {
4839 int64_t i = (int64_t)tupregs[0].dttk_value;
4840 uint64_t val, digit;
4841 uint64_t size = 65; /* enough room for 2^64 in binary */
4842 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4843 int base = 10;
4844
4845 if (nargs > 1) {
4846 if ((base = tupregs[1].dttk_value) <= 1 ||
4847 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4848 *flags |= CPU_DTRACE_ILLOP;
4849 break;
4850 }
4851 }
4852
4853 val = (base == 10 && i < 0) ? i * -1 : i;
4854
4855 if (!DTRACE_INSCRATCH(mstate, size)) {
4856 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4857 regs[rd] = NULL;
4858 break;
4859 }
4860
4861 for (*end-- = '\0'; val; val /= base) {
4862 if ((digit = val % base) <= '9' - '0') {
4863 *end-- = '0' + digit;
4864 } else {
4865 *end-- = 'a' + (digit - ('9' - '0') - 1);
4866 }
4867 }
4868
4869 if (i == 0 && base == 16)
4870 *end-- = '0';
4871
4872 if (base == 16)
4873 *end-- = 'x';
4874
4875 if (i == 0 || base == 8 || base == 16)
4876 *end-- = '0';
4877
4878 if (i < 0 && base == 10)
4879 *end-- = '-';
4880
4881 regs[rd] = (uintptr_t)end + 1;
4882 mstate->dtms_scratch_ptr += size;
4883 break;
4884 }
4885
4886 case DIF_SUBR_HTONS:
4887 case DIF_SUBR_NTOHS:
4888 #ifdef _BIG_ENDIAN
4889 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4890 #else
4891 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4892 #endif
4893 break;
4894
4895
4896 case DIF_SUBR_HTONL:
4897 case DIF_SUBR_NTOHL:
4898 #ifdef _BIG_ENDIAN
4899 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4900 #else
4901 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4902 #endif
4903 break;
4904
4905
4906 case DIF_SUBR_HTONLL:
4907 case DIF_SUBR_NTOHLL:
4908 #ifdef _BIG_ENDIAN
4909 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4910 #else
4911 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4912 #endif
4913 break;
4914
4915
4916 case DIF_SUBR_DIRNAME:
4917 case DIF_SUBR_BASENAME: {
4918 char *dest = (char *)mstate->dtms_scratch_ptr;
4919 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4920 uintptr_t src = tupregs[0].dttk_value;
4921 int i, j, len = dtrace_strlen((char *)src, size);
4922 int lastbase = -1, firstbase = -1, lastdir = -1;
4923 int start, end;
4924
4925 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4926 regs[rd] = NULL;
4927 break;
4928 }
4929
4930 if (!DTRACE_INSCRATCH(mstate, size)) {
4931 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4932 regs[rd] = NULL;
4933 break;
4934 }
4935
4936 /*
4937 * The basename and dirname for a zero-length string is
4938 * defined to be "."
4939 */
4940 if (len == 0) {
4941 len = 1;
4942 src = (uintptr_t)".";
4943 }
4944
4945 /*
4946 * Start from the back of the string, moving back toward the
4947 * front until we see a character that isn't a slash. That
4948 * character is the last character in the basename.
4949 */
4950 for (i = len - 1; i >= 0; i--) {
4951 if (dtrace_load8(src + i) != '/')
4952 break;
4953 }
4954
4955 if (i >= 0)
4956 lastbase = i;
4957
4958 /*
4959 * Starting from the last character in the basename, move
4960 * towards the front until we find a slash. The character
4961 * that we processed immediately before that is the first
4962 * character in the basename.
4963 */
4964 for (; i >= 0; i--) {
4965 if (dtrace_load8(src + i) == '/')
4966 break;
4967 }
4968
4969 if (i >= 0)
4970 firstbase = i + 1;
4971
4972 /*
4973 * Now keep going until we find a non-slash character. That
4974 * character is the last character in the dirname.
4975 */
4976 for (; i >= 0; i--) {
4977 if (dtrace_load8(src + i) != '/')
4978 break;
4979 }
4980
4981 if (i >= 0)
4982 lastdir = i;
4983
4984 ASSERT(!(lastbase == -1 && firstbase != -1));
4985 ASSERT(!(firstbase == -1 && lastdir != -1));
4986
4987 if (lastbase == -1) {
4988 /*
4989 * We didn't find a non-slash character. We know that
4990 * the length is non-zero, so the whole string must be
4991 * slashes. In either the dirname or the basename
4992 * case, we return '/'.
4993 */
4994 ASSERT(firstbase == -1);
4995 firstbase = lastbase = lastdir = 0;
4996 }
4997
4998 if (firstbase == -1) {
4999 /*
5000 * The entire string consists only of a basename
5001 * component. If we're looking for dirname, we need
5002 * to change our string to be just "."; if we're
5003 * looking for a basename, we'll just set the first
5004 * character of the basename to be 0.
5005 */
5006 if (subr == DIF_SUBR_DIRNAME) {
5007 ASSERT(lastdir == -1);
5008 src = (uintptr_t)".";
5009 lastdir = 0;
5010 } else {
5011 firstbase = 0;
5012 }
5013 }
5014
5015 if (subr == DIF_SUBR_DIRNAME) {
5016 if (lastdir == -1) {
5017 /*
5018 * We know that we have a slash in the name --
5019 * or lastdir would be set to 0, above. And
5020 * because lastdir is -1, we know that this
5021 * slash must be the first character. (That
5022 * is, the full string must be of the form
5023 * "/basename".) In this case, the last
5024 * character of the directory name is 0.
5025 */
5026 lastdir = 0;
5027 }
5028
5029 start = 0;
5030 end = lastdir;
5031 } else {
5032 ASSERT(subr == DIF_SUBR_BASENAME);
5033 ASSERT(firstbase != -1 && lastbase != -1);
5034 start = firstbase;
5035 end = lastbase;
5036 }
5037
5038 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5039 dest[j] = dtrace_load8(src + i);
5040
5041 dest[j] = '\0';
5042 regs[rd] = (uintptr_t)dest;
5043 mstate->dtms_scratch_ptr += size;
5044 break;
5045 }
5046
5047 case DIF_SUBR_GETF: {
5048 uintptr_t fd = tupregs[0].dttk_value;
5049 uf_info_t *finfo = &curthread->t_procp->p_user.u_finfo;
5050 file_t *fp;
5051
5052 if (!dtrace_priv_proc(state, mstate)) {
5053 regs[rd] = NULL;
5054 break;
5055 }
5056
5057 /*
5058 * This is safe because fi_nfiles only increases, and the
5059 * fi_list array is not freed when the array size doubles.
5060 * (See the comment in flist_grow() for details on the
5061 * management of the u_finfo structure.)
5062 */
5063 fp = fd < finfo->fi_nfiles ? finfo->fi_list[fd].uf_file : NULL;
5064
5065 mstate->dtms_getf = fp;
5066 regs[rd] = (uintptr_t)fp;
5067 break;
5068 }
5069
5070 case DIF_SUBR_CLEANPATH: {
5071 char *dest = (char *)mstate->dtms_scratch_ptr, c;
5072 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5073 uintptr_t src = tupregs[0].dttk_value;
5074 int i = 0, j = 0;
5075 zone_t *z;
5076
5077 if (!dtrace_strcanload(src, size, mstate, vstate)) {
5078 regs[rd] = NULL;
5079 break;
5080 }
5081
5082 if (!DTRACE_INSCRATCH(mstate, size)) {
5083 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5084 regs[rd] = NULL;
5085 break;
5086 }
5087
5088 /*
5089 * Move forward, loading each character.
5090 */
5091 do {
5092 c = dtrace_load8(src + i++);
5093 next:
5094 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
5095 break;
5096
5097 if (c != '/') {
5098 dest[j++] = c;
5099 continue;
5100 }
5101
5102 c = dtrace_load8(src + i++);
5103
5104 if (c == '/') {
5105 /*
5106 * We have two slashes -- we can just advance
5107 * to the next character.
5108 */
5109 goto next;
5110 }
5111
5112 if (c != '.') {
5113 /*
5114 * This is not "." and it's not ".." -- we can
5115 * just store the "/" and this character and
5116 * drive on.
5117 */
5118 dest[j++] = '/';
5119 dest[j++] = c;
5120 continue;
5121 }
5122
5123 c = dtrace_load8(src + i++);
5124
5125 if (c == '/') {
5126 /*
5127 * This is a "/./" component. We're not going
5128 * to store anything in the destination buffer;
5129 * we're just going to go to the next component.
5130 */
5131 goto next;
5132 }
5133
5134 if (c != '.') {
5135 /*
5136 * This is not ".." -- we can just store the
5137 * "/." and this character and continue
5138 * processing.
5139 */
5140 dest[j++] = '/';
5141 dest[j++] = '.';
5142 dest[j++] = c;
5143 continue;
5144 }
5145
5146 c = dtrace_load8(src + i++);
5147
5148 if (c != '/' && c != '\0') {
5149 /*
5150 * This is not ".." -- it's "..[mumble]".
5151 * We'll store the "/.." and this character
5152 * and continue processing.
5153 */
5154 dest[j++] = '/';
5155 dest[j++] = '.';
5156 dest[j++] = '.';
5157 dest[j++] = c;
5158 continue;
5159 }
5160
5161 /*
5162 * This is "/../" or "/..\0". We need to back up
5163 * our destination pointer until we find a "/".
5164 */
5165 i--;
5166 while (j != 0 && dest[--j] != '/')
5167 continue;
5168
5169 if (c == '\0')
5170 dest[++j] = '/';
5171 } while (c != '\0');
5172
5173 dest[j] = '\0';
5174
5175 if (mstate->dtms_getf != NULL &&
5176 !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5177 (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5178 /*
5179 * If we've done a getf() as a part of this ECB and we
5180 * don't have kernel access (and we're not in the global
5181 * zone), check if the path we cleaned up begins with
5182 * the zone's root path, and trim it off if so. Note
5183 * that this is an output cleanliness issue, not a
5184 * security issue: knowing one's zone root path does
5185 * not enable privilege escalation.
5186 */
5187 if (strstr(dest, z->zone_rootpath) == dest)
5188 dest += strlen(z->zone_rootpath) - 1;
5189 }
5190
5191 regs[rd] = (uintptr_t)dest;
5192 mstate->dtms_scratch_ptr += size;
5193 break;
5194 }
5195
5196 case DIF_SUBR_INET_NTOA:
5197 case DIF_SUBR_INET_NTOA6:
5198 case DIF_SUBR_INET_NTOP: {
5199 size_t size;
5200 int af, argi, i;
5201 char *base, *end;
5202
5203 if (subr == DIF_SUBR_INET_NTOP) {
5204 af = (int)tupregs[0].dttk_value;
5205 argi = 1;
5206 } else {
5207 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5208 argi = 0;
5209 }
5210
5211 if (af == AF_INET) {
5212 ipaddr_t ip4;
5213 uint8_t *ptr8, val;
5214
5215 /*
5216 * Safely load the IPv4 address.
5217 */
5218 ip4 = dtrace_load32(tupregs[argi].dttk_value);
5219
5220 /*
5221 * Check an IPv4 string will fit in scratch.
5222 */
5223 size = INET_ADDRSTRLEN;
5224 if (!DTRACE_INSCRATCH(mstate, size)) {
5225 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5226 regs[rd] = NULL;
5227 break;
5228 }
5229 base = (char *)mstate->dtms_scratch_ptr;
5230 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5231
5232 /*
5233 * Stringify as a dotted decimal quad.
5234 */
5235 *end-- = '\0';
5236 ptr8 = (uint8_t *)&ip4;
5237 for (i = 3; i >= 0; i--) {
5238 val = ptr8[i];
5239
5240 if (val == 0) {
5241 *end-- = '0';
5242 } else {
5243 for (; val; val /= 10) {
5244 *end-- = '0' + (val % 10);
5245 }
5246 }
5247
5248 if (i > 0)
5249 *end-- = '.';
5250 }
5251 ASSERT(end + 1 >= base);
5252
5253 } else if (af == AF_INET6) {
5254 struct in6_addr ip6;
5255 int firstzero, tryzero, numzero, v6end;
5256 uint16_t val;
5257 const char digits[] = "0123456789abcdef";
5258
5259 /*
5260 * Stringify using RFC 1884 convention 2 - 16 bit
5261 * hexadecimal values with a zero-run compression.
5262 * Lower case hexadecimal digits are used.
5263 * eg, fe80::214:4fff:fe0b:76c8.
5264 * The IPv4 embedded form is returned for inet_ntop,
5265 * just the IPv4 string is returned for inet_ntoa6.
5266 */
5267
5268 /*
5269 * Safely load the IPv6 address.
5270 */
5271 dtrace_bcopy(
5272 (void *)(uintptr_t)tupregs[argi].dttk_value,
5273 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5274
5275 /*
5276 * Check an IPv6 string will fit in scratch.
5277 */
5278 size = INET6_ADDRSTRLEN;
5279 if (!DTRACE_INSCRATCH(mstate, size)) {
5280 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5281 regs[rd] = NULL;
5282 break;
5283 }
5284 base = (char *)mstate->dtms_scratch_ptr;
5285 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5286 *end-- = '\0';
5287
5288 /*
5289 * Find the longest run of 16 bit zero values
5290 * for the single allowed zero compression - "::".
5291 */
5292 firstzero = -1;
5293 tryzero = -1;
5294 numzero = 1;
5295 for (i = 0; i < sizeof (struct in6_addr); i++) {
5296 if (ip6._S6_un._S6_u8[i] == 0 &&
5297 tryzero == -1 && i % 2 == 0) {
5298 tryzero = i;
5299 continue;
5300 }
5301
5302 if (tryzero != -1 &&
5303 (ip6._S6_un._S6_u8[i] != 0 ||
5304 i == sizeof (struct in6_addr) - 1)) {
5305
5306 if (i - tryzero <= numzero) {
5307 tryzero = -1;
5308 continue;
5309 }
5310
5311 firstzero = tryzero;
5312 numzero = i - i % 2 - tryzero;
5313 tryzero = -1;
5314
5315 if (ip6._S6_un._S6_u8[i] == 0 &&
5316 i == sizeof (struct in6_addr) - 1)
5317 numzero += 2;
5318 }
5319 }
5320 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5321
5322 /*
5323 * Check for an IPv4 embedded address.
5324 */
5325 v6end = sizeof (struct in6_addr) - 2;
5326 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5327 IN6_IS_ADDR_V4COMPAT(&ip6)) {
5328 for (i = sizeof (struct in6_addr) - 1;
5329 i >= DTRACE_V4MAPPED_OFFSET; i--) {
5330 ASSERT(end >= base);
5331
5332 val = ip6._S6_un._S6_u8[i];
5333
5334 if (val == 0) {
5335 *end-- = '0';
5336 } else {
5337 for (; val; val /= 10) {
5338 *end-- = '0' + val % 10;
5339 }
5340 }
5341
5342 if (i > DTRACE_V4MAPPED_OFFSET)
5343 *end-- = '.';
5344 }
5345
5346 if (subr == DIF_SUBR_INET_NTOA6)
5347 goto inetout;
5348
5349 /*
5350 * Set v6end to skip the IPv4 address that
5351 * we have already stringified.
5352 */
5353 v6end = 10;
5354 }
5355
5356 /*
5357 * Build the IPv6 string by working through the
5358 * address in reverse.
5359 */
5360 for (i = v6end; i >= 0; i -= 2) {
5361 ASSERT(end >= base);
5362
5363 if (i == firstzero + numzero - 2) {
5364 *end-- = ':';
5365 *end-- = ':';
5366 i -= numzero - 2;
5367 continue;
5368 }
5369
5370 if (i < 14 && i != firstzero - 2)
5371 *end-- = ':';
5372
5373 val = (ip6._S6_un._S6_u8[i] << 8) +
5374 ip6._S6_un._S6_u8[i + 1];
5375
5376 if (val == 0) {
5377 *end-- = '0';
5378 } else {
5379 for (; val; val /= 16) {
5380 *end-- = digits[val % 16];
5381 }
5382 }
5383 }
5384 ASSERT(end + 1 >= base);
5385
5386 } else {
5387 /*
5388 * The user didn't use AH_INET or AH_INET6.
5389 */
5390 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5391 regs[rd] = NULL;
5392 break;
5393 }
5394
5395 inetout: regs[rd] = (uintptr_t)end + 1;
5396 mstate->dtms_scratch_ptr += size;
5397 break;
5398 }
5399
5400 }
5401 }
5402
5403 /*
5404 * Emulate the execution of DTrace IR instructions specified by the given
5405 * DIF object. This function is deliberately void of assertions as all of
5406 * the necessary checks are handled by a call to dtrace_difo_validate().
5407 */
5408 static uint64_t
5409 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
5410 dtrace_vstate_t *vstate, dtrace_state_t *state)
5411 {
5412 const dif_instr_t *text = difo->dtdo_buf;
5413 const uint_t textlen = difo->dtdo_len;
5414 const char *strtab = difo->dtdo_strtab;
5415 const uint64_t *inttab = difo->dtdo_inttab;
5416
5417 uint64_t rval = 0;
5418 dtrace_statvar_t *svar;
5419 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5420 dtrace_difv_t *v;
5421 volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5422 volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
5423
5424 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5425 uint64_t regs[DIF_DIR_NREGS];
5426 uint64_t *tmp;
5427
5428 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5429 int64_t cc_r;
5430 uint_t pc = 0, id, opc;
5431 uint8_t ttop = 0;
5432 dif_instr_t instr;
5433 uint_t r1, r2, rd;
5434
5435 /*
5436 * We stash the current DIF object into the machine state: we need it
5437 * for subsequent access checking.
5438 */
5439 mstate->dtms_difo = difo;
5440
5441 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
5442
5443 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5444 opc = pc;
5445
5446 instr = text[pc++];
5447 r1 = DIF_INSTR_R1(instr);
5448 r2 = DIF_INSTR_R2(instr);
5449 rd = DIF_INSTR_RD(instr);
5450
5451 switch (DIF_INSTR_OP(instr)) {
5452 case DIF_OP_OR:
5453 regs[rd] = regs[r1] | regs[r2];
5454 break;
5455 case DIF_OP_XOR:
5456 regs[rd] = regs[r1] ^ regs[r2];
5457 break;
5458 case DIF_OP_AND:
5459 regs[rd] = regs[r1] & regs[r2];
5460 break;
5461 case DIF_OP_SLL:
5462 regs[rd] = regs[r1] << regs[r2];
5463 break;
5464 case DIF_OP_SRL:
5465 regs[rd] = regs[r1] >> regs[r2];
5466 break;
5467 case DIF_OP_SUB:
5468 regs[rd] = regs[r1] - regs[r2];
5469 break;
5470 case DIF_OP_ADD:
5471 regs[rd] = regs[r1] + regs[r2];
5472 break;
5473 case DIF_OP_MUL:
5474 regs[rd] = regs[r1] * regs[r2];
5475 break;
5476 case DIF_OP_SDIV:
5477 if (regs[r2] == 0) {
5478 regs[rd] = 0;
5479 *flags |= CPU_DTRACE_DIVZERO;
5480 } else {
5481 regs[rd] = (int64_t)regs[r1] /
5482 (int64_t)regs[r2];
5483 }
5484 break;
5485
5486 case DIF_OP_UDIV:
5487 if (regs[r2] == 0) {
5488 regs[rd] = 0;
5489 *flags |= CPU_DTRACE_DIVZERO;
5490 } else {
5491 regs[rd] = regs[r1] / regs[r2];
5492 }
5493 break;
5494
5495 case DIF_OP_SREM:
5496 if (regs[r2] == 0) {
5497 regs[rd] = 0;
5498 *flags |= CPU_DTRACE_DIVZERO;
5499 } else {
5500 regs[rd] = (int64_t)regs[r1] %
5501 (int64_t)regs[r2];
5502 }
5503 break;
5504
5505 case DIF_OP_UREM:
5506 if (regs[r2] == 0) {
5507 regs[rd] = 0;
5508 *flags |= CPU_DTRACE_DIVZERO;
5509 } else {
5510 regs[rd] = regs[r1] % regs[r2];
5511 }
5512 break;
5513
5514 case DIF_OP_NOT:
5515 regs[rd] = ~regs[r1];
5516 break;
5517 case DIF_OP_MOV:
5518 regs[rd] = regs[r1];
5519 break;
5520 case DIF_OP_CMP:
5521 cc_r = regs[r1] - regs[r2];
5522 cc_n = cc_r < 0;
5523 cc_z = cc_r == 0;
5524 cc_v = 0;
5525 cc_c = regs[r1] < regs[r2];
5526 break;
5527 case DIF_OP_TST:
5528 cc_n = cc_v = cc_c = 0;
5529 cc_z = regs[r1] == 0;
5530 break;
5531 case DIF_OP_BA:
5532 pc = DIF_INSTR_LABEL(instr);
5533 break;
5534 case DIF_OP_BE:
5535 if (cc_z)
5536 pc = DIF_INSTR_LABEL(instr);
5537 break;
5538 case DIF_OP_BNE:
5539 if (cc_z == 0)
5540 pc = DIF_INSTR_LABEL(instr);
5541 break;
5542 case DIF_OP_BG:
5543 if ((cc_z | (cc_n ^ cc_v)) == 0)
5544 pc = DIF_INSTR_LABEL(instr);
5545 break;
5546 case DIF_OP_BGU:
5547 if ((cc_c | cc_z) == 0)
5548 pc = DIF_INSTR_LABEL(instr);
5549 break;
5550 case DIF_OP_BGE:
5551 if ((cc_n ^ cc_v) == 0)
5552 pc = DIF_INSTR_LABEL(instr);
5553 break;
5554 case DIF_OP_BGEU:
5555 if (cc_c == 0)
5556 pc = DIF_INSTR_LABEL(instr);
5557 break;
5558 case DIF_OP_BL:
5559 if (cc_n ^ cc_v)
5560 pc = DIF_INSTR_LABEL(instr);
5561 break;
5562 case DIF_OP_BLU:
5563 if (cc_c)
5564 pc = DIF_INSTR_LABEL(instr);
5565 break;
5566 case DIF_OP_BLE:
5567 if (cc_z | (cc_n ^ cc_v))
5568 pc = DIF_INSTR_LABEL(instr);
5569 break;
5570 case DIF_OP_BLEU:
5571 if (cc_c | cc_z)
5572 pc = DIF_INSTR_LABEL(instr);
5573 break;
5574 case DIF_OP_RLDSB:
5575 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5576 break;
5577 /*FALLTHROUGH*/
5578 case DIF_OP_LDSB:
5579 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5580 break;
5581 case DIF_OP_RLDSH:
5582 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5583 break;
5584 /*FALLTHROUGH*/
5585 case DIF_OP_LDSH:
5586 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5587 break;
5588 case DIF_OP_RLDSW:
5589 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5590 break;
5591 /*FALLTHROUGH*/
5592 case DIF_OP_LDSW:
5593 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5594 break;
5595 case DIF_OP_RLDUB:
5596 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5597 break;
5598 /*FALLTHROUGH*/
5599 case DIF_OP_LDUB:
5600 regs[rd] = dtrace_load8(regs[r1]);
5601 break;
5602 case DIF_OP_RLDUH:
5603 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5604 break;
5605 /*FALLTHROUGH*/
5606 case DIF_OP_LDUH:
5607 regs[rd] = dtrace_load16(regs[r1]);
5608 break;
5609 case DIF_OP_RLDUW:
5610 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5611 break;
5612 /*FALLTHROUGH*/
5613 case DIF_OP_LDUW:
5614 regs[rd] = dtrace_load32(regs[r1]);
5615 break;
5616 case DIF_OP_RLDX:
5617 if (!dtrace_canload(regs[r1], 8, mstate, vstate))
5618 break;
5619 /*FALLTHROUGH*/
5620 case DIF_OP_LDX:
5621 regs[rd] = dtrace_load64(regs[r1]);
5622 break;
5623 case DIF_OP_ULDSB:
5624 regs[rd] = (int8_t)
5625 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5626 break;
5627 case DIF_OP_ULDSH:
5628 regs[rd] = (int16_t)
5629 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5630 break;
5631 case DIF_OP_ULDSW:
5632 regs[rd] = (int32_t)
5633 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5634 break;
5635 case DIF_OP_ULDUB:
5636 regs[rd] =
5637 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5638 break;
5639 case DIF_OP_ULDUH:
5640 regs[rd] =
5641 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5642 break;
5643 case DIF_OP_ULDUW:
5644 regs[rd] =
5645 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5646 break;
5647 case DIF_OP_ULDX:
5648 regs[rd] =
5649 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5650 break;
5651 case DIF_OP_RET:
5652 rval = regs[rd];
5653 pc = textlen;
5654 break;
5655 case DIF_OP_NOP:
5656 break;
5657 case DIF_OP_SETX:
5658 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5659 break;
5660 case DIF_OP_SETS:
5661 regs[rd] = (uint64_t)(uintptr_t)
5662 (strtab + DIF_INSTR_STRING(instr));
5663 break;
5664 case DIF_OP_SCMP: {
5665 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5666 uintptr_t s1 = regs[r1];
5667 uintptr_t s2 = regs[r2];
5668
5669 if (s1 != NULL &&
5670 !dtrace_strcanload(s1, sz, mstate, vstate))
5671 break;
5672 if (s2 != NULL &&
5673 !dtrace_strcanload(s2, sz, mstate, vstate))
5674 break;
5675
5676 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5677
5678 cc_n = cc_r < 0;
5679 cc_z = cc_r == 0;
5680 cc_v = cc_c = 0;
5681 break;
5682 }
5683 case DIF_OP_LDGA:
5684 regs[rd] = dtrace_dif_variable(mstate, state,
5685 r1, regs[r2]);
5686 break;
5687 case DIF_OP_LDGS:
5688 id = DIF_INSTR_VAR(instr);
5689
5690 if (id >= DIF_VAR_OTHER_UBASE) {
5691 uintptr_t a;
5692
5693 id -= DIF_VAR_OTHER_UBASE;
5694 svar = vstate->dtvs_globals[id];
5695 ASSERT(svar != NULL);
5696 v = &svar->dtsv_var;
5697
5698 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5699 regs[rd] = svar->dtsv_data;
5700 break;
5701 }
5702
5703 a = (uintptr_t)svar->dtsv_data;
5704
5705 if (*(uint8_t *)a == UINT8_MAX) {
5706 /*
5707 * If the 0th byte is set to UINT8_MAX
5708 * then this is to be treated as a
5709 * reference to a NULL variable.
5710 */
5711 regs[rd] = NULL;
5712 } else {
5713 regs[rd] = a + sizeof (uint64_t);
5714 }
5715
5716 break;
5717 }
5718
5719 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5720 break;
5721
5722 case DIF_OP_STGS:
5723 id = DIF_INSTR_VAR(instr);
5724
5725 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5726 id -= DIF_VAR_OTHER_UBASE;
5727
5728 svar = vstate->dtvs_globals[id];
5729 ASSERT(svar != NULL);
5730 v = &svar->dtsv_var;
5731
5732 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5733 uintptr_t a = (uintptr_t)svar->dtsv_data;
5734
5735 ASSERT(a != NULL);
5736 ASSERT(svar->dtsv_size != 0);
5737
5738 if (regs[rd] == NULL) {
5739 *(uint8_t *)a = UINT8_MAX;
5740 break;
5741 } else {
5742 *(uint8_t *)a = 0;
5743 a += sizeof (uint64_t);
5744 }
5745 if (!dtrace_vcanload(
5746 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5747 mstate, vstate))
5748 break;
5749
5750 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5751 (void *)a, &v->dtdv_type);
5752 break;
5753 }
5754
5755 svar->dtsv_data = regs[rd];
5756 break;
5757
5758 case DIF_OP_LDTA:
5759 /*
5760 * There are no DTrace built-in thread-local arrays at
5761 * present. This opcode is saved for future work.
5762 */
5763 *flags |= CPU_DTRACE_ILLOP;
5764 regs[rd] = 0;
5765 break;
5766
5767 case DIF_OP_LDLS:
5768 id = DIF_INSTR_VAR(instr);
5769
5770 if (id < DIF_VAR_OTHER_UBASE) {
5771 /*
5772 * For now, this has no meaning.
5773 */
5774 regs[rd] = 0;
5775 break;
5776 }
5777
5778 id -= DIF_VAR_OTHER_UBASE;
5779
5780 ASSERT(id < vstate->dtvs_nlocals);
5781 ASSERT(vstate->dtvs_locals != NULL);
5782
5783 svar = vstate->dtvs_locals[id];
5784 ASSERT(svar != NULL);
5785 v = &svar->dtsv_var;
5786
5787 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5788 uintptr_t a = (uintptr_t)svar->dtsv_data;
5789 size_t sz = v->dtdv_type.dtdt_size;
5790
5791 sz += sizeof (uint64_t);
5792 ASSERT(svar->dtsv_size == NCPU * sz);
5793 a += CPU->cpu_id * sz;
5794
5795 if (*(uint8_t *)a == UINT8_MAX) {
5796 /*
5797 * If the 0th byte is set to UINT8_MAX
5798 * then this is to be treated as a
5799 * reference to a NULL variable.
5800 */
5801 regs[rd] = NULL;
5802 } else {
5803 regs[rd] = a + sizeof (uint64_t);
5804 }
5805
5806 break;
5807 }
5808
5809 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5810 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5811 regs[rd] = tmp[CPU->cpu_id];
5812 break;
5813
5814 case DIF_OP_STLS:
5815 id = DIF_INSTR_VAR(instr);
5816
5817 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5818 id -= DIF_VAR_OTHER_UBASE;
5819 ASSERT(id < vstate->dtvs_nlocals);
5820
5821 ASSERT(vstate->dtvs_locals != NULL);
5822 svar = vstate->dtvs_locals[id];
5823 ASSERT(svar != NULL);
5824 v = &svar->dtsv_var;
5825
5826 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5827 uintptr_t a = (uintptr_t)svar->dtsv_data;
5828 size_t sz = v->dtdv_type.dtdt_size;
5829
5830 sz += sizeof (uint64_t);
5831 ASSERT(svar->dtsv_size == NCPU * sz);
5832 a += CPU->cpu_id * sz;
5833
5834 if (regs[rd] == NULL) {
5835 *(uint8_t *)a = UINT8_MAX;
5836 break;
5837 } else {
5838 *(uint8_t *)a = 0;
5839 a += sizeof (uint64_t);
5840 }
5841
5842 if (!dtrace_vcanload(
5843 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5844 mstate, vstate))
5845 break;
5846
5847 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5848 (void *)a, &v->dtdv_type);
5849 break;
5850 }
5851
5852 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5853 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5854 tmp[CPU->cpu_id] = regs[rd];
5855 break;
5856
5857 case DIF_OP_LDTS: {
5858 dtrace_dynvar_t *dvar;
5859 dtrace_key_t *key;
5860
5861 id = DIF_INSTR_VAR(instr);
5862 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5863 id -= DIF_VAR_OTHER_UBASE;
5864 v = &vstate->dtvs_tlocals[id];
5865
5866 key = &tupregs[DIF_DTR_NREGS];
5867 key[0].dttk_value = (uint64_t)id;
5868 key[0].dttk_size = 0;
5869 DTRACE_TLS_THRKEY(key[1].dttk_value);
5870 key[1].dttk_size = 0;
5871
5872 dvar = dtrace_dynvar(dstate, 2, key,
5873 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5874 mstate, vstate);
5875
5876 if (dvar == NULL) {
5877 regs[rd] = 0;
5878 break;
5879 }
5880
5881 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5882 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5883 } else {
5884 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5885 }
5886
5887 break;
5888 }
5889
5890 case DIF_OP_STTS: {
5891 dtrace_dynvar_t *dvar;
5892 dtrace_key_t *key;
5893
5894 id = DIF_INSTR_VAR(instr);
5895 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5896 id -= DIF_VAR_OTHER_UBASE;
5897
5898 key = &tupregs[DIF_DTR_NREGS];
5899 key[0].dttk_value = (uint64_t)id;
5900 key[0].dttk_size = 0;
5901 DTRACE_TLS_THRKEY(key[1].dttk_value);
5902 key[1].dttk_size = 0;
5903 v = &vstate->dtvs_tlocals[id];
5904
5905 dvar = dtrace_dynvar(dstate, 2, key,
5906 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5907 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5908 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5909 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5910
5911 /*
5912 * Given that we're storing to thread-local data,
5913 * we need to flush our predicate cache.
5914 */
5915 curthread->t_predcache = NULL;
5916
5917 if (dvar == NULL)
5918 break;
5919
5920 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5921 if (!dtrace_vcanload(
5922 (void *)(uintptr_t)regs[rd],
5923 &v->dtdv_type, mstate, vstate))
5924 break;
5925
5926 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5927 dvar->dtdv_data, &v->dtdv_type);
5928 } else {
5929 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5930 }
5931
5932 break;
5933 }
5934
5935 case DIF_OP_SRA:
5936 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5937 break;
5938
5939 case DIF_OP_CALL:
5940 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5941 regs, tupregs, ttop, mstate, state);
5942 break;
5943
5944 case DIF_OP_PUSHTR:
5945 if (ttop == DIF_DTR_NREGS) {
5946 *flags |= CPU_DTRACE_TUPOFLOW;
5947 break;
5948 }
5949
5950 if (r1 == DIF_TYPE_STRING) {
5951 /*
5952 * If this is a string type and the size is 0,
5953 * we'll use the system-wide default string
5954 * size. Note that we are _not_ looking at
5955 * the value of the DTRACEOPT_STRSIZE option;
5956 * had this been set, we would expect to have
5957 * a non-zero size value in the "pushtr".
5958 */
5959 tupregs[ttop].dttk_size =
5960 dtrace_strlen((char *)(uintptr_t)regs[rd],
5961 regs[r2] ? regs[r2] :
5962 dtrace_strsize_default) + 1;
5963 } else {
5964 tupregs[ttop].dttk_size = regs[r2];
5965 }
5966
5967 tupregs[ttop++].dttk_value = regs[rd];
5968 break;
5969
5970 case DIF_OP_PUSHTV:
5971 if (ttop == DIF_DTR_NREGS) {
5972 *flags |= CPU_DTRACE_TUPOFLOW;
5973 break;
5974 }
5975
5976 tupregs[ttop].dttk_value = regs[rd];
5977 tupregs[ttop++].dttk_size = 0;
5978 break;
5979
5980 case DIF_OP_POPTS:
5981 if (ttop != 0)
5982 ttop--;
5983 break;
5984
5985 case DIF_OP_FLUSHTS:
5986 ttop = 0;
5987 break;
5988
5989 case DIF_OP_LDGAA:
5990 case DIF_OP_LDTAA: {
5991 dtrace_dynvar_t *dvar;
5992 dtrace_key_t *key = tupregs;
5993 uint_t nkeys = ttop;
5994
5995 id = DIF_INSTR_VAR(instr);
5996 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5997 id -= DIF_VAR_OTHER_UBASE;
5998
5999 key[nkeys].dttk_value = (uint64_t)id;
6000 key[nkeys++].dttk_size = 0;
6001
6002 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6003 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6004 key[nkeys++].dttk_size = 0;
6005 v = &vstate->dtvs_tlocals[id];
6006 } else {
6007 v = &vstate->dtvs_globals[id]->dtsv_var;
6008 }
6009
6010 dvar = dtrace_dynvar(dstate, nkeys, key,
6011 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6012 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6013 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6014
6015 if (dvar == NULL) {
6016 regs[rd] = 0;
6017 break;
6018 }
6019
6020 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6021 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6022 } else {
6023 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6024 }
6025
6026 break;
6027 }
6028
6029 case DIF_OP_STGAA:
6030 case DIF_OP_STTAA: {
6031 dtrace_dynvar_t *dvar;
6032 dtrace_key_t *key = tupregs;
6033 uint_t nkeys = ttop;
6034
6035 id = DIF_INSTR_VAR(instr);
6036 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6037 id -= DIF_VAR_OTHER_UBASE;
6038
6039 key[nkeys].dttk_value = (uint64_t)id;
6040 key[nkeys++].dttk_size = 0;
6041
6042 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6043 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6044 key[nkeys++].dttk_size = 0;
6045 v = &vstate->dtvs_tlocals[id];
6046 } else {
6047 v = &vstate->dtvs_globals[id]->dtsv_var;
6048 }
6049
6050 dvar = dtrace_dynvar(dstate, nkeys, key,
6051 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6052 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6053 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6054 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6055
6056 if (dvar == NULL)
6057 break;
6058
6059 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6060 if (!dtrace_vcanload(
6061 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6062 mstate, vstate))
6063 break;
6064
6065 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6066 dvar->dtdv_data, &v->dtdv_type);
6067 } else {
6068 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6069 }
6070
6071 break;
6072 }
6073
6074 case DIF_OP_ALLOCS: {
6075 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6076 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6077
6078 /*
6079 * Rounding up the user allocation size could have
6080 * overflowed large, bogus allocations (like -1ULL) to
6081 * 0.
6082 */
6083 if (size < regs[r1] ||
6084 !DTRACE_INSCRATCH(mstate, size)) {
6085 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6086 regs[rd] = NULL;
6087 break;
6088 }
6089
6090 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6091 mstate->dtms_scratch_ptr += size;
6092 regs[rd] = ptr;
6093 break;
6094 }
6095
6096 case DIF_OP_COPYS:
6097 if (!dtrace_canstore(regs[rd], regs[r2],
6098 mstate, vstate)) {
6099 *flags |= CPU_DTRACE_BADADDR;
6100 *illval = regs[rd];
6101 break;
6102 }
6103
6104 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6105 break;
6106
6107 dtrace_bcopy((void *)(uintptr_t)regs[r1],
6108 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6109 break;
6110
6111 case DIF_OP_STB:
6112 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6113 *flags |= CPU_DTRACE_BADADDR;
6114 *illval = regs[rd];
6115 break;
6116 }
6117 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6118 break;
6119
6120 case DIF_OP_STH:
6121 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6122 *flags |= CPU_DTRACE_BADADDR;
6123 *illval = regs[rd];
6124 break;
6125 }
6126 if (regs[rd] & 1) {
6127 *flags |= CPU_DTRACE_BADALIGN;
6128 *illval = regs[rd];
6129 break;
6130 }
6131 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6132 break;
6133
6134 case DIF_OP_STW:
6135 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6136 *flags |= CPU_DTRACE_BADADDR;
6137 *illval = regs[rd];
6138 break;
6139 }
6140 if (regs[rd] & 3) {
6141 *flags |= CPU_DTRACE_BADALIGN;
6142 *illval = regs[rd];
6143 break;
6144 }
6145 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6146 break;
6147
6148 case DIF_OP_STX:
6149 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6150 *flags |= CPU_DTRACE_BADADDR;
6151 *illval = regs[rd];
6152 break;
6153 }
6154 if (regs[rd] & 7) {
6155 *flags |= CPU_DTRACE_BADALIGN;
6156 *illval = regs[rd];
6157 break;
6158 }
6159 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6160 break;
6161 }
6162 }
6163
6164 if (!(*flags & CPU_DTRACE_FAULT))
6165 return (rval);
6166
6167 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6168 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6169
6170 return (0);
6171 }
6172
6173 static void
6174 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6175 {
6176 dtrace_probe_t *probe = ecb->dte_probe;
6177 dtrace_provider_t *prov = probe->dtpr_provider;
6178 char c[DTRACE_FULLNAMELEN + 80], *str;
6179 char *msg = "dtrace: breakpoint action at probe ";
6180 char *ecbmsg = " (ecb ";
6181 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6182 uintptr_t val = (uintptr_t)ecb;
6183 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6184
6185 if (dtrace_destructive_disallow)
6186 return;
6187
6188 /*
6189 * It's impossible to be taking action on the NULL probe.
6190 */
6191 ASSERT(probe != NULL);
6192
6193 /*
6194 * This is a poor man's (destitute man's?) sprintf(): we want to
6195 * print the provider name, module name, function name and name of
6196 * the probe, along with the hex address of the ECB with the breakpoint
6197 * action -- all of which we must place in the character buffer by
6198 * hand.
6199 */
6200 while (*msg != '\0')
6201 c[i++] = *msg++;
6202
6203 for (str = prov->dtpv_name; *str != '\0'; str++)
6204 c[i++] = *str;
6205 c[i++] = ':';
6206
6207 for (str = probe->dtpr_mod; *str != '\0'; str++)
6208 c[i++] = *str;
6209 c[i++] = ':';
6210
6211 for (str = probe->dtpr_func; *str != '\0'; str++)
6212 c[i++] = *str;
6213 c[i++] = ':';
6214
6215 for (str = probe->dtpr_name; *str != '\0'; str++)
6216 c[i++] = *str;
6217
6218 while (*ecbmsg != '\0')
6219 c[i++] = *ecbmsg++;
6220
6221 while (shift >= 0) {
6222 mask = (uintptr_t)0xf << shift;
6223
6224 if (val >= ((uintptr_t)1 << shift))
6225 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6226 shift -= 4;
6227 }
6228
6229 c[i++] = ')';
6230 c[i] = '\0';
6231
6232 debug_enter(c);
6233 }
6234
6235 static void
6236 dtrace_action_panic(dtrace_ecb_t *ecb)
6237 {
6238 dtrace_probe_t *probe = ecb->dte_probe;
6239
6240 /*
6241 * It's impossible to be taking action on the NULL probe.
6242 */
6243 ASSERT(probe != NULL);
6244
6245 if (dtrace_destructive_disallow)
6246 return;
6247
6248 if (dtrace_panicked != NULL)
6249 return;
6250
6251 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6252 return;
6253
6254 /*
6255 * We won the right to panic. (We want to be sure that only one
6256 * thread calls panic() from dtrace_probe(), and that panic() is
6257 * called exactly once.)
6258 */
6259 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6260 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6261 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6262 }
6263
6264 static void
6265 dtrace_action_raise(uint64_t sig)
6266 {
6267 if (dtrace_destructive_disallow)
6268 return;
6269
6270 if (sig >= NSIG) {
6271 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6272 return;
6273 }
6274
6275 /*
6276 * raise() has a queue depth of 1 -- we ignore all subsequent
6277 * invocations of the raise() action.
6278 */
6279 if (curthread->t_dtrace_sig == 0)
6280 curthread->t_dtrace_sig = (uint8_t)sig;
6281
6282 curthread->t_sig_check = 1;
6283 aston(curthread);
6284 }
6285
6286 static void
6287 dtrace_action_stop(void)
6288 {
6289 if (dtrace_destructive_disallow)
6290 return;
6291
6292 if (!curthread->t_dtrace_stop) {
6293 curthread->t_dtrace_stop = 1;
6294 curthread->t_sig_check = 1;
6295 aston(curthread);
6296 }
6297 }
6298
6299 static void
6300 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
6301 {
6302 hrtime_t now;
6303 volatile uint16_t *flags;
6304 cpu_t *cpu = CPU;
6305
6306 if (dtrace_destructive_disallow)
6307 return;
6308
6309 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
6310
6311 now = dtrace_gethrtime();
6312
6313 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
6314 /*
6315 * We need to advance the mark to the current time.
6316 */
6317 cpu->cpu_dtrace_chillmark = now;
6318 cpu->cpu_dtrace_chilled = 0;
6319 }
6320
6321 /*
6322 * Now check to see if the requested chill time would take us over
6323 * the maximum amount of time allowed in the chill interval. (Or
6324 * worse, if the calculation itself induces overflow.)
