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.
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.
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]
25 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
26 * Use is subject to license terms.
29 #pragma ident "%Z%%M% %I% %E% SMI"
32 * DTrace - Dynamic Tracing for Solaris
34 * This is the implementation of the Solaris Dynamic Tracing framework
35 * (DTrace). The user-visible interface to DTrace is described at length in
36 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
37 * library, the in-kernel DTrace framework, and the DTrace providers are
38 * described in the block comments in the <sys/dtrace.h> header file. The
39 * internal architecture of DTrace is described in the block comments in the
40 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
41 * implementation very much assume mastery of all of these sources; if one has
42 * an unanswered question about the implementation, one should consult them
45 * The functions here are ordered roughly as follows:
47 * - Probe context functions
48 * - Probe hashing functions
49 * - Non-probe context utility functions
50 * - Matching functions
51 * - Provider-to-Framework API functions
52 * - Probe management functions
53 * - DIF object functions
55 * - Predicate functions
58 * - Enabling functions
60 * - Anonymous enabling functions
61 * - Consumer state functions
64 * - Driver cookbook functions
66 * Each group of functions begins with a block comment labelled the "DTrace
67 * [Group] Functions", allowing one to find each block by searching forward
68 * on capital-f functions.
70 #include <sys/errno.h>
75 #include <sys/modctl.h>
77 #include <sys/systm.h>
80 #include <sys/sunddi.h>
82 #include <sys/cpuvar.h>
85 #include <sys/strsubr.h>
87 #include <sys/sysmacros.h>
88 #include <sys/dtrace_impl.h>
89 #include <sys/atomic.h>
90 #include <sys/cmn_err.h>
92 #include <sys/mutex_impl.h>
93 #include <sys/rwlock_impl.h>
95 #include <sys/ctf_api.h>
97 #include <sys/panic.h>
98 #include <sys/priv_impl.h>
100 #include <sys/policy.h>
102 #include <sys/cred_impl.h>
103 #include <sys/procfs_isa.h>
105 #include <sys/taskq.h>
107 #include <sys/mkdev.h>
110 #include <sys/zone.h>
111 #include <sys/socket.h>
112 #include <netinet/in.h>
114 /* FreeBSD includes: */
116 #include <sys/callout.h>
117 #include <sys/ctype.h>
118 #include <sys/limits.h>
120 #include <sys/kernel.h>
121 #include <sys/malloc.h>
122 #include <sys/sysctl.h>
123 #include <sys/lock.h>
124 #include <sys/mutex.h>
125 #include <sys/rwlock.h>
127 #include <sys/dtrace_bsd.h>
128 #include <netinet/in.h>
129 #include "dtrace_cddl.h"
130 #include "dtrace_debug.c"
134 * DTrace Tunable Variables
136 * The following variables may be tuned by adding a line to /etc/system that
137 * includes both the name of the DTrace module ("dtrace") and the name of the
138 * variable. For example:
140 * set dtrace:dtrace_destructive_disallow = 1
142 * In general, the only variables that one should be tuning this way are those
143 * that affect system-wide DTrace behavior, and for which the default behavior
144 * is undesirable. Most of these variables are tunable on a per-consumer
145 * basis using DTrace options, and need not be tuned on a system-wide basis.
146 * When tuning these variables, avoid pathological values; while some attempt
147 * is made to verify the integrity of these variables, they are not considered
148 * part of the supported interface to DTrace, and they are therefore not
149 * checked comprehensively. Further, these variables should not be tuned
150 * dynamically via "mdb -kw" or other means; they should only be tuned via
153 int dtrace_destructive_disallow = 0;
154 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
155 size_t dtrace_difo_maxsize = (256 * 1024);
156 dtrace_optval_t dtrace_dof_maxsize = (256 * 1024);
157 size_t dtrace_global_maxsize = (16 * 1024);
158 size_t dtrace_actions_max = (16 * 1024);
159 size_t dtrace_retain_max = 1024;
160 dtrace_optval_t dtrace_helper_actions_max = 32;
161 dtrace_optval_t dtrace_helper_providers_max = 32;
162 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
163 size_t dtrace_strsize_default = 256;
164 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
165 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
166 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
167 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
168 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
169 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
170 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
171 dtrace_optval_t dtrace_nspec_default = 1;
172 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
173 dtrace_optval_t dtrace_stackframes_default = 20;
174 dtrace_optval_t dtrace_ustackframes_default = 20;
175 dtrace_optval_t dtrace_jstackframes_default = 50;
176 dtrace_optval_t dtrace_jstackstrsize_default = 512;
177 int dtrace_msgdsize_max = 128;
178 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
179 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
180 int dtrace_devdepth_max = 32;
181 int dtrace_err_verbose;
182 hrtime_t dtrace_deadman_interval = NANOSEC;
183 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
184 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
187 * DTrace External Variables
189 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
190 * available to DTrace consumers via the backtick (`) syntax. One of these,
191 * dtrace_zero, is made deliberately so: it is provided as a source of
192 * well-known, zero-filled memory. While this variable is not documented,
193 * it is used by some translators as an implementation detail.
195 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
198 * DTrace Internal Variables
201 static dev_info_t *dtrace_devi; /* device info */
204 static vmem_t *dtrace_arena; /* probe ID arena */
205 static vmem_t *dtrace_minor; /* minor number arena */
206 static taskq_t *dtrace_taskq; /* task queue */
208 static struct unrhdr *dtrace_arena; /* Probe ID number. */
210 static dtrace_probe_t **dtrace_probes; /* array of all probes */
211 static int dtrace_nprobes; /* number of probes */
212 static dtrace_provider_t *dtrace_provider; /* provider list */
213 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
214 static int dtrace_opens; /* number of opens */
215 static int dtrace_helpers; /* number of helpers */
217 static void *dtrace_softstate; /* softstate pointer */
219 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
220 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
221 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
222 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
223 static int dtrace_toxranges; /* number of toxic ranges */
224 static int dtrace_toxranges_max; /* size of toxic range array */
225 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
226 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
227 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
228 static kthread_t *dtrace_panicked; /* panicking thread */
229 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
230 static dtrace_genid_t dtrace_probegen; /* current probe generation */
231 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
232 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
233 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
235 static struct mtx dtrace_unr_mtx;
236 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
237 int dtrace_in_probe; /* non-zero if executing a probe */
238 #if defined(__i386__) || defined(__amd64__)
239 uintptr_t dtrace_in_probe_addr; /* Address of invop when already in probe */
245 * DTrace is protected by three (relatively coarse-grained) locks:
247 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
248 * including enabling state, probes, ECBs, consumer state, helper state,
249 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
250 * probe context is lock-free -- synchronization is handled via the
251 * dtrace_sync() cross call mechanism.
253 * (2) dtrace_provider_lock is required when manipulating provider state, or
254 * when provider state must be held constant.
256 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
257 * when meta provider state must be held constant.
259 * The lock ordering between these three locks is dtrace_meta_lock before
260 * dtrace_provider_lock before dtrace_lock. (In particular, there are
261 * several places where dtrace_provider_lock is held by the framework as it
262 * calls into the providers -- which then call back into the framework,
263 * grabbing dtrace_lock.)
265 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
266 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
267 * role as a coarse-grained lock; it is acquired before both of these locks.
268 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
269 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
270 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
271 * acquired _between_ dtrace_provider_lock and dtrace_lock.
273 static kmutex_t dtrace_lock; /* probe state lock */
274 static kmutex_t dtrace_provider_lock; /* provider state lock */
275 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
278 /* XXX FreeBSD hacks. */
279 static kmutex_t mod_lock;
281 #define cr_suid cr_svuid
282 #define cr_sgid cr_svgid
283 #define ipaddr_t in_addr_t
284 #define mod_modname pathname
285 #define vuprintf vprintf
286 #define ttoproc(_a) ((_a)->td_proc)
287 #define crgetzoneid(_a) 0
290 #define CPU_ON_INTR(_a) 0
292 #define PRIV_EFFECTIVE (1 << 0)
293 #define PRIV_DTRACE_KERNEL (1 << 1)
294 #define PRIV_DTRACE_PROC (1 << 2)
295 #define PRIV_DTRACE_USER (1 << 3)
296 #define PRIV_PROC_OWNER (1 << 4)
297 #define PRIV_PROC_ZONE (1 << 5)
300 SYSCTL_NODE(_debug, OID_AUTO, dtrace, CTLFLAG_RD, 0, "DTrace Information");
304 #define curcpu CPU->cpu_id
309 * DTrace Provider Variables
311 * These are the variables relating to DTrace as a provider (that is, the
312 * provider of the BEGIN, END, and ERROR probes).
314 static dtrace_pattr_t dtrace_provider_attr = {
315 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
316 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
317 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
318 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
319 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
326 static dtrace_pops_t dtrace_provider_ops = {
327 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
328 (void (*)(void *, modctl_t *))dtrace_nullop,
329 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
330 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
331 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
332 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
336 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
339 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
340 static dtrace_id_t dtrace_probeid_end; /* special END probe */
341 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
344 * DTrace Helper Tracing Variables
346 uint32_t dtrace_helptrace_next = 0;
347 uint32_t dtrace_helptrace_nlocals;
348 char *dtrace_helptrace_buffer;
349 int dtrace_helptrace_bufsize = 512 * 1024;
352 int dtrace_helptrace_enabled = 1;
354 int dtrace_helptrace_enabled = 0;
358 * DTrace Error Hashing
360 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
361 * table. This is very useful for checking coverage of tests that are
362 * expected to induce DIF or DOF processing errors, and may be useful for
363 * debugging problems in the DIF code generator or in DOF generation . The
364 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
367 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
368 static const char *dtrace_errlast;
369 static kthread_t *dtrace_errthread;
370 static kmutex_t dtrace_errlock;
374 * DTrace Macros and Constants
376 * These are various macros that are useful in various spots in the
377 * implementation, along with a few random constants that have no meaning
378 * outside of the implementation. There is no real structure to this cpp
379 * mishmash -- but is there ever?
381 #define DTRACE_HASHSTR(hash, probe) \
382 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
384 #define DTRACE_HASHNEXT(hash, probe) \
385 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
387 #define DTRACE_HASHPREV(hash, probe) \
388 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
390 #define DTRACE_HASHEQ(hash, lhs, rhs) \
391 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
392 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
394 #define DTRACE_AGGHASHSIZE_SLEW 17
396 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
399 * The key for a thread-local variable consists of the lower 61 bits of the
400 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
401 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
402 * equal to a variable identifier. This is necessary (but not sufficient) to
403 * assure that global associative arrays never collide with thread-local
404 * variables. To guarantee that they cannot collide, we must also define the
405 * order for keying dynamic variables. That order is:
407 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
409 * Because the variable-key and the tls-key are in orthogonal spaces, there is
410 * no way for a global variable key signature to match a thread-local key
414 #define DTRACE_TLS_THRKEY(where) { \
416 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
417 for (; actv; actv >>= 1) \
419 ASSERT(intr < (1 << 3)); \
420 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
421 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
424 #define DTRACE_TLS_THRKEY(where) { \
425 solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
427 uint_t actv = _c->cpu_intr_actv; \
428 for (; actv; actv >>= 1) \
430 ASSERT(intr < (1 << 3)); \
431 (where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
432 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
436 #define DT_BSWAP_8(x) ((x) & 0xff)
437 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
438 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
439 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
441 #define DT_MASK_LO 0x00000000FFFFFFFFULL
443 #define DTRACE_STORE(type, tomax, offset, what) \
444 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
447 #define DTRACE_ALIGNCHECK(addr, size, flags) \
448 if (addr & (size - 1)) { \
449 *flags |= CPU_DTRACE_BADALIGN; \
450 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
454 #define DTRACE_ALIGNCHECK(addr, size, flags)
458 * Test whether a range of memory starting at testaddr of size testsz falls
459 * within the range of memory described by addr, sz. We take care to avoid
460 * problems with overflow and underflow of the unsigned quantities, and
461 * disallow all negative sizes. Ranges of size 0 are allowed.