6325 */
6326 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
6327 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
6328 *flags |= CPU_DTRACE_ILLOP;
6329 return;
6330 }
6331
6332 while (dtrace_gethrtime() - now < val)
6333 continue;
6334
6335 /*
6336 * Normally, we assure that the value of the variable "timestamp" does
6337 * not change within an ECB. The presence of chill() represents an
6338 * exception to this rule, however.
6339 */
6340 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
6341 cpu->cpu_dtrace_chilled += val;
6342 }
6343
6344 static void
6345 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
6346 uint64_t *buf, uint64_t arg)
6347 {
6348 int nframes = DTRACE_USTACK_NFRAMES(arg);
6349 int strsize = DTRACE_USTACK_STRSIZE(arg);
6350 uint64_t *pcs = &buf[1], *fps;
6351 char *str = (char *)&pcs[nframes];
6352 int size, offs = 0, i, j;
6353 uintptr_t old = mstate->dtms_scratch_ptr, saved;
6354 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6355 char *sym;
6356
6357 /*
6358 * Should be taking a faster path if string space has not been
6359 * allocated.
6360 */
6361 ASSERT(strsize != 0);
6362
6363 /*
6364 * We will first allocate some temporary space for the frame pointers.
6365 */
6366 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6367 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
6368 (nframes * sizeof (uint64_t));
6369
6370 if (!DTRACE_INSCRATCH(mstate, size)) {
6371 /*
6372 * Not enough room for our frame pointers -- need to indicate
6373 * that we ran out of scratch space.
6374 */
6375 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6376 return;
6377 }
6378
6379 mstate->dtms_scratch_ptr += size;
6380 saved = mstate->dtms_scratch_ptr;
6381
6382 /*
6383 * Now get a stack with both program counters and frame pointers.
6384 */
6385 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6386 dtrace_getufpstack(buf, fps, nframes + 1);
6387 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6388
6389 /*
6390 * If that faulted, we're cooked.
6391 */
6392 if (*flags & CPU_DTRACE_FAULT)
6393 goto out;
6394
6395 /*
6396 * Now we want to walk up the stack, calling the USTACK helper. For
6397 * each iteration, we restore the scratch pointer.
6398 */
6399 for (i = 0; i < nframes; i++) {
6400 mstate->dtms_scratch_ptr = saved;
6401
6402 if (offs >= strsize)
6403 break;
6404
6405 sym = (char *)(uintptr_t)dtrace_helper(
6406 DTRACE_HELPER_ACTION_USTACK,
6407 mstate, state, pcs[i], fps[i]);
6408
6409 /*
6410 * If we faulted while running the helper, we're going to
6411 * clear the fault and null out the corresponding string.
6412 */
6413 if (*flags & CPU_DTRACE_FAULT) {
6414 *flags &= ~CPU_DTRACE_FAULT;
6415 str[offs++] = '\0';
6416 continue;
6417 }
6418
6419 if (sym == NULL) {
6420 str[offs++] = '\0';
6421 continue;
6422 }
6423
6424 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6425
6426 /*
6427 * Now copy in the string that the helper returned to us.
6428 */
6429 for (j = 0; offs + j < strsize; j++) {
6430 if ((str[offs + j] = sym[j]) == '\0')
6431 break;
6432 }
6433
6434 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6435
6436 offs += j + 1;
6437 }
6438
6439 if (offs >= strsize) {
6440 /*
6441 * If we didn't have room for all of the strings, we don't
6442 * abort processing -- this needn't be a fatal error -- but we
6443 * still want to increment a counter (dts_stkstroverflows) to
6444 * allow this condition to be warned about. (If this is from
6445 * a jstack() action, it is easily tuned via jstackstrsize.)
6446 */
6447 dtrace_error(&state->dts_stkstroverflows);
6448 }
6449
6450 while (offs < strsize)
6451 str[offs++] = '\0';
6452
6453 out:
6454 mstate->dtms_scratch_ptr = old;
6455 }
6456
6457 /*
6458 * If you're looking for the epicenter of DTrace, you just found it. This
6459 * is the function called by the provider to fire a probe -- from which all
6460 * subsequent probe-context DTrace activity emanates.
6461 */
6462 void
6463 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6464 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6465 {
6466 processorid_t cpuid;
6467 dtrace_icookie_t cookie;
6468 dtrace_probe_t *probe;
6469 dtrace_mstate_t mstate;
6470 dtrace_ecb_t *ecb;
6471 dtrace_action_t *act;
6472 intptr_t offs;
6473 size_t size;
6474 int vtime, onintr;
6475 volatile uint16_t *flags;
6476 hrtime_t now;
6477
6478 /*
6479 * Kick out immediately if this CPU is still being born (in which case
6480 * curthread will be set to -1) or the current thread can't allow
6481 * probes in its current context.
6482 */
6483 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6484 return;
6485
6486 cookie = dtrace_interrupt_disable();
6487 probe = dtrace_probes[id - 1];
6488 cpuid = CPU->cpu_id;
6489 onintr = CPU_ON_INTR(CPU);
6490
6491 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6492 probe->dtpr_predcache == curthread->t_predcache) {
6493 /*
6494 * We have hit in the predicate cache; we know that
6495 * this predicate would evaluate to be false.
6496 */
6497 dtrace_interrupt_enable(cookie);
6498 return;
6499 }
6500
6501 if (panic_quiesce) {
6502 /*
6503 * We don't trace anything if we're panicking.
6504 */
6505 dtrace_interrupt_enable(cookie);
6506 return;
6507 }
6508
6509 now = dtrace_gethrtime();
6510 vtime = dtrace_vtime_references != 0;
6511
6512 if (vtime && curthread->t_dtrace_start)
6513 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6514
6515 mstate.dtms_difo = NULL;
6516 mstate.dtms_probe = probe;
6517 mstate.dtms_strtok = NULL;
6518 mstate.dtms_arg[0] = arg0;
6519 mstate.dtms_arg[1] = arg1;
6520 mstate.dtms_arg[2] = arg2;
6521 mstate.dtms_arg[3] = arg3;
6522 mstate.dtms_arg[4] = arg4;
6523
6524 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6525
6526 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6527 dtrace_predicate_t *pred = ecb->dte_predicate;
6528 dtrace_state_t *state = ecb->dte_state;
6529 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6530 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6531 dtrace_vstate_t *vstate = &state->dts_vstate;
6532 dtrace_provider_t *prov = probe->dtpr_provider;
6533 uint64_t tracememsize = 0;
6534 int committed = 0;
6535 caddr_t tomax;
6536
6537 /*
6538 * A little subtlety with the following (seemingly innocuous)
6539 * declaration of the automatic 'val': by looking at the
6540 * code, you might think that it could be declared in the
6541 * action processing loop, below. (That is, it's only used in
6542 * the action processing loop.) However, it must be declared
6543 * out of that scope because in the case of DIF expression
6544 * arguments to aggregating actions, one iteration of the
6545 * action loop will use the last iteration's value.
6546 */
6547 #ifdef lint
6548 uint64_t val = 0;
6549 #else
6550 uint64_t val;
6551 #endif
6552
6553 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6554 mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
6555 mstate.dtms_getf = NULL;
6556
6557 *flags &= ~CPU_DTRACE_ERROR;
6558
6559 if (prov == dtrace_provider) {
6560 /*
6561 * If dtrace itself is the provider of this probe,
6562 * we're only going to continue processing the ECB if
6563 * arg0 (the dtrace_state_t) is equal to the ECB's
6564 * creating state. (This prevents disjoint consumers
6565 * from seeing one another's metaprobes.)
6566 */
6567 if (arg0 != (uint64_t)(uintptr_t)state)
6568 continue;
6569 }
6570
6571 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6572 /*
6573 * We're not currently active. If our provider isn't
6574 * the dtrace pseudo provider, we're not interested.
6575 */
6576 if (prov != dtrace_provider)
6577 continue;
6578
6579 /*
6580 * Now we must further check if we are in the BEGIN
6581 * probe. If we are, we will only continue processing
6582 * if we're still in WARMUP -- if one BEGIN enabling
6583 * has invoked the exit() action, we don't want to
6584 * evaluate subsequent BEGIN enablings.
6585 */
6586 if (probe->dtpr_id == dtrace_probeid_begin &&
6587 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6588 ASSERT(state->dts_activity ==
6589 DTRACE_ACTIVITY_DRAINING);
6590 continue;
6591 }
6592 }
6593
6594 if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
6595 continue;
6596
6597 if (now - state->dts_alive > dtrace_deadman_timeout) {
6598 /*
6599 * We seem to be dead. Unless we (a) have kernel
6600 * destructive permissions (b) have explicitly enabled
6601 * destructive actions and (c) destructive actions have
6602 * not been disabled, we're going to transition into
6603 * the KILLED state, from which no further processing
6604 * on this state will be performed.
6605 */
6606 if (!dtrace_priv_kernel_destructive(state) ||
6607 !state->dts_cred.dcr_destructive ||
6608 dtrace_destructive_disallow) {
6609 void *activity = &state->dts_activity;
6610 dtrace_activity_t current;
6611
6612 do {
6613 current = state->dts_activity;
6614 } while (dtrace_cas32(activity, current,
6615 DTRACE_ACTIVITY_KILLED) != current);
6616
6617 continue;
6618 }
6619 }
6620
6621 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6622 ecb->dte_alignment, state, &mstate)) < 0)
6623 continue;
6624
6625 tomax = buf->dtb_tomax;
6626 ASSERT(tomax != NULL);
6627
6628 if (ecb->dte_size != 0) {
6629 dtrace_rechdr_t dtrh;
6630 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6631 mstate.dtms_timestamp = dtrace_gethrtime();
6632 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6633 }
6634 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6635 dtrh.dtrh_epid = ecb->dte_epid;
6636 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6637 mstate.dtms_timestamp);
6638 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6639 }
6640
6641 mstate.dtms_epid = ecb->dte_epid;
6642 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6643
6644 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6645 mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
6646
6647 if (pred != NULL) {
6648 dtrace_difo_t *dp = pred->dtp_difo;
6649 int rval;
6650
6651 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6652
6653 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6654 dtrace_cacheid_t cid = probe->dtpr_predcache;
6655
6656 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6657 /*
6658 * Update the predicate cache...
6659 */
6660 ASSERT(cid == pred->dtp_cacheid);
6661 curthread->t_predcache = cid;
6662 }
6663
6664 continue;
6665 }
6666 }
6667
6668 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6669 act != NULL; act = act->dta_next) {
6670 size_t valoffs;
6671 dtrace_difo_t *dp;
6672 dtrace_recdesc_t *rec = &act->dta_rec;
6673
6674 size = rec->dtrd_size;
6675 valoffs = offs + rec->dtrd_offset;
6676
6677 if (DTRACEACT_ISAGG(act->dta_kind)) {
6678 uint64_t v = 0xbad;
6679 dtrace_aggregation_t *agg;
6680
6681 agg = (dtrace_aggregation_t *)act;
6682
6683 if ((dp = act->dta_difo) != NULL)
6684 v = dtrace_dif_emulate(dp,
6685 &mstate, vstate, state);
6686
6687 if (*flags & CPU_DTRACE_ERROR)
6688 continue;
6689
6690 /*
6691 * Note that we always pass the expression
6692 * value from the previous iteration of the
6693 * action loop. This value will only be used
6694 * if there is an expression argument to the
6695 * aggregating action, denoted by the
6696 * dtag_hasarg field.
6697 */
6698 dtrace_aggregate(agg, buf,
6699 offs, aggbuf, v, val);
6700 continue;
6701 }
6702
6703 switch (act->dta_kind) {
6704 case DTRACEACT_STOP:
6705 if (dtrace_priv_proc_destructive(state,
6706 &mstate))
6707 dtrace_action_stop();
6708 continue;
6709
6710 case DTRACEACT_BREAKPOINT:
6711 if (dtrace_priv_kernel_destructive(state))
6712 dtrace_action_breakpoint(ecb);
6713 continue;
6714
6715 case DTRACEACT_PANIC:
6716 if (dtrace_priv_kernel_destructive(state))
6717 dtrace_action_panic(ecb);
6718 continue;
6719
6720 case DTRACEACT_STACK:
6721 if (!dtrace_priv_kernel(state))
6722 continue;
6723
6724 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6725 size / sizeof (pc_t), probe->dtpr_aframes,
6726 DTRACE_ANCHORED(probe) ? NULL :
6727 (uint32_t *)arg0);
6728
6729 continue;
6730
6731 case DTRACEACT_JSTACK:
6732 case DTRACEACT_USTACK:
6733 if (!dtrace_priv_proc(state, &mstate))
6734 continue;
6735
6736 /*
6737 * See comment in DIF_VAR_PID.
6738 */
6739 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6740 CPU_ON_INTR(CPU)) {
6741 int depth = DTRACE_USTACK_NFRAMES(
6742 rec->dtrd_arg) + 1;
6743
6744 dtrace_bzero((void *)(tomax + valoffs),
6745 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6746 + depth * sizeof (uint64_t));
6747
6748 continue;
6749 }
6750
6751 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6752 curproc->p_dtrace_helpers != NULL) {
6753 /*
6754 * This is the slow path -- we have
6755 * allocated string space, and we're
6756 * getting the stack of a process that
6757 * has helpers. Call into a separate
6758 * routine to perform this processing.
6759 */
6760 dtrace_action_ustack(&mstate, state,
6761 (uint64_t *)(tomax + valoffs),
6762 rec->dtrd_arg);
6763 continue;
6764 }
6765
6766 /*
6767 * Clear the string space, since there's no
6768 * helper to do it for us.
6769 */
6770 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
6771 int depth = DTRACE_USTACK_NFRAMES(
6772 rec->dtrd_arg);
6773 size_t strsize = DTRACE_USTACK_STRSIZE(
6774 rec->dtrd_arg);
6775 uint64_t *buf = (uint64_t *)(tomax +
6776 valoffs);
6777 void *strspace = &buf[depth + 1];
6778
6779 dtrace_bzero(strspace,
6780 MIN(depth, strsize));
6781 }
6782
6783 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6784 dtrace_getupcstack((uint64_t *)
6785 (tomax + valoffs),
6786 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6787 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6788 continue;
6789
6790 default:
6791 break;
6792 }
6793
6794 dp = act->dta_difo;
6795 ASSERT(dp != NULL);
6796
6797 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6798
6799 if (*flags & CPU_DTRACE_ERROR)
6800 continue;
6801
6802 switch (act->dta_kind) {
6803 case DTRACEACT_SPECULATE: {
6804 dtrace_rechdr_t *dtrh;
6805
6806 ASSERT(buf == &state->dts_buffer[cpuid]);
6807 buf = dtrace_speculation_buffer(state,
6808 cpuid, val);
6809
6810 if (buf == NULL) {
6811 *flags |= CPU_DTRACE_DROP;
6812 continue;
6813 }
6814
6815 offs = dtrace_buffer_reserve(buf,
6816 ecb->dte_needed, ecb->dte_alignment,
6817 state, NULL);
6818
6819 if (offs < 0) {
6820 *flags |= CPU_DTRACE_DROP;
6821 continue;
6822 }
6823
6824 tomax = buf->dtb_tomax;
6825 ASSERT(tomax != NULL);
6826
6827 if (ecb->dte_size == 0)
6828 continue;
6829
6830 ASSERT3U(ecb->dte_size, >=,
6831 sizeof (dtrace_rechdr_t));
6832 dtrh = ((void *)(tomax + offs));
6833 dtrh->dtrh_epid = ecb->dte_epid;
6834 /*
6835 * When the speculation is committed, all of
6836 * the records in the speculative buffer will
6837 * have their timestamps set to the commit
6838 * time. Until then, it is set to a sentinel
6839 * value, for debugability.
6840 */
6841 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6842 continue;
6843 }
6844
6845 case DTRACEACT_CHILL:
6846 if (dtrace_priv_kernel_destructive(state))
6847 dtrace_action_chill(&mstate, val);
6848 continue;
6849
6850 case DTRACEACT_RAISE:
6851 if (dtrace_priv_proc_destructive(state,
6852 &mstate))
6853 dtrace_action_raise(val);
6854 continue;
6855
6856 case DTRACEACT_COMMIT:
6857 ASSERT(!committed);
6858
6859 /*
6860 * We need to commit our buffer state.
6861 */
6862 if (ecb->dte_size)
6863 buf->dtb_offset = offs + ecb->dte_size;
6864 buf = &state->dts_buffer[cpuid];
6865 dtrace_speculation_commit(state, cpuid, val);
6866 committed = 1;
6867 continue;
6868
6869 case DTRACEACT_DISCARD:
6870 dtrace_speculation_discard(state, cpuid, val);
6871 continue;
6872
6873 case DTRACEACT_DIFEXPR:
6874 case DTRACEACT_LIBACT:
6875 case DTRACEACT_PRINTF:
6876 case DTRACEACT_PRINTA:
6877 case DTRACEACT_SYSTEM:
6878 case DTRACEACT_FREOPEN:
6879 case DTRACEACT_TRACEMEM:
6880 break;
6881
6882 case DTRACEACT_TRACEMEM_DYNSIZE:
6883 tracememsize = val;
6884 break;
6885
6886 case DTRACEACT_SYM:
6887 case DTRACEACT_MOD:
6888 if (!dtrace_priv_kernel(state))
6889 continue;
6890 break;
6891
6892 case DTRACEACT_USYM:
6893 case DTRACEACT_UMOD:
6894 case DTRACEACT_UADDR: {
6895 struct pid *pid = curthread->t_procp->p_pidp;
6896
6897 if (!dtrace_priv_proc(state, &mstate))
6898 continue;
6899
6900 DTRACE_STORE(uint64_t, tomax,
6901 valoffs, (uint64_t)pid->pid_id);
6902 DTRACE_STORE(uint64_t, tomax,
6903 valoffs + sizeof (uint64_t), val);
6904
6905 continue;
6906 }
6907
6908 case DTRACEACT_EXIT: {
6909 /*
6910 * For the exit action, we are going to attempt
6911 * to atomically set our activity to be
6912 * draining. If this fails (either because
6913 * another CPU has beat us to the exit action,
6914 * or because our current activity is something
6915 * other than ACTIVE or WARMUP), we will
6916 * continue. This assures that the exit action
6917 * can be successfully recorded at most once
6918 * when we're in the ACTIVE state. If we're
6919 * encountering the exit() action while in
6920 * COOLDOWN, however, we want to honor the new
6921 * status code. (We know that we're the only
6922 * thread in COOLDOWN, so there is no race.)
6923 */
6924 void *activity = &state->dts_activity;
6925 dtrace_activity_t current = state->dts_activity;
6926
6927 if (current == DTRACE_ACTIVITY_COOLDOWN)
6928 break;
6929
6930 if (current != DTRACE_ACTIVITY_WARMUP)
6931 current = DTRACE_ACTIVITY_ACTIVE;
6932
6933 if (dtrace_cas32(activity, current,
6934 DTRACE_ACTIVITY_DRAINING) != current) {
6935 *flags |= CPU_DTRACE_DROP;
6936 continue;
6937 }
6938
6939 break;
6940 }
6941
6942 default:
6943 ASSERT(0);
6944 }
6945
6946 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6947 uintptr_t end = valoffs + size;
6948
6949 if (tracememsize != 0 &&
6950 valoffs + tracememsize < end) {
6951 end = valoffs + tracememsize;
6952 tracememsize = 0;
6953 }
6954
6955 if (!dtrace_vcanload((void *)(uintptr_t)val,
6956 &dp->dtdo_rtype, &mstate, vstate))
6957 continue;
6958
6959 /*
6960 * If this is a string, we're going to only
6961 * load until we find the zero byte -- after
6962 * which we'll store zero bytes.
6963 */
6964 if (dp->dtdo_rtype.dtdt_kind ==
6965 DIF_TYPE_STRING) {
6966 char c = '\0' + 1;
6967 int intuple = act->dta_intuple;
6968 size_t s;
6969
6970 for (s = 0; s < size; s++) {
6971 if (c != '\0')
6972 c = dtrace_load8(val++);
6973
6974 DTRACE_STORE(uint8_t, tomax,
6975 valoffs++, c);
6976
6977 if (c == '\0' && intuple)
6978 break;
6979 }
6980
6981 continue;
6982 }
6983
6984 while (valoffs < end) {
6985 DTRACE_STORE(uint8_t, tomax, valoffs++,
6986 dtrace_load8(val++));
6987 }
6988
6989 continue;
6990 }
6991
6992 switch (size) {
6993 case 0:
6994 break;
6995
6996 case sizeof (uint8_t):
6997 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6998 break;
6999 case sizeof (uint16_t):
7000 DTRACE_STORE(uint16_t, tomax, valoffs, val);
7001 break;
7002 case sizeof (uint32_t):
7003 DTRACE_STORE(uint32_t, tomax, valoffs, val);
7004 break;
7005 case sizeof (uint64_t):
7006 DTRACE_STORE(uint64_t, tomax, valoffs, val);
7007 break;
7008 default:
7009 /*
7010 * Any other size should have been returned by
7011 * reference, not by value.
7012 */
7013 ASSERT(0);
7014 break;
7015 }
7016 }
7017
7018 if (*flags & CPU_DTRACE_DROP)
7019 continue;
7020
7021 if (*flags & CPU_DTRACE_FAULT) {
7022 int ndx;
7023 dtrace_action_t *err;
7024
7025 buf->dtb_errors++;
7026
7027 if (probe->dtpr_id == dtrace_probeid_error) {
7028 /*
7029 * There's nothing we can do -- we had an
7030 * error on the error probe. We bump an
7031 * error counter to at least indicate that
7032 * this condition happened.
7033 */
7034 dtrace_error(&state->dts_dblerrors);
7035 continue;
7036 }
7037
7038 if (vtime) {
7039 /*
7040 * Before recursing on dtrace_probe(), we
7041 * need to explicitly clear out our start
7042 * time to prevent it from being accumulated
7043 * into t_dtrace_vtime.
7044 */
7045 curthread->t_dtrace_start = 0;
7046 }
7047
7048 /*
7049 * Iterate over the actions to figure out which action
7050 * we were processing when we experienced the error.
7051 * Note that act points _past_ the faulting action; if
7052 * act is ecb->dte_action, the fault was in the
7053 * predicate, if it's ecb->dte_action->dta_next it's
7054 * in action #1, and so on.
7055 */
7056 for (err = ecb->dte_action, ndx = 0;
7057 err != act; err = err->dta_next, ndx++)
7058 continue;
7059
7060 dtrace_probe_error(state, ecb->dte_epid, ndx,
7061 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7062 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7063 cpu_core[cpuid].cpuc_dtrace_illval);
7064
7065 continue;
7066 }
7067
7068 if (!committed)
7069 buf->dtb_offset = offs + ecb->dte_size;
7070 }
7071
7072 if (vtime)
7073 curthread->t_dtrace_start = dtrace_gethrtime();
7074
7075 dtrace_interrupt_enable(cookie);
7076 }
7077
7078 /*
7079 * DTrace Probe Hashing Functions
7080 *
7081 * The functions in this section (and indeed, the functions in remaining
7082 * sections) are not _called_ from probe context. (Any exceptions to this are
7083 * marked with a "Note:".) Rather, they are called from elsewhere in the
7084 * DTrace framework to look-up probes in, add probes to and remove probes from
7085 * the DTrace probe hashes. (Each probe is hashed by each element of the
7086 * probe tuple -- allowing for fast lookups, regardless of what was
7087 * specified.)
7088 */
7089 static uint_t
7090 dtrace_hash_str(char *p)
7091 {
7092 unsigned int g;
7093 uint_t hval = 0;
7094
7095 while (*p) {
7096 hval = (hval << 4) + *p++;
7097 if ((g = (hval & 0xf0000000)) != 0)
7098 hval ^= g >> 24;
7099 hval &= ~g;
7100 }
7101 return (hval);
7102 }
7103
7104 static dtrace_hash_t *
7105 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
7106 {
7107 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
7108
7109 hash->dth_stroffs = stroffs;
7110 hash->dth_nextoffs = nextoffs;
7111 hash->dth_prevoffs = prevoffs;
7112
7113 hash->dth_size = 1;
7114 hash->dth_mask = hash->dth_size - 1;
7115
7116 hash->dth_tab = kmem_zalloc(hash->dth_size *
7117 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
7118
7119 return (hash);
7120 }
7121
7122 static void
7123 dtrace_hash_destroy(dtrace_hash_t *hash)
7124 {
7125 #ifdef DEBUG
7126 int i;
7127
7128 for (i = 0; i < hash->dth_size; i++)
7129 ASSERT(hash->dth_tab[i] == NULL);
7130 #endif
7131
7132 kmem_free(hash->dth_tab,
7133 hash->dth_size * sizeof (dtrace_hashbucket_t *));
7134 kmem_free(hash, sizeof (dtrace_hash_t));
7135 }
7136
7137 static void
7138 dtrace_hash_resize(dtrace_hash_t *hash)
7139 {
7140 int size = hash->dth_size, i, ndx;
7141 int new_size = hash->dth_size << 1;
7142 int new_mask = new_size - 1;
7143 dtrace_hashbucket_t **new_tab, *bucket, *next;
7144
7145 ASSERT((new_size & new_mask) == 0);
7146
7147 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
7148
7149 for (i = 0; i < size; i++) {
7150 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
7151 dtrace_probe_t *probe = bucket->dthb_chain;
7152
7153 ASSERT(probe != NULL);
7154 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
7155
7156 next = bucket->dthb_next;
7157 bucket->dthb_next = new_tab[ndx];
7158 new_tab[ndx] = bucket;
7159 }
7160 }
7161
7162 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
7163 hash->dth_tab = new_tab;
7164 hash->dth_size = new_size;
7165 hash->dth_mask = new_mask;
7166 }
7167
7168 static void
7169 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
7170 {
7171 int hashval = DTRACE_HASHSTR(hash, new);
7172 int ndx = hashval & hash->dth_mask;
7173 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7174 dtrace_probe_t **nextp, **prevp;
7175
7176 for (; bucket != NULL; bucket = bucket->dthb_next) {
7177 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
7178 goto add;
7179 }
7180
7181 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
7182 dtrace_hash_resize(hash);
7183 dtrace_hash_add(hash, new);
7184 return;
7185 }
7186
7187 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
7188 bucket->dthb_next = hash->dth_tab[ndx];
7189 hash->dth_tab[ndx] = bucket;
7190 hash->dth_nbuckets++;
7191
7192 add:
7193 nextp = DTRACE_HASHNEXT(hash, new);
7194 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
7195 *nextp = bucket->dthb_chain;
7196
7197 if (bucket->dthb_chain != NULL) {
7198 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
7199 ASSERT(*prevp == NULL);
7200 *prevp = new;
7201 }
7202
7203 bucket->dthb_chain = new;
7204 bucket->dthb_len++;
7205 }
7206
7207 static dtrace_probe_t *
7208 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
7209 {
7210 int hashval = DTRACE_HASHSTR(hash, template);
7211 int ndx = hashval & hash->dth_mask;
7212 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7213
7214 for (; bucket != NULL; bucket = bucket->dthb_next) {
7215 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7216 return (bucket->dthb_chain);
7217 }
7218
7219 return (NULL);
7220 }
7221
7222 static int
7223 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
7224 {
7225 int hashval = DTRACE_HASHSTR(hash, template);
7226 int ndx = hashval & hash->dth_mask;
7227 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7228
7229 for (; bucket != NULL; bucket = bucket->dthb_next) {
7230 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7231 return (bucket->dthb_len);
7232 }
7233
7234 return (NULL);
7235 }
7236
7237 static void
7238 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7239 {
7240 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7241 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7242
7243 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7244 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7245
7246 /*
7247 * Find the bucket that we're removing this probe from.
7248 */
7249 for (; bucket != NULL; bucket = bucket->dthb_next) {
7250 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7251 break;
7252 }
7253
7254 ASSERT(bucket != NULL);
7255
7256 if (*prevp == NULL) {
7257 if (*nextp == NULL) {
7258 /*
7259 * The removed probe was the only probe on this
7260 * bucket; we need to remove the bucket.
7261 */
7262 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7263
7264 ASSERT(bucket->dthb_chain == probe);
7265 ASSERT(b != NULL);
7266
7267 if (b == bucket) {
7268 hash->dth_tab[ndx] = bucket->dthb_next;
7269 } else {
7270 while (b->dthb_next != bucket)
7271 b = b->dthb_next;
7272 b->dthb_next = bucket->dthb_next;
7273 }
7274
7275 ASSERT(hash->dth_nbuckets > 0);
7276 hash->dth_nbuckets--;
7277 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7278 return;
7279 }
7280
7281 bucket->dthb_chain = *nextp;
7282 } else {
7283 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7284 }
7285
7286 if (*nextp != NULL)
7287 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7288 }
7289
7290 /*
7291 * DTrace Utility Functions
7292 *
7293 * These are random utility functions that are _not_ called from probe context.
7294 */
7295 static int
7296 dtrace_badattr(const dtrace_attribute_t *a)
7297 {
7298 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7299 a->dtat_data > DTRACE_STABILITY_MAX ||
7300 a->dtat_class > DTRACE_CLASS_MAX);
7301 }
7302
7303 /*
7304 * Return a duplicate copy of a string. If the specified string is NULL,
7305 * this function returns a zero-length string.
7306 */
7307 static char *
7308 dtrace_strdup(const char *str)
7309 {
7310 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7311
7312 if (str != NULL)
7313 (void) strcpy(new, str);
7314
7315 return (new);
7316 }
7317
7318 #define DTRACE_ISALPHA(c) \
7319 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7320
7321 static int
7322 dtrace_badname(const char *s)
7323 {
7324 char c;
7325
7326 if (s == NULL || (c = *s++) == '\0')
7327 return (0);
7328
7329 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7330 return (1);
7331
7332 while ((c = *s++) != '\0') {
7333 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7334 c != '-' && c != '_' && c != '.' && c != '`')
7335 return (1);
7336 }
7337
7338 return (0);
7339 }
7340
7341 static void
7342 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7343 {
7344 uint32_t priv;
7345
7346 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7347 /*
7348 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7349 */
7350 priv = DTRACE_PRIV_ALL;
7351 } else {
7352 *uidp = crgetuid(cr);
7353 *zoneidp = crgetzonedid(cr);
7354
7355 priv = 0;
7356 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7357 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7358 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7359 priv |= DTRACE_PRIV_USER;
7360 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7361 priv |= DTRACE_PRIV_PROC;
7362 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7363 priv |= DTRACE_PRIV_OWNER;
7364 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7365 priv |= DTRACE_PRIV_ZONEOWNER;
7366 }
7367
7368 *privp = priv;
7369 }
7370
7371 #ifdef DTRACE_ERRDEBUG
7372 static void
7373 dtrace_errdebug(const char *str)
7374 {
7375 int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
7376 int occupied = 0;
7377
7378 mutex_enter(&dtrace_errlock);
7379 dtrace_errlast = str;
7380 dtrace_errthread = curthread;
7381
7382 while (occupied++ < DTRACE_ERRHASHSZ) {
7383 if (dtrace_errhash[hval].dter_msg == str) {
7384 dtrace_errhash[hval].dter_count++;
7385 goto out;
7386 }
7387
7388 if (dtrace_errhash[hval].dter_msg != NULL) {
7389 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7390 continue;
7391 }
7392
7393 dtrace_errhash[hval].dter_msg = str;
7394 dtrace_errhash[hval].dter_count = 1;
7395 goto out;
7396 }
7397
7398 panic("dtrace: undersized error hash");
7399 out:
7400 mutex_exit(&dtrace_errlock);
7401 }
7402 #endif
7403
7404 /*
7405 * DTrace Matching Functions
7406 *
7407 * These functions are used to match groups of probes, given some elements of
7408 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7409 */
7410 static int
7411 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7412 zoneid_t zoneid)
7413 {
7414 if (priv != DTRACE_PRIV_ALL) {
7415 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7416 uint32_t match = priv & ppriv;
7417
7418 /*
7419 * No PRIV_DTRACE_* privileges...
7420 */
7421 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7422 DTRACE_PRIV_KERNEL)) == 0)
7423 return (0);
7424
7425 /*
7426 * No matching bits, but there were bits to match...
7427 */
7428 if (match == 0 && ppriv != 0)
7429 return (0);
7430
7431 /*
7432 * Need to have permissions to the process, but don't...
7433 */
7434 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7435 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7436 return (0);
7437 }
7438
7439 /*
7440 * Need to be in the same zone unless we possess the
7441 * privilege to examine all zones.
7442 */
7443 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7444 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7445 return (0);
7446 }
7447 }
7448
7449 return (1);
7450 }
7451
7452 /*
7453 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7454 * consists of input pattern strings and an ops-vector to evaluate them.
7455 * This function returns >0 for match, 0 for no match, and <0 for error.
7456 */
7457 static int
7458 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7459 uint32_t priv, uid_t uid, zoneid_t zoneid)
7460 {
7461 dtrace_provider_t *pvp = prp->dtpr_provider;
7462 int rv;
7463
7464 if (pvp->dtpv_defunct)
7465 return (0);
7466
7467 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7468 return (rv);
7469
7470 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7471 return (rv);
7472
7473 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7474 return (rv);
7475
7476 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7477 return (rv);
7478
7479 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7480 return (0);
7481
7482 return (rv);
7483 }
7484
7485 /*
7486 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7487 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7488 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7489 * In addition, all of the recursion cases except for '*' matching have been
7490 * unwound. For '*', we still implement recursive evaluation, but a depth
7491 * counter is maintained and matching is aborted if we recurse too deep.
7492 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7493 */
7494 static int
7495 dtrace_match_glob(const char *s, const char *p, int depth)
7496 {
7497 const char *olds;
7498 char s1, c;
7499 int gs;
7500
7501 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7502 return (-1);
7503
7504 if (s == NULL)
7505 s = ""; /* treat NULL as empty string */
7506
7507 top:
7508 olds = s;
7509 s1 = *s++;
7510
7511 if (p == NULL)
7512 return (0);
7513
7514 if ((c = *p++) == '\0')
7515 return (s1 == '\0');
7516
7517 switch (c) {
7518 case '[': {
7519 int ok = 0, notflag = 0;
7520 char lc = '\0';
7521
7522 if (s1 == '\0')
7523 return (0);
7524
7525 if (*p == '!') {
7526 notflag = 1;
7527 p++;
7528 }
7529
7530 if ((c = *p++) == '\0')
7531 return (0);
7532
7533 do {
7534 if (c == '-' && lc != '\0' && *p != ']') {
7535 if ((c = *p++) == '\0')
7536 return (0);
7537 if (c == '\\' && (c = *p++) == '\0')
7538 return (0);
7539
7540 if (notflag) {
7541 if (s1 < lc || s1 > c)
7542 ok++;
7543 else
7544 return (0);
7545 } else if (lc <= s1 && s1 <= c)
7546 ok++;
7547
7548 } else if (c == '\\' && (c = *p++) == '\0')
7549 return (0);
7550
7551 lc = c; /* save left-hand 'c' for next iteration */
7552
7553 if (notflag) {
7554 if (s1 != c)
7555 ok++;
7556 else
7557 return (0);
7558 } else if (s1 == c)
7559 ok++;
7560
7561 if ((c = *p++) == '\0')
7562 return (0);
7563
7564 } while (c != ']');
7565
7566 if (ok)
7567 goto top;
7568
7569 return (0);
7570 }
7571
7572 case '\\':
7573 if ((c = *p++) == '\0')
7574 return (0);
7575 /*FALLTHRU*/
7576
7577 default:
7578 if (c != s1)
7579 return (0);
7580 /*FALLTHRU*/
7581
7582 case '?':
7583 if (s1 != '\0')
7584 goto top;
7585 return (0);
7586
7587 case '*':
7588 while (*p == '*')
7589 p++; /* consecutive *'s are identical to a single one */
7590
7591 if (*p == '\0')
7592 return (1);
7593
7594 for (s = olds; *s != '\0'; s++) {
7595 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7596 return (gs);
7597 }
7598
7599 return (0);
7600 }
7601 }
7602
7603 /*ARGSUSED*/
7604 static int
7605 dtrace_match_string(const char *s, const char *p, int depth)
7606 {
7607 return (s != NULL && strcmp(s, p) == 0);
7608 }
7609
7610 /*ARGSUSED*/
7611 static int
7612 dtrace_match_nul(const char *s, const char *p, int depth)
7613 {
7614 return (1); /* always match the empty pattern */
7615 }
7616
7617 /*ARGSUSED*/
7618 static int
7619 dtrace_match_nonzero(const char *s, const char *p, int depth)
7620 {
7621 return (s != NULL && s[0] != '\0');
7622 }
7623
7624 static int
7625 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7626 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7627 {
7628 dtrace_probe_t template, *probe;
7629 dtrace_hash_t *hash = NULL;
7630 int len, rc, best = INT_MAX, nmatched = 0;
7631 dtrace_id_t i;
7632
7633 ASSERT(MUTEX_HELD(&dtrace_lock));
7634
7635 /*
7636 * If the probe ID is specified in the key, just lookup by ID and
7637 * invoke the match callback once if a matching probe is found.
7638 */
7639 if (pkp->dtpk_id != DTRACE_IDNONE) {
7640 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7641 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7642 if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
7643 return (DTRACE_MATCH_FAIL);
7644 nmatched++;
7645 }
7646 return (nmatched);
7647 }
7648
7649 template.dtpr_mod = (char *)pkp->dtpk_mod;
7650 template.dtpr_func = (char *)pkp->dtpk_func;
7651 template.dtpr_name = (char *)pkp->dtpk_name;
7652
7653 /*
7654 * We want to find the most distinct of the module name, function
7655 * name, and name. So for each one that is not a glob pattern or
7656 * empty string, we perform a lookup in the corresponding hash and
7657 * use the hash table with the fewest collisions to do our search.
7658 */
7659 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7660 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7661 best = len;
7662 hash = dtrace_bymod;
7663 }
7664
7665 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7666 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7667 best = len;
7668 hash = dtrace_byfunc;
7669 }
7670
7671 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7672 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7673 best = len;
7674 hash = dtrace_byname;
7675 }
7676
7677 /*
7678 * If we did not select a hash table, iterate over every probe and
7679 * invoke our callback for each one that matches our input probe key.
7680 */
7681 if (hash == NULL) {
7682 for (i = 0; i < dtrace_nprobes; i++) {
7683 if ((probe = dtrace_probes[i]) == NULL ||
7684 dtrace_match_probe(probe, pkp, priv, uid,
7685 zoneid) <= 0)
7686 continue;
7687
7688 nmatched++;
7689
7690 if ((rc = (*matched)(probe, arg)) !=
7691 DTRACE_MATCH_NEXT) {
7692 if (rc == DTRACE_MATCH_FAIL)
7693 return (DTRACE_MATCH_FAIL);
7694 break;
7695 }
7696 }
7697
7698 return (nmatched);
7699 }
7700
7701 /*
7702 * If we selected a hash table, iterate over each probe of the same key
7703 * name and invoke the callback for every probe that matches the other
7704 * attributes of our input probe key.
7705 */
7706 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7707 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7708
7709 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7710 continue;
7711
7712 nmatched++;
7713
7714 if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
7715 if (rc == DTRACE_MATCH_FAIL)
7716 return (DTRACE_MATCH_FAIL);
7717 break;
7718 }
7719 }
7720
7721 return (nmatched);
7722 }
7723
7724 /*
7725 * Return the function pointer dtrace_probecmp() should use to compare the
7726 * specified pattern with a string. For NULL or empty patterns, we select
7727 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7728 * For non-empty non-glob strings, we use dtrace_match_string().
7729 */
7730 static dtrace_probekey_f *
7731 dtrace_probekey_func(const char *p)
7732 {
7733 char c;
7734
7735 if (p == NULL || *p == '\0')
7736 return (&dtrace_match_nul);
7737
7738 while ((c = *p++) != '\0') {
7739 if (c == '[' || c == '?' || c == '*' || c == '\\')
7740 return (&dtrace_match_glob);
7741 }
7742
7743 return (&dtrace_match_string);
7744 }
7745
7746 /*
7747 * Build a probe comparison key for use with dtrace_match_probe() from the
7748 * given probe description. By convention, a null key only matches anchored
7749 * probes: if each field is the empty string, reset dtpk_fmatch to
7750 * dtrace_match_nonzero().
7751 */
7752 static void
7753 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7754 {
7755 pkp->dtpk_prov = pdp->dtpd_provider;
7756 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7757
7758 pkp->dtpk_mod = pdp->dtpd_mod;
7759 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7760
7761 pkp->dtpk_func = pdp->dtpd_func;
7762 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7763
7764 pkp->dtpk_name = pdp->dtpd_name;
7765 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7766
7767 pkp->dtpk_id = pdp->dtpd_id;
7768
7769 if (pkp->dtpk_id == DTRACE_IDNONE &&
7770 pkp->dtpk_pmatch == &dtrace_match_nul &&
7771 pkp->dtpk_mmatch == &dtrace_match_nul &&
7772 pkp->dtpk_fmatch == &dtrace_match_nul &&
7773 pkp->dtpk_nmatch == &dtrace_match_nul)
7774 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7775 }
7776
7777 /*
7778 * DTrace Provider-to-Framework API Functions
7779 *
7780 * These functions implement much of the Provider-to-Framework API, as
7781 * described in <sys/dtrace.h>. The parts of the API not in this section are
7782 * the functions in the API for probe management (found below), and
7783 * dtrace_probe() itself (found above).
7784 */
7785
7786 /*
7787 * Register the calling provider with the DTrace framework. This should
7788 * generally be called by DTrace providers in their attach(9E) entry point.
7789 */
7790 int
7791 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7792 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7793 {
7794 dtrace_provider_t *provider;
7795
7796 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7797 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7798 "arguments", name ? name : "<NULL>");
7799 return (EINVAL);
7800 }
7801
7802 if (name[0] == '\0' || dtrace_badname(name)) {
7803 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7804 "provider name", name);
7805 return (EINVAL);
7806 }
7807
7808 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7809 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7810 pops->dtps_destroy == NULL ||
7811 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7812 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7813 "provider ops", name);
7814 return (EINVAL);
7815 }
7816
7817 if (dtrace_badattr(&pap->dtpa_provider) ||
7818 dtrace_badattr(&pap->dtpa_mod) ||
7819 dtrace_badattr(&pap->dtpa_func) ||
7820 dtrace_badattr(&pap->dtpa_name) ||
7821 dtrace_badattr(&pap->dtpa_args)) {
7822 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7823 "provider attributes", name);
7824 return (EINVAL);
7825 }
7826
7827 if (priv & ~DTRACE_PRIV_ALL) {
7828 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7829 "privilege attributes", name);
7830 return (EINVAL);
7831 }
7832
7833 if ((priv & DTRACE_PRIV_KERNEL) &&
7834 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7835 pops->dtps_mode == NULL) {
7836 cmn_err(CE_WARN, "failed to register provider '%s': need "
7837 "dtps_mode() op for given privilege attributes", name);
7838 return (EINVAL);
7839 }
7840
7841 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7842 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7843 (void) strcpy(provider->dtpv_name, name);
7844
7845 provider->dtpv_attr = *pap;
7846 provider->dtpv_priv.dtpp_flags = priv;
7847 if (cr != NULL) {
7848 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7849 provider->dtpv_priv.dtpp_zoneid = crgetzonedid(cr);
7850 }
7851 provider->dtpv_pops = *pops;
7852
7853 if (pops->dtps_provide == NULL) {
7854 ASSERT(pops->dtps_provide_module != NULL);
7855 provider->dtpv_pops.dtps_provide =
7856 (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
7857 }
7858
7859 if (pops->dtps_provide_module == NULL) {
7860 ASSERT(pops->dtps_provide != NULL);
7861 provider->dtpv_pops.dtps_provide_module =
7862 (void (*)(void *, struct modctl *))dtrace_nullop;
7863 }
7864
7865 if (pops->dtps_suspend == NULL) {
7866 ASSERT(pops->dtps_resume == NULL);
7867 provider->dtpv_pops.dtps_suspend =
7868 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7869 provider->dtpv_pops.dtps_resume =
7870 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7871 }
7872
7873 provider->dtpv_arg = arg;
7874 *idp = (dtrace_provider_id_t)provider;
7875
7876 if (pops == &dtrace_provider_ops) {
7877 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7878 ASSERT(MUTEX_HELD(&dtrace_lock));
7879 ASSERT(dtrace_anon.dta_enabling == NULL);
7880
7881 /*
7882 * We make sure that the DTrace provider is at the head of
7883 * the provider chain.