463 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
464 ((testaddr) - (baseaddr) < (basesz) && \
465 (testaddr) + (testsz) - (baseaddr) <= (basesz) && \
466 (testaddr) + (testsz) >= (testaddr))
469 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
470 * alloc_sz on the righthand side of the comparison in order to avoid overflow
471 * or underflow in the comparison with it. This is simpler than the INRANGE
472 * check above, because we know that the dtms_scratch_ptr is valid in the
473 * range. Allocations of size zero are allowed.
475 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
476 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
477 (mstate)->dtms_scratch_ptr >= (alloc_sz))
479 #define DTRACE_LOADFUNC(bits) \
482 dtrace_load##bits(uintptr_t addr) \
484 size_t size = bits / NBBY; \
486 uint##bits##_t rval; \
488 volatile uint16_t *flags = (volatile uint16_t *) \
489 &cpu_core[curcpu].cpuc_dtrace_flags; \
491 DTRACE_ALIGNCHECK(addr, size, flags); \
493 for (i = 0; i < dtrace_toxranges; i++) { \
494 if (addr >= dtrace_toxrange[i].dtt_limit) \
497 if (addr + size <= dtrace_toxrange[i].dtt_base) \
501 * This address falls within a toxic region; return 0. \
503 *flags |= CPU_DTRACE_BADADDR; \
504 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
508 *flags |= CPU_DTRACE_NOFAULT; \
510 rval = *((volatile uint##bits##_t *)addr); \
511 *flags &= ~CPU_DTRACE_NOFAULT; \
513 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
517 #define dtrace_loadptr dtrace_load64
519 #define dtrace_loadptr dtrace_load32
522 #define DTRACE_DYNHASH_FREE 0
523 #define DTRACE_DYNHASH_SINK 1
524 #define DTRACE_DYNHASH_VALID 2
526 #define DTRACE_MATCH_NEXT 0
527 #define DTRACE_MATCH_DONE 1
528 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
529 #define DTRACE_STATE_ALIGN 64
531 #define DTRACE_FLAGS2FLT(flags) \
532 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
533 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
534 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
535 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
536 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
537 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
538 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
539 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
540 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
543 #define DTRACEACT_ISSTRING(act) \
544 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
545 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
547 /* Function prototype definitions: */
548 static size_t dtrace_strlen(const char *, size_t);
549 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
550 static void dtrace_enabling_provide(dtrace_provider_t *);
551 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
552 static void dtrace_enabling_matchall(void);
553 static dtrace_state_t *dtrace_anon_grab(void);
554 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
555 dtrace_state_t *, uint64_t, uint64_t);
556 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
557 static void dtrace_buffer_drop(dtrace_buffer_t *);
558 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
559 dtrace_state_t *, dtrace_mstate_t *);
560 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
562 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
563 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
564 uint16_t dtrace_load16(uintptr_t);
565 uint32_t dtrace_load32(uintptr_t);
566 uint64_t dtrace_load64(uintptr_t);
567 uint8_t dtrace_load8(uintptr_t);
568 void dtrace_dynvar_clean(dtrace_dstate_t *);
569 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
570 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
571 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
574 * DTrace Probe Context Functions
576 * These functions are called from probe context. Because probe context is
577 * any context in which C may be called, arbitrarily locks may be held,
578 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
579 * As a result, functions called from probe context may only call other DTrace
580 * support functions -- they may not interact at all with the system at large.
581 * (Note that the ASSERT macro is made probe-context safe by redefining it in
582 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
583 * loads are to be performed from probe context, they _must_ be in terms of
584 * the safe dtrace_load*() variants.
586 * Some functions in this block are not actually called from probe context;
587 * for these functions, there will be a comment above the function reading
588 * "Note: not called from probe context."
591 dtrace_panic(const char *format, ...)
595 va_start(alist, format);
596 dtrace_vpanic(format, alist);
601 dtrace_assfail(const char *a, const char *f, int l)
603 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
606 * We just need something here that even the most clever compiler
607 * cannot optimize away.
609 return (a[(uintptr_t)f]);
613 * Atomically increment a specified error counter from probe context.
616 dtrace_error(uint32_t *counter)
619 * Most counters stored to in probe context are per-CPU counters.
620 * However, there are some error conditions that are sufficiently
621 * arcane that they don't merit per-CPU storage. If these counters
622 * are incremented concurrently on different CPUs, scalability will be
623 * adversely affected -- but we don't expect them to be white-hot in a
624 * correctly constructed enabling...
631 if ((nval = oval + 1) == 0) {
633 * If the counter would wrap, set it to 1 -- assuring
634 * that the counter is never zero when we have seen
635 * errors. (The counter must be 32-bits because we
636 * aren't guaranteed a 64-bit compare&swap operation.)
637 * To save this code both the infamy of being fingered
638 * by a priggish news story and the indignity of being
639 * the target of a neo-puritan witch trial, we're
640 * carefully avoiding any colorful description of the
641 * likelihood of this condition -- but suffice it to
642 * say that it is only slightly more likely than the
643 * overflow of predicate cache IDs, as discussed in
644 * dtrace_predicate_create().
648 } while (dtrace_cas32(counter, oval, nval) != oval);
652 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
653 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
661 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
663 if (dest < mstate->dtms_scratch_base)
666 if (dest + size < dest)
669 if (dest + size > mstate->dtms_scratch_ptr)
676 dtrace_canstore_statvar(uint64_t addr, size_t sz,
677 dtrace_statvar_t **svars, int nsvars)
681 for (i = 0; i < nsvars; i++) {
682 dtrace_statvar_t *svar = svars[i];
684 if (svar == NULL || svar->dtsv_size == 0)
687 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
695 * Check to see if the address is within a memory region to which a store may
696 * be issued. This includes the DTrace scratch areas, and any DTrace variable
697 * region. The caller of dtrace_canstore() is responsible for performing any
698 * alignment checks that are needed before stores are actually executed.
701 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
702 dtrace_vstate_t *vstate)
705 * First, check to see if the address is in scratch space...
707 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
708 mstate->dtms_scratch_size))
712 * Now check to see if it's a dynamic variable. This check will pick
713 * up both thread-local variables and any global dynamically-allocated
716 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
717 vstate->dtvs_dynvars.dtds_size)) {
718 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
719 uintptr_t base = (uintptr_t)dstate->dtds_base +
720 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
724 * Before we assume that we can store here, we need to make
725 * sure that it isn't in our metadata -- storing to our
726 * dynamic variable metadata would corrupt our state. For
727 * the range to not include any dynamic variable metadata,
730 * (1) Start above the hash table that is at the base of
731 * the dynamic variable space
733 * (2) Have a starting chunk offset that is beyond the
734 * dtrace_dynvar_t that is at the base of every chunk
736 * (3) Not span a chunk boundary
742 chunkoffs = (addr - base) % dstate->dtds_chunksize;
744 if (chunkoffs < sizeof (dtrace_dynvar_t))
747 if (chunkoffs + sz > dstate->dtds_chunksize)
754 * Finally, check the static local and global variables. These checks
755 * take the longest, so we perform them last.
757 if (dtrace_canstore_statvar(addr, sz,
758 vstate->dtvs_locals, vstate->dtvs_nlocals))
761 if (dtrace_canstore_statvar(addr, sz,
762 vstate->dtvs_globals, vstate->dtvs_nglobals))
770 * Convenience routine to check to see if the address is within a memory
771 * region in which a load may be issued given the user's privilege level;
772 * if not, it sets the appropriate error flags and loads 'addr' into the
773 * illegal value slot.
775 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
776 * appropriate memory access protection.
779 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
780 dtrace_vstate_t *vstate)
782 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
785 * If we hold the privilege to read from kernel memory, then
786 * everything is readable.
788 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
792 * You can obviously read that which you can store.
794 if (dtrace_canstore(addr, sz, mstate, vstate))
798 * We're allowed to read from our own string table.
800 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
801 mstate->dtms_difo->dtdo_strlen))
804 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
810 * Convenience routine to check to see if a given string is within a memory
811 * region in which a load may be issued given the user's privilege level;
812 * this exists so that we don't need to issue unnecessary dtrace_strlen()
813 * calls in the event that the user has all privileges.
816 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
817 dtrace_vstate_t *vstate)
822 * If we hold the privilege to read from kernel memory, then
823 * everything is readable.
825 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
828 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
829 if (dtrace_canload(addr, strsz, mstate, vstate))
836 * Convenience routine to check to see if a given variable is within a memory
837 * region in which a load may be issued given the user's privilege level.
840 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
841 dtrace_vstate_t *vstate)
844 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
847 * If we hold the privilege to read from kernel memory, then
848 * everything is readable.
850 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
853 if (type->dtdt_kind == DIF_TYPE_STRING)
854 sz = dtrace_strlen(src,
855 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
857 sz = type->dtdt_size;
859 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
863 * Compare two strings using safe loads.
866 dtrace_strncmp(char *s1, char *s2, size_t limit)
869 volatile uint16_t *flags;
871 if (s1 == s2 || limit == 0)
874 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
880 c1 = dtrace_load8((uintptr_t)s1++);
886 c2 = dtrace_load8((uintptr_t)s2++);
891 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
897 * Compute strlen(s) for a string using safe memory accesses. The additional
898 * len parameter is used to specify a maximum length to ensure completion.
901 dtrace_strlen(const char *s, size_t lim)
905 for (len = 0; len != lim; len++) {
906 if (dtrace_load8((uintptr_t)s++) == '\0')
914 * Check if an address falls within a toxic region.
917 dtrace_istoxic(uintptr_t kaddr, size_t size)
919 uintptr_t taddr, tsize;
922 for (i = 0; i < dtrace_toxranges; i++) {
923 taddr = dtrace_toxrange[i].dtt_base;
924 tsize = dtrace_toxrange[i].dtt_limit - taddr;
926 if (kaddr - taddr < tsize) {
927 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
928 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
932 if (taddr - kaddr < size) {
933 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
934 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
943 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
944 * memory specified by the DIF program. The dst is assumed to be safe memory
945 * that we can store to directly because it is managed by DTrace. As with
946 * standard bcopy, overlapping copies are handled properly.
949 dtrace_bcopy(const void *src, void *dst, size_t len)
953 const uint8_t *s2 = src;
957 *s1++ = dtrace_load8((uintptr_t)s2++);
958 } while (--len != 0);
964 *--s1 = dtrace_load8((uintptr_t)--s2);
965 } while (--len != 0);
971 * Copy src to dst using safe memory accesses, up to either the specified
972 * length, or the point that a nul byte is encountered. The src is assumed to
973 * be unsafe memory specified by the DIF program. The dst is assumed to be
974 * safe memory that we can store to directly because it is managed by DTrace.
975 * Unlike dtrace_bcopy(), overlapping regions are not handled.
978 dtrace_strcpy(const void *src, void *dst, size_t len)
981 uint8_t *s1 = dst, c;
982 const uint8_t *s2 = src;
985 *s1++ = c = dtrace_load8((uintptr_t)s2++);
986 } while (--len != 0 && c != '\0');
991 * Copy src to dst, deriving the size and type from the specified (BYREF)
992 * variable type. The src is assumed to be unsafe memory specified by the DIF
993 * program. The dst is assumed to be DTrace variable memory that is of the
994 * specified type; we assume that we can store to directly.
997 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
999 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1001 if (type->dtdt_kind == DIF_TYPE_STRING) {
1002 dtrace_strcpy(src, dst, type->dtdt_size);
1004 dtrace_bcopy(src, dst, type->dtdt_size);
1009 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1010 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1011 * safe memory that we can access directly because it is managed by DTrace.
1014 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1016 volatile uint16_t *flags;
1018 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1023 if (s1 == NULL || s2 == NULL)
1026 if (s1 != s2 && len != 0) {
1027 const uint8_t *ps1 = s1;
1028 const uint8_t *ps2 = s2;
1031 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1033 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1039 * Zero the specified region using a simple byte-by-byte loop. Note that this
1040 * is for safe DTrace-managed memory only.
1043 dtrace_bzero(void *dst, size_t len)
1047 for (cp = dst; len != 0; len--)
1052 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1056 result[0] = addend1[0] + addend2[0];
1057 result[1] = addend1[1] + addend2[1] +
1058 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1065 * Shift the 128-bit value in a by b. If b is positive, shift left.