7884 */
7885 provider->dtpv_next = dtrace_provider;
7886 dtrace_provider = provider;
7887 return (0);
7888 }
7889
7890 mutex_enter(&dtrace_provider_lock);
7891 mutex_enter(&dtrace_lock);
7892
7893 /*
7894 * If there is at least one provider registered, we'll add this
7895 * provider after the first provider.
7896 */
7897 if (dtrace_provider != NULL) {
7898 provider->dtpv_next = dtrace_provider->dtpv_next;
7899 dtrace_provider->dtpv_next = provider;
7900 } else {
7901 dtrace_provider = provider;
7902 }
7903
7904 if (dtrace_retained != NULL) {
7905 dtrace_enabling_provide(provider);
7906
7907 /*
7908 * Now we need to call dtrace_enabling_matchall() -- which
7909 * will acquire cpu_lock and dtrace_lock. We therefore need
7910 * to drop all of our locks before calling into it...
7911 */
7912 mutex_exit(&dtrace_lock);
7913 mutex_exit(&dtrace_provider_lock);
7914 dtrace_enabling_matchall();
7915
7916 return (0);
7917 }
7918
7919 mutex_exit(&dtrace_lock);
7920 mutex_exit(&dtrace_provider_lock);
7921
7922 return (0);
7923 }
7924
7925 /*
7926 * Unregister the specified provider from the DTrace framework. This should
7927 * generally be called by DTrace providers in their detach(9E) entry point.
7928 */
7929 int
7930 dtrace_unregister(dtrace_provider_id_t id)
7931 {
7932 dtrace_provider_t *old = (dtrace_provider_t *)id;
7933 dtrace_provider_t *prev = NULL;
7934 int i, self = 0, noreap = 0;
7935 dtrace_probe_t *probe, *first = NULL;
7936
7937 if (old->dtpv_pops.dtps_enable ==
7938 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
7939 /*
7940 * If DTrace itself is the provider, we're called with locks
7941 * already held.
7942 */
7943 ASSERT(old == dtrace_provider);
7944 ASSERT(dtrace_devi != NULL);
7945 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7946 ASSERT(MUTEX_HELD(&dtrace_lock));
7947 self = 1;
7948
7949 if (dtrace_provider->dtpv_next != NULL) {
7950 /*
7951 * There's another provider here; return failure.
7952 */
7953 return (EBUSY);
7954 }
7955 } else {
7956 mutex_enter(&dtrace_provider_lock);
7957 mutex_enter(&mod_lock);
7958 mutex_enter(&dtrace_lock);
7959 }
7960
7961 /*
7962 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7963 * probes, we refuse to let providers slither away, unless this
7964 * provider has already been explicitly invalidated.
7965 */
7966 if (!old->dtpv_defunct &&
7967 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7968 dtrace_anon.dta_state->dts_necbs > 0))) {
7969 if (!self) {
7970 mutex_exit(&dtrace_lock);
7971 mutex_exit(&mod_lock);
7972 mutex_exit(&dtrace_provider_lock);
7973 }
7974 return (EBUSY);
7975 }
7976
7977 /*
7978 * Attempt to destroy the probes associated with this provider.
7979 */
7980 for (i = 0; i < dtrace_nprobes; i++) {
7981 if ((probe = dtrace_probes[i]) == NULL)
7982 continue;
7983
7984 if (probe->dtpr_provider != old)
7985 continue;
7986
7987 if (probe->dtpr_ecb == NULL)
7988 continue;
7989
7990 /*
7991 * If we are trying to unregister a defunct provider, and the
7992 * provider was made defunct within the interval dictated by
7993 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7994 * attempt to reap our enablings. To denote that the provider
7995 * should reattempt to unregister itself at some point in the
7996 * future, we will return a differentiable error code (EAGAIN
7997 * instead of EBUSY) in this case.
7998 */
7999 if (dtrace_gethrtime() - old->dtpv_defunct >
8000 dtrace_unregister_defunct_reap)
8001 noreap = 1;
8002
8003 if (!self) {
8004 mutex_exit(&dtrace_lock);
8005 mutex_exit(&mod_lock);
8006 mutex_exit(&dtrace_provider_lock);
8007 }
8008
8009 if (noreap)
8010 return (EBUSY);
8011
8012 (void) taskq_dispatch(dtrace_taskq,
8013 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8014
8015 return (EAGAIN);
8016 }
8017
8018 /*
8019 * All of the probes for this provider are disabled; we can safely
8020 * remove all of them from their hash chains and from the probe array.
8021 */
8022 for (i = 0; i < dtrace_nprobes; i++) {
8023 if ((probe = dtrace_probes[i]) == NULL)
8024 continue;
8025
8026 if (probe->dtpr_provider != old)
8027 continue;
8028
8029 dtrace_probes[i] = NULL;
8030
8031 dtrace_hash_remove(dtrace_bymod, probe);
8032 dtrace_hash_remove(dtrace_byfunc, probe);
8033 dtrace_hash_remove(dtrace_byname, probe);
8034
8035 if (first == NULL) {
8036 first = probe;
8037 probe->dtpr_nextmod = NULL;
8038 } else {
8039 probe->dtpr_nextmod = first;
8040 first = probe;
8041 }
8042 }
8043
8044 /*
8045 * The provider's probes have been removed from the hash chains and
8046 * from the probe array. Now issue a dtrace_sync() to be sure that
8047 * everyone has cleared out from any probe array processing.
8048 */
8049 dtrace_sync();
8050
8051 for (probe = first; probe != NULL; probe = first) {
8052 first = probe->dtpr_nextmod;
8053
8054 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8055 probe->dtpr_arg);
8056 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8057 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8058 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8059 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8060 kmem_free(probe, sizeof (dtrace_probe_t));
8061 }
8062
8063 if ((prev = dtrace_provider) == old) {
8064 ASSERT(self || dtrace_devi == NULL);
8065 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8066 dtrace_provider = old->dtpv_next;
8067 } else {
8068 while (prev != NULL && prev->dtpv_next != old)
8069 prev = prev->dtpv_next;
8070
8071 if (prev == NULL) {
8072 panic("attempt to unregister non-existent "
8073 "dtrace provider %p\n", (void *)id);
8074 }
8075
8076 prev->dtpv_next = old->dtpv_next;
8077 }
8078
8079 if (!self) {
8080 mutex_exit(&dtrace_lock);
8081 mutex_exit(&mod_lock);
8082 mutex_exit(&dtrace_provider_lock);
8083 }
8084
8085 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
8086 kmem_free(old, sizeof (dtrace_provider_t));
8087
8088 return (0);
8089 }
8090
8091 /*
8092 * Invalidate the specified provider. All subsequent probe lookups for the
8093 * specified provider will fail, but its probes will not be removed.
8094 */
8095 void
8096 dtrace_invalidate(dtrace_provider_id_t id)
8097 {
8098 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
8099
8100 ASSERT(pvp->dtpv_pops.dtps_enable !=
8101 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8102
8103 mutex_enter(&dtrace_provider_lock);
8104 mutex_enter(&dtrace_lock);
8105
8106 pvp->dtpv_defunct = dtrace_gethrtime();
8107
8108 mutex_exit(&dtrace_lock);
8109 mutex_exit(&dtrace_provider_lock);
8110 }
8111
8112 /*
8113 * Indicate whether or not DTrace has attached.
8114 */
8115 int
8116 dtrace_attached(void)
8117 {
8118 /*
8119 * dtrace_provider will be non-NULL iff the DTrace driver has
8120 * attached. (It's non-NULL because DTrace is always itself a
8121 * provider.)
8122 */
8123 return (dtrace_provider != NULL);
8124 }
8125
8126 /*
8127 * Remove all the unenabled probes for the given provider. This function is
8128 * not unlike dtrace_unregister(), except that it doesn't remove the provider
8129 * -- just as many of its associated probes as it can.
8130 */
8131 int
8132 dtrace_condense(dtrace_provider_id_t id)
8133 {
8134 dtrace_provider_t *prov = (dtrace_provider_t *)id;
8135 int i;
8136 dtrace_probe_t *probe;
8137
8138 /*
8139 * Make sure this isn't the dtrace provider itself.
8140 */
8141 ASSERT(prov->dtpv_pops.dtps_enable !=
8142 (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8143
8144 mutex_enter(&dtrace_provider_lock);
8145 mutex_enter(&dtrace_lock);
8146
8147 /*
8148 * Attempt to destroy the probes associated with this provider.
8149 */
8150 for (i = 0; i < dtrace_nprobes; i++) {
8151 if ((probe = dtrace_probes[i]) == NULL)
8152 continue;
8153
8154 if (probe->dtpr_provider != prov)
8155 continue;
8156
8157 if (probe->dtpr_ecb != NULL)
8158 continue;
8159
8160 dtrace_probes[i] = NULL;
8161
8162 dtrace_hash_remove(dtrace_bymod, probe);
8163 dtrace_hash_remove(dtrace_byfunc, probe);
8164 dtrace_hash_remove(dtrace_byname, probe);
8165
8166 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
8167 probe->dtpr_arg);
8168 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8169 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8170 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8171 kmem_free(probe, sizeof (dtrace_probe_t));
8172 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
8173 }
8174
8175 mutex_exit(&dtrace_lock);
8176 mutex_exit(&dtrace_provider_lock);
8177
8178 return (0);
8179 }
8180
8181 /*
8182 * DTrace Probe Management Functions
8183 *
8184 * The functions in this section perform the DTrace probe management,
8185 * including functions to create probes, look-up probes, and call into the
8186 * providers to request that probes be provided. Some of these functions are
8187 * in the Provider-to-Framework API; these functions can be identified by the
8188 * fact that they are not declared "static".
8189 */
8190
8191 /*
8192 * Create a probe with the specified module name, function name, and name.
8193 */
8194 dtrace_id_t
8195 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
8196 const char *func, const char *name, int aframes, void *arg)
8197 {
8198 dtrace_probe_t *probe, **probes;
8199 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
8200 dtrace_id_t id;
8201
8202 if (provider == dtrace_provider) {
8203 ASSERT(MUTEX_HELD(&dtrace_lock));
8204 } else {
8205 mutex_enter(&dtrace_lock);
8206 }
8207
8208 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
8209 VM_BESTFIT | VM_SLEEP);
8210 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8211
8212 probe->dtpr_id = id;
8213 probe->dtpr_gen = dtrace_probegen++;
8214 probe->dtpr_mod = dtrace_strdup(mod);
8215 probe->dtpr_func = dtrace_strdup(func);
8216 probe->dtpr_name = dtrace_strdup(name);
8217 probe->dtpr_arg = arg;
8218 probe->dtpr_aframes = aframes;
8219 probe->dtpr_provider = provider;
8220
8221 dtrace_hash_add(dtrace_bymod, probe);
8222 dtrace_hash_add(dtrace_byfunc, probe);
8223 dtrace_hash_add(dtrace_byname, probe);
8224
8225 if (id - 1 >= dtrace_nprobes) {
8226 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8227 size_t nsize = osize << 1;
8228
8229 if (nsize == 0) {
8230 ASSERT(osize == 0);
8231 ASSERT(dtrace_probes == NULL);
8232 nsize = sizeof (dtrace_probe_t *);
8233 }
8234
8235 probes = kmem_zalloc(nsize, KM_SLEEP);
8236
8237 if (dtrace_probes == NULL) {
8238 ASSERT(osize == 0);
8239 dtrace_probes = probes;
8240 dtrace_nprobes = 1;
8241 } else {
8242 dtrace_probe_t **oprobes = dtrace_probes;
8243
8244 bcopy(oprobes, probes, osize);
8245 dtrace_membar_producer();
8246 dtrace_probes = probes;
8247
8248 dtrace_sync();
8249
8250 /*
8251 * All CPUs are now seeing the new probes array; we can
8252 * safely free the old array.
8253 */
8254 kmem_free(oprobes, osize);
8255 dtrace_nprobes <<= 1;
8256 }
8257
8258 ASSERT(id - 1 < dtrace_nprobes);
8259 }
8260
8261 ASSERT(dtrace_probes[id - 1] == NULL);
8262 dtrace_probes[id - 1] = probe;
8263
8264 if (provider != dtrace_provider)
8265 mutex_exit(&dtrace_lock);
8266
8267 return (id);
8268 }
8269
8270 static dtrace_probe_t *
8271 dtrace_probe_lookup_id(dtrace_id_t id)
8272 {
8273 ASSERT(MUTEX_HELD(&dtrace_lock));
8274
8275 if (id == 0 || id > dtrace_nprobes)
8276 return (NULL);
8277
8278 return (dtrace_probes[id - 1]);
8279 }
8280
8281 static int
8282 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8283 {
8284 *((dtrace_id_t *)arg) = probe->dtpr_id;
8285
8286 return (DTRACE_MATCH_DONE);
8287 }
8288
8289 /*
8290 * Look up a probe based on provider and one or more of module name, function
8291 * name and probe name.
8292 */
8293 dtrace_id_t
8294 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
8295 const char *func, const char *name)
8296 {
8297 dtrace_probekey_t pkey;
8298 dtrace_id_t id;
8299 int match;
8300
8301 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8302 pkey.dtpk_pmatch = &dtrace_match_string;
8303 pkey.dtpk_mod = mod;
8304 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8305 pkey.dtpk_func = func;
8306 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8307 pkey.dtpk_name = name;
8308 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8309 pkey.dtpk_id = DTRACE_IDNONE;
8310
8311 mutex_enter(&dtrace_lock);
8312 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8313 dtrace_probe_lookup_match, &id);
8314 mutex_exit(&dtrace_lock);
8315
8316 ASSERT(match == 1 || match == 0);
8317 return (match ? id : 0);
8318 }
8319
8320 /*
8321 * Returns the probe argument associated with the specified probe.
8322 */
8323 void *
8324 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8325 {
8326 dtrace_probe_t *probe;
8327 void *rval = NULL;
8328
8329 mutex_enter(&dtrace_lock);
8330
8331 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8332 probe->dtpr_provider == (dtrace_provider_t *)id)
8333 rval = probe->dtpr_arg;
8334
8335 mutex_exit(&dtrace_lock);
8336
8337 return (rval);
8338 }
8339
8340 /*
8341 * Copy a probe into a probe description.
8342 */
8343 static void
8344 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8345 {
8346 bzero(pdp, sizeof (dtrace_probedesc_t));
8347 pdp->dtpd_id = prp->dtpr_id;
8348
8349 (void) strncpy(pdp->dtpd_provider,
8350 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8351
8352 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8353 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8354 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8355 }
8356
8357 /*
8358 * Called to indicate that a probe -- or probes -- should be provided by a
8359 * specfied provider. If the specified description is NULL, the provider will
8360 * be told to provide all of its probes. (This is done whenever a new
8361 * consumer comes along, or whenever a retained enabling is to be matched.) If
8362 * the specified description is non-NULL, the provider is given the
8363 * opportunity to dynamically provide the specified probe, allowing providers
8364 * to support the creation of probes on-the-fly. (So-called _autocreated_
8365 * probes.) If the provider is NULL, the operations will be applied to all
8366 * providers; if the provider is non-NULL the operations will only be applied
8367 * to the specified provider. The dtrace_provider_lock must be held, and the
8368 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8369 * will need to grab the dtrace_lock when it reenters the framework through
8370 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8371 */
8372 static void
8373 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8374 {
8375 struct modctl *ctl;
8376 int all = 0;
8377
8378 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8379
8380 if (prv == NULL) {
8381 all = 1;
8382 prv = dtrace_provider;
8383 }
8384
8385 do {
8386 /*
8387 * First, call the blanket provide operation.
8388 */
8389 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8390
8391 /*
8392 * Now call the per-module provide operation. We will grab
8393 * mod_lock to prevent the list from being modified. Note
8394 * that this also prevents the mod_busy bits from changing.
8395 * (mod_busy can only be changed with mod_lock held.)
8396 */
8397 mutex_enter(&mod_lock);
8398
8399 ctl = &modules;
8400 do {
8401 if (ctl->mod_busy || ctl->mod_mp == NULL)
8402 continue;
8403
8404 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8405
8406 } while ((ctl = ctl->mod_next) != &modules);
8407
8408 mutex_exit(&mod_lock);
8409 } while (all && (prv = prv->dtpv_next) != NULL);
8410 }
8411
8412 /*
8413 * Iterate over each probe, and call the Framework-to-Provider API function
8414 * denoted by offs.
8415 */
8416 static void
8417 dtrace_probe_foreach(uintptr_t offs)
8418 {
8419 dtrace_provider_t *prov;
8420 void (*func)(void *, dtrace_id_t, void *);
8421 dtrace_probe_t *probe;
8422 dtrace_icookie_t cookie;
8423 int i;
8424
8425 /*
8426 * We disable interrupts to walk through the probe array. This is
8427 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8428 * won't see stale data.
8429 */
8430 cookie = dtrace_interrupt_disable();
8431
8432 for (i = 0; i < dtrace_nprobes; i++) {
8433 if ((probe = dtrace_probes[i]) == NULL)
8434 continue;
8435
8436 if (probe->dtpr_ecb == NULL) {
8437 /*
8438 * This probe isn't enabled -- don't call the function.
8439 */
8440 continue;
8441 }
8442
8443 prov = probe->dtpr_provider;
8444 func = *((void(**)(void *, dtrace_id_t, void *))
8445 ((uintptr_t)&prov->dtpv_pops + offs));
8446
8447 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8448 }
8449
8450 dtrace_interrupt_enable(cookie);
8451 }
8452
8453 static int
8454 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8455 {
8456 dtrace_probekey_t pkey;
8457 uint32_t priv;
8458 uid_t uid;
8459 zoneid_t zoneid;
8460 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
8461
8462 ASSERT(MUTEX_HELD(&dtrace_lock));
8463 dtrace_ecb_create_cache = NULL;
8464
8465 if (desc == NULL) {
8466 /*
8467 * If we're passed a NULL description, we're being asked to
8468 * create an ECB with a NULL probe.
8469 */
8470 (void) dtrace_ecb_create_enable(NULL, enab);
8471 return (0);
8472 }
8473
8474 dtrace_probekey(desc, &pkey);
8475 dtrace_cred2priv(state->dts_cred.dcr_cred, &priv, &uid, &zoneid);
8476
8477 if ((priv & DTRACE_PRIV_ZONEOWNER) &&
8478 state->dts_options[DTRACEOPT_ZONE] != DTRACEOPT_UNSET) {
8479 /*
8480 * If we have the privilege of instrumenting all zones but we
8481 * have been told to instrument but one, we will spoof this up
8482 * depriving ourselves of DTRACE_PRIV_ZONEOWNER for purposes
8483 * of dtrace_match(). (Note that DTRACEOPT_ZONE is not for
8484 * security but rather for performance: it allows the global
8485 * zone to instrument USDT probes in a local zone without
8486 * requiring all zones to be instrumented.)
8487 */
8488 priv &= ~DTRACE_PRIV_ZONEOWNER;
8489 zoneid = state->dts_options[DTRACEOPT_ZONE];
8490 }
8491
8492 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8493 enab));
8494 }
8495
8496 /*
8497 * DTrace Helper Provider Functions
8498 */
8499 static void
8500 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8501 {
8502 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8503 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8504 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8505 }
8506
8507 static void
8508 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8509 const dof_provider_t *dofprov, char *strtab)
8510 {
8511 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8512 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8513 dofprov->dofpv_provattr);
8514 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8515 dofprov->dofpv_modattr);
8516 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8517 dofprov->dofpv_funcattr);
8518 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8519 dofprov->dofpv_nameattr);
8520 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8521 dofprov->dofpv_argsattr);
8522 }
8523
8524 static void
8525 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8526 {
8527 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8528 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8529 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8530 dof_provider_t *provider;
8531 dof_probe_t *probe;
8532 uint32_t *off, *enoff;
8533 uint8_t *arg;
8534 char *strtab;
8535 uint_t i, nprobes;
8536 dtrace_helper_provdesc_t dhpv;
8537 dtrace_helper_probedesc_t dhpb;
8538 dtrace_meta_t *meta = dtrace_meta_pid;
8539 dtrace_mops_t *mops = &meta->dtm_mops;
8540 void *parg;
8541
8542 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8543 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8544 provider->dofpv_strtab * dof->dofh_secsize);
8545 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8546 provider->dofpv_probes * dof->dofh_secsize);
8547 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8548 provider->dofpv_prargs * dof->dofh_secsize);
8549 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8550 provider->dofpv_proffs * dof->dofh_secsize);
8551
8552 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8553 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8554 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8555 enoff = NULL;
8556
8557 /*
8558 * See dtrace_helper_provider_validate().
8559 */
8560 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8561 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8562 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8563 provider->dofpv_prenoffs * dof->dofh_secsize);
8564 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8565 }
8566
8567 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8568
8569 /*
8570 * Create the provider.
8571 */
8572 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8573
8574 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8575 return;
8576
8577 meta->dtm_count++;
8578
8579 /*
8580 * Create the probes.
8581 */
8582 for (i = 0; i < nprobes; i++) {
8583 probe = (dof_probe_t *)(uintptr_t)(daddr +
8584 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8585
8586 dhpb.dthpb_mod = dhp->dofhp_mod;
8587 dhpb.dthpb_func = strtab + probe->dofpr_func;
8588 dhpb.dthpb_name = strtab + probe->dofpr_name;
8589 dhpb.dthpb_base = probe->dofpr_addr;
8590 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8591 dhpb.dthpb_noffs = probe->dofpr_noffs;
8592 if (enoff != NULL) {
8593 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8594 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8595 } else {
8596 dhpb.dthpb_enoffs = NULL;
8597 dhpb.dthpb_nenoffs = 0;
8598 }
8599 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8600 dhpb.dthpb_nargc = probe->dofpr_nargc;
8601 dhpb.dthpb_xargc = probe->dofpr_xargc;
8602 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8603 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8604
8605 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8606 }
8607 }
8608
8609 static void
8610 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8611 {
8612 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8613 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8614 int i;
8615
8616 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8617
8618 for (i = 0; i < dof->dofh_secnum; i++) {
8619 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8620 dof->dofh_secoff + i * dof->dofh_secsize);
8621
8622 if (sec->dofs_type != DOF_SECT_PROVIDER)
8623 continue;
8624
8625 dtrace_helper_provide_one(dhp, sec, pid);
8626 }
8627
8628 /*
8629 * We may have just created probes, so we must now rematch against
8630 * any retained enablings. Note that this call will acquire both
8631 * cpu_lock and dtrace_lock; the fact that we are holding
8632 * dtrace_meta_lock now is what defines the ordering with respect to
8633 * these three locks.
8634 */
8635 dtrace_enabling_matchall();
8636 }
8637
8638 static void
8639 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8640 {
8641 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8642 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8643 dof_sec_t *str_sec;
8644 dof_provider_t *provider;
8645 char *strtab;
8646 dtrace_helper_provdesc_t dhpv;
8647 dtrace_meta_t *meta = dtrace_meta_pid;
8648 dtrace_mops_t *mops = &meta->dtm_mops;
8649
8650 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8651 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8652 provider->dofpv_strtab * dof->dofh_secsize);
8653
8654 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8655
8656 /*
8657 * Create the provider.
8658 */
8659 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8660
8661 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8662
8663 meta->dtm_count--;
8664 }
8665
8666 static void
8667 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8668 {
8669 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8670 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8671 int i;
8672
8673 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8674
8675 for (i = 0; i < dof->dofh_secnum; i++) {
8676 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8677 dof->dofh_secoff + i * dof->dofh_secsize);
8678
8679 if (sec->dofs_type != DOF_SECT_PROVIDER)
8680 continue;
8681
8682 dtrace_helper_provider_remove_one(dhp, sec, pid);
8683 }
8684 }
8685
8686 /*
8687 * DTrace Meta Provider-to-Framework API Functions
8688 *
8689 * These functions implement the Meta Provider-to-Framework API, as described
8690 * in <sys/dtrace.h>.
8691 */
8692 int
8693 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8694 dtrace_meta_provider_id_t *idp)
8695 {
8696 dtrace_meta_t *meta;
8697 dtrace_helpers_t *help, *next;
8698 int i;
8699
8700 *idp = DTRACE_METAPROVNONE;
8701
8702 /*
8703 * We strictly don't need the name, but we hold onto it for
8704 * debuggability. All hail error queues!
8705 */
8706 if (name == NULL) {
8707 cmn_err(CE_WARN, "failed to register meta-provider: "
8708 "invalid name");
8709 return (EINVAL);
8710 }
8711
8712 if (mops == NULL ||
8713 mops->dtms_create_probe == NULL ||
8714 mops->dtms_provide_pid == NULL ||
8715 mops->dtms_remove_pid == NULL) {
8716 cmn_err(CE_WARN, "failed to register meta-register %s: "
8717 "invalid ops", name);
8718 return (EINVAL);
8719 }
8720
8721 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8722 meta->dtm_mops = *mops;
8723 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8724 (void) strcpy(meta->dtm_name, name);
8725 meta->dtm_arg = arg;
8726
8727 mutex_enter(&dtrace_meta_lock);
8728 mutex_enter(&dtrace_lock);
8729
8730 if (dtrace_meta_pid != NULL) {
8731 mutex_exit(&dtrace_lock);
8732 mutex_exit(&dtrace_meta_lock);
8733 cmn_err(CE_WARN, "failed to register meta-register %s: "
8734 "user-land meta-provider exists", name);
8735 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8736 kmem_free(meta, sizeof (dtrace_meta_t));
8737 return (EINVAL);
8738 }
8739
8740 dtrace_meta_pid = meta;
8741 *idp = (dtrace_meta_provider_id_t)meta;
8742
8743 /*
8744 * If there are providers and probes ready to go, pass them
8745 * off to the new meta provider now.
8746 */
8747
8748 help = dtrace_deferred_pid;
8749 dtrace_deferred_pid = NULL;
8750
8751 mutex_exit(&dtrace_lock);
8752
8753 while (help != NULL) {
8754 for (i = 0; i < help->dthps_nprovs; i++) {
8755 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8756 help->dthps_pid);
8757 }
8758
8759 next = help->dthps_next;
8760 help->dthps_next = NULL;
8761 help->dthps_prev = NULL;
8762 help->dthps_deferred = 0;
8763 help = next;
8764 }
8765
8766 mutex_exit(&dtrace_meta_lock);
8767
8768 return (0);
8769 }
8770
8771 int
8772 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8773 {
8774 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8775
8776 mutex_enter(&dtrace_meta_lock);
8777 mutex_enter(&dtrace_lock);
8778
8779 if (old == dtrace_meta_pid) {
8780 pp = &dtrace_meta_pid;
8781 } else {
8782 panic("attempt to unregister non-existent "
8783 "dtrace meta-provider %p\n", (void *)old);
8784 }
8785
8786 if (old->dtm_count != 0) {
8787 mutex_exit(&dtrace_lock);
8788 mutex_exit(&dtrace_meta_lock);
8789 return (EBUSY);
8790 }
8791
8792 *pp = NULL;
8793
8794 mutex_exit(&dtrace_lock);
8795 mutex_exit(&dtrace_meta_lock);
8796
8797 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8798 kmem_free(old, sizeof (dtrace_meta_t));
8799
8800 return (0);
8801 }
8802
8803
8804 /*
8805 * DTrace DIF Object Functions
8806 */
8807 static int
8808 dtrace_difo_err(uint_t pc, const char *format, ...)
8809 {
8810 if (dtrace_err_verbose) {
8811 va_list alist;
8812
8813 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8814 va_start(alist, format);
8815 (void) vuprintf(format, alist);
8816 va_end(alist);
8817 }
8818
8819 #ifdef DTRACE_ERRDEBUG
8820 dtrace_errdebug(format);
8821 #endif
8822 return (1);
8823 }
8824
8825 /*
8826 * Validate a DTrace DIF object by checking the IR instructions. The following
8827 * rules are currently enforced by dtrace_difo_validate():
8828 *
8829 * 1. Each instruction must have a valid opcode
8830 * 2. Each register, string, variable, or subroutine reference must be valid
8831 * 3. No instruction can modify register %r0 (must be zero)
8832 * 4. All instruction reserved bits must be set to zero
8833 * 5. The last instruction must be a "ret" instruction
8834 * 6. All branch targets must reference a valid instruction _after_ the branch
8835 */
8836 static int
8837 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8838 cred_t *cr)
8839 {
8840 int err = 0, i;
8841 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8842 int kcheckload;
8843 uint_t pc;
8844
8845 kcheckload = cr == NULL ||
8846 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8847
8848 dp->dtdo_destructive = 0;
8849
8850 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8851 dif_instr_t instr = dp->dtdo_buf[pc];
8852
8853 uint_t r1 = DIF_INSTR_R1(instr);
8854 uint_t r2 = DIF_INSTR_R2(instr);
8855 uint_t rd = DIF_INSTR_RD(instr);
8856 uint_t rs = DIF_INSTR_RS(instr);
8857 uint_t label = DIF_INSTR_LABEL(instr);
8858 uint_t v = DIF_INSTR_VAR(instr);
8859 uint_t subr = DIF_INSTR_SUBR(instr);
8860 uint_t type = DIF_INSTR_TYPE(instr);
8861 uint_t op = DIF_INSTR_OP(instr);
8862
8863 switch (op) {
8864 case DIF_OP_OR:
8865 case DIF_OP_XOR:
8866 case DIF_OP_AND:
8867 case DIF_OP_SLL:
8868 case DIF_OP_SRL:
8869 case DIF_OP_SRA:
8870 case DIF_OP_SUB:
8871 case DIF_OP_ADD:
8872 case DIF_OP_MUL:
8873 case DIF_OP_SDIV:
8874 case DIF_OP_UDIV:
8875 case DIF_OP_SREM:
8876 case DIF_OP_UREM:
8877 case DIF_OP_COPYS:
8878 if (r1 >= nregs)
8879 err += efunc(pc, "invalid register %u\n", r1);
8880 if (r2 >= nregs)
8881 err += efunc(pc, "invalid register %u\n", r2);
8882 if (rd >= nregs)
8883 err += efunc(pc, "invalid register %u\n", rd);
8884 if (rd == 0)
8885 err += efunc(pc, "cannot write to %r0\n");
8886 break;
8887 case DIF_OP_NOT:
8888 case DIF_OP_MOV:
8889 case DIF_OP_ALLOCS:
8890 if (r1 >= nregs)
8891 err += efunc(pc, "invalid register %u\n", r1);
8892 if (r2 != 0)
8893 err += efunc(pc, "non-zero reserved bits\n");
8894 if (rd >= nregs)
8895 err += efunc(pc, "invalid register %u\n", rd);
8896 if (rd == 0)
8897 err += efunc(pc, "cannot write to %r0\n");
8898 break;
8899 case DIF_OP_LDSB:
8900 case DIF_OP_LDSH:
8901 case DIF_OP_LDSW:
8902 case DIF_OP_LDUB:
8903 case DIF_OP_LDUH:
8904 case DIF_OP_LDUW:
8905 case DIF_OP_LDX:
8906 if (r1 >= nregs)
8907 err += efunc(pc, "invalid register %u\n", r1);
8908 if (r2 != 0)
8909 err += efunc(pc, "non-zero reserved bits\n");
8910 if (rd >= nregs)
8911 err += efunc(pc, "invalid register %u\n", rd);
8912 if (rd == 0)
8913 err += efunc(pc, "cannot write to %r0\n");
8914 if (kcheckload)
8915 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8916 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8917 break;
8918 case DIF_OP_RLDSB:
8919 case DIF_OP_RLDSH:
8920 case DIF_OP_RLDSW:
8921 case DIF_OP_RLDUB:
8922 case DIF_OP_RLDUH:
8923 case DIF_OP_RLDUW:
8924 case DIF_OP_RLDX:
8925 if (r1 >= nregs)
8926 err += efunc(pc, "invalid register %u\n", r1);
8927 if (r2 != 0)
8928 err += efunc(pc, "non-zero reserved bits\n");
8929 if (rd >= nregs)
8930 err += efunc(pc, "invalid register %u\n", rd);
8931 if (rd == 0)
8932 err += efunc(pc, "cannot write to %r0\n");
8933 break;
8934 case DIF_OP_ULDSB:
8935 case DIF_OP_ULDSH:
8936 case DIF_OP_ULDSW:
8937 case DIF_OP_ULDUB:
8938 case DIF_OP_ULDUH:
8939 case DIF_OP_ULDUW:
8940 case DIF_OP_ULDX:
8941 if (r1 >= nregs)
8942 err += efunc(pc, "invalid register %u\n", r1);
8943 if (r2 != 0)
8944 err += efunc(pc, "non-zero reserved bits\n");
8945 if (rd >= nregs)
8946 err += efunc(pc, "invalid register %u\n", rd);
8947 if (rd == 0)
8948 err += efunc(pc, "cannot write to %r0\n");
8949 break;
8950 case DIF_OP_STB:
8951 case DIF_OP_STH:
8952 case DIF_OP_STW:
8953 case DIF_OP_STX:
8954 if (r1 >= nregs)
8955 err += efunc(pc, "invalid register %u\n", r1);
8956 if (r2 != 0)
8957 err += efunc(pc, "non-zero reserved bits\n");
8958 if (rd >= nregs)
8959 err += efunc(pc, "invalid register %u\n", rd);
8960 if (rd == 0)
8961 err += efunc(pc, "cannot write to 0 address\n");
8962 break;
8963 case DIF_OP_CMP:
8964 case DIF_OP_SCMP:
8965 if (r1 >= nregs)
8966 err += efunc(pc, "invalid register %u\n", r1);
8967 if (r2 >= nregs)
8968 err += efunc(pc, "invalid register %u\n", r2);
8969 if (rd != 0)
8970 err += efunc(pc, "non-zero reserved bits\n");
8971 break;
8972 case DIF_OP_TST:
8973 if (r1 >= nregs)
8974 err += efunc(pc, "invalid register %u\n", r1);
8975 if (r2 != 0 || rd != 0)
8976 err += efunc(pc, "non-zero reserved bits\n");
8977 break;
8978 case DIF_OP_BA:
8979 case DIF_OP_BE:
8980 case DIF_OP_BNE:
8981 case DIF_OP_BG:
8982 case DIF_OP_BGU:
8983 case DIF_OP_BGE:
8984 case DIF_OP_BGEU:
8985 case DIF_OP_BL:
8986 case DIF_OP_BLU:
8987 case DIF_OP_BLE:
8988 case DIF_OP_BLEU:
8989 if (label >= dp->dtdo_len) {
8990 err += efunc(pc, "invalid branch target %u\n",
8991 label);
8992 }
8993 if (label <= pc) {
8994 err += efunc(pc, "backward branch to %u\n",
8995 label);
8996 }
8997 break;
8998 case DIF_OP_RET:
8999 if (r1 != 0 || r2 != 0)
9000 err += efunc(pc, "non-zero reserved bits\n");
9001 if (rd >= nregs)
9002 err += efunc(pc, "invalid register %u\n", rd);
9003 break;
9004 case DIF_OP_NOP:
9005 case DIF_OP_POPTS:
9006 case DIF_OP_FLUSHTS:
9007 if (r1 != 0 || r2 != 0 || rd != 0)
9008 err += efunc(pc, "non-zero reserved bits\n");
9009 break;
9010 case DIF_OP_SETX:
9011 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9012 err += efunc(pc, "invalid integer ref %u\n",
9013 DIF_INSTR_INTEGER(instr));
9014 }
9015 if (rd >= nregs)
9016 err += efunc(pc, "invalid register %u\n", rd);
9017 if (rd == 0)
9018 err += efunc(pc, "cannot write to %r0\n");
9019 break;
9020 case DIF_OP_SETS:
9021 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9022 err += efunc(pc, "invalid string ref %u\n",
9023 DIF_INSTR_STRING(instr));
9024 }
9025 if (rd >= nregs)
9026 err += efunc(pc, "invalid register %u\n", rd);
9027 if (rd == 0)
9028 err += efunc(pc, "cannot write to %r0\n");
9029 break;
9030 case DIF_OP_LDGA:
9031 case DIF_OP_LDTA:
9032 if (r1 > DIF_VAR_ARRAY_MAX)
9033 err += efunc(pc, "invalid array %u\n", r1);
9034 if (r2 >= nregs)
9035 err += efunc(pc, "invalid register %u\n", r2);
9036 if (rd >= nregs)
9037 err += efunc(pc, "invalid register %u\n", rd);
9038 if (rd == 0)
9039 err += efunc(pc, "cannot write to %r0\n");
9040 break;
9041 case DIF_OP_LDGS:
9042 case DIF_OP_LDTS:
9043 case DIF_OP_LDLS:
9044 case DIF_OP_LDGAA:
9045 case DIF_OP_LDTAA:
9046 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9047 err += efunc(pc, "invalid variable %u\n", v);
9048 if (rd >= nregs)
9049 err += efunc(pc, "invalid register %u\n", rd);
9050 if (rd == 0)
9051 err += efunc(pc, "cannot write to %r0\n");
9052 break;
9053 case DIF_OP_STGS:
9054 case DIF_OP_STTS:
9055 case DIF_OP_STLS:
9056 case DIF_OP_STGAA:
9057 case DIF_OP_STTAA:
9058 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9059 err += efunc(pc, "invalid variable %u\n", v);
9060 if (rs >= nregs)
9061 err += efunc(pc, "invalid register %u\n", rd);
9062 break;
9063 case DIF_OP_CALL:
9064 if (subr > DIF_SUBR_MAX)
9065 err += efunc(pc, "invalid subr %u\n", subr);
9066 if (rd >= nregs)
9067 err += efunc(pc, "invalid register %u\n", rd);
9068 if (rd == 0)
9069 err += efunc(pc, "cannot write to %r0\n");
9070
9071 if (subr == DIF_SUBR_COPYOUT ||
9072 subr == DIF_SUBR_COPYOUTSTR) {
9073 dp->dtdo_destructive = 1;
9074 }
9075
9076 if (subr == DIF_SUBR_GETF) {
9077 /*
9078 * If we have a getf() we need to record that
9079 * in our state. Note that our state can be
9080 * NULL if this is a helper -- but in that
9081 * case, the call to getf() is itself illegal,
9082 * and will be caught (slightly later) when
9083 * the helper is validated.
9084 */
9085 if (vstate->dtvs_state != NULL)
9086 vstate->dtvs_state->dts_getf++;
9087 }
9088
9089 break;
9090 case DIF_OP_PUSHTR:
9091 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
9092 err += efunc(pc, "invalid ref type %u\n", type);
9093 if (r2 >= nregs)
9094 err += efunc(pc, "invalid register %u\n", r2);
9095 if (rs >= nregs)
9096 err += efunc(pc, "invalid register %u\n", rs);
9097 break;
9098 case DIF_OP_PUSHTV:
9099 if (type != DIF_TYPE_CTF)
9100 err += efunc(pc, "invalid val type %u\n", type);
9101 if (r2 >= nregs)
9102 err += efunc(pc, "invalid register %u\n", r2);
9103 if (rs >= nregs)
9104 err += efunc(pc, "invalid register %u\n", rs);
9105 break;
9106 default:
9107 err += efunc(pc, "invalid opcode %u\n",
9108 DIF_INSTR_OP(instr));
9109 }
9110 }
9111
9112 if (dp->dtdo_len != 0 &&
9113 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
9114 err += efunc(dp->dtdo_len - 1,
9115 "expected 'ret' as last DIF instruction\n");
9116 }
9117
9118 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
9119 /*
9120 * If we're not returning by reference, the size must be either
9121 * 0 or the size of one of the base types.
9122 */
9123 switch (dp->dtdo_rtype.dtdt_size) {
9124 case 0:
9125 case sizeof (uint8_t):
9126 case sizeof (uint16_t):
9127 case sizeof (uint32_t):
9128 case sizeof (uint64_t):
9129 break;
9130
9131 default:
9132 err += efunc(dp->dtdo_len - 1, "bad return size\n");
9133 }
9134 }
9135
9136 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
9137 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
9138 dtrace_diftype_t *vt, *et;
9139 uint_t id, ndx;
9140
9141 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
9142 v->dtdv_scope != DIFV_SCOPE_THREAD &&
9143 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
9144 err += efunc(i, "unrecognized variable scope %d\n",
9145 v->dtdv_scope);
9146 break;
9147 }
9148
9149 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
9150 v->dtdv_kind != DIFV_KIND_SCALAR) {
9151 err += efunc(i, "unrecognized variable type %d\n",
9152 v->dtdv_kind);
9153 break;
9154 }
9155
9156 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
9157 err += efunc(i, "%d exceeds variable id limit\n", id);
9158 break;
9159 }
9160
9161 if (id < DIF_VAR_OTHER_UBASE)
9162 continue;
9163
9164 /*
9165 * For user-defined variables, we need to check that this
9166 * definition is identical to any previous definition that we
9167 * encountered.
9168 */
9169 ndx = id - DIF_VAR_OTHER_UBASE;
9170
9171 switch (v->dtdv_scope) {
9172 case DIFV_SCOPE_GLOBAL:
9173 if (ndx < vstate->dtvs_nglobals) {
9174 dtrace_statvar_t *svar;
9175
9176 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
9177 existing = &svar->dtsv_var;
9178 }
9179
9180 break;
9181
9182 case DIFV_SCOPE_THREAD:
9183 if (ndx < vstate->dtvs_ntlocals)
9184 existing = &vstate->dtvs_tlocals[ndx];
9185 break;
9186
9187 case DIFV_SCOPE_LOCAL:
9188 if (ndx < vstate->dtvs_nlocals) {
9189 dtrace_statvar_t *svar;
9190
9191 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
9192 existing = &svar->dtsv_var;
9193 }
9194
9195 break;
9196 }
9197
9198 vt = &v->dtdv_type;
9199
9200 if (vt->dtdt_flags & DIF_TF_BYREF) {
9201 if (vt->dtdt_size == 0) {
9202 err += efunc(i, "zero-sized variable\n");
9203 break;
9204 }
9205
9206 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
9207 vt->dtdt_size > dtrace_global_maxsize) {
9208 err += efunc(i, "oversized by-ref global\n");
9209 break;
9210 }
9211 }
9212
9213 if (existing == NULL || existing->dtdv_id == 0)
9214 continue;
9215
9216 ASSERT(existing->dtdv_id == v->dtdv_id);
9217 ASSERT(existing->dtdv_scope == v->dtdv_scope);
9218
9219 if (existing->dtdv_kind != v->dtdv_kind)
9220 err += efunc(i, "%d changed variable kind\n", id);
9221
9222 et = &existing->dtdv_type;
9223
9224 if (vt->dtdt_flags != et->dtdt_flags) {
9225 err += efunc(i, "%d changed variable type flags\n", id);
9226 break;
9227 }
9228
9229 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
9230 err += efunc(i, "%d changed variable type size\n", id);
9231 break;
9232 }
9233 }
9234
9235 return (err);
9236 }
9237
9238 /*
9239 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
9240 * are much more constrained than normal DIFOs. Specifically, they may
9241 * not:
9242 *
9243 * 1. Make calls to subroutines other than copyin(), copyinstr() or
9244 * miscellaneous string routines
9245 * 2. Access DTrace variables other than the args[] array, and the
9246 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9247 * 3. Have thread-local variables.