1066 * If b is negative, shift right.
1069 dtrace_shift_128(uint64_t *a, int b)
1079 a[0] = a[1] >> (b - 64);
1083 mask = 1LL << (64 - b);
1085 a[0] |= ((a[1] & mask) << (64 - b));
1090 a[1] = a[0] << (b - 64);
1094 mask = a[0] >> (64 - b);
1102 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1103 * use native multiplication on those, and then re-combine into the
1104 * resulting 128-bit value.
1106 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1113 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1115 uint64_t hi1, hi2, lo1, lo2;
1118 hi1 = factor1 >> 32;
1119 hi2 = factor2 >> 32;
1121 lo1 = factor1 & DT_MASK_LO;
1122 lo2 = factor2 & DT_MASK_LO;
1124 product[0] = lo1 * lo2;
1125 product[1] = hi1 * hi2;
1129 dtrace_shift_128(tmp, 32);
1130 dtrace_add_128(product, tmp, product);
1134 dtrace_shift_128(tmp, 32);
1135 dtrace_add_128(product, tmp, product);
1139 * This privilege check should be used by actions and subroutines to
1140 * verify that the user credentials of the process that enabled the
1141 * invoking ECB match the target credentials
1144 dtrace_priv_proc_common_user(dtrace_state_t *state)
1146 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1149 * We should always have a non-NULL state cred here, since if cred
1150 * is null (anonymous tracing), we fast-path bypass this routine.
1152 ASSERT(s_cr != NULL);
1154 if ((cr = CRED()) != NULL &&
1155 s_cr->cr_uid == cr->cr_uid &&
1156 s_cr->cr_uid == cr->cr_ruid &&
1157 s_cr->cr_uid == cr->cr_suid &&
1158 s_cr->cr_gid == cr->cr_gid &&
1159 s_cr->cr_gid == cr->cr_rgid &&
1160 s_cr->cr_gid == cr->cr_sgid)
1167 * This privilege check should be used by actions and subroutines to
1168 * verify that the zone of the process that enabled the invoking ECB
1169 * matches the target credentials
1172 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1175 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1178 * We should always have a non-NULL state cred here, since if cred
1179 * is null (anonymous tracing), we fast-path bypass this routine.
1181 ASSERT(s_cr != NULL);
1183 if ((cr = CRED()) != NULL &&
1184 s_cr->cr_zone == cr->cr_zone)
1194 * This privilege check should be used by actions and subroutines to
1195 * verify that the process has not setuid or changed credentials.
1198 dtrace_priv_proc_common_nocd(void)
1202 if ((proc = ttoproc(curthread)) != NULL &&
1203 !(proc->p_flag & SNOCD))
1210 dtrace_priv_proc_destructive(dtrace_state_t *state)
1212 int action = state->dts_cred.dcr_action;
1214 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1215 dtrace_priv_proc_common_zone(state) == 0)
1218 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1219 dtrace_priv_proc_common_user(state) == 0)
1222 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1223 dtrace_priv_proc_common_nocd() == 0)
1229 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1235 dtrace_priv_proc_control(dtrace_state_t *state)
1237 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1240 if (dtrace_priv_proc_common_zone(state) &&
1241 dtrace_priv_proc_common_user(state) &&
1242 dtrace_priv_proc_common_nocd())
1245 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1251 dtrace_priv_proc(dtrace_state_t *state)
1253 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1256 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1262 dtrace_priv_kernel(dtrace_state_t *state)
1264 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1267 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1273 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1275 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1278 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1284 * Note: not called from probe context. This function is called
1285 * asynchronously (and at a regular interval) from outside of probe context to
1286 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1287 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1290 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1292 dtrace_dynvar_t *dirty;
1293 dtrace_dstate_percpu_t *dcpu;
1296 for (i = 0; i < NCPU; i++) {
1297 dcpu = &dstate->dtds_percpu[i];
1299 ASSERT(dcpu->dtdsc_rinsing == NULL);
1302 * If the dirty list is NULL, there is no dirty work to do.
1304 if (dcpu->dtdsc_dirty == NULL)
1308 * If the clean list is non-NULL, then we're not going to do
1309 * any work for this CPU -- it means that there has not been
1310 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1311 * since the last time we cleaned house.
1313 if (dcpu->dtdsc_clean != NULL)
1319 * Atomically move the dirty list aside.
1322 dirty = dcpu->dtdsc_dirty;
1325 * Before we zap the dirty list, set the rinsing list.
1326 * (This allows for a potential assertion in
1327 * dtrace_dynvar(): if a free dynamic variable appears
1328 * on a hash chain, either the dirty list or the
1329 * rinsing list for some CPU must be non-NULL.)
1331 dcpu->dtdsc_rinsing = dirty;
1332 dtrace_membar_producer();
1333 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1334 dirty, NULL) != dirty);
1339 * We have no work to do; we can simply return.
1346 for (i = 0; i < NCPU; i++) {
1347 dcpu = &dstate->dtds_percpu[i];
1349 if (dcpu->dtdsc_rinsing == NULL)
1353 * We are now guaranteed that no hash chain contains a pointer
1354 * into this dirty list; we can make it clean.
1356 ASSERT(dcpu->dtdsc_clean == NULL);
1357 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1358 dcpu->dtdsc_rinsing = NULL;
1362 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1363 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1364 * This prevents a race whereby a CPU incorrectly decides that
1365 * the state should be something other than DTRACE_DSTATE_CLEAN
1366 * after dtrace_dynvar_clean() has completed.
1370 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1374 * Depending on the value of the op parameter, this function looks-up,
1375 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1376 * allocation is requested, this function will return a pointer to a
1377 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1378 * variable can be allocated. If NULL is returned, the appropriate counter
1379 * will be incremented.
1382 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1383 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1384 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1386 uint64_t hashval = DTRACE_DYNHASH_VALID;
1387 dtrace_dynhash_t *hash = dstate->dtds_hash;
1388 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1389 processorid_t me = curcpu, cpu = me;
1390 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1391 size_t bucket, ksize;
1392 size_t chunksize = dstate->dtds_chunksize;
1393 uintptr_t kdata, lock, nstate;
1399 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1400 * algorithm. For the by-value portions, we perform the algorithm in
1401 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1402 * bit, and seems to have only a minute effect on distribution. For
1403 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1404 * over each referenced byte. It's painful to do this, but it's much
1405 * better than pathological hash distribution. The efficacy of the
1406 * hashing algorithm (and a comparison with other algorithms) may be
1407 * found by running the ::dtrace_dynstat MDB dcmd.
1409 for (i = 0; i < nkeys; i++) {
1410 if (key[i].dttk_size == 0) {
1411 uint64_t val = key[i].dttk_value;
1413 hashval += (val >> 48) & 0xffff;
1414 hashval += (hashval << 10);
1415 hashval ^= (hashval >> 6);
1417 hashval += (val >> 32) & 0xffff;
1418 hashval += (hashval << 10);
1419 hashval ^= (hashval >> 6);
1421 hashval += (val >> 16) & 0xffff;
1422 hashval += (hashval << 10);
1423 hashval ^= (hashval >> 6);
1425 hashval += val & 0xffff;
1426 hashval += (hashval << 10);
1427 hashval ^= (hashval >> 6);
1430 * This is incredibly painful, but it beats the hell
1431 * out of the alternative.
1433 uint64_t j, size = key[i].dttk_size;
1434 uintptr_t base = (uintptr_t)key[i].dttk_value;
1436 if (!dtrace_canload(base, size, mstate, vstate))
1439 for (j = 0; j < size; j++) {
1440 hashval += dtrace_load8(base + j);
1441 hashval += (hashval << 10);
1442 hashval ^= (hashval >> 6);
1447 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1450 hashval += (hashval << 3);
1451 hashval ^= (hashval >> 11);
1452 hashval += (hashval << 15);
1455 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1456 * comes out to be one of our two sentinel hash values. If this
1457 * actually happens, we set the hashval to be a value known to be a
1458 * non-sentinel value.
1460 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1461 hashval = DTRACE_DYNHASH_VALID;
1464 * Yes, it's painful to do a divide here. If the cycle count becomes
1465 * important here, tricks can be pulled to reduce it. (However, it's
1466 * critical that hash collisions be kept to an absolute minimum;
1467 * they're much more painful than a divide.) It's better to have a
1468 * solution that generates few collisions and still keeps things
1469 * relatively simple.
1471 bucket = hashval % dstate->dtds_hashsize;
1473 if (op == DTRACE_DYNVAR_DEALLOC) {
1474 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1477 while ((lock = *lockp) & 1)
1480 if (dtrace_casptr((volatile void *)lockp,
1481 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1485 dtrace_membar_producer();
1490 lock = hash[bucket].dtdh_lock;
1492 dtrace_membar_consumer();
1494 start = hash[bucket].dtdh_chain;
1495 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1496 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1497 op != DTRACE_DYNVAR_DEALLOC));
1499 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1500 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1501 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1503 if (dvar->dtdv_hashval != hashval) {
1504 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1506 * We've reached the sink, and therefore the
1507 * end of the hash chain; we can kick out of
1508 * the loop knowing that we have seen a valid
1509 * snapshot of state.
1511 ASSERT(dvar->dtdv_next == NULL);
1512 ASSERT(dvar == &dtrace_dynhash_sink);
1516 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1518 * We've gone off the rails: somewhere along
1519 * the line, one of the members of this hash
1520 * chain was deleted. Note that we could also
1521 * detect this by simply letting this loop run
1522 * to completion, as we would eventually hit
1523 * the end of the dirty list. However, we
1524 * want to avoid running the length of the
1525 * dirty list unnecessarily (it might be quite
1526 * long), so we catch this as early as
1527 * possible by detecting the hash marker. In
1528 * this case, we simply set dvar to NULL and
1529 * break; the conditional after the loop will
1530 * send us back to top.
1539 if (dtuple->dtt_nkeys != nkeys)
1542 for (i = 0; i < nkeys; i++, dkey++) {
1543 if (dkey->dttk_size != key[i].dttk_size)
1544 goto next; /* size or type mismatch */
1546 if (dkey->dttk_size != 0) {
1548 (void *)(uintptr_t)key[i].dttk_value,
1549 (void *)(uintptr_t)dkey->dttk_value,
1553 if (dkey->dttk_value != key[i].dttk_value)
1558 if (op != DTRACE_DYNVAR_DEALLOC)
1561 ASSERT(dvar->dtdv_next == NULL ||
1562 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1565 ASSERT(hash[bucket].dtdh_chain != dvar);
1566 ASSERT(start != dvar);
1567 ASSERT(prev->dtdv_next == dvar);
1568 prev->dtdv_next = dvar->dtdv_next;
1570 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1571 start, dvar->dtdv_next) != start) {
1573 * We have failed to atomically swing the
1574 * hash table head pointer, presumably because
1575 * of a conflicting allocation on another CPU.
1576 * We need to reread the hash chain and try
1583 dtrace_membar_producer();
1586 * Now set the hash value to indicate that it's free.
1588 ASSERT(hash[bucket].dtdh_chain != dvar);
1589 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1591 dtrace_membar_producer();
1594 * Set the next pointer to point at the dirty list, and
1595 * atomically swing the dirty pointer to the newly freed dvar.
1598 next = dcpu->dtdsc_dirty;
1599 dvar->dtdv_next = next;
1600 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1603 * Finally, unlock this hash bucket.
1605 ASSERT(hash[bucket].dtdh_lock == lock);
1607 hash[bucket].dtdh_lock++;
1617 * If dvar is NULL, it is because we went off the rails:
1618 * one of the elements that we traversed in the hash chain
1619 * was deleted while we were traversing it. In this case,
1620 * we assert that we aren't doing a dealloc (deallocs lock
1621 * the hash bucket to prevent themselves from racing with
1622 * one another), and retry the hash chain traversal.
1624 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1628 if (op != DTRACE_DYNVAR_ALLOC) {
1630 * If we are not to allocate a new variable, we want to
1631 * return NULL now. Before we return, check that the value
1632 * of the lock word hasn't changed. If it has, we may have
1633 * seen an inconsistent snapshot.
1635 if (op == DTRACE_DYNVAR_NOALLOC) {
1636 if (hash[bucket].dtdh_lock != lock)
1639 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1640 ASSERT(hash[bucket].dtdh_lock == lock);
1642 hash[bucket].dtdh_lock++;
1649 * We need to allocate a new dynamic variable. The size we need is the
1650 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1651 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1652 * the size of any referred-to data (dsize). We then round the final
1653 * size up to the chunksize for allocation.
1655 for (ksize = 0, i = 0; i < nkeys; i++)
1656 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1659 * This should be pretty much impossible, but could happen if, say,
1660 * strange DIF specified the tuple. Ideally, this should be an
1661 * assertion and not an error condition -- but that requires that the
1662 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1663 * bullet-proof. (That is, it must not be able to be fooled by
1664 * malicious DIF.) Given the lack of backwards branches in DIF,
1665 * solving this would presumably not amount to solving the Halting
1666 * Problem -- but it still seems awfully hard.
1668 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1669 ksize + dsize > chunksize) {
1670 dcpu->dtdsc_drops++;
1674 nstate = DTRACE_DSTATE_EMPTY;
1678 free = dcpu->dtdsc_free;
1681 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1684 if (clean == NULL) {
1686 * We're out of dynamic variable space on
1687 * this CPU. Unless we have tried all CPUs,
1688 * we'll try to allocate from a different
1691 switch (dstate->dtds_state) {
1692 case DTRACE_DSTATE_CLEAN: {
1693 void *sp = &dstate->dtds_state;
1698 if (dcpu->dtdsc_dirty != NULL &&
1699 nstate == DTRACE_DSTATE_EMPTY)
1700 nstate = DTRACE_DSTATE_DIRTY;
1702 if (dcpu->dtdsc_rinsing != NULL)
1703 nstate = DTRACE_DSTATE_RINSING;
1705 dcpu = &dstate->dtds_percpu[cpu];
1710 (void) dtrace_cas32(sp,
1711 DTRACE_DSTATE_CLEAN, nstate);
1714 * To increment the correct bean
1715 * counter, take another lap.
1720 case DTRACE_DSTATE_DIRTY:
1721 dcpu->dtdsc_dirty_drops++;
1724 case DTRACE_DSTATE_RINSING:
1725 dcpu->dtdsc_rinsing_drops++;
1728 case DTRACE_DSTATE_EMPTY:
1729 dcpu->dtdsc_drops++;
1733 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1738 * The clean list appears to be non-empty. We want to
1739 * move the clean list to the free list; we start by
1740 * moving the clean pointer aside.
1742 if (dtrace_casptr(&dcpu->dtdsc_clean,
1743 clean, NULL) != clean) {
1745 * We are in one of two situations:
1747 * (a) The clean list was switched to the
1748 * free list by another CPU.
1750 * (b) The clean list was added to by the
1753 * In either of these situations, we can
1754 * just reattempt the free list allocation.
1759 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1762 * Now we'll move the clean list to the free list.
1763 * It's impossible for this to fail: the only way
1764 * the free list can be updated is through this
1765 * code path, and only one CPU can own the clean list.
1766 * Thus, it would only be possible for this to fail if
1767 * this code were racing with dtrace_dynvar_clean().
1768 * (That is, if dtrace_dynvar_clean() updated the clean
1769 * list, and we ended up racing to update the free
1770 * list.) This race is prevented by the dtrace_sync()
1771 * in dtrace_dynvar_clean() -- which flushes the
1772 * owners of the clean lists out before resetting
1775 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1776 ASSERT(rval == NULL);
1781 new_free = dvar->dtdv_next;
1782 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1785 * We have now allocated a new chunk. We copy the tuple keys into the
1786 * tuple array and copy any referenced key data into the data space
1787 * following the tuple array. As we do this, we relocate dttk_value
1788 * in the final tuple to point to the key data address in the chunk.
1790 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1791 dvar->dtdv_data = (void *)(kdata + ksize);
1792 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1794 for (i = 0; i < nkeys; i++) {
1795 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1796 size_t kesize = key[i].dttk_size;
1800 (const void *)(uintptr_t)key[i].dttk_value,
1801 (void *)kdata, kesize);
1802 dkey->dttk_value = kdata;
1803 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1805 dkey->dttk_value = key[i].dttk_value;
1808 dkey->dttk_size = kesize;
1811 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1812 dvar->dtdv_hashval = hashval;
1813 dvar->dtdv_next = start;
1815 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1819 * The cas has failed. Either another CPU is adding an element to
1820 * this hash chain, or another CPU is deleting an element from this
1821 * hash chain. The simplest way to deal with both of these cases
1822 * (though not necessarily the most efficient) is to free our
1823 * allocated block and tail-call ourselves. Note that the free is
1824 * to the dirty list and _not_ to the free list. This is to prevent
1825 * races with allocators, above.
1827 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1829 dtrace_membar_producer();
1832 free = dcpu->dtdsc_dirty;
1833 dvar->dtdv_next = free;
1834 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1836 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1841 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1843 if ((int64_t)nval < (int64_t)*oval)
1849 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1851 if ((int64_t)nval > (int64_t)*oval)
1856 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1858 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1859 int64_t val = (int64_t)nval;
1862 for (i = 0; i < zero; i++) {
1863 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1869 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1870 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1871 quanta[i - 1] += incr;
1876 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1884 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1886 uint64_t arg = *lquanta++;
1887 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1888 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1889 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1890 int32_t val = (int32_t)nval, level;
1893 ASSERT(levels != 0);
1897 * This is an underflow.
1903 level = (val - base) / step;
1905 if (level < levels) {
1906 lquanta[level + 1] += incr;
1911 * This is an overflow.
1913 lquanta[levels + 1] += incr;
1917 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1918 uint16_t high, uint16_t nsteps, int64_t value)
1920 int64_t this = 1, last, next;
1921 int base = 1, order;
1923 ASSERT(factor <= nsteps);
1924 ASSERT(nsteps % factor == 0);
1926 for (order = 0; order < low; order++)
1930 * If our value is less than our factor taken to the power of the
1931 * low order of magnitude, it goes into the zeroth bucket.
1933 if (value < (last = this))
1936 for (this *= factor; order <= high; order++) {
1937 int nbuckets = this > nsteps ? nsteps : this;
1939 if ((next = this * factor) < this) {
1941 * We should not generally get log/linear quantizations
1942 * with a high magnitude that allows 64-bits to
1943 * overflow, but we nonetheless protect against this
1944 * by explicitly checking for overflow, and clamping
1945 * our value accordingly.
1952 * If our value lies within this order of magnitude,
1953 * determine its position by taking the offset within
1954 * the order of magnitude, dividing by the bucket
1955 * width, and adding to our (accumulated) base.
1957 return (base + (value - last) / (this / nbuckets));
1960 base += nbuckets - (nbuckets / factor);
1966 * Our value is greater than or equal to our factor taken to the
1967 * power of one plus the high magnitude -- return the top bucket.
1973 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
1975 uint64_t arg = *llquanta++;
1976 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
1977 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
1978 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
1979 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
1981 llquanta[dtrace_aggregate_llquantize_bucket(factor,
1982 low, high, nsteps, nval)] += incr;
1987 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
1995 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
1997 int64_t snval = (int64_t)nval;
2004 * What we want to say here is:
2006 * data[2] += nval * nval;
2008 * But given that nval is 64-bit, we could easily overflow, so
2009 * we do this as 128-bit arithmetic.
2014 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2015 dtrace_add_128(data + 2, tmp, data + 2);
2020 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2027 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2033 * Aggregate given the tuple in the principal data buffer, and the aggregating
2034 * action denoted by the specified dtrace_aggregation_t. The aggregation
2035 * buffer is specified as the buf parameter. This routine does not return
2036 * failure; if there is no space in the aggregation buffer, the data will be
2037 * dropped, and a corresponding counter incremented.
2040 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2041 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2043 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2044 uint32_t i, ndx, size, fsize;
2045 uint32_t align = sizeof (uint64_t) - 1;
2046 dtrace_aggbuffer_t *agb;
2047 dtrace_aggkey_t *key;
2048 uint32_t hashval = 0, limit, isstr;
2049 caddr_t tomax, data, kdata;
2050 dtrace_actkind_t action;
2051 dtrace_action_t *act;
2057 if (!agg->dtag_hasarg) {
2059 * Currently, only quantize() and lquantize() take additional
2060 * arguments, and they have the same semantics: an increment
2061 * value that defaults to 1 when not present. If additional
2062 * aggregating actions take arguments, the setting of the
2063 * default argument value will presumably have to become more
2069 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2070 size = rec->dtrd_offset - agg->dtag_base;
2071 fsize = size + rec->dtrd_size;
2073 ASSERT(dbuf->dtb_tomax != NULL);
2074 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2076 if ((tomax = buf->dtb_tomax) == NULL) {
2077 dtrace_buffer_drop(buf);
2082 * The metastructure is always at the bottom of the buffer.
2084 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2085 sizeof (dtrace_aggbuffer_t));
2087 if (buf->dtb_offset == 0) {
2089 * We just kludge up approximately 1/8th of the size to be
2090 * buckets. If this guess ends up being routinely
2091 * off-the-mark, we may need to dynamically readjust this
2092 * based on past performance.
2094 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2096 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2097 (uintptr_t)tomax || hashsize == 0) {
2099 * We've been given a ludicrously small buffer;
2100 * increment our drop count and leave.
2102 dtrace_buffer_drop(buf);
2107 * And now, a pathetic attempt to try to get a an odd (or
2108 * perchance, a prime) hash size for better hash distribution.
2110 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2111 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2113 agb->dtagb_hashsize = hashsize;
2114 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2115 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2116 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2118 for (i = 0; i < agb->dtagb_hashsize; i++)
2119 agb->dtagb_hash[i] = NULL;
2122 ASSERT(agg->dtag_first != NULL);
2123 ASSERT(agg->dtag_first->dta_intuple);
2126 * Calculate the hash value based on the key. Note that we _don't_
2127 * include the aggid in the hashing (but we will store it as part of
2128 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2129 * algorithm: a simple, quick algorithm that has no known funnels, and
2130 * gets good distribution in practice. The efficacy of the hashing
2131 * algorithm (and a comparison with other algorithms) may be found by
2132 * running the ::dtrace_aggstat MDB dcmd.
2134 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2135 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2136 limit = i + act->dta_rec.dtrd_size;
2137 ASSERT(limit <= size);
2138 isstr = DTRACEACT_ISSTRING(act);
2140 for (; i < limit; i++) {
2142 hashval += (hashval << 10);
2143 hashval ^= (hashval >> 6);
2145 if (isstr && data[i] == '\0')
2150 hashval += (hashval << 3);
2151 hashval ^= (hashval >> 11);
2152 hashval += (hashval << 15);
2155 * Yes, the divide here is expensive -- but it's generally the least
2156 * of the performance issues given the amount of data that we iterate
2157 * over to compute hash values, compare data, etc.
2159 ndx = hashval % agb->dtagb_hashsize;
2161 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2162 ASSERT((caddr_t)key >= tomax);
2163 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2165 if (hashval != key->dtak_hashval || key->dtak_size != size)
2168 kdata = key->dtak_data;
2169 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2171 for (act = agg->dtag_first; act->dta_intuple;
2172 act = act->dta_next) {
2173 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2174 limit = i + act->dta_rec.dtrd_size;
2175 ASSERT(limit <= size);
2176 isstr = DTRACEACT_ISSTRING(act);
2178 for (; i < limit; i++) {
2179 if (kdata[i] != data[i])
2182 if (isstr && data[i] == '\0')
2187 if (action != key->dtak_action) {
2189 * We are aggregating on the same value in the same
2190 * aggregation with two different aggregating actions.
2191 * (This should have been picked up in the compiler,
2192 * so we may be dealing with errant or devious DIF.)
2193 * This is an error condition; we indicate as much,
2196 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2201 * This is a hit: we need to apply the aggregator to
2202 * the value at this key.
2204 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2211 * We didn't find it. We need to allocate some zero-filled space,
2212 * link it into the hash table appropriately, and apply the aggregator
2213 * to the (zero-filled) value.