9248 * 4. Have dynamic variables.
9249 */
9250 static int
9251 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9252 {
9253 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9254 int err = 0;
9255 uint_t pc;
9256
9257 for (pc = 0; pc < dp->dtdo_len; pc++) {
9258 dif_instr_t instr = dp->dtdo_buf[pc];
9259
9260 uint_t v = DIF_INSTR_VAR(instr);
9261 uint_t subr = DIF_INSTR_SUBR(instr);
9262 uint_t op = DIF_INSTR_OP(instr);
9263
9264 switch (op) {
9265 case DIF_OP_OR:
9266 case DIF_OP_XOR:
9267 case DIF_OP_AND:
9268 case DIF_OP_SLL:
9269 case DIF_OP_SRL:
9270 case DIF_OP_SRA:
9271 case DIF_OP_SUB:
9272 case DIF_OP_ADD:
9273 case DIF_OP_MUL:
9274 case DIF_OP_SDIV:
9275 case DIF_OP_UDIV:
9276 case DIF_OP_SREM:
9277 case DIF_OP_UREM:
9278 case DIF_OP_COPYS:
9279 case DIF_OP_NOT:
9280 case DIF_OP_MOV:
9281 case DIF_OP_RLDSB:
9282 case DIF_OP_RLDSH:
9283 case DIF_OP_RLDSW:
9284 case DIF_OP_RLDUB:
9285 case DIF_OP_RLDUH:
9286 case DIF_OP_RLDUW:
9287 case DIF_OP_RLDX:
9288 case DIF_OP_ULDSB:
9289 case DIF_OP_ULDSH:
9290 case DIF_OP_ULDSW:
9291 case DIF_OP_ULDUB:
9292 case DIF_OP_ULDUH:
9293 case DIF_OP_ULDUW:
9294 case DIF_OP_ULDX:
9295 case DIF_OP_STB:
9296 case DIF_OP_STH:
9297 case DIF_OP_STW:
9298 case DIF_OP_STX:
9299 case DIF_OP_ALLOCS:
9300 case DIF_OP_CMP:
9301 case DIF_OP_SCMP:
9302 case DIF_OP_TST:
9303 case DIF_OP_BA:
9304 case DIF_OP_BE:
9305 case DIF_OP_BNE:
9306 case DIF_OP_BG:
9307 case DIF_OP_BGU:
9308 case DIF_OP_BGE:
9309 case DIF_OP_BGEU:
9310 case DIF_OP_BL:
9311 case DIF_OP_BLU:
9312 case DIF_OP_BLE:
9313 case DIF_OP_BLEU:
9314 case DIF_OP_RET:
9315 case DIF_OP_NOP:
9316 case DIF_OP_POPTS:
9317 case DIF_OP_FLUSHTS:
9318 case DIF_OP_SETX:
9319 case DIF_OP_SETS:
9320 case DIF_OP_LDGA:
9321 case DIF_OP_LDLS:
9322 case DIF_OP_STGS:
9323 case DIF_OP_STLS:
9324 case DIF_OP_PUSHTR:
9325 case DIF_OP_PUSHTV:
9326 break;
9327
9328 case DIF_OP_LDGS:
9329 if (v >= DIF_VAR_OTHER_UBASE)
9330 break;
9331
9332 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9333 break;
9334
9335 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9336 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9337 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9338 v == DIF_VAR_UID || v == DIF_VAR_GID)
9339 break;
9340
9341 err += efunc(pc, "illegal variable %u\n", v);
9342 break;
9343
9344 case DIF_OP_LDTA:
9345 case DIF_OP_LDTS:
9346 case DIF_OP_LDGAA:
9347 case DIF_OP_LDTAA:
9348 err += efunc(pc, "illegal dynamic variable load\n");
9349 break;
9350
9351 case DIF_OP_STTS:
9352 case DIF_OP_STGAA:
9353 case DIF_OP_STTAA:
9354 err += efunc(pc, "illegal dynamic variable store\n");
9355 break;
9356
9357 case DIF_OP_CALL:
9358 if (subr == DIF_SUBR_ALLOCA ||
9359 subr == DIF_SUBR_BCOPY ||
9360 subr == DIF_SUBR_COPYIN ||
9361 subr == DIF_SUBR_COPYINTO ||
9362 subr == DIF_SUBR_COPYINSTR ||
9363 subr == DIF_SUBR_INDEX ||
9364 subr == DIF_SUBR_INET_NTOA ||
9365 subr == DIF_SUBR_INET_NTOA6 ||
9366 subr == DIF_SUBR_INET_NTOP ||
9367 subr == DIF_SUBR_LLTOSTR ||
9368 subr == DIF_SUBR_STRTOLL ||
9369 subr == DIF_SUBR_RINDEX ||
9370 subr == DIF_SUBR_STRCHR ||
9371 subr == DIF_SUBR_STRJOIN ||
9372 subr == DIF_SUBR_STRRCHR ||
9373 subr == DIF_SUBR_STRSTR ||
9374 subr == DIF_SUBR_HTONS ||
9375 subr == DIF_SUBR_HTONL ||
9376 subr == DIF_SUBR_HTONLL ||
9377 subr == DIF_SUBR_NTOHS ||
9378 subr == DIF_SUBR_NTOHL ||
9379 subr == DIF_SUBR_NTOHLL)
9380 break;
9381
9382 err += efunc(pc, "invalid subr %u\n", subr);
9383 break;
9384
9385 default:
9386 err += efunc(pc, "invalid opcode %u\n",
9387 DIF_INSTR_OP(instr));
9388 }
9389 }
9390
9391 return (err);
9392 }
9393
9394 /*
9395 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9396 * basis; 0 if not.
9397 */
9398 static int
9399 dtrace_difo_cacheable(dtrace_difo_t *dp)
9400 {
9401 int i;
9402
9403 if (dp == NULL)
9404 return (0);
9405
9406 for (i = 0; i < dp->dtdo_varlen; i++) {
9407 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9408
9409 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9410 continue;
9411
9412 switch (v->dtdv_id) {
9413 case DIF_VAR_CURTHREAD:
9414 case DIF_VAR_PID:
9415 case DIF_VAR_TID:
9416 case DIF_VAR_EXECNAME:
9417 case DIF_VAR_ZONENAME:
9418 break;
9419
9420 default:
9421 return (0);
9422 }
9423 }
9424
9425 /*
9426 * This DIF object may be cacheable. Now we need to look for any
9427 * array loading instructions, any memory loading instructions, or
9428 * any stores to thread-local variables.
9429 */
9430 for (i = 0; i < dp->dtdo_len; i++) {
9431 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9432
9433 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9434 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9435 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9436 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9437 return (0);
9438 }
9439
9440 return (1);
9441 }
9442
9443 static void
9444 dtrace_difo_hold(dtrace_difo_t *dp)
9445 {
9446 int i;
9447
9448 ASSERT(MUTEX_HELD(&dtrace_lock));
9449
9450 dp->dtdo_refcnt++;
9451 ASSERT(dp->dtdo_refcnt != 0);
9452
9453 /*
9454 * We need to check this DIF object for references to the variable
9455 * DIF_VAR_VTIMESTAMP.
9456 */
9457 for (i = 0; i < dp->dtdo_varlen; i++) {
9458 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9459
9460 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9461 continue;
9462
9463 if (dtrace_vtime_references++ == 0)
9464 dtrace_vtime_enable();
9465 }
9466 }
9467
9468 /*
9469 * This routine calculates the dynamic variable chunksize for a given DIF
9470 * object. The calculation is not fool-proof, and can probably be tricked by
9471 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9472 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9473 * if a dynamic variable size exceeds the chunksize.
9474 */
9475 static void
9476 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9477 {
9478 uint64_t sval;
9479 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9480 const dif_instr_t *text = dp->dtdo_buf;
9481 uint_t pc, srd = 0;
9482 uint_t ttop = 0;
9483 size_t size, ksize;
9484 uint_t id, i;
9485
9486 for (pc = 0; pc < dp->dtdo_len; pc++) {
9487 dif_instr_t instr = text[pc];
9488 uint_t op = DIF_INSTR_OP(instr);
9489 uint_t rd = DIF_INSTR_RD(instr);
9490 uint_t r1 = DIF_INSTR_R1(instr);
9491 uint_t nkeys = 0;
9492 uchar_t scope;
9493
9494 dtrace_key_t *key = tupregs;
9495
9496 switch (op) {
9497 case DIF_OP_SETX:
9498 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9499 srd = rd;
9500 continue;
9501
9502 case DIF_OP_STTS:
9503 key = &tupregs[DIF_DTR_NREGS];
9504 key[0].dttk_size = 0;
9505 key[1].dttk_size = 0;
9506 nkeys = 2;
9507 scope = DIFV_SCOPE_THREAD;
9508 break;
9509
9510 case DIF_OP_STGAA:
9511 case DIF_OP_STTAA:
9512 nkeys = ttop;
9513
9514 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9515 key[nkeys++].dttk_size = 0;
9516
9517 key[nkeys++].dttk_size = 0;
9518
9519 if (op == DIF_OP_STTAA) {
9520 scope = DIFV_SCOPE_THREAD;
9521 } else {
9522 scope = DIFV_SCOPE_GLOBAL;
9523 }
9524
9525 break;
9526
9527 case DIF_OP_PUSHTR:
9528 if (ttop == DIF_DTR_NREGS)
9529 return;
9530
9531 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9532 /*
9533 * If the register for the size of the "pushtr"
9534 * is %r0 (or the value is 0) and the type is
9535 * a string, we'll use the system-wide default
9536 * string size.
9537 */
9538 tupregs[ttop++].dttk_size =
9539 dtrace_strsize_default;
9540 } else {
9541 if (srd == 0)
9542 return;
9543
9544 tupregs[ttop++].dttk_size = sval;
9545 }
9546
9547 break;
9548
9549 case DIF_OP_PUSHTV:
9550 if (ttop == DIF_DTR_NREGS)
9551 return;
9552
9553 tupregs[ttop++].dttk_size = 0;
9554 break;
9555
9556 case DIF_OP_FLUSHTS:
9557 ttop = 0;
9558 break;
9559
9560 case DIF_OP_POPTS:
9561 if (ttop != 0)
9562 ttop--;
9563 break;
9564 }
9565
9566 sval = 0;
9567 srd = 0;
9568
9569 if (nkeys == 0)
9570 continue;
9571
9572 /*
9573 * We have a dynamic variable allocation; calculate its size.
9574 */
9575 for (ksize = 0, i = 0; i < nkeys; i++)
9576 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9577
9578 size = sizeof (dtrace_dynvar_t);
9579 size += sizeof (dtrace_key_t) * (nkeys - 1);
9580 size += ksize;
9581
9582 /*
9583 * Now we need to determine the size of the stored data.
9584 */
9585 id = DIF_INSTR_VAR(instr);
9586
9587 for (i = 0; i < dp->dtdo_varlen; i++) {
9588 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9589
9590 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9591 size += v->dtdv_type.dtdt_size;
9592 break;
9593 }
9594 }
9595
9596 if (i == dp->dtdo_varlen)
9597 return;
9598
9599 /*
9600 * We have the size. If this is larger than the chunk size
9601 * for our dynamic variable state, reset the chunk size.
9602 */
9603 size = P2ROUNDUP(size, sizeof (uint64_t));
9604
9605 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9606 vstate->dtvs_dynvars.dtds_chunksize = size;
9607 }
9608 }
9609
9610 static void
9611 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9612 {
9613 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9614 uint_t id;
9615
9616 ASSERT(MUTEX_HELD(&dtrace_lock));
9617 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9618
9619 for (i = 0; i < dp->dtdo_varlen; i++) {
9620 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9621 dtrace_statvar_t *svar, ***svarp;
9622 size_t dsize = 0;
9623 uint8_t scope = v->dtdv_scope;
9624 int *np;
9625
9626 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9627 continue;
9628
9629 id -= DIF_VAR_OTHER_UBASE;
9630
9631 switch (scope) {
9632 case DIFV_SCOPE_THREAD:
9633 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9634 dtrace_difv_t *tlocals;
9635
9636 if ((ntlocals = (otlocals << 1)) == 0)
9637 ntlocals = 1;
9638
9639 osz = otlocals * sizeof (dtrace_difv_t);
9640 nsz = ntlocals * sizeof (dtrace_difv_t);
9641
9642 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9643
9644 if (osz != 0) {
9645 bcopy(vstate->dtvs_tlocals,
9646 tlocals, osz);
9647 kmem_free(vstate->dtvs_tlocals, osz);
9648 }
9649
9650 vstate->dtvs_tlocals = tlocals;
9651 vstate->dtvs_ntlocals = ntlocals;
9652 }
9653
9654 vstate->dtvs_tlocals[id] = *v;
9655 continue;
9656
9657 case DIFV_SCOPE_LOCAL:
9658 np = &vstate->dtvs_nlocals;
9659 svarp = &vstate->dtvs_locals;
9660
9661 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9662 dsize = NCPU * (v->dtdv_type.dtdt_size +
9663 sizeof (uint64_t));
9664 else
9665 dsize = NCPU * sizeof (uint64_t);
9666
9667 break;
9668
9669 case DIFV_SCOPE_GLOBAL:
9670 np = &vstate->dtvs_nglobals;
9671 svarp = &vstate->dtvs_globals;
9672
9673 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9674 dsize = v->dtdv_type.dtdt_size +
9675 sizeof (uint64_t);
9676
9677 break;
9678
9679 default:
9680 ASSERT(0);
9681 }
9682
9683 while (id >= (oldsvars = *np)) {
9684 dtrace_statvar_t **statics;
9685 int newsvars, oldsize, newsize;
9686
9687 if ((newsvars = (oldsvars << 1)) == 0)
9688 newsvars = 1;
9689
9690 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9691 newsize = newsvars * sizeof (dtrace_statvar_t *);
9692
9693 statics = kmem_zalloc(newsize, KM_SLEEP);
9694
9695 if (oldsize != 0) {
9696 bcopy(*svarp, statics, oldsize);
9697 kmem_free(*svarp, oldsize);
9698 }
9699
9700 *svarp = statics;
9701 *np = newsvars;
9702 }
9703
9704 if ((svar = (*svarp)[id]) == NULL) {
9705 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9706 svar->dtsv_var = *v;
9707
9708 if ((svar->dtsv_size = dsize) != 0) {
9709 svar->dtsv_data = (uint64_t)(uintptr_t)
9710 kmem_zalloc(dsize, KM_SLEEP);
9711 }
9712
9713 (*svarp)[id] = svar;
9714 }
9715
9716 svar->dtsv_refcnt++;
9717 }
9718
9719 dtrace_difo_chunksize(dp, vstate);
9720 dtrace_difo_hold(dp);
9721 }
9722
9723 static dtrace_difo_t *
9724 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9725 {
9726 dtrace_difo_t *new;
9727 size_t sz;
9728
9729 ASSERT(dp->dtdo_buf != NULL);
9730 ASSERT(dp->dtdo_refcnt != 0);
9731
9732 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9733
9734 ASSERT(dp->dtdo_buf != NULL);
9735 sz = dp->dtdo_len * sizeof (dif_instr_t);
9736 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9737 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9738 new->dtdo_len = dp->dtdo_len;
9739
9740 if (dp->dtdo_strtab != NULL) {
9741 ASSERT(dp->dtdo_strlen != 0);
9742 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9743 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9744 new->dtdo_strlen = dp->dtdo_strlen;
9745 }
9746
9747 if (dp->dtdo_inttab != NULL) {
9748 ASSERT(dp->dtdo_intlen != 0);
9749 sz = dp->dtdo_intlen * sizeof (uint64_t);
9750 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9751 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9752 new->dtdo_intlen = dp->dtdo_intlen;
9753 }
9754
9755 if (dp->dtdo_vartab != NULL) {
9756 ASSERT(dp->dtdo_varlen != 0);
9757 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9758 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9759 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9760 new->dtdo_varlen = dp->dtdo_varlen;
9761 }
9762
9763 dtrace_difo_init(new, vstate);
9764 return (new);
9765 }
9766
9767 static void
9768 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9769 {
9770 int i;
9771
9772 ASSERT(dp->dtdo_refcnt == 0);
9773
9774 for (i = 0; i < dp->dtdo_varlen; i++) {
9775 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9776 dtrace_statvar_t *svar, **svarp;
9777 uint_t id;
9778 uint8_t scope = v->dtdv_scope;
9779 int *np;
9780
9781 switch (scope) {
9782 case DIFV_SCOPE_THREAD:
9783 continue;
9784
9785 case DIFV_SCOPE_LOCAL:
9786 np = &vstate->dtvs_nlocals;
9787 svarp = vstate->dtvs_locals;
9788 break;
9789
9790 case DIFV_SCOPE_GLOBAL:
9791 np = &vstate->dtvs_nglobals;
9792 svarp = vstate->dtvs_globals;
9793 break;
9794
9795 default:
9796 ASSERT(0);
9797 }
9798
9799 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9800 continue;
9801
9802 id -= DIF_VAR_OTHER_UBASE;
9803 ASSERT(id < *np);
9804
9805 svar = svarp[id];
9806 ASSERT(svar != NULL);
9807 ASSERT(svar->dtsv_refcnt > 0);
9808
9809 if (--svar->dtsv_refcnt > 0)
9810 continue;
9811
9812 if (svar->dtsv_size != 0) {
9813 ASSERT(svar->dtsv_data != NULL);
9814 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9815 svar->dtsv_size);
9816 }
9817
9818 kmem_free(svar, sizeof (dtrace_statvar_t));
9819 svarp[id] = NULL;
9820 }
9821
9822 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9823 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9824 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9825 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9826
9827 kmem_free(dp, sizeof (dtrace_difo_t));
9828 }
9829
9830 static void
9831 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9832 {
9833 int i;
9834
9835 ASSERT(MUTEX_HELD(&dtrace_lock));
9836 ASSERT(dp->dtdo_refcnt != 0);
9837
9838 for (i = 0; i < dp->dtdo_varlen; i++) {
9839 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9840
9841 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9842 continue;
9843
9844 ASSERT(dtrace_vtime_references > 0);
9845 if (--dtrace_vtime_references == 0)
9846 dtrace_vtime_disable();
9847 }
9848
9849 if (--dp->dtdo_refcnt == 0)
9850 dtrace_difo_destroy(dp, vstate);
9851 }
9852
9853 /*
9854 * DTrace Format Functions
9855 */
9856 static uint16_t
9857 dtrace_format_add(dtrace_state_t *state, char *str)
9858 {
9859 char *fmt, **new;
9860 uint16_t ndx, len = strlen(str) + 1;
9861
9862 fmt = kmem_zalloc(len, KM_SLEEP);
9863 bcopy(str, fmt, len);
9864
9865 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9866 if (state->dts_formats[ndx] == NULL) {
9867 state->dts_formats[ndx] = fmt;
9868 return (ndx + 1);
9869 }
9870 }
9871
9872 if (state->dts_nformats == USHRT_MAX) {
9873 /*
9874 * This is only likely if a denial-of-service attack is being
9875 * attempted. As such, it's okay to fail silently here.
9876 */
9877 kmem_free(fmt, len);
9878 return (0);
9879 }
9880
9881 /*
9882 * For simplicity, we always resize the formats array to be exactly the
9883 * number of formats.
9884 */
9885 ndx = state->dts_nformats++;
9886 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9887
9888 if (state->dts_formats != NULL) {
9889 ASSERT(ndx != 0);
9890 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9891 kmem_free(state->dts_formats, ndx * sizeof (char *));
9892 }
9893
9894 state->dts_formats = new;
9895 state->dts_formats[ndx] = fmt;
9896
9897 return (ndx + 1);
9898 }
9899
9900 static void
9901 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9902 {
9903 char *fmt;
9904
9905 ASSERT(state->dts_formats != NULL);
9906 ASSERT(format <= state->dts_nformats);
9907 ASSERT(state->dts_formats[format - 1] != NULL);
9908
9909 fmt = state->dts_formats[format - 1];
9910 kmem_free(fmt, strlen(fmt) + 1);
9911 state->dts_formats[format - 1] = NULL;
9912 }
9913
9914 static void
9915 dtrace_format_destroy(dtrace_state_t *state)
9916 {
9917 int i;
9918
9919 if (state->dts_nformats == 0) {
9920 ASSERT(state->dts_formats == NULL);
9921 return;
9922 }
9923
9924 ASSERT(state->dts_formats != NULL);
9925
9926 for (i = 0; i < state->dts_nformats; i++) {
9927 char *fmt = state->dts_formats[i];
9928
9929 if (fmt == NULL)
9930 continue;
9931
9932 kmem_free(fmt, strlen(fmt) + 1);
9933 }
9934
9935 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9936 state->dts_nformats = 0;
9937 state->dts_formats = NULL;
9938 }
9939
9940 /*
9941 * DTrace Predicate Functions
9942 */
9943 static dtrace_predicate_t *
9944 dtrace_predicate_create(dtrace_difo_t *dp)
9945 {
9946 dtrace_predicate_t *pred;
9947
9948 ASSERT(MUTEX_HELD(&dtrace_lock));
9949 ASSERT(dp->dtdo_refcnt != 0);
9950
9951 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9952 pred->dtp_difo = dp;
9953 pred->dtp_refcnt = 1;
9954
9955 if (!dtrace_difo_cacheable(dp))
9956 return (pred);
9957
9958 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9959 /*
9960 * This is only theoretically possible -- we have had 2^32
9961 * cacheable predicates on this machine. We cannot allow any
9962 * more predicates to become cacheable: as unlikely as it is,
9963 * there may be a thread caching a (now stale) predicate cache
9964 * ID. (N.B.: the temptation is being successfully resisted to
9965 * have this cmn_err() "Holy shit -- we executed this code!")
9966 */
9967 return (pred);
9968 }
9969
9970 pred->dtp_cacheid = dtrace_predcache_id++;
9971
9972 return (pred);
9973 }
9974
9975 static void
9976 dtrace_predicate_hold(dtrace_predicate_t *pred)
9977 {
9978 ASSERT(MUTEX_HELD(&dtrace_lock));
9979 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9980 ASSERT(pred->dtp_refcnt > 0);
9981
9982 pred->dtp_refcnt++;
9983 }
9984
9985 static void
9986 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9987 {
9988 dtrace_difo_t *dp = pred->dtp_difo;
9989
9990 ASSERT(MUTEX_HELD(&dtrace_lock));
9991 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9992 ASSERT(pred->dtp_refcnt > 0);
9993
9994 if (--pred->dtp_refcnt == 0) {
9995 dtrace_difo_release(pred->dtp_difo, vstate);
9996 kmem_free(pred, sizeof (dtrace_predicate_t));
9997 }
9998 }
9999
10000 /*
10001 * DTrace Action Description Functions
10002 */
10003 static dtrace_actdesc_t *
10004 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
10005 uint64_t uarg, uint64_t arg)
10006 {
10007 dtrace_actdesc_t *act;
10008
10009 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
10010 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
10011
10012 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
10013 act->dtad_kind = kind;
10014 act->dtad_ntuple = ntuple;
10015 act->dtad_uarg = uarg;
10016 act->dtad_arg = arg;
10017 act->dtad_refcnt = 1;
10018
10019 return (act);
10020 }
10021
10022 static void
10023 dtrace_actdesc_hold(dtrace_actdesc_t *act)
10024 {
10025 ASSERT(act->dtad_refcnt >= 1);
10026 act->dtad_refcnt++;
10027 }
10028
10029 static void
10030 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
10031 {
10032 dtrace_actkind_t kind = act->dtad_kind;
10033 dtrace_difo_t *dp;
10034
10035 ASSERT(act->dtad_refcnt >= 1);
10036
10037 if (--act->dtad_refcnt != 0)
10038 return;
10039
10040 if ((dp = act->dtad_difo) != NULL)
10041 dtrace_difo_release(dp, vstate);
10042
10043 if (DTRACEACT_ISPRINTFLIKE(kind)) {
10044 char *str = (char *)(uintptr_t)act->dtad_arg;
10045
10046 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
10047 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
10048
10049 if (str != NULL)
10050 kmem_free(str, strlen(str) + 1);
10051 }
10052
10053 kmem_free(act, sizeof (dtrace_actdesc_t));
10054 }
10055
10056 /*
10057 * DTrace ECB Functions
10058 */
10059 static dtrace_ecb_t *
10060 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
10061 {
10062 dtrace_ecb_t *ecb;
10063 dtrace_epid_t epid;
10064
10065 ASSERT(MUTEX_HELD(&dtrace_lock));
10066
10067 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
10068 ecb->dte_predicate = NULL;
10069 ecb->dte_probe = probe;
10070
10071 /*
10072 * The default size is the size of the default action: recording
10073 * the header.
10074 */
10075 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
10076 ecb->dte_alignment = sizeof (dtrace_epid_t);
10077
10078 epid = state->dts_epid++;
10079
10080 if (epid - 1 >= state->dts_necbs) {
10081 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
10082 int necbs = state->dts_necbs << 1;
10083
10084 ASSERT(epid == state->dts_necbs + 1);
10085
10086 if (necbs == 0) {
10087 ASSERT(oecbs == NULL);
10088 necbs = 1;
10089 }
10090
10091 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
10092
10093 if (oecbs != NULL)
10094 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
10095
10096 dtrace_membar_producer();
10097 state->dts_ecbs = ecbs;
10098
10099 if (oecbs != NULL) {
10100 /*
10101 * If this state is active, we must dtrace_sync()
10102 * before we can free the old dts_ecbs array: we're
10103 * coming in hot, and there may be active ring
10104 * buffer processing (which indexes into the dts_ecbs
10105 * array) on another CPU.
10106 */
10107 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
10108 dtrace_sync();
10109
10110 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
10111 }
10112
10113 dtrace_membar_producer();
10114 state->dts_necbs = necbs;
10115 }
10116
10117 ecb->dte_state = state;
10118
10119 ASSERT(state->dts_ecbs[epid - 1] == NULL);
10120 dtrace_membar_producer();
10121 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
10122
10123 return (ecb);
10124 }
10125
10126 static int
10127 dtrace_ecb_enable(dtrace_ecb_t *ecb)
10128 {
10129 dtrace_probe_t *probe = ecb->dte_probe;
10130
10131 ASSERT(MUTEX_HELD(&cpu_lock));
10132 ASSERT(MUTEX_HELD(&dtrace_lock));
10133 ASSERT(ecb->dte_next == NULL);
10134
10135 if (probe == NULL) {
10136 /*
10137 * This is the NULL probe -- there's nothing to do.
10138 */
10139 return (0);
10140 }
10141
10142 if (probe->dtpr_ecb == NULL) {
10143 dtrace_provider_t *prov = probe->dtpr_provider;
10144
10145 /*
10146 * We're the first ECB on this probe.
10147 */
10148 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
10149
10150 if (ecb->dte_predicate != NULL)
10151 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
10152
10153 return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
10154 probe->dtpr_id, probe->dtpr_arg));
10155 } else {
10156 /*
10157 * This probe is already active. Swing the last pointer to
10158 * point to the new ECB, and issue a dtrace_sync() to assure
10159 * that all CPUs have seen the change.
10160 */
10161 ASSERT(probe->dtpr_ecb_last != NULL);
10162 probe->dtpr_ecb_last->dte_next = ecb;
10163 probe->dtpr_ecb_last = ecb;
10164 probe->dtpr_predcache = 0;
10165
10166 dtrace_sync();
10167 return (0);
10168 }
10169 }
10170
10171 static void
10172 dtrace_ecb_resize(dtrace_ecb_t *ecb)
10173 {
10174 dtrace_action_t *act;
10175 uint32_t curneeded = UINT32_MAX;
10176 uint32_t aggbase = UINT32_MAX;
10177
10178 /*
10179 * If we record anything, we always record the dtrace_rechdr_t. (And
10180 * we always record it first.)
10181 */
10182 ecb->dte_size = sizeof (dtrace_rechdr_t);
10183 ecb->dte_alignment = sizeof (dtrace_epid_t);
10184
10185 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10186 dtrace_recdesc_t *rec = &act->dta_rec;
10187 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
10188
10189 ecb->dte_alignment = MAX(ecb->dte_alignment,
10190 rec->dtrd_alignment);
10191
10192 if (DTRACEACT_ISAGG(act->dta_kind)) {
10193 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10194
10195 ASSERT(rec->dtrd_size != 0);
10196 ASSERT(agg->dtag_first != NULL);
10197 ASSERT(act->dta_prev->dta_intuple);
10198 ASSERT(aggbase != UINT32_MAX);
10199 ASSERT(curneeded != UINT32_MAX);
10200
10201 agg->dtag_base = aggbase;
10202
10203 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10204 rec->dtrd_offset = curneeded;
10205 curneeded += rec->dtrd_size;
10206 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
10207
10208 aggbase = UINT32_MAX;
10209 curneeded = UINT32_MAX;
10210 } else if (act->dta_intuple) {
10211 if (curneeded == UINT32_MAX) {
10212 /*
10213 * This is the first record in a tuple. Align
10214 * curneeded to be at offset 4 in an 8-byte
10215 * aligned block.
10216 */
10217 ASSERT(act->dta_prev == NULL ||
10218 !act->dta_prev->dta_intuple);
10219 ASSERT3U(aggbase, ==, UINT32_MAX);
10220 curneeded = P2PHASEUP(ecb->dte_size,
10221 sizeof (uint64_t), sizeof (dtrace_aggid_t));
10222
10223 aggbase = curneeded - sizeof (dtrace_aggid_t);
10224 ASSERT(IS_P2ALIGNED(aggbase,
10225 sizeof (uint64_t)));
10226 }
10227 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10228 rec->dtrd_offset = curneeded;
10229 curneeded += rec->dtrd_size;
10230 } else {
10231 /* tuples must be followed by an aggregation */
10232 ASSERT(act->dta_prev == NULL ||
10233 !act->dta_prev->dta_intuple);
10234
10235 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10236 rec->dtrd_alignment);
10237 rec->dtrd_offset = ecb->dte_size;
10238 ecb->dte_size += rec->dtrd_size;
10239 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10240 }
10241 }
10242
10243 if ((act = ecb->dte_action) != NULL &&
10244 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10245 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10246 /*
10247 * If the size is still sizeof (dtrace_rechdr_t), then all
10248 * actions store no data; set the size to 0.
10249 */
10250 ecb->dte_size = 0;
10251 }
10252
10253 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10254 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10255 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10256 ecb->dte_needed);
10257 }
10258
10259 static dtrace_action_t *
10260 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10261 {
10262 dtrace_aggregation_t *agg;
10263 size_t size = sizeof (uint64_t);
10264 int ntuple = desc->dtad_ntuple;
10265 dtrace_action_t *act;
10266 dtrace_recdesc_t *frec;
10267 dtrace_aggid_t aggid;
10268 dtrace_state_t *state = ecb->dte_state;
10269
10270 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10271 agg->dtag_ecb = ecb;
10272
10273 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10274
10275 switch (desc->dtad_kind) {
10276 case DTRACEAGG_MIN:
10277 agg->dtag_initial = INT64_MAX;
10278 agg->dtag_aggregate = dtrace_aggregate_min;
10279 break;
10280
10281 case DTRACEAGG_MAX:
10282 agg->dtag_initial = INT64_MIN;
10283 agg->dtag_aggregate = dtrace_aggregate_max;
10284 break;
10285
10286 case DTRACEAGG_COUNT:
10287 agg->dtag_aggregate = dtrace_aggregate_count;
10288 break;
10289
10290 case DTRACEAGG_QUANTIZE:
10291 agg->dtag_aggregate = dtrace_aggregate_quantize;
10292 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10293 sizeof (uint64_t);
10294 break;
10295
10296 case DTRACEAGG_LQUANTIZE: {
10297 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10298 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10299
10300 agg->dtag_initial = desc->dtad_arg;
10301 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10302
10303 if (step == 0 || levels == 0)
10304 goto err;
10305
10306 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10307 break;
10308 }
10309
10310 case DTRACEAGG_LLQUANTIZE: {
10311 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10312 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10313 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10314 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10315 int64_t v;
10316
10317 agg->dtag_initial = desc->dtad_arg;
10318 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10319
10320 if (factor < 2 || low >= high || nsteps < factor)
10321 goto err;
10322
10323 /*
10324 * Now check that the number of steps evenly divides a power
10325 * of the factor. (This assures both integer bucket size and
10326 * linearity within each magnitude.)
10327 */
10328 for (v = factor; v < nsteps; v *= factor)
10329 continue;
10330
10331 if ((v % nsteps) || (nsteps % factor))
10332 goto err;
10333
10334 size = (dtrace_aggregate_llquantize_bucket(factor,
10335 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10336 break;
10337 }
10338
10339 case DTRACEAGG_AVG:
10340 agg->dtag_aggregate = dtrace_aggregate_avg;
10341 size = sizeof (uint64_t) * 2;
10342 break;
10343
10344 case DTRACEAGG_STDDEV:
10345 agg->dtag_aggregate = dtrace_aggregate_stddev;
10346 size = sizeof (uint64_t) * 4;
10347 break;
10348
10349 case DTRACEAGG_SUM:
10350 agg->dtag_aggregate = dtrace_aggregate_sum;
10351 break;
10352
10353 default:
10354 goto err;
10355 }
10356
10357 agg->dtag_action.dta_rec.dtrd_size = size;
10358
10359 if (ntuple == 0)
10360 goto err;
10361
10362 /*
10363 * We must make sure that we have enough actions for the n-tuple.
10364 */
10365 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10366 if (DTRACEACT_ISAGG(act->dta_kind))
10367 break;
10368
10369 if (--ntuple == 0) {
10370 /*
10371 * This is the action with which our n-tuple begins.
10372 */
10373 agg->dtag_first = act;
10374 goto success;
10375 }
10376 }
10377
10378 /*
10379 * This n-tuple is short by ntuple elements. Return failure.
10380 */
10381 ASSERT(ntuple != 0);
10382 err:
10383 kmem_free(agg, sizeof (dtrace_aggregation_t));
10384 return (NULL);
10385
10386 success:
10387 /*
10388 * If the last action in the tuple has a size of zero, it's actually
10389 * an expression argument for the aggregating action.
10390 */
10391 ASSERT(ecb->dte_action_last != NULL);
10392 act = ecb->dte_action_last;
10393
10394 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10395 ASSERT(act->dta_difo != NULL);
10396
10397 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10398 agg->dtag_hasarg = 1;
10399 }
10400
10401 /*
10402 * We need to allocate an id for this aggregation.
10403 */
10404 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10405 VM_BESTFIT | VM_SLEEP);
10406
10407 if (aggid - 1 >= state->dts_naggregations) {
10408 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10409 dtrace_aggregation_t **aggs;
10410 int naggs = state->dts_naggregations << 1;
10411 int onaggs = state->dts_naggregations;
10412
10413 ASSERT(aggid == state->dts_naggregations + 1);
10414
10415 if (naggs == 0) {
10416 ASSERT(oaggs == NULL);
10417 naggs = 1;
10418 }
10419
10420 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10421
10422 if (oaggs != NULL) {
10423 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10424 kmem_free(oaggs, onaggs * sizeof (*aggs));
10425 }
10426
10427 state->dts_aggregations = aggs;
10428 state->dts_naggregations = naggs;
10429 }
10430
10431 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10432 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10433
10434 frec = &agg->dtag_first->dta_rec;
10435 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10436 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10437
10438 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10439 ASSERT(!act->dta_intuple);
10440 act->dta_intuple = 1;
10441 }
10442
10443 return (&agg->dtag_action);
10444 }
10445
10446 static void
10447 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10448 {
10449 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10450 dtrace_state_t *state = ecb->dte_state;
10451 dtrace_aggid_t aggid = agg->dtag_id;
10452
10453 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10454 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10455
10456 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10457 state->dts_aggregations[aggid - 1] = NULL;
10458
10459 kmem_free(agg, sizeof (dtrace_aggregation_t));
10460 }
10461
10462 static int
10463 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10464 {
10465 dtrace_action_t *action, *last;
10466 dtrace_difo_t *dp = desc->dtad_difo;
10467 uint32_t size = 0, align = sizeof (uint8_t), mask;
10468 uint16_t format = 0;
10469 dtrace_recdesc_t *rec;
10470 dtrace_state_t *state = ecb->dte_state;
10471 dtrace_optval_t *opt = state->dts_options, nframes, strsize;
10472 uint64_t arg = desc->dtad_arg;
10473
10474 ASSERT(MUTEX_HELD(&dtrace_lock));
10475 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10476
10477 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10478 /*
10479 * If this is an aggregating action, there must be neither
10480 * a speculate nor a commit on the action chain.
10481 */
10482 dtrace_action_t *act;
10483
10484 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10485 if (act->dta_kind == DTRACEACT_COMMIT)
10486 return (EINVAL);
10487
10488 if (act->dta_kind == DTRACEACT_SPECULATE)
10489 return (EINVAL);
10490 }
10491
10492 action = dtrace_ecb_aggregation_create(ecb, desc);
10493
10494 if (action == NULL)
10495 return (EINVAL);
10496 } else {
10497 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10498 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10499 dp != NULL && dp->dtdo_destructive)) {
10500 state->dts_destructive = 1;
10501 }
10502
10503 switch (desc->dtad_kind) {
10504 case DTRACEACT_PRINTF:
10505 case DTRACEACT_PRINTA:
10506 case DTRACEACT_SYSTEM:
10507 case DTRACEACT_FREOPEN:
10508 case DTRACEACT_DIFEXPR:
10509 /*
10510 * We know that our arg is a string -- turn it into a
10511 * format.
10512 */
10513 if (arg == NULL) {
10514 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10515 desc->dtad_kind == DTRACEACT_DIFEXPR);
10516 format = 0;
10517 } else {
10518 ASSERT(arg != NULL);
10519 ASSERT(arg > KERNELBASE);
10520 format = dtrace_format_add(state,
10521 (char *)(uintptr_t)arg);
10522 }
10523
10524 /*FALLTHROUGH*/
10525 case DTRACEACT_LIBACT:
10526 case DTRACEACT_TRACEMEM:
10527 case DTRACEACT_TRACEMEM_DYNSIZE:
10528 if (dp == NULL)
10529 return (EINVAL);
10530
10531 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10532 break;
10533
10534 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10535 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10536 return (EINVAL);
10537
10538 size = opt[DTRACEOPT_STRSIZE];
10539 }
10540
10541 break;
10542
10543 case DTRACEACT_STACK:
10544 if ((nframes = arg) == 0) {
10545 nframes = opt[DTRACEOPT_STACKFRAMES];
10546 ASSERT(nframes > 0);
10547 arg = nframes;
10548 }
10549
10550 size = nframes * sizeof (pc_t);
10551 break;
10552
10553 case DTRACEACT_JSTACK:
10554 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10555 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10556
10557 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10558 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10559
10560 arg = DTRACE_USTACK_ARG(nframes, strsize);
10561
10562 /*FALLTHROUGH*/
10563 case DTRACEACT_USTACK:
10564 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10565 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10566 strsize = DTRACE_USTACK_STRSIZE(arg);
10567 nframes = opt[DTRACEOPT_USTACKFRAMES];
10568 ASSERT(nframes > 0);
10569 arg = DTRACE_USTACK_ARG(nframes, strsize);
10570 }
10571
10572 /*
10573 * Save a slot for the pid.
10574 */
10575 size = (nframes + 1) * sizeof (uint64_t);
10576 size += DTRACE_USTACK_STRSIZE(arg);
10577 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10578
10579 break;
10580
10581 case DTRACEACT_SYM:
10582 case DTRACEACT_MOD:
10583 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10584 sizeof (uint64_t)) ||
10585 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10586 return (EINVAL);
10587 break;
10588
10589 case DTRACEACT_USYM:
10590 case DTRACEACT_UMOD:
10591 case DTRACEACT_UADDR:
10592 if (dp == NULL ||
10593 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10594 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10595 return (EINVAL);
10596
10597 /*
10598 * We have a slot for the pid, plus a slot for the
10599 * argument. To keep things simple (aligned with
10600 * bitness-neutral sizing), we store each as a 64-bit
10601 * quantity.
10602 */
10603 size = 2 * sizeof (uint64_t);
10604 break;
10605
10606 case DTRACEACT_STOP:
10607 case DTRACEACT_BREAKPOINT:
10608 case DTRACEACT_PANIC:
10609 break;
10610
10611 case DTRACEACT_CHILL:
10612 case DTRACEACT_DISCARD:
10613 case DTRACEACT_RAISE:
10614 if (dp == NULL)
10615 return (EINVAL);
10616 break;
10617
10618 case DTRACEACT_EXIT:
10619 if (dp == NULL ||
10620 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10621 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10622 return (EINVAL);
10623 break;
10624
10625 case DTRACEACT_SPECULATE:
10626 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10627 return (EINVAL);
10628
10629 if (dp == NULL)
10630 return (EINVAL);
10631
10632 state->dts_speculates = 1;
10633 break;
10634
10635 case DTRACEACT_COMMIT: {
10636 dtrace_action_t *act = ecb->dte_action;
10637
10638 for (; act != NULL; act = act->dta_next) {
10639 if (act->dta_kind == DTRACEACT_COMMIT)
10640 return (EINVAL);
10641 }
10642
10643 if (dp == NULL)
10644 return (EINVAL);
10645 break;
10646 }
10647
10648 default:
10649 return (EINVAL);
10650 }
10651
10652 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10653 /*
10654 * If this is a data-storing action or a speculate,
10655 * we must be sure that there isn't a commit on the
10656 * action chain.
10657 */
10658 dtrace_action_t *act = ecb->dte_action;
10659
10660 for (; act != NULL; act = act->dta_next) {
10661 if (act->dta_kind == DTRACEACT_COMMIT)
10662 return (EINVAL);
10663 }
10664 }
10665
10666 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10667 action->dta_rec.dtrd_size = size;
10668 }
10669
10670 action->dta_refcnt = 1;
10671 rec = &action->dta_rec;
10672 size = rec->dtrd_size;
10673
10674 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10675 if (!(size & mask)) {
10676 align = mask + 1;
10677 break;
10678 }
10679 }
10680
10681 action->dta_kind = desc->dtad_kind;
10682
10683 if ((action->dta_difo = dp) != NULL)
10684 dtrace_difo_hold(dp);
10685
10686 rec->dtrd_action = action->dta_kind;
10687 rec->dtrd_arg = arg;
10688 rec->dtrd_uarg = desc->dtad_uarg;
10689 rec->dtrd_alignment = (uint16_t)align;
10690 rec->dtrd_format = format;
10691
10692 if ((last = ecb->dte_action_last) != NULL) {
10693 ASSERT(ecb->dte_action != NULL);
10694 action->dta_prev = last;
10695 last->dta_next = action;
10696 } else {
10697 ASSERT(ecb->dte_action == NULL);
10698 ecb->dte_action = action;
10699 }
10700
10701 ecb->dte_action_last = action;
10702
10703 return (0);
10704 }
10705
10706 static void
10707 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10708 {
10709 dtrace_action_t *act = ecb->dte_action, *next;
10710 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10711 dtrace_difo_t *dp;
10712 uint16_t format;
10713
10714 if (act != NULL && act->dta_refcnt > 1) {
10715 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10716 act->dta_refcnt--;
10717 } else {
10718 for (; act != NULL; act = next) {
10719 next = act->dta_next;
10720 ASSERT(next != NULL || act == ecb->dte_action_last);
10721 ASSERT(act->dta_refcnt == 1);
10722
10723 if ((format = act->dta_rec.dtrd_format) != 0)
10724 dtrace_format_remove(ecb->dte_state, format);
10725
10726 if ((dp = act->dta_difo) != NULL)
10727 dtrace_difo_release(dp, vstate);
10728
10729 if (DTRACEACT_ISAGG(act->dta_kind)) {
10730 dtrace_ecb_aggregation_destroy(ecb, act);
10731 } else {
10732 kmem_free(act, sizeof (dtrace_action_t));
10733 }
10734 }
10735 }
10736
10737 ecb->dte_action = NULL;
10738 ecb->dte_action_last = NULL;
10739 ecb->dte_size = 0;
10740 }
10741
10742 static void
10743 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10744 {
10745 /*
10746 * We disable the ECB by removing it from its probe.
10747 */
10748 dtrace_ecb_t *pecb, *prev = NULL;
10749 dtrace_probe_t *probe = ecb->dte_probe;
10750
10751 ASSERT(MUTEX_HELD(&dtrace_lock));
10752
10753 if (probe == NULL) {
10754 /*
10755 * This is the NULL probe; there is nothing to disable.
10756 */
10757 return;
10758 }
10759
10760 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10761 if (pecb == ecb)
10762 break;
10763 prev = pecb;
10764 }
10765
10766 ASSERT(pecb != NULL);
10767
10768 if (prev == NULL) {
10769 probe->dtpr_ecb = ecb->dte_next;
10770 } else {
10771 prev->dte_next = ecb->dte_next;
10772 }
10773
10774 if (ecb == probe->dtpr_ecb_last) {
10775 ASSERT(ecb->dte_next == NULL);
10776 probe->dtpr_ecb_last = prev;
10777 }
10778
10779 /*
10780 * The ECB has been disconnected from the probe; now sync to assure
10781 * that all CPUs have seen the change before returning.
10782 */
10783 dtrace_sync();
10784
10785 if (probe->dtpr_ecb == NULL) {
10786 /*
10787 * That was the last ECB on the probe; clear the predicate
10788 * cache ID for the probe, disable it and sync one more time
10789 * to assure that we'll never hit it again.
10790 */
10791 dtrace_provider_t *prov = probe->dtpr_provider;
10792
10793 ASSERT(ecb->dte_next == NULL);
10794 ASSERT(probe->dtpr_ecb_last == NULL);
10795 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10796 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10797 probe->dtpr_id, probe->dtpr_arg);
10798 dtrace_sync();
10799 } else {
10800 /*
10801 * There is at least one ECB remaining on the probe. If there
10802 * is _exactly_ one, set the probe's predicate cache ID to be
10803 * the predicate cache ID of the remaining ECB.
10804 */
10805 ASSERT(probe->dtpr_ecb_last != NULL);
10806 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10807
10808 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10809 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10810
10811 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10812
10813 if (p != NULL)
10814 probe->dtpr_predcache = p->dtp_cacheid;
10815 }
10816
10817 ecb->dte_next = NULL;
10818 }
10819 }
10820
10821 static void
10822 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10823 {
10824 dtrace_state_t *state = ecb->dte_state;
10825 dtrace_vstate_t *vstate = &state->dts_vstate;
10826 dtrace_predicate_t *pred;
10827 dtrace_epid_t epid = ecb->dte_epid;
10828
10829 ASSERT(MUTEX_HELD(&dtrace_lock));
10830 ASSERT(ecb->dte_next == NULL);
10831 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10832
10833 if ((pred = ecb->dte_predicate) != NULL)
10834 dtrace_predicate_release(pred, vstate);
10835
10836 dtrace_ecb_action_remove(ecb);
10837
10838 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10839 state->dts_ecbs[epid - 1] = NULL;
10840
10841 kmem_free(ecb, sizeof (dtrace_ecb_t));
10842 }
10843
10844 static dtrace_ecb_t *
10845 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10846 dtrace_enabling_t *enab)
10847 {
10848 dtrace_ecb_t *ecb;
10849 dtrace_predicate_t *pred;
10850 dtrace_actdesc_t *act;
10851 dtrace_provider_t *prov;
10852 dtrace_ecbdesc_t *desc = enab->dten_current;
10853
10854 ASSERT(MUTEX_HELD(&dtrace_lock));
10855 ASSERT(state != NULL);
10856
10857 ecb = dtrace_ecb_add(state, probe);
10858 ecb->dte_uarg = desc->dted_uarg;
10859
10860 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10861 dtrace_predicate_hold(pred);
10862 ecb->dte_predicate = pred;
10863 }
10864
10865 if (probe != NULL) {
10866 /*
10867 * If the provider shows more leg than the consumer is old
10868 * enough to see, we need to enable the appropriate implicit
10869 * predicate bits to prevent the ecb from activating at
10870 * revealing times.
10871 *
10872 * Providers specifying DTRACE_PRIV_USER at register time
10873 * are stating that they need the /proc-style privilege
10874 * model to be enforced, and this is what DTRACE_COND_OWNER
10875 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10876 */
10877 prov = probe->dtpr_provider;
10878 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10879 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10880 ecb->dte_cond |= DTRACE_COND_OWNER;
10881
10882 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10883 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10884 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10885
10886 /*
10887 * If the provider shows us kernel innards and the user
10888 * is lacking sufficient privilege, enable the
10889 * DTRACE_COND_USERMODE implicit predicate.
10890 */
10891 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10892 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10893 ecb->dte_cond |= DTRACE_COND_USERMODE;
10894 }
10895
10896 if (dtrace_ecb_create_cache != NULL) {
10897 /*
10898 * If we have a cached ecb, we'll use its action list instead
10899 * of creating our own (saving both time and space).
10900 */
10901 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10902 dtrace_action_t *act = cached->dte_action;
10903
10904 if (act != NULL) {
10905 ASSERT(act->dta_refcnt > 0);
10906 act->dta_refcnt++;
10907 ecb->dte_action = act;
10908 ecb->dte_action_last = cached->dte_action_last;
10909 ecb->dte_needed = cached->dte_needed;
10910 ecb->dte_size = cached->dte_size;
10911 ecb->dte_alignment = cached->dte_alignment;
10912 }
10913
10914 return (ecb);
10915 }
10916
10917 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10918 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10919 dtrace_ecb_destroy(ecb);
10920 return (NULL);
10921 }
10922 }
10923
10924 dtrace_ecb_resize(ecb);
10925
10926 return (dtrace_ecb_create_cache = ecb);
10927 }
10928
10929 static int
10930 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10931 {
10932 dtrace_ecb_t *ecb;
10933 dtrace_enabling_t *enab = arg;
10934 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10935
10936 ASSERT(state != NULL);
10937
10938 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10939 /*
10940 * This probe was created in a generation for which this
10941 * enabling has previously created ECBs; we don't want to
10942 * enable it again, so just kick out.
10943 */
10944 return (DTRACE_MATCH_NEXT);
10945 }
10946
10947 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10948 return (DTRACE_MATCH_DONE);
10949
10950 if (dtrace_ecb_enable(ecb) < 0)
10951 return (DTRACE_MATCH_FAIL);
10952
10953 return (DTRACE_MATCH_NEXT);
10954 }
10955
10956 static dtrace_ecb_t *
10957 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10958 {
10959 dtrace_ecb_t *ecb;
10960
10961 ASSERT(MUTEX_HELD(&dtrace_lock));
10962
10963 if (id == 0 || id > state->dts_necbs)
10964 return (NULL);
10965
10966 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10967 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10968
10969 return (state->dts_ecbs[id - 1]);
10970 }
10971
10972 static dtrace_aggregation_t *
10973 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10974 {
10975 dtrace_aggregation_t *agg;
10976
10977 ASSERT(MUTEX_HELD(&dtrace_lock));
10978
10979 if (id == 0 || id > state->dts_naggregations)
10980 return (NULL);
10981
10982 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10983 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10984 agg->dtag_id == id);
10985
10986 return (state->dts_aggregations[id - 1]);
10987 }
10988
10989 /*
10990 * DTrace Buffer Functions
10991 *
10992 * The following functions manipulate DTrace buffers. Most of these functions
10993 * are called in the context of establishing or processing consumer state;
10994 * exceptions are explicitly noted.
10995 */
10996
10997 /*
10998 * Note: called from cross call context. This function switches the two
10999 * buffers on a given CPU. The atomicity of this operation is assured by
11000 * disabling interrupts while the actual switch takes place; the disabling of
11001 * interrupts serializes the execution with any execution of dtrace_probe() on
11002 * the same CPU.
11003 */
11004 static void
11005 dtrace_buffer_switch(dtrace_buffer_t *buf)
11006 {
11007 caddr_t tomax = buf->dtb_tomax;
11008 caddr_t xamot = buf->dtb_xamot;
11009 dtrace_icookie_t cookie;
11010 hrtime_t now;
11011
11012 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11013 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
11014
11015 cookie = dtrace_interrupt_disable();
11016 now = dtrace_gethrtime();
11017 buf->dtb_tomax = xamot;
11018 buf->dtb_xamot = tomax;
11019 buf->dtb_xamot_drops = buf->dtb_drops;
11020 buf->dtb_xamot_offset = buf->dtb_offset;
11021 buf->dtb_xamot_errors = buf->dtb_errors;
11022 buf->dtb_xamot_flags = buf->dtb_flags;
11023 buf->dtb_offset = 0;
11024 buf->dtb_drops = 0;
11025 buf->dtb_errors = 0;
11026 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
11027 buf->dtb_interval = now - buf->dtb_switched;
11028 buf->dtb_switched = now;
11029 dtrace_interrupt_enable(cookie);
11030 }
11031
11032 /*
11033 * Note: called from cross call context. This function activates a buffer
11034 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
11035 * is guaranteed by the disabling of interrupts.
11036 */
11037 static void
11038 dtrace_buffer_activate(dtrace_state_t *state)
11039 {
11040 dtrace_buffer_t *buf;
11041 dtrace_icookie_t cookie = dtrace_interrupt_disable();
11042
11043 buf = &state->dts_buffer[CPU->cpu_id];
11044
11045 if (buf->dtb_tomax != NULL) {
11046 /*
11047 * We might like to assert that the buffer is marked inactive,
11048 * but this isn't necessarily true: the buffer for the CPU
11049 * that processes the BEGIN probe has its buffer activated
11050 * manually. In this case, we take the (harmless) action
11051 * re-clearing the bit INACTIVE bit.
11052 */
11053 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
11054 }
11055
11056 dtrace_interrupt_enable(cookie);
11057 }
11058
11059 static int
11060 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
11061 processorid_t cpu, int *factor)
11062 {
11063 cpu_t *cp;
11064 dtrace_buffer_t *buf;
11065 int allocated = 0, desired = 0;
11066
11067 ASSERT(MUTEX_HELD(&cpu_lock));
11068 ASSERT(MUTEX_HELD(&dtrace_lock));
11069
11070 *factor = 1;
11071
11072 if (size > dtrace_nonroot_maxsize &&
11073 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
11074 return (EFBIG);
11075
11076 cp = cpu_list;
11077
11078 do {
11079 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11080 continue;
11081
11082 buf = &bufs[cp->cpu_id];
11083
11084 /*
11085 * If there is already a buffer allocated for this CPU, it
11086 * is only possible that this is a DR event. In this case,
11087 * the buffer size must match our specified size.
11088 */
11089 if (buf->dtb_tomax != NULL) {
11090 ASSERT(buf->dtb_size == size);
11091 continue;
11092 }
11093
11094 ASSERT(buf->dtb_xamot == NULL);
11095
11096 if ((buf->dtb_tomax = kmem_zalloc(size,
11097 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11098 goto err;
11099
11100 buf->dtb_size = size;
11101 buf->dtb_flags = flags;
11102 buf->dtb_offset = 0;
11103 buf->dtb_drops = 0;
11104
11105 if (flags & DTRACEBUF_NOSWITCH)
11106 continue;
11107
11108 if ((buf->dtb_xamot = kmem_zalloc(size,
11109 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11110 goto err;
11111 } while ((cp = cp->cpu_next) != cpu_list);
11112
11113 return (0);
11114
11115 err:
11116 cp = cpu_list;
11117
11118 do {
11119 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11120 continue;
11121
11122 buf = &bufs[cp->cpu_id];
11123 desired += 2;
11124
11125 if (buf->dtb_xamot != NULL) {
11126 ASSERT(buf->dtb_tomax != NULL);
11127 ASSERT(buf->dtb_size == size);
11128 kmem_free(buf->dtb_xamot, size);
11129 allocated++;
11130 }
11131
11132 if (buf->dtb_tomax != NULL) {
11133 ASSERT(buf->dtb_size == size);
11134 kmem_free(buf->dtb_tomax, size);
11135 allocated++;
11136 }
11137
11138 buf->dtb_tomax = NULL;
11139 buf->dtb_xamot = NULL;
11140 buf->dtb_size = 0;
11141 } while ((cp = cp->cpu_next) != cpu_list);
11142
11143 *factor = desired / (allocated > 0 ? allocated : 1);
11144
11145 return (ENOMEM);
11146 }
11147
11148 /*
11149 * Note: called from probe context. This function just increments the drop
11150 * count on a buffer. It has been made a function to allow for the
11151 * possibility of understanding the source of mysterious drop counts. (A
11152 * problem for which one may be particularly disappointed that DTrace cannot
11153 * be used to understand DTrace.)
11154 */
11155 static void
11156 dtrace_buffer_drop(dtrace_buffer_t *buf)
11157 {
11158 buf->dtb_drops++;
11159 }
11160
11161 /*
11162 * Note: called from probe context. This function is called to reserve space
11163 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
11164 * mstate. Returns the new offset in the buffer, or a negative value if an
11165 * error has occurred.
11166 */
11167 static intptr_t
11168 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11169 dtrace_state_t *state, dtrace_mstate_t *mstate)
11170 {
11171 intptr_t offs = buf->dtb_offset, soffs;
11172 intptr_t woffs;
11173 caddr_t tomax;
11174 size_t total;
11175
11176 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11177 return (-1);
11178
11179 if ((tomax = buf->dtb_tomax) == NULL) {
11180 dtrace_buffer_drop(buf);
11181 return (-1);
11182 }
11183
11184 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11185 while (offs & (align - 1)) {
11186 /*
11187 * Assert that our alignment is off by a number which
11188 * is itself sizeof (uint32_t) aligned.
11189 */
11190 ASSERT(!((align - (offs & (align - 1))) &
11191 (sizeof (uint32_t) - 1)));
11192 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11193 offs += sizeof (uint32_t);
11194 }
11195
11196 if ((soffs = offs + needed) > buf->dtb_size) {
11197 dtrace_buffer_drop(buf);
11198 return (-1);
11199 }
11200
11201 if (mstate == NULL)
11202 return (offs);
11203
11204 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11205 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11206 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11207
11208 return (offs);
11209 }
11210
11211 if (buf->dtb_flags & DTRACEBUF_FILL) {
11212 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11213 (buf->dtb_flags & DTRACEBUF_FULL))
11214 return (-1);
11215 goto out;
11216 }
11217
11218 total = needed + (offs & (align - 1));
11219
11220 /*
11221 * For a ring buffer, life is quite a bit more complicated. Before
11222 * we can store any padding, we need to adjust our wrapping offset.
11223 * (If we've never before wrapped or we're not about to, no adjustment
11224 * is required.)
11225 */
11226 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11227 offs + total > buf->dtb_size) {
11228 woffs = buf->dtb_xamot_offset;
11229
11230 if (offs + total > buf->dtb_size) {
11231 /*
11232 * We can't fit in the end of the buffer. First, a
11233 * sanity check that we can fit in the buffer at all.
11234 */
11235 if (total > buf->dtb_size) {
11236 dtrace_buffer_drop(buf);
11237 return (-1);
11238 }
11239
11240 /*
11241 * We're going to be storing at the top of the buffer,
11242 * so now we need to deal with the wrapped offset. We
11243 * only reset our wrapped offset to 0 if it is
11244 * currently greater than the current offset. If it
11245 * is less than the current offset, it is because a
11246 * previous allocation induced a wrap -- but the
11247 * allocation didn't subsequently take the space due
11248 * to an error or false predicate evaluation. In this
11249 * case, we'll just leave the wrapped offset alone: if
11250 * the wrapped offset hasn't been advanced far enough
11251 * for this allocation, it will be adjusted in the
11252 * lower loop.
11253 */
11254 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11255 if (woffs >= offs)
11256 woffs = 0;
11257 } else {
11258 woffs = 0;
11259 }
11260
11261 /*
11262 * Now we know that we're going to be storing to the
11263 * top of the buffer and that there is room for us
11264 * there. We need to clear the buffer from the current
11265 * offset to the end (there may be old gunk there).
11266 */
11267 while (offs < buf->dtb_size)
11268 tomax[offs++] = 0;
11269
11270 /*
11271 * We need to set our offset to zero. And because we
11272 * are wrapping, we need to set the bit indicating as
11273 * much. We can also adjust our needed space back
11274 * down to the space required by the ECB -- we know
11275 * that the top of the buffer is aligned.
11276 */
11277 offs = 0;
11278 total = needed;
11279 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11280 } else {
11281 /*
11282 * There is room for us in the buffer, so we simply
11283 * need to check the wrapped offset.
11284 */
11285 if (woffs < offs) {
11286 /*
11287 * The wrapped offset is less than the offset.
11288 * This can happen if we allocated buffer space
11289 * that induced a wrap, but then we didn't
11290 * subsequently take the space due to an error
11291 * or false predicate evaluation. This is
11292 * okay; we know that _this_ allocation isn't
11293 * going to induce a wrap. We still can't
11294 * reset the wrapped offset to be zero,
11295 * however: the space may have been trashed in
11296 * the previous failed probe attempt. But at
11297 * least the wrapped offset doesn't need to
11298 * be adjusted at all...
11299 */
11300 goto out;
11301 }
11302 }
11303
11304 while (offs + total > woffs) {
11305 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11306 size_t size;
11307
11308 if (epid == DTRACE_EPIDNONE) {
11309 size = sizeof (uint32_t);
11310 } else {
11311 ASSERT3U(epid, <=, state->dts_necbs);
11312 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11313
11314 size = state->dts_ecbs[epid - 1]->dte_size;
11315 }
11316
11317 ASSERT(woffs + size <= buf->dtb_size);
11318 ASSERT(size != 0);
11319
11320 if (woffs + size == buf->dtb_size) {
11321 /*
11322 * We've reached the end of the buffer; we want
11323 * to set the wrapped offset to 0 and break
11324 * out. However, if the offs is 0, then we're
11325 * in a strange edge-condition: the amount of
11326 * space that we want to reserve plus the size
11327 * of the record that we're overwriting is
11328 * greater than the size of the buffer. This
11329 * is problematic because if we reserve the
11330 * space but subsequently don't consume it (due
11331 * to a failed predicate or error) the wrapped
11332 * offset will be 0 -- yet the EPID at offset 0
11333 * will not be committed. This situation is
11334 * relatively easy to deal with: if we're in
11335 * this case, the buffer is indistinguishable
11336 * from one that hasn't wrapped; we need only
11337 * finish the job by clearing the wrapped bit,
11338 * explicitly setting the offset to be 0, and
11339 * zero'ing out the old data in the buffer.
11340 */
11341 if (offs == 0) {
11342 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11343 buf->dtb_offset = 0;
11344 woffs = total;
11345
11346 while (woffs < buf->dtb_size)
11347 tomax[woffs++] = 0;
11348 }
11349
11350 woffs = 0;
11351 break;
11352 }
11353
11354 woffs += size;
11355 }
11356
11357 /*
11358 * We have a wrapped offset. It may be that the wrapped offset
11359 * has become zero -- that's okay.
11360 */
11361 buf->dtb_xamot_offset = woffs;
11362 }
11363
11364 out:
11365 /*
11366 * Now we can plow the buffer with any necessary padding.
11367 */
11368 while (offs & (align - 1)) {
11369 /*
11370 * Assert that our alignment is off by a number which
11371 * is itself sizeof (uint32_t) aligned.
11372 */
11373 ASSERT(!((align - (offs & (align - 1))) &
11374 (sizeof (uint32_t) - 1)));
11375 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11376 offs += sizeof (uint32_t);
11377 }
11378
11379 if (buf->dtb_flags & DTRACEBUF_FILL) {
11380 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11381 buf->dtb_flags |= DTRACEBUF_FULL;
11382 return (-1);
11383 }
11384 }
11385
11386 if (mstate == NULL)
11387 return (offs);
11388
11389 /*
11390 * For ring buffers and fill buffers, the scratch space is always
11391 * the inactive buffer.
11392 */
11393 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11394 mstate->dtms_scratch_size = buf->dtb_size;
11395 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11396
11397 return (offs);
11398 }
11399
11400 static void
11401 dtrace_buffer_polish(dtrace_buffer_t *buf)
11402 {
11403 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11404 ASSERT(MUTEX_HELD(&dtrace_lock));
11405
11406 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11407 return;
11408
11409 /*
11410 * We need to polish the ring buffer. There are three cases:
11411 *
11412 * - The first (and presumably most common) is that there is no gap
11413 * between the buffer offset and the wrapped offset. In this case,
11414 * there is nothing in the buffer that isn't valid data; we can
11415 * mark the buffer as polished and return.
11416 *
11417 * - The second (less common than the first but still more common
11418 * than the third) is that there is a gap between the buffer offset
11419 * and the wrapped offset, and the wrapped offset is larger than the
11420 * buffer offset. This can happen because of an alignment issue, or
11421 * can happen because of a call to dtrace_buffer_reserve() that
11422 * didn't subsequently consume the buffer space. In this case,
11423 * we need to zero the data from the buffer offset to the wrapped
11424 * offset.
11425 *
11426 * - The third (and least common) is that there is a gap between the
11427 * buffer offset and the wrapped offset, but the wrapped offset is
11428 * _less_ than the buffer offset. This can only happen because a
11429 * call to dtrace_buffer_reserve() induced a wrap, but the space
11430 * was not subsequently consumed. In this case, we need to zero the
11431 * space from the offset to the end of the buffer _and_ from the
11432 * top of the buffer to the wrapped offset.
11433 */
11434 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11435 bzero(buf->dtb_tomax + buf->dtb_offset,
11436 buf->dtb_xamot_offset - buf->dtb_offset);
11437 }
11438
11439 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11440 bzero(buf->dtb_tomax + buf->dtb_offset,
11441 buf->dtb_size - buf->dtb_offset);
11442 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11443 }
11444 }
11445
11446 /*
11447 * This routine determines if data generated at the specified time has likely
11448 * been entirely consumed at user-level. This routine is called to determine
11449 * if an ECB on a defunct probe (but for an active enabling) can be safely
11450 * disabled and destroyed.
11451 */
11452 static int
11453 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11454 {
11455 int i;
11456
11457 for (i = 0; i < NCPU; i++) {
11458 dtrace_buffer_t *buf = &bufs[i];
11459
11460 if (buf->dtb_size == 0)
11461 continue;
11462
11463 if (buf->dtb_flags & DTRACEBUF_RING)
11464 return (0);
11465
11466 if (!buf->dtb_switched && buf->dtb_offset != 0)
11467 return (0);
11468
11469 if (buf->dtb_switched - buf->dtb_interval < when)
11470 return (0);
11471 }
11472
11473 return (1);
11474 }
11475
11476 static void
11477 dtrace_buffer_free(dtrace_buffer_t *bufs)
11478 {
11479 int i;
11480
11481 for (i = 0; i < NCPU; i++) {
11482 dtrace_buffer_t *buf = &bufs[i];
11483
11484 if (buf->dtb_tomax == NULL) {
11485 ASSERT(buf->dtb_xamot == NULL);
11486 ASSERT(buf->dtb_size == 0);
11487 continue;
11488 }
11489
11490 if (buf->dtb_xamot != NULL) {
11491 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11492 kmem_free(buf->dtb_xamot, buf->dtb_size);
11493 }
11494
11495 kmem_free(buf->dtb_tomax, buf->dtb_size);
11496 buf->dtb_size = 0;
11497 buf->dtb_tomax = NULL;
11498 buf->dtb_xamot = NULL;
11499 }
11500 }
11501
11502 /*
11503 * DTrace Enabling Functions
11504 */
11505 static dtrace_enabling_t *
11506 dtrace_enabling_create(dtrace_vstate_t *vstate)
11507 {
11508 dtrace_enabling_t *enab;
11509
11510 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11511 enab->dten_vstate = vstate;
11512
11513 return (enab);
11514 }
11515
11516 static void
11517 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11518 {
11519 dtrace_ecbdesc_t **ndesc;
11520 size_t osize, nsize;
11521
11522 /*
11523 * We can't add to enablings after we've enabled them, or after we've
11524 * retained them.
11525 */
11526 ASSERT(enab->dten_probegen == 0);
11527 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11528
11529 if (enab->dten_ndesc < enab->dten_maxdesc) {
11530 enab->dten_desc[enab->dten_ndesc++] = ecb;
11531 return;
11532 }
11533
11534 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11535
11536 if (enab->dten_maxdesc == 0) {
11537 enab->dten_maxdesc = 1;
11538 } else {
11539 enab->dten_maxdesc <<= 1;
11540 }
11541
11542 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11543
11544 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11545 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11546 bcopy(enab->dten_desc, ndesc, osize);
11547 kmem_free(enab->dten_desc, osize);
11548
11549 enab->dten_desc = ndesc;
11550 enab->dten_desc[enab->dten_ndesc++] = ecb;
11551 }
11552
11553 static void
11554 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11555 dtrace_probedesc_t *pd)
11556 {
11557 dtrace_ecbdesc_t *new;
11558 dtrace_predicate_t *pred;
11559 dtrace_actdesc_t *act;
11560
11561 /*
11562 * We're going to create a new ECB description that matches the
11563 * specified ECB in every way, but has the specified probe description.
11564 */
11565 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11566
11567 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11568 dtrace_predicate_hold(pred);
11569
11570 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11571 dtrace_actdesc_hold(act);
11572
11573 new->dted_action = ecb->dted_action;
11574 new->dted_pred = ecb->dted_pred;
11575 new->dted_probe = *pd;
11576 new->dted_uarg = ecb->dted_uarg;
11577
11578 dtrace_enabling_add(enab, new);
11579 }
11580
11581 static void
11582 dtrace_enabling_dump(dtrace_enabling_t *enab)
11583 {
11584 int i;
11585
11586 for (i = 0; i < enab->dten_ndesc; i++) {
11587 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11588
11589 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11590 desc->dtpd_provider, desc->dtpd_mod,
11591 desc->dtpd_func, desc->dtpd_name);
11592 }
11593 }
11594
11595 static void
11596 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11597 {
11598 int i;
11599 dtrace_ecbdesc_t *ep;
11600 dtrace_vstate_t *vstate = enab->dten_vstate;
11601
11602 ASSERT(MUTEX_HELD(&dtrace_lock));
11603
11604 for (i = 0; i < enab->dten_ndesc; i++) {
11605 dtrace_actdesc_t *act, *next;
11606 dtrace_predicate_t *pred;
11607
11608 ep = enab->dten_desc[i];
11609
11610 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11611 dtrace_predicate_release(pred, vstate);
11612
11613 for (act = ep->dted_action; act != NULL; act = next) {
11614 next = act->dtad_next;
11615 dtrace_actdesc_release(act, vstate);
11616 }
11617
11618 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11619 }
11620
11621 kmem_free(enab->dten_desc,
11622 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11623
11624 /*
11625 * If this was a retained enabling, decrement the dts_nretained count
11626 * and take it off of the dtrace_retained list.
11627 */
11628 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11629 dtrace_retained == enab) {
11630 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11631 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11632 enab->dten_vstate->dtvs_state->dts_nretained--;
11633 dtrace_retained_gen++;
11634 }
11635
11636 if (enab->dten_prev == NULL) {
11637 if (dtrace_retained == enab) {
11638 dtrace_retained = enab->dten_next;
11639
11640 if (dtrace_retained != NULL)
11641 dtrace_retained->dten_prev = NULL;
11642 }
11643 } else {
11644 ASSERT(enab != dtrace_retained);
11645 ASSERT(dtrace_retained != NULL);
11646 enab->dten_prev->dten_next = enab->dten_next;
11647 }
11648
11649 if (enab->dten_next != NULL) {
11650 ASSERT(dtrace_retained != NULL);
11651 enab->dten_next->dten_prev = enab->dten_prev;
11652 }
11653
11654 kmem_free(enab, sizeof (dtrace_enabling_t));
11655 }
11656
11657 static int
11658 dtrace_enabling_retain(dtrace_enabling_t *enab)
11659 {
11660 dtrace_state_t *state;
11661
11662 ASSERT(MUTEX_HELD(&dtrace_lock));
11663 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11664 ASSERT(enab->dten_vstate != NULL);
11665
11666 state = enab->dten_vstate->dtvs_state;
11667 ASSERT(state != NULL);
11668
11669 /*
11670 * We only allow each state to retain dtrace_retain_max enablings.
11671 */
11672 if (state->dts_nretained >= dtrace_retain_max)
11673 return (ENOSPC);
11674
11675 state->dts_nretained++;
11676 dtrace_retained_gen++;
11677
11678 if (dtrace_retained == NULL) {
11679 dtrace_retained = enab;
11680 return (0);
11681 }
11682
11683 enab->dten_next = dtrace_retained;
11684 dtrace_retained->dten_prev = enab;
11685 dtrace_retained = enab;
11686
11687 return (0);
11688 }
11689
11690 static int
11691 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11692 dtrace_probedesc_t *create)
11693 {
11694 dtrace_enabling_t *new, *enab;
11695 int found = 0, err = ENOENT;
11696
11697 ASSERT(MUTEX_HELD(&dtrace_lock));
11698 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11699 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11700 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11701 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11702
11703 new = dtrace_enabling_create(&state->dts_vstate);
11704
11705 /*
11706 * Iterate over all retained enablings, looking for enablings that
11707 * match the specified state.
11708 */
11709 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11710 int i;
11711
11712 /*
11713 * dtvs_state can only be NULL for helper enablings -- and
11714 * helper enablings can't be retained.
11715 */
11716 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11717
11718 if (enab->dten_vstate->dtvs_state != state)
11719 continue;
11720
11721 /*
11722 * Now iterate over each probe description; we're looking for
11723 * an exact match to the specified probe description.
11724 */
11725 for (i = 0; i < enab->dten_ndesc; i++) {
11726 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11727 dtrace_probedesc_t *pd = &ep->dted_probe;
11728
11729 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11730 continue;
11731
11732 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11733 continue;
11734
11735 if (strcmp(pd->dtpd_func, match->dtpd_func))
11736 continue;
11737
11738 if (strcmp(pd->dtpd_name, match->dtpd_name))
11739 continue;
11740
11741 /*
11742 * We have a winning probe! Add it to our growing
11743 * enabling.
11744 */
11745 found = 1;
11746 dtrace_enabling_addlike(new, ep, create);
11747 }
11748 }
11749
11750 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11751 dtrace_enabling_destroy(new);
11752 return (err);
11753 }
11754
11755 return (0);
11756 }
11757
11758 static void
11759 dtrace_enabling_retract(dtrace_state_t *state)
11760 {
11761 dtrace_enabling_t *enab, *next;
11762
11763 ASSERT(MUTEX_HELD(&dtrace_lock));
11764
11765 /*
11766 * Iterate over all retained enablings, destroy the enablings retained
11767 * for the specified state.
11768 */
11769 for (enab = dtrace_retained; enab != NULL; enab = next) {
11770 next = enab->dten_next;
11771
11772 /*
11773 * dtvs_state can only be NULL for helper enablings -- and
11774 * helper enablings can't be retained.
11775 */
11776 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11777
11778 if (enab->dten_vstate->dtvs_state == state) {
11779 ASSERT(state->dts_nretained > 0);
11780 dtrace_enabling_destroy(enab);
11781 }
11782 }
11783
11784 ASSERT(state->dts_nretained == 0);
11785 }
11786
11787 static int
11788 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11789 {
11790 int i = 0;
11791 int total_matched = 0, matched = 0;
11792
11793 ASSERT(MUTEX_HELD(&cpu_lock));
11794 ASSERT(MUTEX_HELD(&dtrace_lock));
11795
11796 for (i = 0; i < enab->dten_ndesc; i++) {
11797 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11798
11799 enab->dten_current = ep;
11800 enab->dten_error = 0;
11801
11802 /*
11803 * If a provider failed to enable a probe then get out and
11804 * let the consumer know we failed.
11805 */
11806 if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
11807 return (EBUSY);
11808
11809 total_matched += matched;
11810
11811 if (enab->dten_error != 0) {
11812 /*
11813 * If we get an error half-way through enabling the
11814 * probes, we kick out -- perhaps with some number of
11815 * them enabled. Leaving enabled probes enabled may
11816 * be slightly confusing for user-level, but we expect
11817 * that no one will attempt to actually drive on in
11818 * the face of such errors. If this is an anonymous
11819 * enabling (indicated with a NULL nmatched pointer),
11820 * we cmn_err() a message. We aren't expecting to
11821 * get such an error -- such as it can exist at all,
11822 * it would be a result of corrupted DOF in the driver
11823 * properties.
11824 */
11825 if (nmatched == NULL) {
11826 cmn_err(CE_WARN, "dtrace_enabling_match() "
11827 "error on %p: %d", (void *)ep,
11828 enab->dten_error);
11829 }
11830
11831 return (enab->dten_error);
11832 }
11833 }
11834
11835 enab->dten_probegen = dtrace_probegen;
11836 if (nmatched != NULL)
11837 *nmatched = total_matched;
11838
11839 return (0);
11840 }
11841
11842 static void
11843 dtrace_enabling_matchall(void)
11844 {
11845 dtrace_enabling_t *enab;
11846
11847 mutex_enter(&cpu_lock);
11848 mutex_enter(&dtrace_lock);
11849
11850 /*
11851 * Iterate over all retained enablings to see if any probes match
11852 * against them. We only perform this operation on enablings for which
11853 * we have sufficient permissions by virtue of being in the global zone
11854 * or in the same zone as the DTrace client. Because we can be called
11855 * after dtrace_detach() has been called, we cannot assert that there
11856 * are retained enablings. We can safely load from dtrace_retained,
11857 * however: the taskq_destroy() at the end of dtrace_detach() will
11858 * block pending our completion.
11859 */
11860 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11861 dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
11862 cred_t *cr = dcr->dcr_cred;
11863 zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
11864
11865 if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
11866 (zone == GLOBAL_ZONEID || getzoneid() == zone)))
11867 (void) dtrace_enabling_match(enab, NULL);
11868 }
11869
11870 mutex_exit(&dtrace_lock);
11871 mutex_exit(&cpu_lock);
11872 }
11873
11874 /*
11875 * If an enabling is to be enabled without having matched probes (that is, if
11876 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11877 * enabling must be _primed_ by creating an ECB for every ECB description.
11878 * This must be done to assure that we know the number of speculations, the
11879 * number of aggregations, the minimum buffer size needed, etc. before we
11880 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11881 * enabling any probes, we create ECBs for every ECB decription, but with a
11882 * NULL probe -- which is exactly what this function does.
11883 */
11884 static void
11885 dtrace_enabling_prime(dtrace_state_t *state)
11886 {
11887 dtrace_enabling_t *enab;
11888 int i;
11889
11890 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11891 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11892
11893 if (enab->dten_vstate->dtvs_state != state)
11894 continue;
11895
11896 /*
11897 * We don't want to prime an enabling more than once, lest
11898 * we allow a malicious user to induce resource exhaustion.
11899 * (The ECBs that result from priming an enabling aren't
11900 * leaked -- but they also aren't deallocated until the
11901 * consumer state is destroyed.)
11902 */
11903 if (enab->dten_primed)
11904 continue;
11905
11906 for (i = 0; i < enab->dten_ndesc; i++) {
11907 enab->dten_current = enab->dten_desc[i];
11908 (void) dtrace_probe_enable(NULL, enab);
11909 }
11910
11911 enab->dten_primed = 1;
11912 }
11913 }
11914
11915 /*
11916 * Called to indicate that probes should be provided due to retained
11917 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11918 * must take an initial lap through the enabling calling the dtps_provide()
11919 * entry point explicitly to allow for autocreated probes.
11920 */
11921 static void
11922 dtrace_enabling_provide(dtrace_provider_t *prv)
11923 {
11924 int i, all = 0;
11925 dtrace_probedesc_t desc;
11926 dtrace_genid_t gen;
11927
11928 ASSERT(MUTEX_HELD(&dtrace_lock));
11929 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11930
11931 if (prv == NULL) {
11932 all = 1;
11933 prv = dtrace_provider;
11934 }
11935
11936 do {
11937 dtrace_enabling_t *enab;
11938 void *parg = prv->dtpv_arg;
11939
11940 retry:
11941 gen = dtrace_retained_gen;
11942 for (enab = dtrace_retained; enab != NULL;
11943 enab = enab->dten_next) {
11944 for (i = 0; i < enab->dten_ndesc; i++) {
11945 desc = enab->dten_desc[i]->dted_probe;
11946 mutex_exit(&dtrace_lock);
11947 prv->dtpv_pops.dtps_provide(parg, &desc);
11948 mutex_enter(&dtrace_lock);
11949 /*
11950 * Process the retained enablings again if
11951 * they have changed while we weren't holding
11952 * dtrace_lock.
11953 */
11954 if (gen != dtrace_retained_gen)
11955 goto retry;
11956 }
11957 }
11958 } while (all && (prv = prv->dtpv_next) != NULL);
11959
11960 mutex_exit(&dtrace_lock);
11961 dtrace_probe_provide(NULL, all ? NULL : prv);
11962 mutex_enter(&dtrace_lock);
11963 }
11964
11965 /*
11966 * Called to reap ECBs that are attached to probes from defunct providers.
11967 */
11968 static void
11969 dtrace_enabling_reap(void)
11970 {
11971 dtrace_provider_t *prov;
11972 dtrace_probe_t *probe;
11973 dtrace_ecb_t *ecb;
11974 hrtime_t when;
11975 int i;
11976
11977 mutex_enter(&cpu_lock);
11978 mutex_enter(&dtrace_lock);
11979
11980 for (i = 0; i < dtrace_nprobes; i++) {
11981 if ((probe = dtrace_probes[i]) == NULL)
11982 continue;
11983
11984 if (probe->dtpr_ecb == NULL)
11985 continue;
11986
11987 prov = probe->dtpr_provider;
11988
11989 if ((when = prov->dtpv_defunct) == 0)
11990 continue;
11991
11992 /*
11993 * We have ECBs on a defunct provider: we want to reap these
11994 * ECBs to allow the provider to unregister. The destruction
11995 * of these ECBs must be done carefully: if we destroy the ECB
11996 * and the consumer later wishes to consume an EPID that
11997 * corresponds to the destroyed ECB (and if the EPID metadata
11998 * has not been previously consumed), the consumer will abort
11999 * processing on the unknown EPID. To reduce (but not, sadly,
12000 * eliminate) the possibility of this, we will only destroy an
12001 * ECB for a defunct provider if, for the state that
12002 * corresponds to the ECB:
12003 *
12004 * (a) There is no speculative tracing (which can effectively
12005 * cache an EPID for an arbitrary amount of time).
12006 *
12007 * (b) The principal buffers have been switched twice since the
12008 * provider became defunct.
12009 *
12010 * (c) The aggregation buffers are of zero size or have been
12011 * switched twice since the provider became defunct.
12012 *
12013 * We use dts_speculates to determine (a) and call a function
12014 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
12015 * that as soon as we've been unable to destroy one of the ECBs
12016 * associated with the probe, we quit trying -- reaping is only
12017 * fruitful in as much as we can destroy all ECBs associated
12018 * with the defunct provider's probes.
12019 */
12020 while ((ecb = probe->dtpr_ecb) != NULL) {
12021 dtrace_state_t *state = ecb->dte_state;
12022 dtrace_buffer_t *buf = state->dts_buffer;
12023 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
12024
12025 if (state->dts_speculates)
12026 break;
12027
12028 if (!dtrace_buffer_consumed(buf, when))
12029 break;
12030
12031 if (!dtrace_buffer_consumed(aggbuf, when))
12032 break;
12033
12034 dtrace_ecb_disable(ecb);
12035 ASSERT(probe->dtpr_ecb != ecb);
12036 dtrace_ecb_destroy(ecb);
12037 }
12038 }
12039
12040 mutex_exit(&dtrace_lock);
12041 mutex_exit(&cpu_lock);
12042 }
12043
12044 /*
12045 * DTrace DOF Functions
12046 */
12047 /*ARGSUSED*/
12048 static void
12049 dtrace_dof_error(dof_hdr_t *dof, const char *str)
12050 {
12051 if (dtrace_err_verbose)
12052 cmn_err(CE_WARN, "failed to process DOF: %s", str);
12053
12054 #ifdef DTRACE_ERRDEBUG
12055 dtrace_errdebug(str);
12056 #endif
12057 }
12058
12059 /*
12060 * Create DOF out of a currently enabled state. Right now, we only create
12061 * DOF containing the run-time options -- but this could be expanded to create
12062 * complete DOF representing the enabled state.
12063 */
12064 static dof_hdr_t *
12065 dtrace_dof_create(dtrace_state_t *state)
12066 {
12067 dof_hdr_t *dof;
12068 dof_sec_t *sec;
12069 dof_optdesc_t *opt;
12070 int i, len = sizeof (dof_hdr_t) +
12071 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
12072 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12073
12074 ASSERT(MUTEX_HELD(&dtrace_lock));
12075
12076 dof = kmem_zalloc(len, KM_SLEEP);
12077 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
12078 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
12079 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
12080 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
12081
12082 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
12083 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
12084 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
12085 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
12086 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
12087 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
12088
12089 dof->dofh_flags = 0;
12090 dof->dofh_hdrsize = sizeof (dof_hdr_t);
12091 dof->dofh_secsize = sizeof (dof_sec_t);
12092 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
12093 dof->dofh_secoff = sizeof (dof_hdr_t);
12094 dof->dofh_loadsz = len;
12095 dof->dofh_filesz = len;
12096 dof->dofh_pad = 0;
12097
12098 /*
12099 * Fill in the option section header...
12100 */
12101 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
12102 sec->dofs_type = DOF_SECT_OPTDESC;
12103 sec->dofs_align = sizeof (uint64_t);
12104 sec->dofs_flags = DOF_SECF_LOAD;
12105 sec->dofs_entsize = sizeof (dof_optdesc_t);
12106
12107 opt = (dof_optdesc_t *)((uintptr_t)sec +
12108 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
12109
12110 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
12111 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12112
12113 for (i = 0; i < DTRACEOPT_MAX; i++) {
12114 opt[i].dofo_option = i;
12115 opt[i].dofo_strtab = DOF_SECIDX_NONE;
12116 opt[i].dofo_value = state->dts_options[i];
12117 }
12118
12119 return (dof);
12120 }
12121
12122 static dof_hdr_t *
12123 dtrace_dof_copyin(uintptr_t uarg, int *errp)
12124 {
12125 dof_hdr_t hdr, *dof;
12126
12127 ASSERT(!MUTEX_HELD(&dtrace_lock));
12128
12129 /*
12130 * First, we're going to copyin() the sizeof (dof_hdr_t).
12131 */
12132 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
12133 dtrace_dof_error(NULL, "failed to copyin DOF header");
12134 *errp = EFAULT;
12135 return (NULL);
12136 }
12137
12138 /*
12139 * Now we'll allocate the entire DOF and copy it in -- provided
12140 * that the length isn't outrageous.