2215 offs = buf->dtb_offset;
2216 while (offs & (align - 1))
2217 offs += sizeof (uint32_t);
2220 * If we don't have enough room to both allocate a new key _and_
2221 * its associated data, increment the drop count and return.
2223 if ((uintptr_t)tomax + offs + fsize >
2224 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2225 dtrace_buffer_drop(buf);
2230 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2231 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2232 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2234 key->dtak_data = kdata = tomax + offs;
2235 buf->dtb_offset = offs + fsize;
2238 * Now copy the data across.
2240 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2242 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2246 * Because strings are not zeroed out by default, we need to iterate
2247 * looking for actions that store strings, and we need to explicitly
2248 * pad these strings out with zeroes.
2250 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2253 if (!DTRACEACT_ISSTRING(act))
2256 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2257 limit = i + act->dta_rec.dtrd_size;
2258 ASSERT(limit <= size);
2260 for (nul = 0; i < limit; i++) {
2266 if (data[i] != '\0')
2273 for (i = size; i < fsize; i++)
2276 key->dtak_hashval = hashval;
2277 key->dtak_size = size;
2278 key->dtak_action = action;
2279 key->dtak_next = agb->dtagb_hash[ndx];
2280 agb->dtagb_hash[ndx] = key;
2283 * Finally, apply the aggregator.
2285 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2286 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2290 * Given consumer state, this routine finds a speculation in the INACTIVE
2291 * state and transitions it into the ACTIVE state. If there is no speculation
2292 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2293 * incremented -- it is up to the caller to take appropriate action.
2296 dtrace_speculation(dtrace_state_t *state)
2299 dtrace_speculation_state_t current;
2300 uint32_t *stat = &state->dts_speculations_unavail, count;
2302 while (i < state->dts_nspeculations) {
2303 dtrace_speculation_t *spec = &state->dts_speculations[i];
2305 current = spec->dtsp_state;
2307 if (current != DTRACESPEC_INACTIVE) {
2308 if (current == DTRACESPEC_COMMITTINGMANY ||
2309 current == DTRACESPEC_COMMITTING ||
2310 current == DTRACESPEC_DISCARDING)
2311 stat = &state->dts_speculations_busy;
2316 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2317 current, DTRACESPEC_ACTIVE) == current)
2322 * We couldn't find a speculation. If we found as much as a single
2323 * busy speculation buffer, we'll attribute this failure as "busy"
2324 * instead of "unavail".
2328 } while (dtrace_cas32(stat, count, count + 1) != count);
2334 * This routine commits an active speculation. If the specified speculation
2335 * is not in a valid state to perform a commit(), this routine will silently do
2336 * nothing. The state of the specified speculation is transitioned according
2337 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2340 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2341 dtrace_specid_t which)
2343 dtrace_speculation_t *spec;
2344 dtrace_buffer_t *src, *dest;
2345 uintptr_t daddr, saddr, dlimit;
2346 dtrace_speculation_state_t current, new = 0;
2352 if (which > state->dts_nspeculations) {
2353 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2357 spec = &state->dts_speculations[which - 1];
2358 src = &spec->dtsp_buffer[cpu];
2359 dest = &state->dts_buffer[cpu];
2362 current = spec->dtsp_state;
2364 if (current == DTRACESPEC_COMMITTINGMANY)
2368 case DTRACESPEC_INACTIVE:
2369 case DTRACESPEC_DISCARDING:
2372 case DTRACESPEC_COMMITTING:
2374 * This is only possible if we are (a) commit()'ing
2375 * without having done a prior speculate() on this CPU
2376 * and (b) racing with another commit() on a different
2377 * CPU. There's nothing to do -- we just assert that
2380 ASSERT(src->dtb_offset == 0);
2383 case DTRACESPEC_ACTIVE:
2384 new = DTRACESPEC_COMMITTING;
2387 case DTRACESPEC_ACTIVEONE:
2389 * This speculation is active on one CPU. If our
2390 * buffer offset is non-zero, we know that the one CPU
2391 * must be us. Otherwise, we are committing on a
2392 * different CPU from the speculate(), and we must
2393 * rely on being asynchronously cleaned.
2395 if (src->dtb_offset != 0) {
2396 new = DTRACESPEC_COMMITTING;
2401 case DTRACESPEC_ACTIVEMANY:
2402 new = DTRACESPEC_COMMITTINGMANY;
2408 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2409 current, new) != current);
2412 * We have set the state to indicate that we are committing this
2413 * speculation. Now reserve the necessary space in the destination
2416 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2417 sizeof (uint64_t), state, NULL)) < 0) {
2418 dtrace_buffer_drop(dest);
2423 * We have the space; copy the buffer across. (Note that this is a
2424 * highly subobtimal bcopy(); in the unlikely event that this becomes
2425 * a serious performance issue, a high-performance DTrace-specific
2426 * bcopy() should obviously be invented.)
2428 daddr = (uintptr_t)dest->dtb_tomax + offs;
2429 dlimit = daddr + src->dtb_offset;
2430 saddr = (uintptr_t)src->dtb_tomax;
2433 * First, the aligned portion.
2435 while (dlimit - daddr >= sizeof (uint64_t)) {
2436 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2438 daddr += sizeof (uint64_t);
2439 saddr += sizeof (uint64_t);
2443 * Now any left-over bit...
2445 while (dlimit - daddr)
2446 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2449 * Finally, commit the reserved space in the destination buffer.
2451 dest->dtb_offset = offs + src->dtb_offset;
2455 * If we're lucky enough to be the only active CPU on this speculation
2456 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2458 if (current == DTRACESPEC_ACTIVE ||
2459 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2460 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2461 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2463 ASSERT(rval == DTRACESPEC_COMMITTING);
2466 src->dtb_offset = 0;
2467 src->dtb_xamot_drops += src->dtb_drops;
2472 * This routine discards an active speculation. If the specified speculation
2473 * is not in a valid state to perform a discard(), this routine will silently
2474 * do nothing. The state of the specified speculation is transitioned
2475 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2478 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2479 dtrace_specid_t which)
2481 dtrace_speculation_t *spec;
2482 dtrace_speculation_state_t current, new = 0;
2483 dtrace_buffer_t *buf;
2488 if (which > state->dts_nspeculations) {
2489 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2493 spec = &state->dts_speculations[which - 1];
2494 buf = &spec->dtsp_buffer[cpu];
2497 current = spec->dtsp_state;
2500 case DTRACESPEC_INACTIVE:
2501 case DTRACESPEC_COMMITTINGMANY:
2502 case DTRACESPEC_COMMITTING:
2503 case DTRACESPEC_DISCARDING:
2506 case DTRACESPEC_ACTIVE:
2507 case DTRACESPEC_ACTIVEMANY:
2508 new = DTRACESPEC_DISCARDING;
2511 case DTRACESPEC_ACTIVEONE:
2512 if (buf->dtb_offset != 0) {
2513 new = DTRACESPEC_INACTIVE;
2515 new = DTRACESPEC_DISCARDING;
2522 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2523 current, new) != current);
2525 buf->dtb_offset = 0;
2530 * Note: not called from probe context. This function is called
2531 * asynchronously from cross call context to clean any speculations that are
2532 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2533 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2537 dtrace_speculation_clean_here(dtrace_state_t *state)
2539 dtrace_icookie_t cookie;
2540 processorid_t cpu = curcpu;
2541 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2544 cookie = dtrace_interrupt_disable();
2546 if (dest->dtb_tomax == NULL) {
2547 dtrace_interrupt_enable(cookie);
2551 for (i = 0; i < state->dts_nspeculations; i++) {
2552 dtrace_speculation_t *spec = &state->dts_speculations[i];
2553 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2555 if (src->dtb_tomax == NULL)
2558 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2559 src->dtb_offset = 0;
2563 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2566 if (src->dtb_offset == 0)
2569 dtrace_speculation_commit(state, cpu, i + 1);
2572 dtrace_interrupt_enable(cookie);
2576 * Note: not called from probe context. This function is called
2577 * asynchronously (and at a regular interval) to clean any speculations that
2578 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2579 * is work to be done, it cross calls all CPUs to perform that work;
2580 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2581 * INACTIVE state until they have been cleaned by all CPUs.
2584 dtrace_speculation_clean(dtrace_state_t *state)
2589 for (i = 0; i < state->dts_nspeculations; i++) {
2590 dtrace_speculation_t *spec = &state->dts_speculations[i];
2592 ASSERT(!spec->dtsp_cleaning);
2594 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2595 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2599 spec->dtsp_cleaning = 1;
2605 dtrace_xcall(DTRACE_CPUALL,
2606 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2609 * We now know that all CPUs have committed or discarded their
2610 * speculation buffers, as appropriate. We can now set the state
2613 for (i = 0; i < state->dts_nspeculations; i++) {
2614 dtrace_speculation_t *spec = &state->dts_speculations[i];
2615 dtrace_speculation_state_t current, new;
2617 if (!spec->dtsp_cleaning)
2620 current = spec->dtsp_state;
2621 ASSERT(current == DTRACESPEC_DISCARDING ||
2622 current == DTRACESPEC_COMMITTINGMANY);
2624 new = DTRACESPEC_INACTIVE;
2626 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2627 ASSERT(rv == current);
2628 spec->dtsp_cleaning = 0;
2633 * Called as part of a speculate() to get the speculative buffer associated
2634 * with a given speculation. Returns NULL if the specified speculation is not
2635 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2636 * the active CPU is not the specified CPU -- the speculation will be
2637 * atomically transitioned into the ACTIVEMANY state.
2639 static dtrace_buffer_t *
2640 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2641 dtrace_specid_t which)
2643 dtrace_speculation_t *spec;
2644 dtrace_speculation_state_t current, new = 0;
2645 dtrace_buffer_t *buf;
2650 if (which > state->dts_nspeculations) {
2651 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2655 spec = &state->dts_speculations[which - 1];
2656 buf = &spec->dtsp_buffer[cpuid];
2659 current = spec->dtsp_state;
2662 case DTRACESPEC_INACTIVE:
2663 case DTRACESPEC_COMMITTINGMANY:
2664 case DTRACESPEC_DISCARDING:
2667 case DTRACESPEC_COMMITTING:
2668 ASSERT(buf->dtb_offset == 0);
2671 case DTRACESPEC_ACTIVEONE:
2673 * This speculation is currently active on one CPU.
2674 * Check the offset in the buffer; if it's non-zero,
2675 * that CPU must be us (and we leave the state alone).
2676 * If it's zero, assume that we're starting on a new
2677 * CPU -- and change the state to indicate that the
2678 * speculation is active on more than one CPU.
2680 if (buf->dtb_offset != 0)
2683 new = DTRACESPEC_ACTIVEMANY;
2686 case DTRACESPEC_ACTIVEMANY:
2689 case DTRACESPEC_ACTIVE:
2690 new = DTRACESPEC_ACTIVEONE;
2696 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2697 current, new) != current);
2699 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2704 * Return a string. In the event that the user lacks the privilege to access
2705 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2706 * don't fail access checking.
2708 * dtrace_dif_variable() uses this routine as a helper for various
2709 * builtin values such as 'execname' and 'probefunc.'
2712 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2713 dtrace_mstate_t *mstate)
2715 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2720 * The easy case: this probe is allowed to read all of memory, so
2721 * we can just return this as a vanilla pointer.
2723 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2727 * This is the tougher case: we copy the string in question from
2728 * kernel memory into scratch memory and return it that way: this
2729 * ensures that we won't trip up when access checking tests the
2730 * BYREF return value.
2732 strsz = dtrace_strlen((char *)addr, size) + 1;
2734 if (mstate->dtms_scratch_ptr + strsz >
2735 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2736 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2740 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2742 ret = mstate->dtms_scratch_ptr;
2743 mstate->dtms_scratch_ptr += strsz;
2748 * Return a string from a memoy address which is known to have one or
2749 * more concatenated, individually zero terminated, sub-strings.
2750 * In the event that the user lacks the privilege to access
2751 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2752 * don't fail access checking.
2754 * dtrace_dif_variable() uses this routine as a helper for various
2755 * builtin values such as 'execargs'.