12141 */
12142 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
12143 dtrace_dof_error(&hdr, "load size exceeds maximum");
12144 *errp = E2BIG;
12145 return (NULL);
12146 }
12147
12148 if (hdr.dofh_loadsz < sizeof (hdr)) {
12149 dtrace_dof_error(&hdr, "invalid load size");
12150 *errp = EINVAL;
12151 return (NULL);
12152 }
12153
12154 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12155
12156 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
12157 dof->dofh_loadsz != hdr.dofh_loadsz) {
12158 kmem_free(dof, hdr.dofh_loadsz);
12159 *errp = EFAULT;
12160 return (NULL);
12161 }
12162
12163 return (dof);
12164 }
12165
12166 static dof_hdr_t *
12167 dtrace_dof_property(const char *name)
12168 {
12169 uchar_t *buf;
12170 uint64_t loadsz;
12171 unsigned int len, i;
12172 dof_hdr_t *dof;
12173
12174 /*
12175 * Unfortunately, array of values in .conf files are always (and
12176 * only) interpreted to be integer arrays. We must read our DOF
12177 * as an integer array, and then squeeze it into a byte array.
12178 */
12179 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12180 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12181 return (NULL);
12182
12183 for (i = 0; i < len; i++)
12184 buf[i] = (uchar_t)(((int *)buf)[i]);
12185
12186 if (len < sizeof (dof_hdr_t)) {
12187 ddi_prop_free(buf);
12188 dtrace_dof_error(NULL, "truncated header");
12189 return (NULL);
12190 }
12191
12192 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12193 ddi_prop_free(buf);
12194 dtrace_dof_error(NULL, "truncated DOF");
12195 return (NULL);
12196 }
12197
12198 if (loadsz >= dtrace_dof_maxsize) {
12199 ddi_prop_free(buf);
12200 dtrace_dof_error(NULL, "oversized DOF");
12201 return (NULL);
12202 }
12203
12204 dof = kmem_alloc(loadsz, KM_SLEEP);
12205 bcopy(buf, dof, loadsz);
12206 ddi_prop_free(buf);
12207
12208 return (dof);
12209 }
12210
12211 static void
12212 dtrace_dof_destroy(dof_hdr_t *dof)
12213 {
12214 kmem_free(dof, dof->dofh_loadsz);
12215 }
12216
12217 /*
12218 * Return the dof_sec_t pointer corresponding to a given section index. If the
12219 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12220 * a type other than DOF_SECT_NONE is specified, the header is checked against
12221 * this type and NULL is returned if the types do not match.
12222 */
12223 static dof_sec_t *
12224 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12225 {
12226 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12227 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12228
12229 if (i >= dof->dofh_secnum) {
12230 dtrace_dof_error(dof, "referenced section index is invalid");
12231 return (NULL);
12232 }
12233
12234 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12235 dtrace_dof_error(dof, "referenced section is not loadable");
12236 return (NULL);
12237 }
12238
12239 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12240 dtrace_dof_error(dof, "referenced section is the wrong type");
12241 return (NULL);
12242 }
12243
12244 return (sec);
12245 }
12246
12247 static dtrace_probedesc_t *
12248 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12249 {
12250 dof_probedesc_t *probe;
12251 dof_sec_t *strtab;
12252 uintptr_t daddr = (uintptr_t)dof;
12253 uintptr_t str;
12254 size_t size;
12255
12256 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12257 dtrace_dof_error(dof, "invalid probe section");
12258 return (NULL);
12259 }
12260
12261 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12262 dtrace_dof_error(dof, "bad alignment in probe description");
12263 return (NULL);
12264 }
12265
12266 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12267 dtrace_dof_error(dof, "truncated probe description");
12268 return (NULL);
12269 }
12270
12271 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12272 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12273
12274 if (strtab == NULL)
12275 return (NULL);
12276
12277 str = daddr + strtab->dofs_offset;
12278 size = strtab->dofs_size;
12279
12280 if (probe->dofp_provider >= strtab->dofs_size) {
12281 dtrace_dof_error(dof, "corrupt probe provider");
12282 return (NULL);
12283 }
12284
12285 (void) strncpy(desc->dtpd_provider,
12286 (char *)(str + probe->dofp_provider),
12287 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12288
12289 if (probe->dofp_mod >= strtab->dofs_size) {
12290 dtrace_dof_error(dof, "corrupt probe module");
12291 return (NULL);
12292 }
12293
12294 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12295 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12296
12297 if (probe->dofp_func >= strtab->dofs_size) {
12298 dtrace_dof_error(dof, "corrupt probe function");
12299 return (NULL);
12300 }
12301
12302 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12303 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12304
12305 if (probe->dofp_name >= strtab->dofs_size) {
12306 dtrace_dof_error(dof, "corrupt probe name");
12307 return (NULL);
12308 }
12309
12310 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12311 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12312
12313 return (desc);
12314 }
12315
12316 static dtrace_difo_t *
12317 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12318 cred_t *cr)
12319 {
12320 dtrace_difo_t *dp;
12321 size_t ttl = 0;
12322 dof_difohdr_t *dofd;
12323 uintptr_t daddr = (uintptr_t)dof;
12324 size_t max = dtrace_difo_maxsize;
12325 int i, l, n;
12326
12327 static const struct {
12328 int section;
12329 int bufoffs;
12330 int lenoffs;
12331 int entsize;
12332 int align;
12333 const char *msg;
12334 } difo[] = {
12335 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12336 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12337 sizeof (dif_instr_t), "multiple DIF sections" },
12338
12339 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12340 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12341 sizeof (uint64_t), "multiple integer tables" },
12342
12343 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12344 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12345 sizeof (char), "multiple string tables" },
12346
12347 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12348 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12349 sizeof (uint_t), "multiple variable tables" },
12350
12351 { DOF_SECT_NONE, 0, 0, 0, NULL }
12352 };
12353
12354 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12355 dtrace_dof_error(dof, "invalid DIFO header section");
12356 return (NULL);
12357 }
12358
12359 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12360 dtrace_dof_error(dof, "bad alignment in DIFO header");
12361 return (NULL);
12362 }
12363
12364 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12365 sec->dofs_size % sizeof (dof_secidx_t)) {
12366 dtrace_dof_error(dof, "bad size in DIFO header");
12367 return (NULL);
12368 }
12369
12370 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12371 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12372
12373 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12374 dp->dtdo_rtype = dofd->dofd_rtype;
12375
12376 for (l = 0; l < n; l++) {
12377 dof_sec_t *subsec;
12378 void **bufp;
12379 uint32_t *lenp;
12380
12381 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12382 dofd->dofd_links[l])) == NULL)
12383 goto err; /* invalid section link */
12384
12385 if (ttl + subsec->dofs_size > max) {
12386 dtrace_dof_error(dof, "exceeds maximum size");
12387 goto err;
12388 }
12389
12390 ttl += subsec->dofs_size;
12391
12392 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12393 if (subsec->dofs_type != difo[i].section)
12394 continue;
12395
12396 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12397 dtrace_dof_error(dof, "section not loaded");
12398 goto err;
12399 }
12400
12401 if (subsec->dofs_align != difo[i].align) {
12402 dtrace_dof_error(dof, "bad alignment");
12403 goto err;
12404 }
12405
12406 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12407 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12408
12409 if (*bufp != NULL) {
12410 dtrace_dof_error(dof, difo[i].msg);
12411 goto err;
12412 }
12413
12414 if (difo[i].entsize != subsec->dofs_entsize) {
12415 dtrace_dof_error(dof, "entry size mismatch");
12416 goto err;
12417 }
12418
12419 if (subsec->dofs_entsize != 0 &&
12420 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12421 dtrace_dof_error(dof, "corrupt entry size");
12422 goto err;
12423 }
12424
12425 *lenp = subsec->dofs_size;
12426 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12427 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12428 *bufp, subsec->dofs_size);
12429
12430 if (subsec->dofs_entsize != 0)
12431 *lenp /= subsec->dofs_entsize;
12432
12433 break;
12434 }
12435
12436 /*
12437 * If we encounter a loadable DIFO sub-section that is not
12438 * known to us, assume this is a broken program and fail.
12439 */
12440 if (difo[i].section == DOF_SECT_NONE &&
12441 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12442 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12443 goto err;
12444 }
12445 }
12446
12447 if (dp->dtdo_buf == NULL) {
12448 /*
12449 * We can't have a DIF object without DIF text.
12450 */
12451 dtrace_dof_error(dof, "missing DIF text");
12452 goto err;
12453 }
12454
12455 /*
12456 * Before we validate the DIF object, run through the variable table
12457 * looking for the strings -- if any of their size are under, we'll set
12458 * their size to be the system-wide default string size. Note that
12459 * this should _not_ happen if the "strsize" option has been set --
12460 * in this case, the compiler should have set the size to reflect the
12461 * setting of the option.
12462 */
12463 for (i = 0; i < dp->dtdo_varlen; i++) {
12464 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12465 dtrace_diftype_t *t = &v->dtdv_type;
12466
12467 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12468 continue;
12469
12470 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12471 t->dtdt_size = dtrace_strsize_default;
12472 }
12473
12474 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12475 goto err;
12476
12477 dtrace_difo_init(dp, vstate);
12478 return (dp);
12479
12480 err:
12481 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12482 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12483 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12484 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12485
12486 kmem_free(dp, sizeof (dtrace_difo_t));
12487 return (NULL);
12488 }
12489
12490 static dtrace_predicate_t *
12491 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12492 cred_t *cr)
12493 {
12494 dtrace_difo_t *dp;
12495
12496 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12497 return (NULL);
12498
12499 return (dtrace_predicate_create(dp));
12500 }
12501
12502 static dtrace_actdesc_t *
12503 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12504 cred_t *cr)
12505 {
12506 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12507 dof_actdesc_t *desc;
12508 dof_sec_t *difosec;
12509 size_t offs;
12510 uintptr_t daddr = (uintptr_t)dof;
12511 uint64_t arg;
12512 dtrace_actkind_t kind;
12513
12514 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12515 dtrace_dof_error(dof, "invalid action section");
12516 return (NULL);
12517 }
12518
12519 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12520 dtrace_dof_error(dof, "truncated action description");
12521 return (NULL);
12522 }
12523
12524 if (sec->dofs_align != sizeof (uint64_t)) {
12525 dtrace_dof_error(dof, "bad alignment in action description");
12526 return (NULL);
12527 }
12528
12529 if (sec->dofs_size < sec->dofs_entsize) {
12530 dtrace_dof_error(dof, "section entry size exceeds total size");
12531 return (NULL);
12532 }
12533
12534 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12535 dtrace_dof_error(dof, "bad entry size in action description");
12536 return (NULL);
12537 }
12538
12539 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12540 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12541 return (NULL);
12542 }
12543
12544 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12545 desc = (dof_actdesc_t *)(daddr +
12546 (uintptr_t)sec->dofs_offset + offs);
12547 kind = (dtrace_actkind_t)desc->dofa_kind;
12548
12549 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12550 (kind != DTRACEACT_PRINTA ||
12551 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12552 (kind == DTRACEACT_DIFEXPR &&
12553 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12554 dof_sec_t *strtab;
12555 char *str, *fmt;
12556 uint64_t i;
12557
12558 /*
12559 * The argument to these actions is an index into the
12560 * DOF string table. For printf()-like actions, this
12561 * is the format string. For print(), this is the
12562 * CTF type of the expression result.
12563 */
12564 if ((strtab = dtrace_dof_sect(dof,
12565 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12566 goto err;
12567
12568 str = (char *)((uintptr_t)dof +
12569 (uintptr_t)strtab->dofs_offset);
12570
12571 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12572 if (str[i] == '\0')
12573 break;
12574 }
12575
12576 if (i >= strtab->dofs_size) {
12577 dtrace_dof_error(dof, "bogus format string");
12578 goto err;
12579 }
12580
12581 if (i == desc->dofa_arg) {
12582 dtrace_dof_error(dof, "empty format string");
12583 goto err;
12584 }
12585
12586 i -= desc->dofa_arg;
12587 fmt = kmem_alloc(i + 1, KM_SLEEP);
12588 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12589 arg = (uint64_t)(uintptr_t)fmt;
12590 } else {
12591 if (kind == DTRACEACT_PRINTA) {
12592 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12593 arg = 0;
12594 } else {
12595 arg = desc->dofa_arg;
12596 }
12597 }
12598
12599 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12600 desc->dofa_uarg, arg);
12601
12602 if (last != NULL) {
12603 last->dtad_next = act;
12604 } else {
12605 first = act;
12606 }
12607
12608 last = act;
12609
12610 if (desc->dofa_difo == DOF_SECIDX_NONE)
12611 continue;
12612
12613 if ((difosec = dtrace_dof_sect(dof,
12614 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12615 goto err;
12616
12617 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12618
12619 if (act->dtad_difo == NULL)
12620 goto err;
12621 }
12622
12623 ASSERT(first != NULL);
12624 return (first);
12625
12626 err:
12627 for (act = first; act != NULL; act = next) {
12628 next = act->dtad_next;
12629 dtrace_actdesc_release(act, vstate);
12630 }
12631
12632 return (NULL);
12633 }
12634
12635 static dtrace_ecbdesc_t *
12636 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12637 cred_t *cr)
12638 {
12639 dtrace_ecbdesc_t *ep;
12640 dof_ecbdesc_t *ecb;
12641 dtrace_probedesc_t *desc;
12642 dtrace_predicate_t *pred = NULL;
12643
12644 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12645 dtrace_dof_error(dof, "truncated ECB description");
12646 return (NULL);
12647 }
12648
12649 if (sec->dofs_align != sizeof (uint64_t)) {
12650 dtrace_dof_error(dof, "bad alignment in ECB description");
12651 return (NULL);
12652 }
12653
12654 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12655 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12656
12657 if (sec == NULL)
12658 return (NULL);
12659
12660 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12661 ep->dted_uarg = ecb->dofe_uarg;
12662 desc = &ep->dted_probe;
12663
12664 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12665 goto err;
12666
12667 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12668 if ((sec = dtrace_dof_sect(dof,
12669 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12670 goto err;
12671
12672 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12673 goto err;
12674
12675 ep->dted_pred.dtpdd_predicate = pred;
12676 }
12677
12678 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12679 if ((sec = dtrace_dof_sect(dof,
12680 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12681 goto err;
12682
12683 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12684
12685 if (ep->dted_action == NULL)
12686 goto err;
12687 }
12688
12689 return (ep);
12690
12691 err:
12692 if (pred != NULL)
12693 dtrace_predicate_release(pred, vstate);
12694 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12695 return (NULL);
12696 }
12697
12698 /*
12699 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12700 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12701 * site of any user SETX relocations to account for load object base address.
12702 * In the future, if we need other relocations, this function can be extended.
12703 */
12704 static int
12705 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12706 {
12707 uintptr_t daddr = (uintptr_t)dof;
12708 dof_relohdr_t *dofr =
12709 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12710 dof_sec_t *ss, *rs, *ts;
12711 dof_relodesc_t *r;
12712 uint_t i, n;
12713
12714 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12715 sec->dofs_align != sizeof (dof_secidx_t)) {
12716 dtrace_dof_error(dof, "invalid relocation header");
12717 return (-1);
12718 }
12719
12720 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12721 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12722 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12723
12724 if (ss == NULL || rs == NULL || ts == NULL)
12725 return (-1); /* dtrace_dof_error() has been called already */
12726
12727 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12728 rs->dofs_align != sizeof (uint64_t)) {
12729 dtrace_dof_error(dof, "invalid relocation section");
12730 return (-1);
12731 }
12732
12733 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12734 n = rs->dofs_size / rs->dofs_entsize;
12735
12736 for (i = 0; i < n; i++) {
12737 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12738
12739 switch (r->dofr_type) {
12740 case DOF_RELO_NONE:
12741 break;
12742 case DOF_RELO_SETX:
12743 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12744 sizeof (uint64_t) > ts->dofs_size) {
12745 dtrace_dof_error(dof, "bad relocation offset");
12746 return (-1);
12747 }
12748
12749 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12750 dtrace_dof_error(dof, "misaligned setx relo");
12751 return (-1);
12752 }
12753
12754 *(uint64_t *)taddr += ubase;
12755 break;
12756 default:
12757 dtrace_dof_error(dof, "invalid relocation type");
12758 return (-1);
12759 }
12760
12761 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12762 }
12763
12764 return (0);
12765 }
12766
12767 /*
12768 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12769 * header: it should be at the front of a memory region that is at least
12770 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12771 * size. It need not be validated in any other way.
12772 */
12773 static int
12774 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12775 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12776 {
12777 uint64_t len = dof->dofh_loadsz, seclen;
12778 uintptr_t daddr = (uintptr_t)dof;
12779 dtrace_ecbdesc_t *ep;
12780 dtrace_enabling_t *enab;
12781 uint_t i;
12782
12783 ASSERT(MUTEX_HELD(&dtrace_lock));
12784 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12785
12786 /*
12787 * Check the DOF header identification bytes. In addition to checking
12788 * valid settings, we also verify that unused bits/bytes are zeroed so
12789 * we can use them later without fear of regressing existing binaries.
12790 */
12791 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12792 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12793 dtrace_dof_error(dof, "DOF magic string mismatch");
12794 return (-1);
12795 }
12796
12797 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12798 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12799 dtrace_dof_error(dof, "DOF has invalid data model");
12800 return (-1);
12801 }
12802
12803 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12804 dtrace_dof_error(dof, "DOF encoding mismatch");
12805 return (-1);
12806 }
12807
12808 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12809 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12810 dtrace_dof_error(dof, "DOF version mismatch");
12811 return (-1);
12812 }
12813
12814 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12815 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12816 return (-1);
12817 }
12818
12819 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12820 dtrace_dof_error(dof, "DOF uses too many integer registers");
12821 return (-1);
12822 }
12823
12824 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12825 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12826 return (-1);
12827 }
12828
12829 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12830 if (dof->dofh_ident[i] != 0) {
12831 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12832 return (-1);
12833 }
12834 }
12835
12836 if (dof->dofh_flags & ~DOF_FL_VALID) {
12837 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12838 return (-1);
12839 }
12840
12841 if (dof->dofh_secsize == 0) {
12842 dtrace_dof_error(dof, "zero section header size");
12843 return (-1);
12844 }
12845
12846 /*
12847 * Check that the section headers don't exceed the amount of DOF
12848 * data. Note that we cast the section size and number of sections
12849 * to uint64_t's to prevent possible overflow in the multiplication.
12850 */
12851 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12852
12853 if (dof->dofh_secoff > len || seclen > len ||
12854 dof->dofh_secoff + seclen > len) {
12855 dtrace_dof_error(dof, "truncated section headers");
12856 return (-1);
12857 }
12858
12859 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12860 dtrace_dof_error(dof, "misaligned section headers");
12861 return (-1);
12862 }
12863
12864 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12865 dtrace_dof_error(dof, "misaligned section size");
12866 return (-1);
12867 }
12868
12869 /*
12870 * Take an initial pass through the section headers to be sure that
12871 * the headers don't have stray offsets. If the 'noprobes' flag is
12872 * set, do not permit sections relating to providers, probes, or args.
12873 */
12874 for (i = 0; i < dof->dofh_secnum; i++) {
12875 dof_sec_t *sec = (dof_sec_t *)(daddr +
12876 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12877
12878 if (noprobes) {
12879 switch (sec->dofs_type) {
12880 case DOF_SECT_PROVIDER:
12881 case DOF_SECT_PROBES:
12882 case DOF_SECT_PRARGS:
12883 case DOF_SECT_PROFFS:
12884 dtrace_dof_error(dof, "illegal sections "
12885 "for enabling");
12886 return (-1);
12887 }
12888 }
12889
12890 if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
12891 !(sec->dofs_flags & DOF_SECF_LOAD)) {
12892 dtrace_dof_error(dof, "loadable section with load "
12893 "flag unset");
12894 return (-1);
12895 }
12896
12897 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12898 continue; /* just ignore non-loadable sections */
12899
12900 if (sec->dofs_align & (sec->dofs_align - 1)) {
12901 dtrace_dof_error(dof, "bad section alignment");
12902 return (-1);
12903 }
12904
12905 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12906 dtrace_dof_error(dof, "misaligned section");
12907 return (-1);
12908 }
12909
12910 if (sec->dofs_offset > len || sec->dofs_size > len ||
12911 sec->dofs_offset + sec->dofs_size > len) {
12912 dtrace_dof_error(dof, "corrupt section header");
12913 return (-1);
12914 }
12915
12916 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12917 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12918 dtrace_dof_error(dof, "non-terminating string table");
12919 return (-1);
12920 }
12921 }
12922
12923 /*
12924 * Take a second pass through the sections and locate and perform any
12925 * relocations that are present. We do this after the first pass to
12926 * be sure that all sections have had their headers validated.
12927 */
12928 for (i = 0; i < dof->dofh_secnum; i++) {
12929 dof_sec_t *sec = (dof_sec_t *)(daddr +
12930 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12931
12932 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12933 continue; /* skip sections that are not loadable */
12934
12935 switch (sec->dofs_type) {
12936 case DOF_SECT_URELHDR:
12937 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12938 return (-1);
12939 break;
12940 }
12941 }
12942
12943 if ((enab = *enabp) == NULL)
12944 enab = *enabp = dtrace_enabling_create(vstate);
12945
12946 for (i = 0; i < dof->dofh_secnum; i++) {
12947 dof_sec_t *sec = (dof_sec_t *)(daddr +
12948 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12949
12950 if (sec->dofs_type != DOF_SECT_ECBDESC)
12951 continue;
12952
12953 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12954 dtrace_enabling_destroy(enab);
12955 *enabp = NULL;
12956 return (-1);
12957 }
12958
12959 dtrace_enabling_add(enab, ep);
12960 }
12961
12962 return (0);
12963 }
12964
12965 /*
12966 * Process DOF for any options. This routine assumes that the DOF has been
12967 * at least processed by dtrace_dof_slurp().
12968 */
12969 static int
12970 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12971 {
12972 int i, rval;
12973 uint32_t entsize;
12974 size_t offs;
12975 dof_optdesc_t *desc;
12976
12977 for (i = 0; i < dof->dofh_secnum; i++) {
12978 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12979 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12980
12981 if (sec->dofs_type != DOF_SECT_OPTDESC)
12982 continue;
12983
12984 if (sec->dofs_align != sizeof (uint64_t)) {
12985 dtrace_dof_error(dof, "bad alignment in "
12986 "option description");
12987 return (EINVAL);
12988 }
12989
12990 if ((entsize = sec->dofs_entsize) == 0) {
12991 dtrace_dof_error(dof, "zeroed option entry size");
12992 return (EINVAL);
12993 }
12994
12995 if (entsize < sizeof (dof_optdesc_t)) {
12996 dtrace_dof_error(dof, "bad option entry size");
12997 return (EINVAL);
12998 }
12999
13000 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
13001 desc = (dof_optdesc_t *)((uintptr_t)dof +
13002 (uintptr_t)sec->dofs_offset + offs);
13003
13004 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
13005 dtrace_dof_error(dof, "non-zero option string");
13006 return (EINVAL);
13007 }
13008
13009 if (desc->dofo_value == DTRACEOPT_UNSET) {
13010 dtrace_dof_error(dof, "unset option");
13011 return (EINVAL);
13012 }
13013
13014 if ((rval = dtrace_state_option(state,
13015 desc->dofo_option, desc->dofo_value)) != 0) {
13016 dtrace_dof_error(dof, "rejected option");
13017 return (rval);
13018 }
13019 }
13020 }
13021
13022 return (0);
13023 }
13024
13025 /*
13026 * DTrace Consumer State Functions
13027 */
13028 int
13029 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
13030 {
13031 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
13032 void *base;
13033 uintptr_t limit;
13034 dtrace_dynvar_t *dvar, *next, *start;
13035 int i;
13036
13037 ASSERT(MUTEX_HELD(&dtrace_lock));
13038 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
13039
13040 bzero(dstate, sizeof (dtrace_dstate_t));
13041
13042 if ((dstate->dtds_chunksize = chunksize) == 0)
13043 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
13044
13045 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
13046 size = min;
13047
13048 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
13049 return (ENOMEM);
13050
13051 dstate->dtds_size = size;
13052 dstate->dtds_base = base;
13053 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
13054 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
13055
13056 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
13057
13058 if (hashsize != 1 && (hashsize & 1))
13059 hashsize--;
13060
13061 dstate->dtds_hashsize = hashsize;
13062 dstate->dtds_hash = dstate->dtds_base;
13063
13064 /*
13065 * Set all of our hash buckets to point to the single sink, and (if
13066 * it hasn't already been set), set the sink's hash value to be the
13067 * sink sentinel value. The sink is needed for dynamic variable
13068 * lookups to know that they have iterated over an entire, valid hash
13069 * chain.
13070 */
13071 for (i = 0; i < hashsize; i++)
13072 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
13073
13074 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
13075 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
13076
13077 /*
13078 * Determine number of active CPUs. Divide free list evenly among
13079 * active CPUs.
13080 */
13081 start = (dtrace_dynvar_t *)
13082 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
13083 limit = (uintptr_t)base + size;
13084
13085 maxper = (limit - (uintptr_t)start) / NCPU;
13086 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13087
13088 for (i = 0; i < NCPU; i++) {
13089 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13090
13091 /*
13092 * If we don't even have enough chunks to make it once through
13093 * NCPUs, we're just going to allocate everything to the first
13094 * CPU. And if we're on the last CPU, we're going to allocate
13095 * whatever is left over. In either case, we set the limit to
13096 * be the limit of the dynamic variable space.
13097 */
13098 if (maxper == 0 || i == NCPU - 1) {
13099 limit = (uintptr_t)base + size;
13100 start = NULL;
13101 } else {
13102 limit = (uintptr_t)start + maxper;
13103 start = (dtrace_dynvar_t *)limit;
13104 }
13105
13106 ASSERT(limit <= (uintptr_t)base + size);
13107
13108 for (;;) {
13109 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13110 dstate->dtds_chunksize);
13111
13112 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13113 break;
13114
13115 dvar->dtdv_next = next;
13116 dvar = next;
13117 }
13118
13119 if (maxper == 0)
13120 break;
13121 }
13122
13123 return (0);
13124 }
13125
13126 void
13127 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13128 {
13129 ASSERT(MUTEX_HELD(&cpu_lock));
13130
13131 if (dstate->dtds_base == NULL)
13132 return;
13133
13134 kmem_free(dstate->dtds_base, dstate->dtds_size);
13135 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13136 }
13137
13138 static void
13139 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13140 {
13141 /*
13142 * Logical XOR, where are you?
13143 */
13144 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13145
13146 if (vstate->dtvs_nglobals > 0) {
13147 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13148 sizeof (dtrace_statvar_t *));
13149 }
13150
13151 if (vstate->dtvs_ntlocals > 0) {
13152 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13153 sizeof (dtrace_difv_t));
13154 }
13155
13156 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13157
13158 if (vstate->dtvs_nlocals > 0) {
13159 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13160 sizeof (dtrace_statvar_t *));
13161 }
13162 }
13163
13164 static void
13165 dtrace_state_clean(dtrace_state_t *state)
13166 {
13167 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13168 return;
13169
13170 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13171 dtrace_speculation_clean(state);
13172 }
13173
13174 static void
13175 dtrace_state_deadman(dtrace_state_t *state)
13176 {
13177 hrtime_t now;
13178
13179 dtrace_sync();
13180
13181 now = dtrace_gethrtime();
13182
13183 if (state != dtrace_anon.dta_state &&
13184 now - state->dts_laststatus >= dtrace_deadman_user)
13185 return;
13186
13187 /*
13188 * We must be sure that dts_alive never appears to be less than the
13189 * value upon entry to dtrace_state_deadman(), and because we lack a
13190 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13191 * store INT64_MAX to it, followed by a memory barrier, followed by
13192 * the new value. This assures that dts_alive never appears to be
13193 * less than its true value, regardless of the order in which the
13194 * stores to the underlying storage are issued.
13195 */
13196 state->dts_alive = INT64_MAX;
13197 dtrace_membar_producer();
13198 state->dts_alive = now;
13199 }
13200
13201 dtrace_state_t *
13202 dtrace_state_create(dev_t *devp, cred_t *cr)
13203 {
13204 minor_t minor;
13205 major_t major;
13206 char c[30];
13207 dtrace_state_t *state;
13208 dtrace_optval_t *opt;
13209 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13210
13211 ASSERT(MUTEX_HELD(&dtrace_lock));
13212 ASSERT(MUTEX_HELD(&cpu_lock));
13213
13214 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13215 VM_BESTFIT | VM_SLEEP);
13216
13217 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13218 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13219 return (NULL);
13220 }
13221
13222 state = ddi_get_soft_state(dtrace_softstate, minor);
13223 state->dts_epid = DTRACE_EPIDNONE + 1;
13224
13225 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
13226 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13227 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13228
13229 if (devp != NULL) {
13230 major = getemajor(*devp);
13231 } else {
13232 major = ddi_driver_major(dtrace_devi);
13233 }
13234
13235 state->dts_dev = makedevice(major, minor);
13236
13237 if (devp != NULL)
13238 *devp = state->dts_dev;
13239
13240 /*
13241 * We allocate NCPU buffers. On the one hand, this can be quite
13242 * a bit of memory per instance (nearly 36K on a Starcat). On the
13243 * other hand, it saves an additional memory reference in the probe
13244 * path.
13245 */
13246 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13247 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13248 state->dts_cleaner = CYCLIC_NONE;
13249 state->dts_deadman = CYCLIC_NONE;
13250 state->dts_vstate.dtvs_state = state;
13251
13252 for (i = 0; i < DTRACEOPT_MAX; i++)
13253 state->dts_options[i] = DTRACEOPT_UNSET;
13254
13255 /*
13256 * Set the default options.
13257 */
13258 opt = state->dts_options;
13259 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13260 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13261 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13262 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13263 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13264 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13265 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13266 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13267 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13268 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13269 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13270 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13271 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13272 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13273
13274 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13275
13276 /*
13277 * Depending on the user credentials, we set flag bits which alter probe
13278 * visibility or the amount of destructiveness allowed. In the case of
13279 * actual anonymous tracing, or the possession of all privileges, all of
13280 * the normal checks are bypassed.
13281 */
13282 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13283 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13284 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13285 } else {
13286 /*
13287 * Set up the credentials for this instantiation. We take a
13288 * hold on the credential to prevent it from disappearing on
13289 * us; this in turn prevents the zone_t referenced by this
13290 * credential from disappearing. This means that we can
13291 * examine the credential and the zone from probe context.
13292 */
13293 crhold(cr);
13294 state->dts_cred.dcr_cred = cr;
13295
13296 /*
13297 * CRA_PROC means "we have *some* privilege for dtrace" and
13298 * unlocks the use of variables like pid, zonename, etc.
13299 */
13300 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13301 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13302 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13303 }
13304
13305 /*
13306 * dtrace_user allows use of syscall and profile providers.
13307 * If the user also has proc_owner and/or proc_zone, we
13308 * extend the scope to include additional visibility and
13309 * destructive power.
13310 */
13311 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13312 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13313 state->dts_cred.dcr_visible |=
13314 DTRACE_CRV_ALLPROC;
13315
13316 state->dts_cred.dcr_action |=
13317 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13318 }
13319
13320 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13321 state->dts_cred.dcr_visible |=
13322 DTRACE_CRV_ALLZONE;
13323
13324 state->dts_cred.dcr_action |=
13325 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13326 }
13327
13328 /*
13329 * If we have all privs in whatever zone this is,
13330 * we can do destructive things to processes which
13331 * have altered credentials.
13332 */
13333 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13334 cr->cr_zone->zone_privset)) {
13335 state->dts_cred.dcr_action |=
13336 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13337 }
13338 }
13339
13340 /*
13341 * Holding the dtrace_kernel privilege also implies that
13342 * the user has the dtrace_user privilege from a visibility
13343 * perspective. But without further privileges, some
13344 * destructive actions are not available.
13345 */
13346 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13347 /*
13348 * Make all probes in all zones visible. However,
13349 * this doesn't mean that all actions become available
13350 * to all zones.
13351 */
13352 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13353 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13354
13355 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13356 DTRACE_CRA_PROC;
13357 /*
13358 * Holding proc_owner means that destructive actions
13359 * for *this* zone are allowed.
13360 */
13361 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13362 state->dts_cred.dcr_action |=
13363 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13364
13365 /*
13366 * Holding proc_zone means that destructive actions
13367 * for this user/group ID in all zones is allowed.
13368 */
13369 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13370 state->dts_cred.dcr_action |=
13371 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13372
13373 /*
13374 * If we have all privs in whatever zone this is,
13375 * we can do destructive things to processes which
13376 * have altered credentials.
13377 */
13378 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13379 cr->cr_zone->zone_privset)) {
13380 state->dts_cred.dcr_action |=
13381 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13382 }
13383 }
13384
13385 /*
13386 * Holding the dtrace_proc privilege gives control over fasttrap
13387 * and pid providers. We need to grant wider destructive
13388 * privileges in the event that the user has proc_owner and/or
13389 * proc_zone.
13390 */
13391 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13392 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13393 state->dts_cred.dcr_action |=
13394 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13395
13396 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13397 state->dts_cred.dcr_action |=
13398 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13399 }
13400 }
13401
13402 return (state);
13403 }
13404
13405 static int
13406 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13407 {
13408 dtrace_optval_t *opt = state->dts_options, size;
13409 processorid_t cpu;
13410 int flags = 0, rval, factor, divisor = 1;
13411
13412 ASSERT(MUTEX_HELD(&dtrace_lock));
13413 ASSERT(MUTEX_HELD(&cpu_lock));
13414 ASSERT(which < DTRACEOPT_MAX);
13415 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13416 (state == dtrace_anon.dta_state &&
13417 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13418
13419 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13420 return (0);
13421
13422 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13423 cpu = opt[DTRACEOPT_CPU];
13424
13425 if (which == DTRACEOPT_SPECSIZE)
13426 flags |= DTRACEBUF_NOSWITCH;
13427
13428 if (which == DTRACEOPT_BUFSIZE) {
13429 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13430 flags |= DTRACEBUF_RING;
13431
13432 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13433 flags |= DTRACEBUF_FILL;
13434
13435 if (state != dtrace_anon.dta_state ||
13436 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13437 flags |= DTRACEBUF_INACTIVE;
13438 }
13439
13440 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
13441 /*
13442 * The size must be 8-byte aligned. If the size is not 8-byte
13443 * aligned, drop it down by the difference.
13444 */
13445 if (size & (sizeof (uint64_t) - 1))
13446 size -= size & (sizeof (uint64_t) - 1);
13447
13448 if (size < state->dts_reserve) {
13449 /*
13450 * Buffers always must be large enough to accommodate
13451 * their prereserved space. We return E2BIG instead
13452 * of ENOMEM in this case to allow for user-level
13453 * software to differentiate the cases.
13454 */
13455 return (E2BIG);
13456 }
13457
13458 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
13459
13460 if (rval != ENOMEM) {
13461 opt[which] = size;
13462 return (rval);
13463 }
13464
13465 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13466 return (rval);
13467
13468 for (divisor = 2; divisor < factor; divisor <<= 1)
13469 continue;
13470 }
13471
13472 return (ENOMEM);
13473 }
13474
13475 static int
13476 dtrace_state_buffers(dtrace_state_t *state)
13477 {
13478 dtrace_speculation_t *spec = state->dts_speculations;
13479 int rval, i;
13480
13481 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13482 DTRACEOPT_BUFSIZE)) != 0)
13483 return (rval);
13484
13485 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13486 DTRACEOPT_AGGSIZE)) != 0)
13487 return (rval);
13488
13489 for (i = 0; i < state->dts_nspeculations; i++) {
13490 if ((rval = dtrace_state_buffer(state,
13491 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13492 return (rval);
13493 }
13494
13495 return (0);
13496 }
13497
13498 static void
13499 dtrace_state_prereserve(dtrace_state_t *state)
13500 {
13501 dtrace_ecb_t *ecb;
13502 dtrace_probe_t *probe;
13503
13504 state->dts_reserve = 0;
13505
13506 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13507 return;
13508
13509 /*
13510 * If our buffer policy is a "fill" buffer policy, we need to set the
13511 * prereserved space to be the space required by the END probes.
13512 */
13513 probe = dtrace_probes[dtrace_probeid_end - 1];
13514 ASSERT(probe != NULL);
13515
13516 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13517 if (ecb->dte_state != state)
13518 continue;
13519
13520 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13521 }
13522 }
13523
13524 static int
13525 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13526 {
13527 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13528 dtrace_speculation_t *spec;
13529 dtrace_buffer_t *buf;
13530 cyc_handler_t hdlr;
13531 cyc_time_t when;
13532 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13533 dtrace_icookie_t cookie;
13534
13535 mutex_enter(&cpu_lock);
13536 mutex_enter(&dtrace_lock);
13537
13538 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13539 rval = EBUSY;
13540 goto out;
13541 }
13542
13543 /*
13544 * Before we can perform any checks, we must prime all of the
13545 * retained enablings that correspond to this state.
13546 */
13547 dtrace_enabling_prime(state);
13548
13549 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13550 rval = EACCES;
13551 goto out;
13552 }
13553
13554 dtrace_state_prereserve(state);
13555
13556 /*
13557 * Now we want to do is try to allocate our speculations.
13558 * We do not automatically resize the number of speculations; if
13559 * this fails, we will fail the operation.
13560 */
13561 nspec = opt[DTRACEOPT_NSPEC];
13562 ASSERT(nspec != DTRACEOPT_UNSET);
13563
13564 if (nspec > INT_MAX) {
13565 rval = ENOMEM;
13566 goto out;
13567 }
13568
13569 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
13570 KM_NOSLEEP | KM_NORMALPRI);
13571
13572 if (spec == NULL) {
13573 rval = ENOMEM;
13574 goto out;
13575 }
13576
13577 state->dts_speculations = spec;
13578 state->dts_nspeculations = (int)nspec;
13579
13580 for (i = 0; i < nspec; i++) {
13581 if ((buf = kmem_zalloc(bufsize,
13582 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
13583 rval = ENOMEM;
13584 goto err;
13585 }
13586
13587 spec[i].dtsp_buffer = buf;
13588 }
13589
13590 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13591 if (dtrace_anon.dta_state == NULL) {
13592 rval = ENOENT;
13593 goto out;
13594 }
13595
13596 if (state->dts_necbs != 0) {
13597 rval = EALREADY;
13598 goto out;
13599 }
13600
13601 state->dts_anon = dtrace_anon_grab();
13602 ASSERT(state->dts_anon != NULL);
13603 state = state->dts_anon;
13604
13605 /*
13606 * We want "grabanon" to be set in the grabbed state, so we'll
13607 * copy that option value from the grabbing state into the
13608 * grabbed state.
13609 */
13610 state->dts_options[DTRACEOPT_GRABANON] =
13611 opt[DTRACEOPT_GRABANON];
13612
13613 *cpu = dtrace_anon.dta_beganon;
13614
13615 /*
13616 * If the anonymous state is active (as it almost certainly
13617 * is if the anonymous enabling ultimately matched anything),
13618 * we don't allow any further option processing -- but we
13619 * don't return failure.
13620 */
13621 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13622 goto out;
13623 }
13624
13625 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13626 opt[DTRACEOPT_AGGSIZE] != 0) {
13627 if (state->dts_aggregations == NULL) {
13628 /*
13629 * We're not going to create an aggregation buffer
13630 * because we don't have any ECBs that contain
13631 * aggregations -- set this option to 0.
13632 */
13633 opt[DTRACEOPT_AGGSIZE] = 0;
13634 } else {
13635 /*
13636 * If we have an aggregation buffer, we must also have
13637 * a buffer to use as scratch.
13638 */
13639 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13640 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13641 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13642 }
13643 }
13644 }
13645
13646 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13647 opt[DTRACEOPT_SPECSIZE] != 0) {
13648 if (!state->dts_speculates) {
13649 /*
13650 * We're not going to create speculation buffers
13651 * because we don't have any ECBs that actually
13652 * speculate -- set the speculation size to 0.
13653 */
13654 opt[DTRACEOPT_SPECSIZE] = 0;
13655 }
13656 }
13657
13658 /*
13659 * The bare minimum size for any buffer that we're actually going to
13660 * do anything to is sizeof (uint64_t).
13661 */
13662 sz = sizeof (uint64_t);
13663
13664 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13665 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13666 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13667 /*
13668 * A buffer size has been explicitly set to 0 (or to a size
13669 * that will be adjusted to 0) and we need the space -- we
13670 * need to return failure. We return ENOSPC to differentiate
13671 * it from failing to allocate a buffer due to failure to meet
13672 * the reserve (for which we return E2BIG).
13673 */
13674 rval = ENOSPC;
13675 goto out;
13676 }
13677
13678 if ((rval = dtrace_state_buffers(state)) != 0)
13679 goto err;
13680
13681 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13682 sz = dtrace_dstate_defsize;
13683
13684 do {
13685 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13686
13687 if (rval == 0)
13688 break;
13689
13690 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13691 goto err;
13692 } while (sz >>= 1);
13693
13694 opt[DTRACEOPT_DYNVARSIZE] = sz;
13695
13696 if (rval != 0)
13697 goto err;
13698
13699 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13700 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13701
13702 if (opt[DTRACEOPT_CLEANRATE] == 0)
13703 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13704
13705 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13706 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13707
13708 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13709 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13710
13711 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13712 hdlr.cyh_arg = state;
13713 hdlr.cyh_level = CY_LOW_LEVEL;
13714
13715 when.cyt_when = 0;
13716 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13717
13718 state->dts_cleaner = cyclic_add(&hdlr, &when);
13719
13720 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13721 hdlr.cyh_arg = state;
13722 hdlr.cyh_level = CY_LOW_LEVEL;
13723
13724 when.cyt_when = 0;
13725 when.cyt_interval = dtrace_deadman_interval;
13726
13727 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13728 state->dts_deadman = cyclic_add(&hdlr, &when);
13729
13730 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13731
13732 if (state->dts_getf != 0 &&
13733 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13734 /*
13735 * We don't have kernel privs but we have at least one call
13736 * to getf(); we need to bump our zone's count, and (if
13737 * this is the first enabling to have an unprivileged call
13738 * to getf()) we need to hook into closef().
13739 */
13740 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
13741
13742 if (dtrace_getf++ == 0) {
13743 ASSERT(dtrace_closef == NULL);
13744 dtrace_closef = dtrace_getf_barrier;
13745 }
13746 }
13747
13748 /*
13749 * Now it's time to actually fire the BEGIN probe. We need to disable
13750 * interrupts here both to record the CPU on which we fired the BEGIN
13751 * probe (the data from this CPU will be processed first at user
13752 * level) and to manually activate the buffer for this CPU.
13753 */
13754 cookie = dtrace_interrupt_disable();
13755 *cpu = CPU->cpu_id;
13756 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13757 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13758
13759 dtrace_probe(dtrace_probeid_begin,
13760 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13761 dtrace_interrupt_enable(cookie);
13762 /*
13763 * We may have had an exit action from a BEGIN probe; only change our
13764 * state to ACTIVE if we're still in WARMUP.
13765 */
13766 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13767 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13768
13769 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13770 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13771
13772 /*
13773 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13774 * want each CPU to transition its principal buffer out of the
13775 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13776 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13777 * atomically transition from processing none of a state's ECBs to
13778 * processing all of them.
13779 */
13780 dtrace_xcall(DTRACE_CPUALL,
13781 (dtrace_xcall_t)dtrace_buffer_activate, state);
13782 goto out;
13783
13784 err:
13785 dtrace_buffer_free(state->dts_buffer);
13786 dtrace_buffer_free(state->dts_aggbuffer);
13787
13788 if ((nspec = state->dts_nspeculations) == 0) {
13789 ASSERT(state->dts_speculations == NULL);
13790 goto out;
13791 }
13792
13793 spec = state->dts_speculations;
13794 ASSERT(spec != NULL);
13795
13796 for (i = 0; i < state->dts_nspeculations; i++) {
13797 if ((buf = spec[i].dtsp_buffer) == NULL)
13798 break;
13799
13800 dtrace_buffer_free(buf);
13801 kmem_free(buf, bufsize);
13802 }
13803
13804 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13805 state->dts_nspeculations = 0;
13806 state->dts_speculations = NULL;
13807
13808 out:
13809 mutex_exit(&dtrace_lock);
13810 mutex_exit(&cpu_lock);
13811
13812 return (rval);
13813 }
13814
13815 static int
13816 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13817 {
13818 dtrace_icookie_t cookie;
13819
13820 ASSERT(MUTEX_HELD(&dtrace_lock));
13821
13822 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13823 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13824 return (EINVAL);
13825
13826 /*
13827 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13828 * to be sure that every CPU has seen it. See below for the details
13829 * on why this is done.
13830 */
13831 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13832 dtrace_sync();
13833
13834 /*
13835 * By this point, it is impossible for any CPU to be still processing
13836 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13837 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13838 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13839 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13840 * iff we're in the END probe.
13841 */
13842 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13843 dtrace_sync();
13844 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13845
13846 /*
13847 * Finally, we can release the reserve and call the END probe. We
13848 * disable interrupts across calling the END probe to allow us to
13849 * return the CPU on which we actually called the END probe. This
13850 * allows user-land to be sure that this CPU's principal buffer is
13851 * processed last.
13852 */
13853 state->dts_reserve = 0;
13854
13855 cookie = dtrace_interrupt_disable();
13856 *cpu = CPU->cpu_id;
13857 dtrace_probe(dtrace_probeid_end,
13858 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13859 dtrace_interrupt_enable(cookie);
13860
13861 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13862 dtrace_sync();
13863
13864 if (state->dts_getf != 0 &&
13865 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
13866 /*
13867 * We don't have kernel privs but we have at least one call
13868 * to getf(); we need to lower our zone's count, and (if
13869 * this is the last enabling to have an unprivileged call
13870 * to getf()) we need to clear the closef() hook.
13871 */
13872 ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
13873 ASSERT(dtrace_closef == dtrace_getf_barrier);
13874 ASSERT(dtrace_getf > 0);
13875
13876 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
13877
13878 if (--dtrace_getf == 0)
13879 dtrace_closef = NULL;
13880 }
13881
13882 return (0);
13883 }
13884
13885 static int
13886 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13887 dtrace_optval_t val)
13888 {
13889 ASSERT(MUTEX_HELD(&dtrace_lock));
13890
13891 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13892 return (EBUSY);
13893
13894 if (option >= DTRACEOPT_MAX)
13895 return (EINVAL);
13896
13897 if (option != DTRACEOPT_CPU && val < 0)
13898 return (EINVAL);
13899
13900 switch (option) {
13901 case DTRACEOPT_DESTRUCTIVE:
13902 if (dtrace_destructive_disallow)
13903 return (EACCES);
13904
13905 state->dts_cred.dcr_destructive = 1;
13906 break;
13907
13908 case DTRACEOPT_BUFSIZE:
13909 case DTRACEOPT_DYNVARSIZE:
13910 case DTRACEOPT_AGGSIZE:
13911 case DTRACEOPT_SPECSIZE:
13912 case DTRACEOPT_STRSIZE:
13913 if (val < 0)
13914 return (EINVAL);
13915
13916 if (val >= LONG_MAX) {
13917 /*
13918 * If this is an otherwise negative value, set it to
13919 * the highest multiple of 128m less than LONG_MAX.
13920 * Technically, we're adjusting the size without
13921 * regard to the buffer resizing policy, but in fact,
13922 * this has no effect -- if we set the buffer size to
13923 * ~LONG_MAX and the buffer policy is ultimately set to
13924 * be "manual", the buffer allocation is guaranteed to
13925 * fail, if only because the allocation requires two
13926 * buffers. (We set the the size to the highest
13927 * multiple of 128m because it ensures that the size
13928 * will remain a multiple of a megabyte when
13929 * repeatedly halved -- all the way down to 15m.)
13930 */
13931 val = LONG_MAX - (1 << 27) + 1;
13932 }
13933 }
13934
13935 state->dts_options[option] = val;
13936
13937 return (0);
13938 }
13939
13940 static void
13941 dtrace_state_destroy(dtrace_state_t *state)
13942 {
13943 dtrace_ecb_t *ecb;
13944 dtrace_vstate_t *vstate = &state->dts_vstate;
13945 minor_t minor = getminor(state->dts_dev);
13946 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13947 dtrace_speculation_t *spec = state->dts_speculations;
13948 int nspec = state->dts_nspeculations;
13949 uint32_t match;
13950
13951 ASSERT(MUTEX_HELD(&dtrace_lock));
13952 ASSERT(MUTEX_HELD(&cpu_lock));
13953
13954 /*
13955 * First, retract any retained enablings for this state.
13956 */
13957 dtrace_enabling_retract(state);
13958 ASSERT(state->dts_nretained == 0);
13959
13960 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13961 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13962 /*
13963 * We have managed to come into dtrace_state_destroy() on a
13964 * hot enabling -- almost certainly because of a disorderly
13965 * shutdown of a consumer. (That is, a consumer that is
13966 * exiting without having called dtrace_stop().) In this case,
13967 * we're going to set our activity to be KILLED, and then
13968 * issue a sync to be sure that everyone is out of probe
13969 * context before we start blowing away ECBs.
13970 */
13971 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13972 dtrace_sync();
13973 }
13974
13975 /*
13976 * Release the credential hold we took in dtrace_state_create().
13977 */
13978 if (state->dts_cred.dcr_cred != NULL)
13979 crfree(state->dts_cred.dcr_cred);
13980
13981 /*
13982 * Now we can safely disable and destroy any enabled probes. Because
13983 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13984 * (especially if they're all enabled), we take two passes through the
13985 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13986 * in the second we disable whatever is left over.
13987 */
13988 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13989 for (i = 0; i < state->dts_necbs; i++) {
13990 if ((ecb = state->dts_ecbs[i]) == NULL)
13991 continue;
13992
13993 if (match && ecb->dte_probe != NULL) {
13994 dtrace_probe_t *probe = ecb->dte_probe;
13995 dtrace_provider_t *prov = probe->dtpr_provider;
13996
13997 if (!(prov->dtpv_priv.dtpp_flags & match))
13998 continue;
13999 }
14000
14001 dtrace_ecb_disable(ecb);
14002 dtrace_ecb_destroy(ecb);
14003 }
14004
14005 if (!match)
14006 break;
14007 }
14008
14009 /*
14010 * Before we free the buffers, perform one more sync to assure that
14011 * every CPU is out of probe context.
14012 */
14013 dtrace_sync();
14014
14015 dtrace_buffer_free(state->dts_buffer);
14016 dtrace_buffer_free(state->dts_aggbuffer);
14017
14018 for (i = 0; i < nspec; i++)
14019 dtrace_buffer_free(spec[i].dtsp_buffer);
14020
14021 if (state->dts_cleaner != CYCLIC_NONE)
14022 cyclic_remove(state->dts_cleaner);
14023
14024 if (state->dts_deadman != CYCLIC_NONE)
14025 cyclic_remove(state->dts_deadman);
14026
14027 dtrace_dstate_fini(&vstate->dtvs_dynvars);
14028 dtrace_vstate_fini(vstate);
14029 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14030
14031 if (state->dts_aggregations != NULL) {
14032 #ifdef DEBUG
14033 for (i = 0; i < state->dts_naggregations; i++)
14034 ASSERT(state->dts_aggregations[i] == NULL);
14035 #endif
14036 ASSERT(state->dts_naggregations > 0);
14037 kmem_free(state->dts_aggregations,
14038 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14039 }
14040
14041 kmem_free(state->dts_buffer, bufsize);
14042 kmem_free(state->dts_aggbuffer, bufsize);
14043
14044 for (i = 0; i < nspec; i++)
14045 kmem_free(spec[i].dtsp_buffer, bufsize);
14046
14047 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14048
14049 dtrace_format_destroy(state);
14050
14051 vmem_destroy(state->dts_aggid_arena);
14052 ddi_soft_state_free(dtrace_softstate, minor);
14053 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14054 }
14055
14056 /*
14057 * DTrace Anonymous Enabling Functions
14058 */
14059 static dtrace_state_t *
14060 dtrace_anon_grab(void)
14061 {
14062 dtrace_state_t *state;
14063
14064 ASSERT(MUTEX_HELD(&dtrace_lock));
14065
14066 if ((state = dtrace_anon.dta_state) == NULL) {
14067 ASSERT(dtrace_anon.dta_enabling == NULL);
14068 return (NULL);
14069 }
14070
14071 ASSERT(dtrace_anon.dta_enabling != NULL);
14072 ASSERT(dtrace_retained != NULL);
14073
14074 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14075 dtrace_anon.dta_enabling = NULL;
14076 dtrace_anon.dta_state = NULL;
14077
14078 return (state);
14079 }
14080
14081 static void
14082 dtrace_anon_property(void)
14083 {
14084 int i, rv;
14085 dtrace_state_t *state;
14086 dof_hdr_t *dof;
14087 char c[32]; /* enough for "dof-data-" + digits */
14088
14089 ASSERT(MUTEX_HELD(&dtrace_lock));
14090 ASSERT(MUTEX_HELD(&cpu_lock));
14091
14092 for (i = 0; ; i++) {
14093 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14094
14095 dtrace_err_verbose = 1;
14096
14097 if ((dof = dtrace_dof_property(c)) == NULL) {
14098 dtrace_err_verbose = 0;
14099 break;
14100 }
14101
14102 /*
14103 * We want to create anonymous state, so we need to transition
14104 * the kernel debugger to indicate that DTrace is active. If
14105 * this fails (e.g. because the debugger has modified text in
14106 * some way), we won't continue with the processing.
14107 */
14108 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14109 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14110 "enabling ignored.");
14111 dtrace_dof_destroy(dof);
14112 break;
14113 }
14114
14115 /*
14116 * If we haven't allocated an anonymous state, we'll do so now.
14117 */
14118 if ((state = dtrace_anon.dta_state) == NULL) {
14119 state = dtrace_state_create(NULL, NULL);
14120 dtrace_anon.dta_state = state;
14121
14122 if (state == NULL) {
14123 /*
14124 * This basically shouldn't happen: the only
14125 * failure mode from dtrace_state_create() is a
14126 * failure of ddi_soft_state_zalloc() that
14127 * itself should never happen. Still, the
14128 * interface allows for a failure mode, and
14129 * we want to fail as gracefully as possible:
14130 * we'll emit an error message and cease
14131 * processing anonymous state in this case.
14132 */
14133 cmn_err(CE_WARN, "failed to create "
14134 "anonymous state");
14135 dtrace_dof_destroy(dof);
14136 break;
14137 }
14138 }
14139
14140 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14141 &dtrace_anon.dta_enabling, 0, B_TRUE);
14142
14143 if (rv == 0)
14144 rv = dtrace_dof_options(dof, state);
14145
14146 dtrace_err_verbose = 0;
14147 dtrace_dof_destroy(dof);
14148
14149 if (rv != 0) {
14150 /*
14151 * This is malformed DOF; chuck any anonymous state
14152 * that we created.
14153 */
14154 ASSERT(dtrace_anon.dta_enabling == NULL);
14155 dtrace_state_destroy(state);
14156 dtrace_anon.dta_state = NULL;
14157 break;
14158 }
14159
14160 ASSERT(dtrace_anon.dta_enabling != NULL);
14161 }
14162
14163 if (dtrace_anon.dta_enabling != NULL) {
14164 int rval;
14165
14166 /*
14167 * dtrace_enabling_retain() can only fail because we are
14168 * trying to retain more enablings than are allowed -- but
14169 * we only have one anonymous enabling, and we are guaranteed
14170 * to be allowed at least one retained enabling; we assert
14171 * that dtrace_enabling_retain() returns success.
14172 */
14173 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14174 ASSERT(rval == 0);
14175
14176 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14177 }
14178 }
14179
14180 /*
14181 * DTrace Helper Functions
14182 */
14183 static void
14184 dtrace_helper_trace(dtrace_helper_action_t *helper,
14185 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14186 {
14187 uint32_t size, next, nnext, i;
14188 dtrace_helptrace_t *ent, *buffer;
14189 uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14190
14191 if ((buffer = dtrace_helptrace_buffer) == NULL)
14192 return;
14193
14194 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14195
14196 /*
14197 * What would a tracing framework be without its own tracing
14198 * framework? (Well, a hell of a lot simpler, for starters...)
14199 */
14200 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14201 sizeof (uint64_t) - sizeof (uint64_t);
14202
14203 /*
14204 * Iterate until we can allocate a slot in the trace buffer.
14205 */
14206 do {
14207 next = dtrace_helptrace_next;
14208
14209 if (next + size < dtrace_helptrace_bufsize) {
14210 nnext = next + size;
14211 } else {
14212 nnext = size;
14213 }
14214 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14215
14216 /*
14217 * We have our slot; fill it in.
14218 */
14219 if (nnext == size) {
14220 dtrace_helptrace_wrapped++;
14221 next = 0;
14222 }
14223
14224 ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
14225 ent->dtht_helper = helper;
14226 ent->dtht_where = where;
14227 ent->dtht_nlocals = vstate->dtvs_nlocals;
14228
14229 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14230 mstate->dtms_fltoffs : -1;
14231 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14232 ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
14233
14234 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14235 dtrace_statvar_t *svar;
14236
14237 if ((svar = vstate->dtvs_locals[i]) == NULL)
14238 continue;
14239
14240 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14241 ent->dtht_locals[i] =
14242 ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
14243 }
14244 }
14245
14246 static uint64_t
14247 dtrace_helper(int which, dtrace_mstate_t *mstate,
14248 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14249 {
14250 uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14251 uint64_t sarg0 = mstate->dtms_arg[0];
14252 uint64_t sarg1 = mstate->dtms_arg[1];
14253 uint64_t rval;
14254 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14255 dtrace_helper_action_t *helper;
14256 dtrace_vstate_t *vstate;
14257 dtrace_difo_t *pred;
14258 int i, trace = dtrace_helptrace_buffer != NULL;
14259
14260 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14261
14262 if (helpers == NULL)
14263 return (0);
14264
14265 if ((helper = helpers->dthps_actions[which]) == NULL)
14266 return (0);
14267
14268 vstate = &helpers->dthps_vstate;
14269 mstate->dtms_arg[0] = arg0;
14270 mstate->dtms_arg[1] = arg1;
14271
14272 /*
14273 * Now iterate over each helper. If its predicate evaluates to 'true',
14274 * we'll call the corresponding actions. Note that the below calls
14275 * to dtrace_dif_emulate() may set faults in machine state. This is
14276 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14277 * the stored DIF offset with its own (which is the desired behavior).
14278 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14279 * from machine state; this is okay, too.
14280 */
14281 for (; helper != NULL; helper = helper->dtha_next) {
14282 if ((pred = helper->dtha_predicate) != NULL) {
14283 if (trace)
14284 dtrace_helper_trace(helper, mstate, vstate, 0);
14285
14286 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14287 goto next;
14288
14289 if (*flags & CPU_DTRACE_FAULT)
14290 goto err;
14291 }
14292
14293 for (i = 0; i < helper->dtha_nactions; i++) {
14294 if (trace)
14295 dtrace_helper_trace(helper,
14296 mstate, vstate, i + 1);
14297
14298 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14299 mstate, vstate, state);
14300
14301 if (*flags & CPU_DTRACE_FAULT)
14302 goto err;
14303 }
14304
14305 next:
14306 if (trace)
14307 dtrace_helper_trace(helper, mstate, vstate,
14308 DTRACE_HELPTRACE_NEXT);
14309 }
14310
14311 if (trace)
14312 dtrace_helper_trace(helper, mstate, vstate,
14313 DTRACE_HELPTRACE_DONE);
14314
14315 /*
14316 * Restore the arg0 that we saved upon entry.
14317 */
14318 mstate->dtms_arg[0] = sarg0;
14319 mstate->dtms_arg[1] = sarg1;
14320
14321 return (rval);
14322
14323 err:
14324 if (trace)
14325 dtrace_helper_trace(helper, mstate, vstate,
14326 DTRACE_HELPTRACE_ERR);
14327
14328 /*
14329 * Restore the arg0 that we saved upon entry.
14330 */
14331 mstate->dtms_arg[0] = sarg0;
14332 mstate->dtms_arg[1] = sarg1;
14333
14334 return (NULL);
14335 }
14336
14337 static void
14338 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14339 dtrace_vstate_t *vstate)
14340 {
14341 int i;
14342
14343 if (helper->dtha_predicate != NULL)
14344 dtrace_difo_release(helper->dtha_predicate, vstate);
14345
14346 for (i = 0; i < helper->dtha_nactions; i++) {
14347 ASSERT(helper->dtha_actions[i] != NULL);
14348 dtrace_difo_release(helper->dtha_actions[i], vstate);
14349 }
14350
14351 kmem_free(helper->dtha_actions,
14352 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14353 kmem_free(helper, sizeof (dtrace_helper_action_t));
14354 }
14355
14356 static int
14357 dtrace_helper_destroygen(int gen)
14358 {
14359 proc_t *p = curproc;
14360 dtrace_helpers_t *help = p->p_dtrace_helpers;
14361 dtrace_vstate_t *vstate;
14362 int i;
14363
14364 ASSERT(MUTEX_HELD(&dtrace_lock));
14365
14366 if (help == NULL || gen > help->dthps_generation)
14367 return (EINVAL);
14368
14369 vstate = &help->dthps_vstate;
14370
14371 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14372 dtrace_helper_action_t *last = NULL, *h, *next;
14373
14374 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14375 next = h->dtha_next;
14376
14377 if (h->dtha_generation == gen) {
14378 if (last != NULL) {
14379 last->dtha_next = next;
14380 } else {
14381 help->dthps_actions[i] = next;
14382 }
14383
14384 dtrace_helper_action_destroy(h, vstate);
14385 } else {
14386 last = h;
14387 }
14388 }
14389 }
14390
14391 /*
14392 * Interate until we've cleared out all helper providers with the
14393 * given generation number.
14394 */
14395 for (;;) {
14396 dtrace_helper_provider_t *prov;
14397
14398 /*
14399 * Look for a helper provider with the right generation. We
14400 * have to start back at the beginning of the list each time
14401 * because we drop dtrace_lock. It's unlikely that we'll make
14402 * more than two passes.
14403 */
14404 for (i = 0; i < help->dthps_nprovs; i++) {
14405 prov = help->dthps_provs[i];
14406
14407 if (prov->dthp_generation == gen)
14408 break;
14409 }
14410
14411 /*
14412 * If there were no matches, we're done.
14413 */
14414 if (i == help->dthps_nprovs)
14415 break;
14416
14417 /*
14418 * Move the last helper provider into this slot.
14419 */
14420 help->dthps_nprovs--;
14421 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14422 help->dthps_provs[help->dthps_nprovs] = NULL;
14423
14424 mutex_exit(&dtrace_lock);
14425
14426 /*
14427 * If we have a meta provider, remove this helper provider.
14428 */
14429 mutex_enter(&dtrace_meta_lock);
14430 if (dtrace_meta_pid != NULL) {
14431 ASSERT(dtrace_deferred_pid == NULL);
14432 dtrace_helper_provider_remove(&prov->dthp_prov,
14433 p->p_pid);
14434 }
14435 mutex_exit(&dtrace_meta_lock);
14436
14437 dtrace_helper_provider_destroy(prov);
14438
14439 mutex_enter(&dtrace_lock);
14440 }
14441
14442 return (0);
14443 }
14444
14445 static int
14446 dtrace_helper_validate(dtrace_helper_action_t *helper)
14447 {
14448 int err = 0, i;
14449 dtrace_difo_t *dp;
14450
14451 if ((dp = helper->dtha_predicate) != NULL)
14452 err += dtrace_difo_validate_helper(dp);
14453
14454 for (i = 0; i < helper->dtha_nactions; i++)
14455 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14456
14457 return (err == 0);
14458 }
14459
14460 static int
14461 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14462 {
14463 dtrace_helpers_t *help;
14464 dtrace_helper_action_t *helper, *last;
14465 dtrace_actdesc_t *act;
14466 dtrace_vstate_t *vstate;
14467 dtrace_predicate_t *pred;
14468 int count = 0, nactions = 0, i;
14469
14470 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14471 return (EINVAL);
14472
14473 help = curproc->p_dtrace_helpers;
14474 last = help->dthps_actions[which];
14475 vstate = &help->dthps_vstate;
14476
14477 for (count = 0; last != NULL; last = last->dtha_next) {
14478 count++;
14479 if (last->dtha_next == NULL)
14480 break;
14481 }
14482
14483 /*
14484 * If we already have dtrace_helper_actions_max helper actions for this
14485 * helper action type, we'll refuse to add a new one.
14486 */
14487 if (count >= dtrace_helper_actions_max)
14488 return (ENOSPC);
14489
14490 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14491 helper->dtha_generation = help->dthps_generation;
14492
14493 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14494 ASSERT(pred->dtp_difo != NULL);
14495 dtrace_difo_hold(pred->dtp_difo);
14496 helper->dtha_predicate = pred->dtp_difo;
14497 }
14498
14499 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14500 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14501 goto err;
14502
14503 if (act->dtad_difo == NULL)
14504 goto err;
14505
14506 nactions++;
14507 }
14508
14509 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14510 (helper->dtha_nactions = nactions), KM_SLEEP);
14511
14512 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14513 dtrace_difo_hold(act->dtad_difo);
14514 helper->dtha_actions[i++] = act->dtad_difo;
14515 }
14516
14517 if (!dtrace_helper_validate(helper))
14518 goto err;
14519
14520 if (last == NULL) {
14521 help->dthps_actions[which] = helper;
14522 } else {
14523 last->dtha_next = helper;
14524 }
14525
14526 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14527 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14528 dtrace_helptrace_next = 0;
14529 }
14530
14531 return (0);
14532 err:
14533 dtrace_helper_action_destroy(helper, vstate);
14534 return (EINVAL);
14535 }
14536
14537 static void
14538 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14539 dof_helper_t *dofhp)
14540 {
14541 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14542
14543 mutex_enter(&dtrace_meta_lock);
14544 mutex_enter(&dtrace_lock);
14545
14546 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14547 /*
14548 * If the dtrace module is loaded but not attached, or if
14549 * there aren't isn't a meta provider registered to deal with
14550 * these provider descriptions, we need to postpone creating
14551 * the actual providers until later.
14552 */
14553
14554 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14555 dtrace_deferred_pid != help) {
14556 help->dthps_deferred = 1;
14557 help->dthps_pid = p->p_pid;
14558 help->dthps_next = dtrace_deferred_pid;
14559 help->dthps_prev = NULL;
14560 if (dtrace_deferred_pid != NULL)
14561 dtrace_deferred_pid->dthps_prev = help;
14562 dtrace_deferred_pid = help;
14563 }
14564
14565 mutex_exit(&dtrace_lock);
14566
14567 } else if (dofhp != NULL) {
14568 /*
14569 * If the dtrace module is loaded and we have a particular
14570 * helper provider description, pass that off to the
14571 * meta provider.
14572 */
14573
14574 mutex_exit(&dtrace_lock);
14575
14576 dtrace_helper_provide(dofhp, p->p_pid);
14577
14578 } else {
14579 /*
14580 * Otherwise, just pass all the helper provider descriptions
14581 * off to the meta provider.
14582 */
14583
14584 int i;
14585 mutex_exit(&dtrace_lock);
14586
14587 for (i = 0; i < help->dthps_nprovs; i++) {
14588 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14589 p->p_pid);
14590 }
14591 }
14592
14593 mutex_exit(&dtrace_meta_lock);
14594 }
14595
14596 static int
14597 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14598 {
14599 dtrace_helpers_t *help;
14600 dtrace_helper_provider_t *hprov, **tmp_provs;
14601 uint_t tmp_maxprovs, i;
14602
14603 ASSERT(MUTEX_HELD(&dtrace_lock));
14604
14605 help = curproc->p_dtrace_helpers;
14606 ASSERT(help != NULL);
14607
14608 /*
14609 * If we already have dtrace_helper_providers_max helper providers,
14610 * we're refuse to add a new one.
14611 */
14612 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14613 return (ENOSPC);
14614
14615 /*
14616 * Check to make sure this isn't a duplicate.
14617 */
14618 for (i = 0; i < help->dthps_nprovs; i++) {
14619 if (dofhp->dofhp_addr ==
14620 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14621 return (EALREADY);
14622 }
14623
14624 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14625 hprov->dthp_prov = *dofhp;
14626 hprov->dthp_ref = 1;
14627 hprov->dthp_generation = gen;
14628
14629 /*
14630 * Allocate a bigger table for helper providers if it's already full.
14631 */
14632 if (help->dthps_maxprovs == help->dthps_nprovs) {
14633 tmp_maxprovs = help->dthps_maxprovs;
14634 tmp_provs = help->dthps_provs;
14635
14636 if (help->dthps_maxprovs == 0)
14637 help->dthps_maxprovs = 2;
14638 else
14639 help->dthps_maxprovs *= 2;
14640 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14641 help->dthps_maxprovs = dtrace_helper_providers_max;
14642
14643 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14644
14645 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14646 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14647
14648 if (tmp_provs != NULL) {
14649 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14650 sizeof (dtrace_helper_provider_t *));
14651 kmem_free(tmp_provs, tmp_maxprovs *
14652 sizeof (dtrace_helper_provider_t *));
14653 }
14654 }
14655
14656 help->dthps_provs[help->dthps_nprovs] = hprov;
14657 help->dthps_nprovs++;
14658
14659 return (0);
14660 }
14661
14662 static void
14663 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14664 {
14665 mutex_enter(&dtrace_lock);
14666
14667 if (--hprov->dthp_ref == 0) {
14668 dof_hdr_t *dof;
14669 mutex_exit(&dtrace_lock);
14670 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14671 dtrace_dof_destroy(dof);
14672 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14673 } else {
14674 mutex_exit(&dtrace_lock);
14675 }
14676 }
14677
14678 static int
14679 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14680 {
14681 uintptr_t daddr = (uintptr_t)dof;
14682 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14683 dof_provider_t *provider;
14684 dof_probe_t *probe;
14685 uint8_t *arg;
14686 char *strtab, *typestr;
14687 dof_stridx_t typeidx;
14688 size_t typesz;
14689 uint_t nprobes, j, k;
14690
14691 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14692
14693 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14694 dtrace_dof_error(dof, "misaligned section offset");
14695 return (-1);
14696 }
14697
14698 /*
14699 * The section needs to be large enough to contain the DOF provider
14700 * structure appropriate for the given version.
14701 */
14702 if (sec->dofs_size <
14703 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14704 offsetof(dof_provider_t, dofpv_prenoffs) :
14705 sizeof (dof_provider_t))) {
14706 dtrace_dof_error(dof, "provider section too small");
14707 return (-1);
14708 }
14709
14710 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14711 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14712 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14713 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14714 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14715
14716 if (str_sec == NULL || prb_sec == NULL ||
14717 arg_sec == NULL || off_sec == NULL)
14718 return (-1);
14719
14720 enoff_sec = NULL;
14721
14722 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14723 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14724 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14725 provider->dofpv_prenoffs)) == NULL)
14726 return (-1);
14727
14728 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14729
14730 if (provider->dofpv_name >= str_sec->dofs_size ||
14731 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14732 dtrace_dof_error(dof, "invalid provider name");
14733 return (-1);
14734 }
14735
14736 if (prb_sec->dofs_entsize == 0 ||
14737 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14738 dtrace_dof_error(dof, "invalid entry size");
14739 return (-1);
14740 }
14741
14742 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14743 dtrace_dof_error(dof, "misaligned entry size");
14744 return (-1);
14745 }
14746
14747 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14748 dtrace_dof_error(dof, "invalid entry size");
14749 return (-1);
14750 }
14751
14752 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14753 dtrace_dof_error(dof, "misaligned section offset");
14754 return (-1);
14755 }
14756
14757 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14758 dtrace_dof_error(dof, "invalid entry size");
14759 return (-1);
14760 }
14761
14762 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14763
14764 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14765
14766 /*
14767 * Take a pass through the probes to check for errors.
14768 */
14769 for (j = 0; j < nprobes; j++) {
14770 probe = (dof_probe_t *)(uintptr_t)(daddr +
14771 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14772
14773 if (probe->dofpr_func >= str_sec->dofs_size) {
14774 dtrace_dof_error(dof, "invalid function name");
14775 return (-1);
14776 }
14777
14778 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14779 dtrace_dof_error(dof, "function name too long");
14780 return (-1);
14781 }
14782
14783 if (probe->dofpr_name >= str_sec->dofs_size ||
14784 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14785 dtrace_dof_error(dof, "invalid probe name");
14786 return (-1);
14787 }
14788
14789 /*
14790 * The offset count must not wrap the index, and the offsets
14791 * must also not overflow the section's data.
14792 */
14793 if (probe->dofpr_offidx + probe->dofpr_noffs <
14794 probe->dofpr_offidx ||
14795 (probe->dofpr_offidx + probe->dofpr_noffs) *
14796 off_sec->dofs_entsize > off_sec->dofs_size) {
14797 dtrace_dof_error(dof, "invalid probe offset");
14798 return (-1);
14799 }
14800
14801 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14802 /*
14803 * If there's no is-enabled offset section, make sure
14804 * there aren't any is-enabled offsets. Otherwise
14805 * perform the same checks as for probe offsets
14806 * (immediately above).
14807 */
14808 if (enoff_sec == NULL) {
14809 if (probe->dofpr_enoffidx != 0 ||
14810 probe->dofpr_nenoffs != 0) {
14811 dtrace_dof_error(dof, "is-enabled "
14812 "offsets with null section");
14813 return (-1);
14814 }
14815 } else if (probe->dofpr_enoffidx +
14816 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14817 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14818 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14819 dtrace_dof_error(dof, "invalid is-enabled "
14820 "offset");
14821 return (-1);
14822 }
14823
14824 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14825 dtrace_dof_error(dof, "zero probe and "
14826 "is-enabled offsets");
14827 return (-1);
14828 }
14829 } else if (probe->dofpr_noffs == 0) {
14830 dtrace_dof_error(dof, "zero probe offsets");
14831 return (-1);
14832 }
14833
14834 if (probe->dofpr_argidx + probe->dofpr_xargc <
14835 probe->dofpr_argidx ||
14836 (probe->dofpr_argidx + probe->dofpr_xargc) *
14837 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14838 dtrace_dof_error(dof, "invalid args");
14839 return (-1);
14840 }
14841
14842 typeidx = probe->dofpr_nargv;
14843 typestr = strtab + probe->dofpr_nargv;
14844 for (k = 0; k < probe->dofpr_nargc; k++) {
14845 if (typeidx >= str_sec->dofs_size) {
14846 dtrace_dof_error(dof, "bad "
14847 "native argument type");
14848 return (-1);
14849 }
14850
14851 typesz = strlen(typestr) + 1;
14852 if (typesz > DTRACE_ARGTYPELEN) {
14853 dtrace_dof_error(dof, "native "
14854 "argument type too long");
14855 return (-1);
14856 }
14857 typeidx += typesz;
14858 typestr += typesz;
14859 }
14860
14861 typeidx = probe->dofpr_xargv;
14862 typestr = strtab + probe->dofpr_xargv;
14863 for (k = 0; k < probe->dofpr_xargc; k++) {
14864 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14865 dtrace_dof_error(dof, "bad "
14866 "native argument index");
14867 return (-1);
14868 }
14869
14870 if (typeidx >= str_sec->dofs_size) {
14871 dtrace_dof_error(dof, "bad "
14872 "translated argument type");
14873 return (-1);
14874 }
14875
14876 typesz = strlen(typestr) + 1;
14877 if (typesz > DTRACE_ARGTYPELEN) {
14878 dtrace_dof_error(dof, "translated argument "
14879 "type too long");
14880 return (-1);
14881 }
14882
14883 typeidx += typesz;
14884 typestr += typesz;
14885 }
14886 }
14887
14888 return (0);
14889 }
14890
14891 static int
14892 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14893 {
14894 dtrace_helpers_t *help;
14895 dtrace_vstate_t *vstate;
14896 dtrace_enabling_t *enab = NULL;
14897 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14898 uintptr_t daddr = (uintptr_t)dof;
14899
14900 ASSERT(MUTEX_HELD(&dtrace_lock));
14901
14902 if ((help = curproc->p_dtrace_helpers) == NULL)
14903 help = dtrace_helpers_create(curproc);
14904
14905 vstate = &help->dthps_vstate;
14906
14907 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14908 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14909 dtrace_dof_destroy(dof);
14910 return (rv);
14911 }
14912
14913 /*
14914 * Look for helper providers and validate their descriptions.
14915 */
14916 if (dhp != NULL) {
14917 for (i = 0; i < dof->dofh_secnum; i++) {
14918 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14919 dof->dofh_secoff + i * dof->dofh_secsize);
14920
14921 if (sec->dofs_type != DOF_SECT_PROVIDER)
14922 continue;
14923
14924 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14925 dtrace_enabling_destroy(enab);
14926 dtrace_dof_destroy(dof);
14927 return (-1);
14928 }
14929
14930 nprovs++;
14931 }
14932 }
14933
14934 /*
14935 * Now we need to walk through the ECB descriptions in the enabling.
14936 */
14937 for (i = 0; i < enab->dten_ndesc; i++) {
14938 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14939 dtrace_probedesc_t *desc = &ep->dted_probe;
14940
14941 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14942 continue;
14943
14944 if (strcmp(desc->dtpd_mod, "helper") != 0)
14945 continue;
14946
14947 if (strcmp(desc->dtpd_func, "ustack") != 0)
14948 continue;
14949
14950 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14951 ep)) != 0) {
14952 /*
14953 * Adding this helper action failed -- we are now going
14954 * to rip out the entire generation and return failure.
14955 */
14956 (void) dtrace_helper_destroygen(help->dthps_generation);
14957 dtrace_enabling_destroy(enab);
14958 dtrace_dof_destroy(dof);
14959 return (-1);
14960 }
14961
14962 nhelpers++;
14963 }
14964
14965 if (nhelpers < enab->dten_ndesc)
14966 dtrace_dof_error(dof, "unmatched helpers");
14967
14968 gen = help->dthps_generation++;
14969 dtrace_enabling_destroy(enab);
14970
14971 if (dhp != NULL && nprovs > 0) {
14972 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14973 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14974 mutex_exit(&dtrace_lock);
14975 dtrace_helper_provider_register(curproc, help, dhp);
14976 mutex_enter(&dtrace_lock);
14977
14978 destroy = 0;
14979 }
14980 }
14981
14982 if (destroy)
14983 dtrace_dof_destroy(dof);
14984
14985 return (gen);
14986 }
14987
14988 static dtrace_helpers_t *
14989 dtrace_helpers_create(proc_t *p)
14990 {
14991 dtrace_helpers_t *help;
14992
14993 ASSERT(MUTEX_HELD(&dtrace_lock));
14994 ASSERT(p->p_dtrace_helpers == NULL);
14995
14996 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14997 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14998 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14999
15000 p->p_dtrace_helpers = help;
15001 dtrace_helpers++;
15002
15003 return (help);
15004 }
15005
15006 static void
15007 dtrace_helpers_destroy(void)
15008 {
15009 dtrace_helpers_t *help;
15010 dtrace_vstate_t *vstate;
15011 proc_t *p = curproc;
15012 int i;
15013
15014 mutex_enter(&dtrace_lock);
15015
15016 ASSERT(p->p_dtrace_helpers != NULL);
15017 ASSERT(dtrace_helpers > 0);
15018
15019 help = p->p_dtrace_helpers;
15020 vstate = &help->dthps_vstate;
15021
15022 /*
15023 * We're now going to lose the help from this process.
15024 */
15025 p->p_dtrace_helpers = NULL;
15026 dtrace_sync();
15027
15028 /*
15029 * Destory the helper actions.
15030 */
15031 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15032 dtrace_helper_action_t *h, *next;
15033
15034 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15035 next = h->dtha_next;
15036 dtrace_helper_action_destroy(h, vstate);
15037 h = next;
15038 }
15039 }
15040
15041 mutex_exit(&dtrace_lock);
15042
15043 /*
15044 * Destroy the helper providers.
15045 */
15046 if (help->dthps_maxprovs > 0) {
15047 mutex_enter(&dtrace_meta_lock);
15048 if (dtrace_meta_pid != NULL) {
15049 ASSERT(dtrace_deferred_pid == NULL);
15050
15051 for (i = 0; i < help->dthps_nprovs; i++) {
15052 dtrace_helper_provider_remove(
15053 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15054 }
15055 } else {
15056 mutex_enter(&dtrace_lock);
15057 ASSERT(help->dthps_deferred == 0 ||
15058 help->dthps_next != NULL ||
15059 help->dthps_prev != NULL ||
15060 help == dtrace_deferred_pid);
15061
15062 /*
15063 * Remove the helper from the deferred list.
15064 */
15065 if (help->dthps_next != NULL)
15066 help->dthps_next->dthps_prev = help->dthps_prev;
15067 if (help->dthps_prev != NULL)
15068 help->dthps_prev->dthps_next = help->dthps_next;
15069 if (dtrace_deferred_pid == help) {
15070 dtrace_deferred_pid = help->dthps_next;
15071 ASSERT(help->dthps_prev == NULL);
15072 }
15073
15074 mutex_exit(&dtrace_lock);
15075 }
15076
15077 mutex_exit(&dtrace_meta_lock);
15078
15079 for (i = 0; i < help->dthps_nprovs; i++) {
15080 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15081 }
15082
15083 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15084 sizeof (dtrace_helper_provider_t *));
15085 }
15086
15087 mutex_enter(&dtrace_lock);
15088
15089 dtrace_vstate_fini(&help->dthps_vstate);
15090 kmem_free(help->dthps_actions,
15091 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15092 kmem_free(help, sizeof (dtrace_helpers_t));
15093
15094 --dtrace_helpers;
15095 mutex_exit(&dtrace_lock);
15096 }
15097
15098 static void
15099 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15100 {
15101 dtrace_helpers_t *help, *newhelp;
15102 dtrace_helper_action_t *helper, *new, *last;
15103 dtrace_difo_t *dp;
15104 dtrace_vstate_t *vstate;
15105 int i, j, sz, hasprovs = 0;
15106
15107 mutex_enter(&dtrace_lock);
15108 ASSERT(from->p_dtrace_helpers != NULL);
15109 ASSERT(dtrace_helpers > 0);
15110
15111 help = from->p_dtrace_helpers;
15112 newhelp = dtrace_helpers_create(to);
15113 ASSERT(to->p_dtrace_helpers != NULL);
15114
15115 newhelp->dthps_generation = help->dthps_generation;
15116 vstate = &newhelp->dthps_vstate;
15117
15118 /*
15119 * Duplicate the helper actions.
15120 */
15121 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15122 if ((helper = help->dthps_actions[i]) == NULL)
15123 continue;
15124
15125 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15126 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15127 KM_SLEEP);
15128 new->dtha_generation = helper->dtha_generation;
15129
15130 if ((dp = helper->dtha_predicate) != NULL) {
15131 dp = dtrace_difo_duplicate(dp, vstate);
15132 new->dtha_predicate = dp;
15133 }
15134
15135 new->dtha_nactions = helper->dtha_nactions;
15136 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15137 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15138
15139 for (j = 0; j < new->dtha_nactions; j++) {
15140 dtrace_difo_t *dp = helper->dtha_actions[j];
15141
15142 ASSERT(dp != NULL);
15143 dp = dtrace_difo_duplicate(dp, vstate);
15144 new->dtha_actions[j] = dp;
15145 }
15146
15147 if (last != NULL) {
15148 last->dtha_next = new;
15149 } else {
15150 newhelp->dthps_actions[i] = new;
15151 }
15152
15153 last = new;
15154 }
15155 }
15156
15157 /*
15158 * Duplicate the helper providers and register them with the
15159 * DTrace framework.
15160 */
15161 if (help->dthps_nprovs > 0) {
15162 newhelp->dthps_nprovs = help->dthps_nprovs;
15163 newhelp->dthps_maxprovs = help->dthps_nprovs;
15164 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15165 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15166 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15167 newhelp->dthps_provs[i] = help->dthps_provs[i];
15168 newhelp->dthps_provs[i]->dthp_ref++;
15169 }
15170
15171 hasprovs = 1;
15172 }
15173
15174 mutex_exit(&dtrace_lock);
15175
15176 if (hasprovs)
15177 dtrace_helper_provider_register(to, newhelp, NULL);
15178 }
15179
15180 /*
15181 * DTrace Hook Functions
15182 */
15183 static void
15184 dtrace_module_loaded(struct modctl *ctl)
15185 {
15186 dtrace_provider_t *prv;
15187
15188 mutex_enter(&dtrace_provider_lock);
15189 mutex_enter(&mod_lock);
15190
15191 ASSERT(ctl->mod_busy);
15192
15193 /*
15194 * We're going to call each providers per-module provide operation
15195 * specifying only this module.
15196 */
15197 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15198 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15199
15200 mutex_exit(&mod_lock);
15201 mutex_exit(&dtrace_provider_lock);
15202
15203 /*
15204 * If we have any retained enablings, we need to match against them.
15205 * Enabling probes requires that cpu_lock be held, and we cannot hold
15206 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15207 * module. (In particular, this happens when loading scheduling
15208 * classes.) So if we have any retained enablings, we need to dispatch
15209 * our task queue to do the match for us.
15210 */
15211 mutex_enter(&dtrace_lock);
15212
15213 if (dtrace_retained == NULL) {
15214 mutex_exit(&dtrace_lock);
15215 return;
15216 }
15217
15218 (void) taskq_dispatch(dtrace_taskq,
15219 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15220
15221 mutex_exit(&dtrace_lock);
15222
15223 /*
15224 * And now, for a little heuristic sleaze: in general, we want to
15225 * match modules as soon as they load. However, we cannot guarantee
15226 * this, because it would lead us to the lock ordering violation
15227 * outlined above. The common case, of course, is that cpu_lock is
15228 * _not_ held -- so we delay here for a clock tick, hoping that that's
15229 * long enough for the task queue to do its work. If it's not, it's
15230 * not a serious problem -- it just means that the module that we
15231 * just loaded may not be immediately instrumentable.
15232 */
15233 delay(1);
15234 }
15235
15236 static void
15237 dtrace_module_unloaded(struct modctl *ctl)
15238 {
15239 dtrace_probe_t template, *probe, *first, *next;
15240 dtrace_provider_t *prov;
15241
15242 template.dtpr_mod = ctl->mod_modname;
15243
15244 mutex_enter(&dtrace_provider_lock);
15245 mutex_enter(&mod_lock);
15246 mutex_enter(&dtrace_lock);
15247
15248 if (dtrace_bymod == NULL) {
15249 /*
15250 * The DTrace module is loaded (obviously) but not attached;
15251 * we don't have any work to do.
15252 */
15253 mutex_exit(&dtrace_provider_lock);
15254 mutex_exit(&mod_lock);
15255 mutex_exit(&dtrace_lock);
15256 return;
15257 }
15258
15259 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15260 probe != NULL; probe = probe->dtpr_nextmod) {
15261 if (probe->dtpr_ecb != NULL) {
15262 mutex_exit(&dtrace_provider_lock);
15263 mutex_exit(&mod_lock);
15264 mutex_exit(&dtrace_lock);
15265
15266 /*
15267 * This shouldn't _actually_ be possible -- we're
15268 * unloading a module that has an enabled probe in it.