2758 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2759 dtrace_mstate_t *mstate)
2765 if (mstate->dtms_scratch_ptr + strsz >
2766 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2767 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2771 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2774 /* Replace sub-string termination characters with a space. */
2775 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2780 ret = mstate->dtms_scratch_ptr;
2781 mstate->dtms_scratch_ptr += strsz;
2786 * This function implements the DIF emulator's variable lookups. The emulator
2787 * passes a reserved variable identifier and optional built-in array index.
2790 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2794 * If we're accessing one of the uncached arguments, we'll turn this
2795 * into a reference in the args array.
2797 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2798 ndx = v - DIF_VAR_ARG0;
2804 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2805 if (ndx >= sizeof (mstate->dtms_arg) /
2806 sizeof (mstate->dtms_arg[0])) {
2807 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2808 dtrace_provider_t *pv;
2811 pv = mstate->dtms_probe->dtpr_provider;
2812 if (pv->dtpv_pops.dtps_getargval != NULL)
2813 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2814 mstate->dtms_probe->dtpr_id,
2815 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2817 val = dtrace_getarg(ndx, aframes);
2820 * This is regrettably required to keep the compiler
2821 * from tail-optimizing the call to dtrace_getarg().
2822 * The condition always evaluates to true, but the
2823 * compiler has no way of figuring that out a priori.
2824 * (None of this would be necessary if the compiler
2825 * could be relied upon to _always_ tail-optimize
2826 * the call to dtrace_getarg() -- but it can't.)
2828 if (mstate->dtms_probe != NULL)
2834 return (mstate->dtms_arg[ndx]);
2837 case DIF_VAR_UREGS: {
2840 if (!dtrace_priv_proc(state))
2843 if ((lwp = curthread->t_lwp) == NULL) {
2844 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2845 cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2849 return (dtrace_getreg(lwp->lwp_regs, ndx));
2853 case DIF_VAR_UREGS: {
2854 struct trapframe *tframe;
2856 if (!dtrace_priv_proc(state))
2859 if ((tframe = curthread->td_frame) == NULL) {
2860 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2861 cpu_core[curcpu].cpuc_dtrace_illval = 0;
2865 return (dtrace_getreg(tframe, ndx));
2869 case DIF_VAR_CURTHREAD:
2870 if (!dtrace_priv_kernel(state))
2872 return ((uint64_t)(uintptr_t)curthread);
2874 case DIF_VAR_TIMESTAMP:
2875 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2876 mstate->dtms_timestamp = dtrace_gethrtime();
2877 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2879 return (mstate->dtms_timestamp);
2881 case DIF_VAR_VTIMESTAMP:
2882 ASSERT(dtrace_vtime_references != 0);
2883 return (curthread->t_dtrace_vtime);
2885 case DIF_VAR_WALLTIMESTAMP:
2886 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2887 mstate->dtms_walltimestamp = dtrace_gethrestime();
2888 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2890 return (mstate->dtms_walltimestamp);
2894 if (!dtrace_priv_kernel(state))
2896 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2897 mstate->dtms_ipl = dtrace_getipl();
2898 mstate->dtms_present |= DTRACE_MSTATE_IPL;
2900 return (mstate->dtms_ipl);
2904 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2905 return (mstate->dtms_epid);
2908 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2909 return (mstate->dtms_probe->dtpr_id);
2911 case DIF_VAR_STACKDEPTH:
2912 if (!dtrace_priv_kernel(state))
2914 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2915 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2917 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2918 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2920 return (mstate->dtms_stackdepth);
2922 case DIF_VAR_USTACKDEPTH:
2923 if (!dtrace_priv_proc(state))
2925 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2927 * See comment in DIF_VAR_PID.
2929 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2931 mstate->dtms_ustackdepth = 0;
2933 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2934 mstate->dtms_ustackdepth =
2935 dtrace_getustackdepth();
2936 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2938 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2940 return (mstate->dtms_ustackdepth);
2942 case DIF_VAR_CALLER:
2943 if (!dtrace_priv_kernel(state))
2945 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2946 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2948 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2950 * If this is an unanchored probe, we are
2951 * required to go through the slow path:
2952 * dtrace_caller() only guarantees correct
2953 * results for anchored probes.
2955 pc_t caller[2] = {0, 0};
2957 dtrace_getpcstack(caller, 2, aframes,
2958 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2959 mstate->dtms_caller = caller[1];
2960 } else if ((mstate->dtms_caller =
2961 dtrace_caller(aframes)) == -1) {
2963 * We have failed to do this the quick way;
2964 * we must resort to the slower approach of
2965 * calling dtrace_getpcstack().
2969 dtrace_getpcstack(&caller, 1, aframes, NULL);
2970 mstate->dtms_caller = caller;
2973 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
2975 return (mstate->dtms_caller);
2977 case DIF_VAR_UCALLER:
2978 if (!dtrace_priv_proc(state))
2981 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
2985 * dtrace_getupcstack() fills in the first uint64_t
2986 * with the current PID. The second uint64_t will
2987 * be the program counter at user-level. The third
2988 * uint64_t will contain the caller, which is what
2992 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2993 dtrace_getupcstack(ustack, 3);
2994 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2995 mstate->dtms_ucaller = ustack[2];
2996 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
2999 return (mstate->dtms_ucaller);
3001 case DIF_VAR_PROBEPROV:
3002 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3003 return (dtrace_dif_varstr(
3004 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3007 case DIF_VAR_PROBEMOD:
3008 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3009 return (dtrace_dif_varstr(
3010 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3013 case DIF_VAR_PROBEFUNC:
3014 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3015 return (dtrace_dif_varstr(
3016 (uintptr_t)mstate->dtms_probe->dtpr_func,
3019 case DIF_VAR_PROBENAME:
3020 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3021 return (dtrace_dif_varstr(
3022 (uintptr_t)mstate->dtms_probe->dtpr_name,
3026 if (!dtrace_priv_proc(state))
3031 * Note that we are assuming that an unanchored probe is
3032 * always due to a high-level interrupt. (And we're assuming
3033 * that there is only a single high level interrupt.)
3035 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3036 return (pid0.pid_id);
3039 * It is always safe to dereference one's own t_procp pointer:
3040 * it always points to a valid, allocated proc structure.
3041 * Further, it is always safe to dereference the p_pidp member
3042 * of one's own proc structure. (These are truisms becuase
3043 * threads and processes don't clean up their own state --
3044 * they leave that task to whomever reaps them.)
3046 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3048 return ((uint64_t)curproc->p_pid);
3052 if (!dtrace_priv_proc(state))
3057 * See comment in DIF_VAR_PID.
3059 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3060 return (pid0.pid_id);
3063 * It is always safe to dereference one's own t_procp pointer:
3064 * it always points to a valid, allocated proc structure.
3065 * (This is true because threads don't clean up their own
3066 * state -- they leave that task to whomever reaps them.)
3068 return ((uint64_t)curthread->t_procp->p_ppid);
3070 return ((uint64_t)curproc->p_pptr->p_pid);
3076 * See comment in DIF_VAR_PID.
3078 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3082 return ((uint64_t)curthread->t_tid);
3084 case DIF_VAR_EXECARGS: {
3085 struct pargs *p_args = curthread->td_proc->p_args;
3090 return (dtrace_dif_varstrz(
3091 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3094 case DIF_VAR_EXECNAME:
3096 if (!dtrace_priv_proc(state))
3100 * See comment in DIF_VAR_PID.
3102 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3103 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3106 * It is always safe to dereference one's own t_procp pointer:
3107 * it always points to a valid, allocated proc structure.
3108 * (This is true because threads don't clean up their own
3109 * state -- they leave that task to whomever reaps them.)
3111 return (dtrace_dif_varstr(
3112 (uintptr_t)curthread->t_procp->p_user.u_comm,
3115 return (dtrace_dif_varstr(
3116 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3119 case DIF_VAR_ZONENAME:
3121 if (!dtrace_priv_proc(state))
3125 * See comment in DIF_VAR_PID.
3127 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3128 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3131 * It is always safe to dereference one's own t_procp pointer:
3132 * it always points to a valid, allocated proc structure.
3133 * (This is true because threads don't clean up their own
3134 * state -- they leave that task to whomever reaps them.)
3136 return (dtrace_dif_varstr(
3137 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3144 if (!dtrace_priv_proc(state))
3149 * See comment in DIF_VAR_PID.
3151 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3152 return ((uint64_t)p0.p_cred->cr_uid);
3156 * It is always safe to dereference one's own t_procp pointer:
3157 * it always points to a valid, allocated proc structure.
3158 * (This is true because threads don't clean up their own
3159 * state -- they leave that task to whomever reaps them.)
3161 * Additionally, it is safe to dereference one's own process
3162 * credential, since this is never NULL after process birth.
3164 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3167 if (!dtrace_priv_proc(state))
3172 * See comment in DIF_VAR_PID.
3174 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3175 return ((uint64_t)p0.p_cred->cr_gid);
3179 * It is always safe to dereference one's own t_procp pointer:
3180 * it always points to a valid, allocated proc structure.
3181 * (This is true because threads don't clean up their own
3182 * state -- they leave that task to whomever reaps them.)
3184 * Additionally, it is safe to dereference one's own process
3185 * credential, since this is never NULL after process birth.
3187 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3189 case DIF_VAR_ERRNO: {
3192 if (!dtrace_priv_proc(state))
3196 * See comment in DIF_VAR_PID.
3198 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3202 * It is always safe to dereference one's own t_lwp pointer in
3203 * the event that this pointer is non-NULL. (This is true
3204 * because threads and lwps don't clean up their own state --
3205 * they leave that task to whomever reaps them.)
3207 if ((lwp = curthread->t_lwp) == NULL)
3210 return ((uint64_t)lwp->lwp_errno);
3212 return (curthread->td_errno);
3221 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3227 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3228 * Notice that we don't bother validating the proper number of arguments or
3229 * their types in the tuple stack. This isn't needed because all argument
3230 * interpretation is safe because of our load safety -- the worst that can
3231 * happen is that a bogus program can obtain bogus results.
3234 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3235 dtrace_key_t *tupregs, int nargs,
3236 dtrace_mstate_t *mstate, dtrace_state_t *state)
3238 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3239 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3240 dtrace_vstate_t *vstate = &state->dts_vstate;
3253 struct thread *lowner;
3255 struct lock_object *li;
3262 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3266 case DIF_SUBR_MUTEX_OWNED:
3267 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3273 m.mx = dtrace_load64(tupregs[0].dttk_value);
3274 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3275 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3277 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3280 case DIF_SUBR_MUTEX_OWNER:
3281 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3287 m.mx = dtrace_load64(tupregs[0].dttk_value);
3288 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3289 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3290 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3295 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3296 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3302 m.mx = dtrace_load64(tupregs[0].dttk_value);
3303 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3306 case DIF_SUBR_MUTEX_TYPE_SPIN:
3307 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3313 m.mx = dtrace_load64(tupregs[0].dttk_value);
3314 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3317 case DIF_SUBR_RW_READ_HELD: {
3320 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3326 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3327 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3331 case DIF_SUBR_RW_WRITE_HELD:
3332 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3338 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3339 regs[rd] = _RW_WRITE_HELD(&r.ri);
3342 case DIF_SUBR_RW_ISWRITER:
3343 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3349 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3350 regs[rd] = _RW_ISWRITER(&r.ri);
3354 case DIF_SUBR_MUTEX_OWNED:
3355 if (!dtrace_canload(tupregs[0].dttk_value,
3356 sizeof (struct lock_object), mstate, vstate)) {
3360 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3361 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3364 case DIF_SUBR_MUTEX_OWNER:
3365 if (!dtrace_canload(tupregs[0].dttk_value,
3366 sizeof (struct lock_object), mstate, vstate)) {
3370 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3371 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3372 regs[rd] = (uintptr_t)lowner;
3375 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3376 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3381 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3382 /* XXX - should be only LC_SLEEPABLE? */
3383 regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3384 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3387 case DIF_SUBR_MUTEX_TYPE_SPIN:
3388 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3393 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3394 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3397 case DIF_SUBR_RW_READ_HELD:
3398 case DIF_SUBR_SX_SHARED_HELD:
3399 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3404 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3405 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3409 case DIF_SUBR_RW_WRITE_HELD:
3410 case DIF_SUBR_SX_EXCLUSIVE_HELD:
3411 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3416 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3417 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3418 regs[rd] = (lowner == curthread);
3421 case DIF_SUBR_RW_ISWRITER:
3422 case DIF_SUBR_SX_ISEXCLUSIVE:
3423 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3428 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3429 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3432 #endif /* ! defined(sun) */
3434 case DIF_SUBR_BCOPY: {
3436 * We need to be sure that the destination is in the scratch
3437 * region -- no other region is allowed.
3439 uintptr_t src = tupregs[0].dttk_value;
3440 uintptr_t dest = tupregs[1].dttk_value;
3441 size_t size = tupregs[2].dttk_value;
3443 if (!dtrace_inscratch(dest, size, mstate)) {
3444 *flags |= CPU_DTRACE_BADADDR;
3449 if (!dtrace_canload(src, size, mstate, vstate)) {
3454 dtrace_bcopy((void *)src, (void *)dest, size);
3458 case DIF_SUBR_ALLOCA:
3459 case DIF_SUBR_COPYIN: {
3460 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3462 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3463 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3466 * This action doesn't require any credential checks since
3467 * probes will not activate in user contexts to which the
3468 * enabling user does not have permissions.
3472 * Rounding up the user allocation size could have overflowed
3473 * a large, bogus allocation (like -1ULL) to 0.
3475 if (scratch_size < size ||
3476 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3477 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3482 if (subr == DIF_SUBR_COPYIN) {
3483 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3484 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3485 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3488 mstate->dtms_scratch_ptr += scratch_size;
3493 case DIF_SUBR_COPYINTO: {
3494 uint64_t size = tupregs[1].dttk_value;
3495 uintptr_t dest = tupregs[2].dttk_value;
3498 * This action doesn't require any credential checks since
3499 * probes will not activate in user contexts to which the
3500 * enabling user does not have permissions.
3502 if (!dtrace_inscratch(dest, size, mstate)) {
3503 *flags |= CPU_DTRACE_BADADDR;
3508 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3509 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3510 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3514 case DIF_SUBR_COPYINSTR: {
3515 uintptr_t dest = mstate->dtms_scratch_ptr;
3516 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3518 if (nargs > 1 && tupregs[1].dttk_value < size)
3519 size = tupregs[1].dttk_value + 1;
3522 * This action doesn't require any credential checks since
3523 * probes will not activate in user contexts to which the
3524 * enabling user does not have permissions.
3526 if (!DTRACE_INSCRATCH(mstate, size)) {
3527 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3532 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3533 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3534 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3536 ((char *)dest)[size - 1] = '\0';
3537 mstate->dtms_scratch_ptr += size;
3543 case DIF_SUBR_MSGSIZE:
3544 case DIF_SUBR_MSGDSIZE: {
3545 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3546 uintptr_t wptr, rptr;
3550 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3552 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3558 wptr = dtrace_loadptr(baddr +
3559 offsetof(mblk_t, b_wptr));
3561 rptr = dtrace_loadptr(baddr +
3562 offsetof(mblk_t, b_rptr));
3565 *flags |= CPU_DTRACE_BADADDR;
3566 *illval = tupregs[0].dttk_value;
3570 daddr = dtrace_loadptr(baddr +
3571 offsetof(mblk_t, b_datap));
3573 baddr = dtrace_loadptr(baddr +
3574 offsetof(mblk_t, b_cont));
3577 * We want to prevent against denial-of-service here,
3578 * so we're only going to search the list for
3579 * dtrace_msgdsize_max mblks.
3581 if (cont++ > dtrace_msgdsize_max) {
3582 *flags |= CPU_DTRACE_ILLOP;
3586 if (subr == DIF_SUBR_MSGDSIZE) {
3587 if (dtrace_load8(daddr +
3588 offsetof(dblk_t, db_type)) != M_DATA)
3592 count += wptr - rptr;
3595 if (!(*flags & CPU_DTRACE_FAULT))
3602 case DIF_SUBR_PROGENYOF: {
3603 pid_t pid = tupregs[0].dttk_value;
3607 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3609 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3611 if (p->p_pidp->pid_id == pid) {
3613 if (p->p_pid == pid) {
3620 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3626 case DIF_SUBR_SPECULATION:
3627 regs[rd] = dtrace_speculation(state);
3630 case DIF_SUBR_COPYOUT: {
3631 uintptr_t kaddr = tupregs[0].dttk_value;
3632 uintptr_t uaddr = tupregs[1].dttk_value;
3633 uint64_t size = tupregs[2].dttk_value;
3635 if (!dtrace_destructive_disallow &&
3636 dtrace_priv_proc_control(state) &&
3637 !dtrace_istoxic(kaddr, size)) {
3638 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3639 dtrace_copyout(kaddr, uaddr, size, flags);
3640 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3645 case DIF_SUBR_COPYOUTSTR: {
3646 uintptr_t kaddr = tupregs[0].dttk_value;
3647 uintptr_t uaddr = tupregs[1].dttk_value;
3648 uint64_t size = tupregs[2].dttk_value;
3650 if (!dtrace_destructive_disallow &&
3651 dtrace_priv_proc_control(state) &&
3652 !dtrace_istoxic(kaddr, size)) {
3653 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3654 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3655 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3660 case DIF_SUBR_STRLEN: {
3662 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3663 sz = dtrace_strlen((char *)addr,
3664 state->dts_options[DTRACEOPT_STRSIZE]);
3666 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3676 case DIF_SUBR_STRCHR:
3677 case DIF_SUBR_STRRCHR: {
3679 * We're going to iterate over the string looking for the
3680 * specified character. We will iterate until we have reached
3681 * the string length or we have found the character. If this
3682 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3683 * of the specified character instead of the first.
3685 uintptr_t saddr = tupregs[0].dttk_value;
3686 uintptr_t addr = tupregs[0].dttk_value;
3687 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3688 char c, target = (char)tupregs[1].dttk_value;
3690 for (regs[rd] = 0; addr < limit; addr++) {
3691 if ((c = dtrace_load8(addr)) == target) {
3694 if (subr == DIF_SUBR_STRCHR)
3702 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3710 case DIF_SUBR_STRSTR:
3711 case DIF_SUBR_INDEX:
3712 case DIF_SUBR_RINDEX: {
3714 * We're going to iterate over the string looking for the
3715 * specified string. We will iterate until we have reached
3716 * the string length or we have found the string. (Yes, this
3717 * is done in the most naive way possible -- but considering
3718 * that the string we're searching for is likely to be
3719 * relatively short, the complexity of Rabin-Karp or similar
3720 * hardly seems merited.)
3722 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3723 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3724 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3725 size_t len = dtrace_strlen(addr, size);
3726 size_t sublen = dtrace_strlen(substr, size);
3727 char *limit = addr + len, *orig = addr;
3728 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3731 regs[rd] = notfound;
3733 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3738 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3745 * strstr() and index()/rindex() have similar semantics if
3746 * both strings are the empty string: strstr() returns a
3747 * pointer to the (empty) string, and index() and rindex()
3748 * both return index 0 (regardless of any position argument).
3750 if (sublen == 0 && len == 0) {
3751 if (subr == DIF_SUBR_STRSTR)
3752 regs[rd] = (uintptr_t)addr;
3758 if (subr != DIF_SUBR_STRSTR) {
3759 if (subr == DIF_SUBR_RINDEX) {
3766 * Both index() and rindex() take an optional position
3767 * argument that denotes the starting position.
3770 int64_t pos = (int64_t)tupregs[2].dttk_value;
3773 * If the position argument to index() is
3774 * negative, Perl implicitly clamps it at
3775 * zero. This semantic is a little surprising
3776 * given the special meaning of negative
3777 * positions to similar Perl functions like
3778 * substr(), but it appears to reflect a
3779 * notion that index() can start from a
3780 * negative index and increment its way up to
3781 * the string. Given this notion, Perl's
3782 * rindex() is at least self-consistent in
3783 * that it implicitly clamps positions greater
3784 * than the string length to be the string
3785 * length. Where Perl completely loses
3786 * coherence, however, is when the specified
3787 * substring is the empty string (""). In
3788 * this case, even if the position is
3789 * negative, rindex() returns 0 -- and even if
3790 * the position is greater than the length,
3791 * index() returns the string length. These
3792 * semantics violate the notion that index()
3793 * should never return a value less than the
3794 * specified position and that rindex() should
3795 * never return a value greater than the
3796 * specified position. (One assumes that
3797 * these semantics are artifacts of Perl's
3798 * implementation and not the results of
3799 * deliberate design -- it beggars belief that
3800 * even Larry Wall could desire such oddness.)
3801 * While in the abstract one would wish for
3802 * consistent position semantics across
3803 * substr(), index() and rindex() -- or at the
3804 * very least self-consistent position
3805 * semantics for index() and rindex() -- we
3806 * instead opt to keep with the extant Perl
3807 * semantics, in all their broken glory. (Do
3808 * we have more desire to maintain Perl's
3809 * semantics than Perl does? Probably.)
3811 if (subr == DIF_SUBR_RINDEX) {
3835 for (regs[rd] = notfound; addr != limit; addr += inc) {
3836 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3837 if (subr != DIF_SUBR_STRSTR) {
3839 * As D index() and rindex() are
3840 * modeled on Perl (and not on awk),
3841 * we return a zero-based (and not a
3842 * one-based) index. (For you Perl
3843 * weenies: no, we're not going to add
3844 * $[ -- and shouldn't you be at a con
3847 regs[rd] = (uintptr_t)(addr - orig);
3851 ASSERT(subr == DIF_SUBR_STRSTR);
3852 regs[rd] = (uintptr_t)addr;
3860 case DIF_SUBR_STRTOK: {
3861 uintptr_t addr = tupregs[0].dttk_value;
3862 uintptr_t tokaddr = tupregs[1].dttk_value;
3863 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3864 uintptr_t limit, toklimit = tokaddr + size;
3865 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
3866 char *dest = (char *)mstate->dtms_scratch_ptr;
3870 * Check both the token buffer and (later) the input buffer,
3871 * since both could be non-scratch addresses.
3873 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3878 if (!DTRACE_INSCRATCH(mstate, size)) {
3879 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3886 * If the address specified is NULL, we use our saved
3887 * strtok pointer from the mstate. Note that this
3888 * means that the saved strtok pointer is _only_
3889 * valid within multiple enablings of the same probe --
3890 * it behaves like an implicit clause-local variable.
3892 addr = mstate->dtms_strtok;
3895 * If the user-specified address is non-NULL we must
3896 * access check it. This is the only time we have
3897 * a chance to do so, since this address may reside
3898 * in the string table of this clause-- future calls
3899 * (when we fetch addr from mstate->dtms_strtok)
3900 * would fail this access check.
3902 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3909 * First, zero the token map, and then process the token
3910 * string -- setting a bit in the map for every character
3911 * found in the token string.
3913 for (i = 0; i < sizeof (tokmap); i++)
3916 for (; tokaddr < toklimit; tokaddr++) {
3917 if ((c = dtrace_load8(tokaddr)) == '\0')
3920 ASSERT((c >> 3) < sizeof (tokmap));
3921 tokmap[c >> 3] |= (1 << (c & 0x7));
3924 for (limit = addr + size; addr < limit; addr++) {
3926 * We're looking for a character that is _not_ contained
3927 * in the token string.
3929 if ((c = dtrace_load8(addr)) == '\0')
3932 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3938 * We reached the end of the string without finding
3939 * any character that was not in the token string.
3940 * We return NULL in this case, and we set the saved
3941 * address to NULL as well.
3944 mstate->dtms_strtok = 0;
3949 * From here on, we're copying into the destination string.