15269 * (It's normally up to the provider to make sure that
15270 * this can't happen.) However, because dtps_enable()
15271 * doesn't have a failure mode, there can be an
15272 * enable/unload race. Upshot: we don't want to
15273 * assert, but we're not going to disable the
15274 * probe, either.
15275 */
15276 if (dtrace_err_verbose) {
15277 cmn_err(CE_WARN, "unloaded module '%s' had "
15278 "enabled probes", ctl->mod_modname);
15279 }
15280
15281 return;
15282 }
15283 }
15284
15285 probe = first;
15286
15287 for (first = NULL; probe != NULL; probe = next) {
15288 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15289
15290 dtrace_probes[probe->dtpr_id - 1] = NULL;
15291
15292 next = probe->dtpr_nextmod;
15293 dtrace_hash_remove(dtrace_bymod, probe);
15294 dtrace_hash_remove(dtrace_byfunc, probe);
15295 dtrace_hash_remove(dtrace_byname, probe);
15296
15297 if (first == NULL) {
15298 first = probe;
15299 probe->dtpr_nextmod = NULL;
15300 } else {
15301 probe->dtpr_nextmod = first;
15302 first = probe;
15303 }
15304 }
15305
15306 /*
15307 * We've removed all of the module's probes from the hash chains and
15308 * from the probe array. Now issue a dtrace_sync() to be sure that
15309 * everyone has cleared out from any probe array processing.
15310 */
15311 dtrace_sync();
15312
15313 for (probe = first; probe != NULL; probe = first) {
15314 first = probe->dtpr_nextmod;
15315 prov = probe->dtpr_provider;
15316 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15317 probe->dtpr_arg);
15318 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15319 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15320 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15321 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15322 kmem_free(probe, sizeof (dtrace_probe_t));
15323 }
15324
15325 mutex_exit(&dtrace_lock);
15326 mutex_exit(&mod_lock);
15327 mutex_exit(&dtrace_provider_lock);
15328 }
15329
15330 void
15331 dtrace_suspend(void)
15332 {
15333 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15334 }
15335
15336 void
15337 dtrace_resume(void)
15338 {
15339 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15340 }
15341
15342 static int
15343 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15344 {
15345 ASSERT(MUTEX_HELD(&cpu_lock));
15346 mutex_enter(&dtrace_lock);
15347
15348 switch (what) {
15349 case CPU_CONFIG: {
15350 dtrace_state_t *state;
15351 dtrace_optval_t *opt, rs, c;
15352
15353 /*
15354 * For now, we only allocate a new buffer for anonymous state.
15355 */
15356 if ((state = dtrace_anon.dta_state) == NULL)
15357 break;
15358
15359 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15360 break;
15361
15362 opt = state->dts_options;
15363 c = opt[DTRACEOPT_CPU];
15364
15365 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15366 break;
15367
15368 /*
15369 * Regardless of what the actual policy is, we're going to
15370 * temporarily set our resize policy to be manual. We're
15371 * also going to temporarily set our CPU option to denote
15372 * the newly configured CPU.
15373 */
15374 rs = opt[DTRACEOPT_BUFRESIZE];
15375 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15376 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15377
15378 (void) dtrace_state_buffers(state);
15379
15380 opt[DTRACEOPT_BUFRESIZE] = rs;
15381 opt[DTRACEOPT_CPU] = c;
15382
15383 break;
15384 }
15385
15386 case CPU_UNCONFIG:
15387 /*
15388 * We don't free the buffer in the CPU_UNCONFIG case. (The
15389 * buffer will be freed when the consumer exits.)
15390 */
15391 break;
15392
15393 default:
15394 break;
15395 }
15396
15397 mutex_exit(&dtrace_lock);
15398 return (0);
15399 }
15400
15401 static void
15402 dtrace_cpu_setup_initial(processorid_t cpu)
15403 {
15404 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15405 }
15406
15407 static void
15408 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15409 {
15410 if (dtrace_toxranges >= dtrace_toxranges_max) {
15411 int osize, nsize;
15412 dtrace_toxrange_t *range;
15413
15414 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15415
15416 if (osize == 0) {
15417 ASSERT(dtrace_toxrange == NULL);
15418 ASSERT(dtrace_toxranges_max == 0);
15419 dtrace_toxranges_max = 1;
15420 } else {
15421 dtrace_toxranges_max <<= 1;
15422 }
15423
15424 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15425 range = kmem_zalloc(nsize, KM_SLEEP);
15426
15427 if (dtrace_toxrange != NULL) {
15428 ASSERT(osize != 0);
15429 bcopy(dtrace_toxrange, range, osize);
15430 kmem_free(dtrace_toxrange, osize);
15431 }
15432
15433 dtrace_toxrange = range;
15434 }
15435
15436 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
15437 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
15438
15439 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15440 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15441 dtrace_toxranges++;
15442 }
15443
15444 static void
15445 dtrace_getf_barrier()
15446 {
15447 /*
15448 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
15449 * that contain calls to getf(), this routine will be called on every
15450 * closef() before either the underlying vnode is released or the
15451 * file_t itself is freed. By the time we are here, it is essential
15452 * that the file_t can no longer be accessed from a call to getf()
15453 * in probe context -- that assures that a dtrace_sync() can be used
15454 * to clear out any enablings referring to the old structures.
15455 */
15456 if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
15457 kcred->cr_zone->zone_dtrace_getf != 0)
15458 dtrace_sync();
15459 }
15460
15461 /*
15462 * DTrace Driver Cookbook Functions
15463 */
15464 /*ARGSUSED*/
15465 static int
15466 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15467 {
15468 dtrace_provider_id_t id;
15469 dtrace_state_t *state = NULL;
15470 dtrace_enabling_t *enab;
15471
15472 mutex_enter(&cpu_lock);
15473 mutex_enter(&dtrace_provider_lock);
15474 mutex_enter(&dtrace_lock);
15475
15476 if (ddi_soft_state_init(&dtrace_softstate,
15477 sizeof (dtrace_state_t), 0) != 0) {
15478 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15479 mutex_exit(&cpu_lock);
15480 mutex_exit(&dtrace_provider_lock);
15481 mutex_exit(&dtrace_lock);
15482 return (DDI_FAILURE);
15483 }
15484
15485 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15486 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15487 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15488 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15489 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15490 ddi_remove_minor_node(devi, NULL);
15491 ddi_soft_state_fini(&dtrace_softstate);
15492 mutex_exit(&cpu_lock);
15493 mutex_exit(&dtrace_provider_lock);
15494 mutex_exit(&dtrace_lock);
15495 return (DDI_FAILURE);
15496 }
15497
15498 ddi_report_dev(devi);
15499 dtrace_devi = devi;
15500
15501 dtrace_modload = dtrace_module_loaded;
15502 dtrace_modunload = dtrace_module_unloaded;
15503 dtrace_cpu_init = dtrace_cpu_setup_initial;
15504 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15505 dtrace_helpers_fork = dtrace_helpers_duplicate;
15506 dtrace_cpustart_init = dtrace_suspend;
15507 dtrace_cpustart_fini = dtrace_resume;
15508 dtrace_debugger_init = dtrace_suspend;
15509 dtrace_debugger_fini = dtrace_resume;
15510
15511 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15512
15513 ASSERT(MUTEX_HELD(&cpu_lock));
15514
15515 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15516 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15517 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15518 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15519 VM_SLEEP | VMC_IDENTIFIER);
15520 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15521 1, INT_MAX, 0);
15522
15523 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15524 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15525 NULL, NULL, NULL, NULL, NULL, 0);
15526
15527 ASSERT(MUTEX_HELD(&cpu_lock));
15528 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15529 offsetof(dtrace_probe_t, dtpr_nextmod),
15530 offsetof(dtrace_probe_t, dtpr_prevmod));
15531
15532 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15533 offsetof(dtrace_probe_t, dtpr_nextfunc),
15534 offsetof(dtrace_probe_t, dtpr_prevfunc));
15535
15536 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15537 offsetof(dtrace_probe_t, dtpr_nextname),
15538 offsetof(dtrace_probe_t, dtpr_prevname));
15539
15540 if (dtrace_retain_max < 1) {
15541 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15542 "setting to 1", dtrace_retain_max);
15543 dtrace_retain_max = 1;
15544 }
15545
15546 /*
15547 * Now discover our toxic ranges.
15548 */
15549 dtrace_toxic_ranges(dtrace_toxrange_add);
15550
15551 /*
15552 * Before we register ourselves as a provider to our own framework,
15553 * we would like to assert that dtrace_provider is NULL -- but that's
15554 * not true if we were loaded as a dependency of a DTrace provider.
15555 * Once we've registered, we can assert that dtrace_provider is our
15556 * pseudo provider.
15557 */
15558 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15559 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15560
15561 ASSERT(dtrace_provider != NULL);
15562 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15563
15564 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15565 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15566 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15567 dtrace_provider, NULL, NULL, "END", 0, NULL);
15568 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15569 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15570
15571 dtrace_anon_property();
15572 mutex_exit(&cpu_lock);
15573
15574 /*
15575 * If there are already providers, we must ask them to provide their
15576 * probes, and then match any anonymous enabling against them. Note
15577 * that there should be no other retained enablings at this time:
15578 * the only retained enablings at this time should be the anonymous
15579 * enabling.
15580 */
15581 if (dtrace_anon.dta_enabling != NULL) {
15582 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15583
15584 dtrace_enabling_provide(NULL);
15585 state = dtrace_anon.dta_state;
15586
15587 /*
15588 * We couldn't hold cpu_lock across the above call to
15589 * dtrace_enabling_provide(), but we must hold it to actually
15590 * enable the probes. We have to drop all of our locks, pick
15591 * up cpu_lock, and regain our locks before matching the
15592 * retained anonymous enabling.
15593 */
15594 mutex_exit(&dtrace_lock);
15595 mutex_exit(&dtrace_provider_lock);
15596
15597 mutex_enter(&cpu_lock);
15598 mutex_enter(&dtrace_provider_lock);
15599 mutex_enter(&dtrace_lock);
15600
15601 if ((enab = dtrace_anon.dta_enabling) != NULL)
15602 (void) dtrace_enabling_match(enab, NULL);
15603
15604 mutex_exit(&cpu_lock);
15605 }
15606
15607 mutex_exit(&dtrace_lock);
15608 mutex_exit(&dtrace_provider_lock);
15609
15610 if (state != NULL) {
15611 /*
15612 * If we created any anonymous state, set it going now.
15613 */
15614 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15615 }
15616
15617 return (DDI_SUCCESS);
15618 }
15619
15620 /*ARGSUSED*/
15621 static int
15622 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15623 {
15624 dtrace_state_t *state;
15625 uint32_t priv;
15626 uid_t uid;
15627 zoneid_t zoneid;
15628
15629 if (getminor(*devp) == DTRACEMNRN_HELPER)
15630 return (0);
15631
15632 /*
15633 * If this wasn't an open with the "helper" minor, then it must be
15634 * the "dtrace" minor.
15635 */
15636 if (getminor(*devp) != DTRACEMNRN_DTRACE)
15637 return (ENXIO);
15638
15639 /*
15640 * If no DTRACE_PRIV_* bits are set in the credential, then the
15641 * caller lacks sufficient permission to do anything with DTrace.
15642 */
15643 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15644 if (priv == DTRACE_PRIV_NONE)
15645 return (EACCES);
15646
15647 /*
15648 * Ask all providers to provide all their probes.
15649 */
15650 mutex_enter(&dtrace_provider_lock);
15651 dtrace_probe_provide(NULL, NULL);
15652 mutex_exit(&dtrace_provider_lock);
15653
15654 mutex_enter(&cpu_lock);
15655 mutex_enter(&dtrace_lock);
15656 dtrace_opens++;
15657 dtrace_membar_producer();
15658
15659 /*
15660 * If the kernel debugger is active (that is, if the kernel debugger
15661 * modified text in some way), we won't allow the open.
15662 */
15663 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15664 dtrace_opens--;
15665 mutex_exit(&cpu_lock);
15666 mutex_exit(&dtrace_lock);
15667 return (EBUSY);
15668 }
15669
15670 if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
15671 /*
15672 * If DTrace helper tracing is enabled, we need to allocate the
15673 * trace buffer and initialize the values.
15674 */
15675 dtrace_helptrace_buffer =
15676 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15677 dtrace_helptrace_next = 0;
15678 dtrace_helptrace_wrapped = 0;
15679 dtrace_helptrace_enable = 0;
15680 }
15681
15682 state = dtrace_state_create(devp, cred_p);
15683 mutex_exit(&cpu_lock);
15684
15685 if (state == NULL) {
15686 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15687 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15688 mutex_exit(&dtrace_lock);
15689 return (EAGAIN);
15690 }
15691
15692 mutex_exit(&dtrace_lock);
15693
15694 return (0);
15695 }
15696
15697 /*ARGSUSED*/
15698 static int
15699 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15700 {
15701 minor_t minor = getminor(dev);
15702 dtrace_state_t *state;
15703 dtrace_helptrace_t *buf = NULL;
15704
15705 if (minor == DTRACEMNRN_HELPER)
15706 return (0);
15707
15708 state = ddi_get_soft_state(dtrace_softstate, minor);
15709
15710 mutex_enter(&cpu_lock);
15711 mutex_enter(&dtrace_lock);
15712
15713 if (state->dts_anon) {
15714 /*
15715 * There is anonymous state. Destroy that first.
15716 */
15717 ASSERT(dtrace_anon.dta_state == NULL);
15718 dtrace_state_destroy(state->dts_anon);
15719 }
15720
15721 if (dtrace_helptrace_disable) {
15722 /*
15723 * If we have been told to disable helper tracing, set the
15724 * buffer to NULL before calling into dtrace_state_destroy();
15725 * we take advantage of its dtrace_sync() to know that no
15726 * CPU is in probe context with enabled helper tracing
15727 * after it returns.
15728 */
15729 buf = dtrace_helptrace_buffer;
15730 dtrace_helptrace_buffer = NULL;
15731 }
15732
15733 dtrace_state_destroy(state);
15734 ASSERT(dtrace_opens > 0);
15735
15736 /*
15737 * Only relinquish control of the kernel debugger interface when there
15738 * are no consumers and no anonymous enablings.
15739 */
15740 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15741 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15742
15743 if (buf != NULL) {
15744 kmem_free(buf, dtrace_helptrace_bufsize);
15745 dtrace_helptrace_disable = 0;
15746 }
15747
15748 mutex_exit(&dtrace_lock);
15749 mutex_exit(&cpu_lock);
15750
15751 return (0);
15752 }
15753
15754 /*ARGSUSED*/
15755 static int
15756 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15757 {
15758 int rval;
15759 dof_helper_t help, *dhp = NULL;
15760
15761 switch (cmd) {
15762 case DTRACEHIOC_ADDDOF:
15763 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15764 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15765 return (EFAULT);
15766 }
15767
15768 dhp = &help;
15769 arg = (intptr_t)help.dofhp_dof;
15770 /*FALLTHROUGH*/
15771
15772 case DTRACEHIOC_ADD: {
15773 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15774
15775 if (dof == NULL)
15776 return (rval);
15777
15778 mutex_enter(&dtrace_lock);
15779
15780 /*
15781 * dtrace_helper_slurp() takes responsibility for the dof --
15782 * it may free it now or it may save it and free it later.
15783 */
15784 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15785 *rv = rval;
15786 rval = 0;
15787 } else {
15788 rval = EINVAL;
15789 }
15790
15791 mutex_exit(&dtrace_lock);
15792 return (rval);
15793 }
15794
15795 case DTRACEHIOC_REMOVE: {
15796 mutex_enter(&dtrace_lock);
15797 rval = dtrace_helper_destroygen(arg);
15798 mutex_exit(&dtrace_lock);
15799
15800 return (rval);
15801 }
15802
15803 default:
15804 break;
15805 }
15806
15807 return (ENOTTY);
15808 }
15809
15810 /*ARGSUSED*/
15811 static int
15812 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15813 {
15814 minor_t minor = getminor(dev);
15815 dtrace_state_t *state;
15816 int rval;
15817
15818 if (minor == DTRACEMNRN_HELPER)
15819 return (dtrace_ioctl_helper(cmd, arg, rv));
15820
15821 state = ddi_get_soft_state(dtrace_softstate, minor);
15822
15823 if (state->dts_anon) {
15824 ASSERT(dtrace_anon.dta_state == NULL);
15825 state = state->dts_anon;
15826 }
15827
15828 switch (cmd) {
15829 case DTRACEIOC_PROVIDER: {
15830 dtrace_providerdesc_t pvd;
15831 dtrace_provider_t *pvp;
15832
15833 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15834 return (EFAULT);
15835
15836 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15837 mutex_enter(&dtrace_provider_lock);
15838
15839 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15840 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15841 break;
15842 }
15843
15844 mutex_exit(&dtrace_provider_lock);
15845
15846 if (pvp == NULL)
15847 return (ESRCH);
15848
15849 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15850 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15851 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15852 return (EFAULT);
15853
15854 return (0);
15855 }
15856
15857 case DTRACEIOC_EPROBE: {
15858 dtrace_eprobedesc_t epdesc;
15859 dtrace_ecb_t *ecb;
15860 dtrace_action_t *act;
15861 void *buf;
15862 size_t size;
15863 uintptr_t dest;
15864 int nrecs;
15865
15866 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15867 return (EFAULT);
15868
15869 mutex_enter(&dtrace_lock);
15870
15871 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15872 mutex_exit(&dtrace_lock);
15873 return (EINVAL);
15874 }
15875
15876 if (ecb->dte_probe == NULL) {
15877 mutex_exit(&dtrace_lock);
15878 return (EINVAL);
15879 }
15880
15881 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15882 epdesc.dtepd_uarg = ecb->dte_uarg;
15883 epdesc.dtepd_size = ecb->dte_size;
15884
15885 nrecs = epdesc.dtepd_nrecs;
15886 epdesc.dtepd_nrecs = 0;
15887 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15888 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15889 continue;
15890
15891 epdesc.dtepd_nrecs++;
15892 }
15893
15894 /*
15895 * Now that we have the size, we need to allocate a temporary
15896 * buffer in which to store the complete description. We need
15897 * the temporary buffer to be able to drop dtrace_lock()
15898 * across the copyout(), below.
15899 */
15900 size = sizeof (dtrace_eprobedesc_t) +
15901 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15902
15903 buf = kmem_alloc(size, KM_SLEEP);
15904 dest = (uintptr_t)buf;
15905
15906 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15907 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15908
15909 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15910 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15911 continue;
15912
15913 if (nrecs-- == 0)
15914 break;
15915
15916 bcopy(&act->dta_rec, (void *)dest,
15917 sizeof (dtrace_recdesc_t));
15918 dest += sizeof (dtrace_recdesc_t);
15919 }
15920
15921 mutex_exit(&dtrace_lock);
15922
15923 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15924 kmem_free(buf, size);
15925 return (EFAULT);
15926 }
15927
15928 kmem_free(buf, size);
15929 return (0);
15930 }
15931
15932 case DTRACEIOC_AGGDESC: {
15933 dtrace_aggdesc_t aggdesc;
15934 dtrace_action_t *act;
15935 dtrace_aggregation_t *agg;
15936 int nrecs;
15937 uint32_t offs;
15938 dtrace_recdesc_t *lrec;
15939 void *buf;
15940 size_t size;
15941 uintptr_t dest;
15942
15943 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15944 return (EFAULT);
15945
15946 mutex_enter(&dtrace_lock);
15947
15948 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15949 mutex_exit(&dtrace_lock);
15950 return (EINVAL);
15951 }
15952
15953 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15954
15955 nrecs = aggdesc.dtagd_nrecs;
15956 aggdesc.dtagd_nrecs = 0;
15957
15958 offs = agg->dtag_base;
15959 lrec = &agg->dtag_action.dta_rec;
15960 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15961
15962 for (act = agg->dtag_first; ; act = act->dta_next) {
15963 ASSERT(act->dta_intuple ||
15964 DTRACEACT_ISAGG(act->dta_kind));
15965
15966 /*
15967 * If this action has a record size of zero, it
15968 * denotes an argument to the aggregating action.
15969 * Because the presence of this record doesn't (or
15970 * shouldn't) affect the way the data is interpreted,
15971 * we don't copy it out to save user-level the
15972 * confusion of dealing with a zero-length record.
15973 */
15974 if (act->dta_rec.dtrd_size == 0) {
15975 ASSERT(agg->dtag_hasarg);
15976 continue;
15977 }
15978
15979 aggdesc.dtagd_nrecs++;
15980
15981 if (act == &agg->dtag_action)
15982 break;
15983 }
15984
15985 /*
15986 * Now that we have the size, we need to allocate a temporary
15987 * buffer in which to store the complete description. We need
15988 * the temporary buffer to be able to drop dtrace_lock()
15989 * across the copyout(), below.
15990 */
15991 size = sizeof (dtrace_aggdesc_t) +
15992 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15993
15994 buf = kmem_alloc(size, KM_SLEEP);
15995 dest = (uintptr_t)buf;
15996
15997 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15998 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15999
16000 for (act = agg->dtag_first; ; act = act->dta_next) {
16001 dtrace_recdesc_t rec = act->dta_rec;
16002
16003 /*
16004 * See the comment in the above loop for why we pass
16005 * over zero-length records.
16006 */
16007 if (rec.dtrd_size == 0) {
16008 ASSERT(agg->dtag_hasarg);
16009 continue;
16010 }
16011
16012 if (nrecs-- == 0)
16013 break;
16014
16015 rec.dtrd_offset -= offs;
16016 bcopy(&rec, (void *)dest, sizeof (rec));
16017 dest += sizeof (dtrace_recdesc_t);
16018
16019 if (act == &agg->dtag_action)
16020 break;
16021 }
16022
16023 mutex_exit(&dtrace_lock);
16024
16025 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16026 kmem_free(buf, size);
16027 return (EFAULT);
16028 }
16029
16030 kmem_free(buf, size);
16031 return (0);
16032 }
16033
16034 case DTRACEIOC_ENABLE: {
16035 dof_hdr_t *dof;
16036 dtrace_enabling_t *enab = NULL;
16037 dtrace_vstate_t *vstate;
16038 int err = 0;
16039
16040 *rv = 0;
16041
16042 /*
16043 * If a NULL argument has been passed, we take this as our
16044 * cue to reevaluate our enablings.
16045 */
16046 if (arg == NULL) {
16047 dtrace_enabling_matchall();
16048
16049 return (0);
16050 }
16051
16052 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16053 return (rval);
16054
16055 mutex_enter(&cpu_lock);
16056 mutex_enter(&dtrace_lock);
16057 vstate = &state->dts_vstate;
16058
16059 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16060 mutex_exit(&dtrace_lock);
16061 mutex_exit(&cpu_lock);
16062 dtrace_dof_destroy(dof);
16063 return (EBUSY);
16064 }
16065
16066 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16067 mutex_exit(&dtrace_lock);
16068 mutex_exit(&cpu_lock);
16069 dtrace_dof_destroy(dof);
16070 return (EINVAL);
16071 }
16072
16073 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16074 dtrace_enabling_destroy(enab);
16075 mutex_exit(&dtrace_lock);
16076 mutex_exit(&cpu_lock);
16077 dtrace_dof_destroy(dof);
16078 return (rval);
16079 }
16080
16081 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16082 err = dtrace_enabling_retain(enab);
16083 } else {
16084 dtrace_enabling_destroy(enab);
16085 }
16086
16087 mutex_exit(&cpu_lock);
16088 mutex_exit(&dtrace_lock);
16089 dtrace_dof_destroy(dof);
16090
16091 return (err);
16092 }
16093
16094 case DTRACEIOC_REPLICATE: {
16095 dtrace_repldesc_t desc;
16096 dtrace_probedesc_t *match = &desc.dtrpd_match;
16097 dtrace_probedesc_t *create = &desc.dtrpd_create;
16098 int err;
16099
16100 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16101 return (EFAULT);
16102
16103 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16104 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16105 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16106 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16107
16108 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16109 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16110 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16111 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16112
16113 mutex_enter(&dtrace_lock);
16114 err = dtrace_enabling_replicate(state, match, create);
16115 mutex_exit(&dtrace_lock);
16116
16117 return (err);
16118 }
16119
16120 case DTRACEIOC_PROBEMATCH:
16121 case DTRACEIOC_PROBES: {
16122 dtrace_probe_t *probe = NULL;
16123 dtrace_probedesc_t desc;
16124 dtrace_probekey_t pkey;
16125 dtrace_id_t i;
16126 int m = 0;
16127 uint32_t priv;
16128 uid_t uid;
16129 zoneid_t zoneid;
16130
16131 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16132 return (EFAULT);
16133
16134 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16135 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16136 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16137 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16138
16139 /*
16140 * Before we attempt to match this probe, we want to give
16141 * all providers the opportunity to provide it.
16142 */
16143 if (desc.dtpd_id == DTRACE_IDNONE) {
16144 mutex_enter(&dtrace_provider_lock);
16145 dtrace_probe_provide(&desc, NULL);
16146 mutex_exit(&dtrace_provider_lock);
16147 desc.dtpd_id++;
16148 }
16149
16150 if (cmd == DTRACEIOC_PROBEMATCH) {
16151 dtrace_probekey(&desc, &pkey);
16152 pkey.dtpk_id = DTRACE_IDNONE;
16153 }
16154
16155 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16156
16157 mutex_enter(&dtrace_lock);
16158
16159 if (cmd == DTRACEIOC_PROBEMATCH) {
16160 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16161 if ((probe = dtrace_probes[i - 1]) != NULL &&
16162 (m = dtrace_match_probe(probe, &pkey,
16163 priv, uid, zoneid)) != 0)
16164 break;
16165 }
16166
16167 if (m < 0) {
16168 mutex_exit(&dtrace_lock);
16169 return (EINVAL);
16170 }
16171
16172 } else {
16173 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16174 if ((probe = dtrace_probes[i - 1]) != NULL &&
16175 dtrace_match_priv(probe, priv, uid, zoneid))
16176 break;
16177 }
16178 }
16179
16180 if (probe == NULL) {
16181 mutex_exit(&dtrace_lock);
16182 return (ESRCH);
16183 }
16184
16185 dtrace_probe_description(probe, &desc);
16186 mutex_exit(&dtrace_lock);
16187
16188 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16189 return (EFAULT);
16190
16191 return (0);
16192 }
16193
16194 case DTRACEIOC_PROBEARG: {
16195 dtrace_argdesc_t desc;
16196 dtrace_probe_t *probe;
16197 dtrace_provider_t *prov;
16198
16199 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16200 return (EFAULT);
16201
16202 if (desc.dtargd_id == DTRACE_IDNONE)
16203 return (EINVAL);
16204
16205 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16206 return (EINVAL);
16207
16208 mutex_enter(&dtrace_provider_lock);
16209 mutex_enter(&mod_lock);
16210 mutex_enter(&dtrace_lock);
16211
16212 if (desc.dtargd_id > dtrace_nprobes) {
16213 mutex_exit(&dtrace_lock);
16214 mutex_exit(&mod_lock);
16215 mutex_exit(&dtrace_provider_lock);
16216 return (EINVAL);
16217 }
16218
16219 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16220 mutex_exit(&dtrace_lock);
16221 mutex_exit(&mod_lock);
16222 mutex_exit(&dtrace_provider_lock);
16223 return (EINVAL);
16224 }
16225
16226 mutex_exit(&dtrace_lock);
16227
16228 prov = probe->dtpr_provider;
16229
16230 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16231 /*
16232 * There isn't any typed information for this probe.
16233 * Set the argument number to DTRACE_ARGNONE.
16234 */
16235 desc.dtargd_ndx = DTRACE_ARGNONE;
16236 } else {
16237 desc.dtargd_native[0] = '\0';
16238 desc.dtargd_xlate[0] = '\0';
16239 desc.dtargd_mapping = desc.dtargd_ndx;
16240
16241 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16242 probe->dtpr_id, probe->dtpr_arg, &desc);
16243 }
16244
16245 mutex_exit(&mod_lock);
16246 mutex_exit(&dtrace_provider_lock);
16247
16248 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16249 return (EFAULT);
16250
16251 return (0);
16252 }
16253
16254 case DTRACEIOC_GO: {
16255 processorid_t cpuid;
16256 rval = dtrace_state_go(state, &cpuid);
16257
16258 if (rval != 0)
16259 return (rval);
16260
16261 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16262 return (EFAULT);
16263
16264 return (0);
16265 }
16266
16267 case DTRACEIOC_STOP: {
16268 processorid_t cpuid;
16269
16270 mutex_enter(&dtrace_lock);
16271 rval = dtrace_state_stop(state, &cpuid);
16272 mutex_exit(&dtrace_lock);
16273
16274 if (rval != 0)
16275 return (rval);
16276
16277 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16278 return (EFAULT);
16279
16280 return (0);
16281 }
16282
16283 case DTRACEIOC_DOFGET: {
16284 dof_hdr_t hdr, *dof;
16285 uint64_t len;
16286
16287 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16288 return (EFAULT);
16289
16290 mutex_enter(&dtrace_lock);
16291 dof = dtrace_dof_create(state);
16292 mutex_exit(&dtrace_lock);
16293
16294 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16295 rval = copyout(dof, (void *)arg, len);
16296 dtrace_dof_destroy(dof);
16297
16298 return (rval == 0 ? 0 : EFAULT);
16299 }
16300
16301 case DTRACEIOC_AGGSNAP:
16302 case DTRACEIOC_BUFSNAP: {
16303 dtrace_bufdesc_t desc;
16304 caddr_t cached;
16305 dtrace_buffer_t *buf;
16306
16307 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16308 return (EFAULT);
16309
16310 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16311 return (EINVAL);
16312
16313 mutex_enter(&dtrace_lock);
16314
16315 if (cmd == DTRACEIOC_BUFSNAP) {
16316 buf = &state->dts_buffer[desc.dtbd_cpu];
16317 } else {
16318 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16319 }
16320
16321 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16322 size_t sz = buf->dtb_offset;
16323
16324 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16325 mutex_exit(&dtrace_lock);
16326 return (EBUSY);
16327 }
16328
16329 /*
16330 * If this buffer has already been consumed, we're
16331 * going to indicate that there's nothing left here
16332 * to consume.
16333 */
16334 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16335 mutex_exit(&dtrace_lock);
16336
16337 desc.dtbd_size = 0;
16338 desc.dtbd_drops = 0;
16339 desc.dtbd_errors = 0;
16340 desc.dtbd_oldest = 0;
16341 sz = sizeof (desc);
16342
16343 if (copyout(&desc, (void *)arg, sz) != 0)
16344 return (EFAULT);
16345
16346 return (0);
16347 }
16348
16349 /*
16350 * If this is a ring buffer that has wrapped, we want
16351 * to copy the whole thing out.
16352 */
16353 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16354 dtrace_buffer_polish(buf);
16355 sz = buf->dtb_size;
16356 }
16357
16358 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16359 mutex_exit(&dtrace_lock);
16360 return (EFAULT);
16361 }
16362
16363 desc.dtbd_size = sz;
16364 desc.dtbd_drops = buf->dtb_drops;
16365 desc.dtbd_errors = buf->dtb_errors;
16366 desc.dtbd_oldest = buf->dtb_xamot_offset;
16367 desc.dtbd_timestamp = dtrace_gethrtime();
16368
16369 mutex_exit(&dtrace_lock);
16370
16371 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16372 return (EFAULT);
16373
16374 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16375
16376 return (0);
16377 }
16378
16379 if (buf->dtb_tomax == NULL) {
16380 ASSERT(buf->dtb_xamot == NULL);
16381 mutex_exit(&dtrace_lock);
16382 return (ENOENT);
16383 }
16384
16385 cached = buf->dtb_tomax;
16386 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16387
16388 dtrace_xcall(desc.dtbd_cpu,
16389 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16390
16391 state->dts_errors += buf->dtb_xamot_errors;
16392
16393 /*
16394 * If the buffers did not actually switch, then the cross call
16395 * did not take place -- presumably because the given CPU is
16396 * not in the ready set. If this is the case, we'll return
16397 * ENOENT.
16398 */
16399 if (buf->dtb_tomax == cached) {
16400 ASSERT(buf->dtb_xamot != cached);
16401 mutex_exit(&dtrace_lock);
16402 return (ENOENT);
16403 }
16404
16405 ASSERT(cached == buf->dtb_xamot);
16406
16407 /*
16408 * We have our snapshot; now copy it out.
16409 */
16410 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16411 buf->dtb_xamot_offset) != 0) {
16412 mutex_exit(&dtrace_lock);
16413 return (EFAULT);
16414 }
16415
16416 desc.dtbd_size = buf->dtb_xamot_offset;
16417 desc.dtbd_drops = buf->dtb_xamot_drops;
16418 desc.dtbd_errors = buf->dtb_xamot_errors;
16419 desc.dtbd_oldest = 0;
16420 desc.dtbd_timestamp = buf->dtb_switched;
16421
16422 mutex_exit(&dtrace_lock);
16423
16424 /*
16425 * Finally, copy out the buffer description.
16426 */
16427 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16428 return (EFAULT);
16429
16430 return (0);
16431 }
16432
16433 case DTRACEIOC_CONF: {
16434 dtrace_conf_t conf;
16435
16436 bzero(&conf, sizeof (conf));
16437 conf.dtc_difversion = DIF_VERSION;
16438 conf.dtc_difintregs = DIF_DIR_NREGS;
16439 conf.dtc_diftupregs = DIF_DTR_NREGS;
16440 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16441
16442 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16443 return (EFAULT);
16444
16445 return (0);
16446 }
16447
16448 case DTRACEIOC_STATUS: {
16449 dtrace_status_t stat;
16450 dtrace_dstate_t *dstate;
16451 int i, j;
16452 uint64_t nerrs;
16453
16454 /*
16455 * See the comment in dtrace_state_deadman() for the reason
16456 * for setting dts_laststatus to INT64_MAX before setting
16457 * it to the correct value.
16458 */
16459 state->dts_laststatus = INT64_MAX;
16460 dtrace_membar_producer();
16461 state->dts_laststatus = dtrace_gethrtime();
16462
16463 bzero(&stat, sizeof (stat));
16464
16465 mutex_enter(&dtrace_lock);
16466
16467 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16468 mutex_exit(&dtrace_lock);
16469 return (ENOENT);
16470 }
16471
16472 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16473 stat.dtst_exiting = 1;
16474
16475 nerrs = state->dts_errors;
16476 dstate = &state->dts_vstate.dtvs_dynvars;
16477
16478 for (i = 0; i < NCPU; i++) {
16479 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16480
16481 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16482 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16483 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16484
16485 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16486 stat.dtst_filled++;
16487
16488 nerrs += state->dts_buffer[i].dtb_errors;
16489
16490 for (j = 0; j < state->dts_nspeculations; j++) {
16491 dtrace_speculation_t *spec;
16492 dtrace_buffer_t *buf;
16493
16494 spec = &state->dts_speculations[j];
16495 buf = &spec->dtsp_buffer[i];
16496 stat.dtst_specdrops += buf->dtb_xamot_drops;
16497 }
16498 }
16499
16500 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16501 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16502 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16503 stat.dtst_dblerrors = state->dts_dblerrors;
16504 stat.dtst_killed =
16505 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16506 stat.dtst_errors = nerrs;
16507
16508 mutex_exit(&dtrace_lock);
16509
16510 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16511 return (EFAULT);
16512
16513 return (0);
16514 }
16515
16516 case DTRACEIOC_FORMAT: {
16517 dtrace_fmtdesc_t fmt;
16518 char *str;
16519 int len;
16520
16521 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16522 return (EFAULT);
16523
16524 mutex_enter(&dtrace_lock);
16525
16526 if (fmt.dtfd_format == 0 ||
16527 fmt.dtfd_format > state->dts_nformats) {
16528 mutex_exit(&dtrace_lock);
16529 return (EINVAL);
16530 }
16531
16532 /*
16533 * Format strings are allocated contiguously and they are
16534 * never freed; if a format index is less than the number
16535 * of formats, we can assert that the format map is non-NULL
16536 * and that the format for the specified index is non-NULL.
16537 */
16538 ASSERT(state->dts_formats != NULL);
16539 str = state->dts_formats[fmt.dtfd_format - 1];
16540 ASSERT(str != NULL);
16541
16542 len = strlen(str) + 1;
16543
16544 if (len > fmt.dtfd_length) {
16545 fmt.dtfd_length = len;
16546
16547 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16548 mutex_exit(&dtrace_lock);
16549 return (EINVAL);
16550 }
16551 } else {
16552 if (copyout(str, fmt.dtfd_string, len) != 0) {
16553 mutex_exit(&dtrace_lock);
16554 return (EINVAL);
16555 }
16556 }
16557
16558 mutex_exit(&dtrace_lock);
16559 return (0);
16560 }
16561
16562 default:
16563 break;
16564 }
16565
16566 return (ENOTTY);
16567 }
16568
16569 /*ARGSUSED*/
16570 static int
16571 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16572 {
16573 dtrace_state_t *state;
16574
16575 switch (cmd) {
16576 case DDI_DETACH:
16577 break;
16578
16579 case DDI_SUSPEND:
16580 return (DDI_SUCCESS);
16581
16582 default:
16583 return (DDI_FAILURE);
16584 }
16585
16586 mutex_enter(&cpu_lock);
16587 mutex_enter(&dtrace_provider_lock);
16588 mutex_enter(&dtrace_lock);
16589
16590 ASSERT(dtrace_opens == 0);
16591
16592 if (dtrace_helpers > 0) {
16593 mutex_exit(&dtrace_provider_lock);
16594 mutex_exit(&dtrace_lock);
16595 mutex_exit(&cpu_lock);
16596 return (DDI_FAILURE);
16597 }
16598
16599 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16600 mutex_exit(&dtrace_provider_lock);
16601 mutex_exit(&dtrace_lock);
16602 mutex_exit(&cpu_lock);
16603 return (DDI_FAILURE);
16604 }
16605
16606 dtrace_provider = NULL;
16607
16608 if ((state = dtrace_anon_grab()) != NULL) {
16609 /*
16610 * If there were ECBs on this state, the provider should
16611 * have not been allowed to detach; assert that there is
16612 * none.
16613 */
16614 ASSERT(state->dts_necbs == 0);
16615 dtrace_state_destroy(state);
16616
16617 /*
16618 * If we're being detached with anonymous state, we need to
16619 * indicate to the kernel debugger that DTrace is now inactive.
16620 */
16621 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16622 }
16623
16624 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16625 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16626 dtrace_cpu_init = NULL;
16627 dtrace_helpers_cleanup = NULL;
16628 dtrace_helpers_fork = NULL;
16629 dtrace_cpustart_init = NULL;
16630 dtrace_cpustart_fini = NULL;
16631 dtrace_debugger_init = NULL;
16632 dtrace_debugger_fini = NULL;
16633 dtrace_modload = NULL;
16634 dtrace_modunload = NULL;
16635
16636 ASSERT(dtrace_getf == 0);
16637 ASSERT(dtrace_closef == NULL);
16638
16639 mutex_exit(&cpu_lock);
16640
16641 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16642 dtrace_probes = NULL;
16643 dtrace_nprobes = 0;
16644
16645 dtrace_hash_destroy(dtrace_bymod);
16646 dtrace_hash_destroy(dtrace_byfunc);
16647 dtrace_hash_destroy(dtrace_byname);
16648 dtrace_bymod = NULL;
16649 dtrace_byfunc = NULL;
16650 dtrace_byname = NULL;
16651
16652 kmem_cache_destroy(dtrace_state_cache);
16653 vmem_destroy(dtrace_minor);
16654 vmem_destroy(dtrace_arena);
16655
16656 if (dtrace_toxrange != NULL) {
16657 kmem_free(dtrace_toxrange,
16658 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16659 dtrace_toxrange = NULL;
16660 dtrace_toxranges = 0;
16661 dtrace_toxranges_max = 0;
16662 }
16663
16664 ddi_remove_minor_node(dtrace_devi, NULL);
16665 dtrace_devi = NULL;
16666
16667 ddi_soft_state_fini(&dtrace_softstate);
16668
16669 ASSERT(dtrace_vtime_references == 0);
16670 ASSERT(dtrace_opens == 0);
16671 ASSERT(dtrace_retained == NULL);
16672
16673 mutex_exit(&dtrace_lock);
16674 mutex_exit(&dtrace_provider_lock);
16675
16676 /*
16677 * We don't destroy the task queue until after we have dropped our
16678 * locks (taskq_destroy() may block on running tasks). To prevent
16679 * attempting to do work after we have effectively detached but before
16680 * the task queue has been destroyed, all tasks dispatched via the
16681 * task queue must check that DTrace is still attached before
16682 * performing any operation.
16683 */
16684 taskq_destroy(dtrace_taskq);
16685 dtrace_taskq = NULL;
16686
16687 return (DDI_SUCCESS);
16688 }
16689
16690 /*ARGSUSED*/
16691 static int
16692 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16693 {
16694 int error;
16695
16696 switch (infocmd) {
16697 case DDI_INFO_DEVT2DEVINFO:
16698 *result = (void *)dtrace_devi;
16699 error = DDI_SUCCESS;
16700 break;
16701 case DDI_INFO_DEVT2INSTANCE:
16702 *result = (void *)0;
16703 error = DDI_SUCCESS;
16704 break;
16705 default:
16706 error = DDI_FAILURE;
16707 }
16708 return (error);
16709 }
16710
16711 static struct cb_ops dtrace_cb_ops = {
16712 dtrace_open, /* open */
16713 dtrace_close, /* close */
16714 nulldev, /* strategy */
16715 nulldev, /* print */
16716 nodev, /* dump */
16717 nodev, /* read */
16718 nodev, /* write */
16719 dtrace_ioctl, /* ioctl */
16720 nodev, /* devmap */
16721 nodev, /* mmap */
16722 nodev, /* segmap */
16723 nochpoll, /* poll */
16724 ddi_prop_op, /* cb_prop_op */
16725 0, /* streamtab */
16726 D_NEW | D_MP /* Driver compatibility flag */
16727 };
16728
16729 static struct dev_ops dtrace_ops = {
16730 DEVO_REV, /* devo_rev */
16731 0, /* refcnt */
16732 dtrace_info, /* get_dev_info */
16733 nulldev, /* identify */
16734 nulldev, /* probe */
16735 dtrace_attach, /* attach */
16736 dtrace_detach, /* detach */
16737 nodev, /* reset */
16738 &dtrace_cb_ops, /* driver operations */
16739 NULL, /* bus operations */
16740 nodev, /* dev power */
16741 ddi_quiesce_not_needed, /* quiesce */
16742 };
16743
16744 static struct modldrv modldrv = {
16745 &mod_driverops, /* module type (this is a pseudo driver) */
16746 "Dynamic Tracing", /* name of module */
16747 &dtrace_ops, /* driver ops */
16748 };
16749
16750 static struct modlinkage modlinkage = {
16751 MODREV_1,
16752 (void *)&modldrv,
16753 NULL
16754 };
16755
16756 int
16757 _init(void)
16758 {
16759 return (mod_install(&modlinkage));
16760 }
16761
16762 int
16763 _info(struct modinfo *modinfop)
16764 {
16765 return (mod_info(&modlinkage, modinfop));
16766 }
16767
16768 int
16769 _fini(void)
16770 {
16771 return (mod_remove(&modlinkage));
16772 }