3951 for (i = 0; addr < limit && i < size - 1; addr++) {
3952 if ((c = dtrace_load8(addr)) == '\0')
3955 if (tokmap[c >> 3] & (1 << (c & 0x7)))
3964 regs[rd] = (uintptr_t)dest;
3965 mstate->dtms_scratch_ptr += size;
3966 mstate->dtms_strtok = addr;
3970 case DIF_SUBR_SUBSTR: {
3971 uintptr_t s = tupregs[0].dttk_value;
3972 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3973 char *d = (char *)mstate->dtms_scratch_ptr;
3974 int64_t index = (int64_t)tupregs[1].dttk_value;
3975 int64_t remaining = (int64_t)tupregs[2].dttk_value;
3976 size_t len = dtrace_strlen((char *)s, size);
3979 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3984 if (!DTRACE_INSCRATCH(mstate, size)) {
3985 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3991 remaining = (int64_t)size;
3996 if (index < 0 && index + remaining > 0) {
4002 if (index >= len || index < 0) {
4004 } else if (remaining < 0) {
4005 remaining += len - index;
4006 } else if (index + remaining > size) {
4007 remaining = size - index;
4010 for (i = 0; i < remaining; i++) {
4011 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4017 mstate->dtms_scratch_ptr += size;
4018 regs[rd] = (uintptr_t)d;
4023 case DIF_SUBR_GETMAJOR:
4025 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4027 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4031 case DIF_SUBR_GETMINOR:
4033 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4035 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4039 case DIF_SUBR_DDI_PATHNAME: {
4041 * This one is a galactic mess. We are going to roughly
4042 * emulate ddi_pathname(), but it's made more complicated
4043 * by the fact that we (a) want to include the minor name and
4044 * (b) must proceed iteratively instead of recursively.
4046 uintptr_t dest = mstate->dtms_scratch_ptr;
4047 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4048 char *start = (char *)dest, *end = start + size - 1;
4049 uintptr_t daddr = tupregs[0].dttk_value;
4050 int64_t minor = (int64_t)tupregs[1].dttk_value;
4052 int i, len, depth = 0;
4055 * Due to all the pointer jumping we do and context we must
4056 * rely upon, we just mandate that the user must have kernel
4057 * read privileges to use this routine.
4059 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4060 *flags |= CPU_DTRACE_KPRIV;
4065 if (!DTRACE_INSCRATCH(mstate, size)) {
4066 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4074 * We want to have a name for the minor. In order to do this,
4075 * we need to walk the minor list from the devinfo. We want
4076 * to be sure that we don't infinitely walk a circular list,
4077 * so we check for circularity by sending a scout pointer
4078 * ahead two elements for every element that we iterate over;
4079 * if the list is circular, these will ultimately point to the
4080 * same element. You may recognize this little trick as the
4081 * answer to a stupid interview question -- one that always
4082 * seems to be asked by those who had to have it laboriously
4083 * explained to them, and who can't even concisely describe
4084 * the conditions under which one would be forced to resort to
4085 * this technique. Needless to say, those conditions are
4086 * found here -- and probably only here. Is this the only use
4087 * of this infamous trick in shipping, production code? If it
4088 * isn't, it probably should be...
4091 uintptr_t maddr = dtrace_loadptr(daddr +
4092 offsetof(struct dev_info, devi_minor));
4094 uintptr_t next = offsetof(struct ddi_minor_data, next);
4095 uintptr_t name = offsetof(struct ddi_minor_data,
4096 d_minor) + offsetof(struct ddi_minor, name);
4097 uintptr_t dev = offsetof(struct ddi_minor_data,
4098 d_minor) + offsetof(struct ddi_minor, dev);
4102 scout = dtrace_loadptr(maddr + next);
4104 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4107 m = dtrace_load64(maddr + dev) & MAXMIN64;
4109 m = dtrace_load32(maddr + dev) & MAXMIN;
4112 maddr = dtrace_loadptr(maddr + next);
4117 scout = dtrace_loadptr(scout + next);
4122 scout = dtrace_loadptr(scout + next);
4127 if (scout == maddr) {
4128 *flags |= CPU_DTRACE_ILLOP;
4136 * We have the minor data. Now we need to
4137 * copy the minor's name into the end of the
4140 s = (char *)dtrace_loadptr(maddr + name);
4141 len = dtrace_strlen(s, size);
4143 if (*flags & CPU_DTRACE_FAULT)
4147 if ((end -= (len + 1)) < start)
4153 for (i = 1; i <= len; i++)
4154 end[i] = dtrace_load8((uintptr_t)s++);
4159 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4160 ddi_node_state_t devi_state;
4162 devi_state = dtrace_load32(daddr +
4163 offsetof(struct dev_info, devi_node_state));
4165 if (*flags & CPU_DTRACE_FAULT)
4168 if (devi_state >= DS_INITIALIZED) {
4169 s = (char *)dtrace_loadptr(daddr +
4170 offsetof(struct dev_info, devi_addr));
4171 len = dtrace_strlen(s, size);
4173 if (*flags & CPU_DTRACE_FAULT)
4177 if ((end -= (len + 1)) < start)
4183 for (i = 1; i <= len; i++)
4184 end[i] = dtrace_load8((uintptr_t)s++);
4188 * Now for the node name...
4190 s = (char *)dtrace_loadptr(daddr +
4191 offsetof(struct dev_info, devi_node_name));
4193 daddr = dtrace_loadptr(daddr +
4194 offsetof(struct dev_info, devi_parent));
4197 * If our parent is NULL (that is, if we're the root
4198 * node), we're going to use the special path
4204 len = dtrace_strlen(s, size);
4205 if (*flags & CPU_DTRACE_FAULT)
4208 if ((end -= (len + 1)) < start)
4211 for (i = 1; i <= len; i++)
4212 end[i] = dtrace_load8((uintptr_t)s++);
4215 if (depth++ > dtrace_devdepth_max) {
4216 *flags |= CPU_DTRACE_ILLOP;
4222 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4225 regs[rd] = (uintptr_t)end;
4226 mstate->dtms_scratch_ptr += size;
4233 case DIF_SUBR_STRJOIN: {
4234 char *d = (char *)mstate->dtms_scratch_ptr;
4235 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4236 uintptr_t s1 = tupregs[0].dttk_value;
4237 uintptr_t s2 = tupregs[1].dttk_value;
4240 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4241 !dtrace_strcanload(s2, size, mstate, vstate)) {
4246 if (!DTRACE_INSCRATCH(mstate, size)) {
4247 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4254 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4259 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4267 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4272 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4277 mstate->dtms_scratch_ptr += i;
4278 regs[rd] = (uintptr_t)d;
4284 case DIF_SUBR_LLTOSTR: {
4285 int64_t i = (int64_t)tupregs[0].dttk_value;
4286 int64_t val = i < 0 ? i * -1 : i;
4287 uint64_t size = 22; /* enough room for 2^64 in decimal */
4288 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4290 if (!DTRACE_INSCRATCH(mstate, size)) {
4291 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4296 for (*end-- = '\0'; val; val /= 10)
4297 *end-- = '0' + (val % 10);
4305 regs[rd] = (uintptr_t)end + 1;
4306 mstate->dtms_scratch_ptr += size;
4310 case DIF_SUBR_HTONS:
4311 case DIF_SUBR_NTOHS:
4312 #if BYTE_ORDER == BIG_ENDIAN
4313 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4315 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4320 case DIF_SUBR_HTONL:
4321 case DIF_SUBR_NTOHL:
4322 #if BYTE_ORDER == BIG_ENDIAN
4323 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4325 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4330 case DIF_SUBR_HTONLL:
4331 case DIF_SUBR_NTOHLL:
4332 #if BYTE_ORDER == BIG_ENDIAN
4333 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4335 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4340 case DIF_SUBR_DIRNAME:
4341 case DIF_SUBR_BASENAME: {
4342 char *dest = (char *)mstate->dtms_scratch_ptr;
4343 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4344 uintptr_t src = tupregs[0].dttk_value;
4345 int i, j, len = dtrace_strlen((char *)src, size);
4346 int lastbase = -1, firstbase = -1, lastdir = -1;
4349 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4354 if (!DTRACE_INSCRATCH(mstate, size)) {
4355 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4361 * The basename and dirname for a zero-length string is
4366 src = (uintptr_t)".";
4370 * Start from the back of the string, moving back toward the
4371 * front until we see a character that isn't a slash. That
4372 * character is the last character in the basename.
4374 for (i = len - 1; i >= 0; i--) {
4375 if (dtrace_load8(src + i) != '/')
4383 * Starting from the last character in the basename, move
4384 * towards the front until we find a slash. The character
4385 * that we processed immediately before that is the first
4386 * character in the basename.
4388 for (; i >= 0; i--) {
4389 if (dtrace_load8(src + i) == '/')
4397 * Now keep going until we find a non-slash character. That
4398 * character is the last character in the dirname.
4400 for (; i >= 0; i--) {
4401 if (dtrace_load8(src + i) != '/')
4408 ASSERT(!(lastbase == -1 && firstbase != -1));
4409 ASSERT(!(firstbase == -1 && lastdir != -1));
4411 if (lastbase == -1) {
4413 * We didn't find a non-slash character. We know that
4414 * the length is non-zero, so the whole string must be
4415 * slashes. In either the dirname or the basename
4416 * case, we return '/'.
4418 ASSERT(firstbase == -1);
4419 firstbase = lastbase = lastdir = 0;
4422 if (firstbase == -1) {
4424 * The entire string consists only of a basename
4425 * component. If we're looking for dirname, we need
4426 * to change our string to be just "."; if we're
4427 * looking for a basename, we'll just set the first
4428 * character of the basename to be 0.
4430 if (subr == DIF_SUBR_DIRNAME) {
4431 ASSERT(lastdir == -1);
4432 src = (uintptr_t)".";
4439 if (subr == DIF_SUBR_DIRNAME) {
4440 if (lastdir == -1) {
4442 * We know that we have a slash in the name --
4443 * or lastdir would be set to 0, above. And
4444 * because lastdir is -1, we know that this
4445 * slash must be the first character. (That
4446 * is, the full string must be of the form
4447 * "/basename".) In this case, the last
4448 * character of the directory name is 0.
4456 ASSERT(subr == DIF_SUBR_BASENAME);
4457 ASSERT(firstbase != -1 && lastbase != -1);
4462 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4463 dest[j] = dtrace_load8(src + i);
4466 regs[rd] = (uintptr_t)dest;
4467 mstate->dtms_scratch_ptr += size;
4471 case DIF_SUBR_CLEANPATH: {
4472 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4473 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4474 uintptr_t src = tupregs[0].dttk_value;
4477 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4482 if (!DTRACE_INSCRATCH(mstate, size)) {
4483 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4489 * Move forward, loading each character.
4492 c = dtrace_load8(src + i++);
4494 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4502 c = dtrace_load8(src + i++);
4506 * We have two slashes -- we can just advance
4507 * to the next character.
4514 * This is not "." and it's not ".." -- we can
4515 * just store the "/" and this character and
4523 c = dtrace_load8(src + i++);
4527 * This is a "/./" component. We're not going
4528 * to store anything in the destination buffer;
4529 * we're just going to go to the next component.
4536 * This is not ".." -- we can just store the
4537 * "/." and this character and continue
4546 c = dtrace_load8(src + i++);
4548 if (c != '/' && c != '\0') {
4550 * This is not ".." -- it's "..[mumble]".
4551 * We'll store the "/.." and this character
4552 * and continue processing.
4562 * This is "/../" or "/..\0". We need to back up
4563 * our destination pointer until we find a "/".
4566 while (j != 0 && dest[--j] != '/')
4571 } while (c != '\0');
4574 regs[rd] = (uintptr_t)dest;
4575 mstate->dtms_scratch_ptr += size;
4579 case DIF_SUBR_INET_NTOA:
4580 case DIF_SUBR_INET_NTOA6:
4581 case DIF_SUBR_INET_NTOP: {
4586 if (subr == DIF_SUBR_INET_NTOP) {
4587 af = (int)tupregs[0].dttk_value;
4590 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4594 if (af == AF_INET) {
4599 * Safely load the IPv4 address.
4601 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4604 * Check an IPv4 string will fit in scratch.
4606 size = INET_ADDRSTRLEN;
4607 if (!DTRACE_INSCRATCH(mstate, size)) {
4608 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4612 base = (char *)mstate->dtms_scratch_ptr;
4613 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4616 * Stringify as a dotted decimal quad.
4619 ptr8 = (uint8_t *)&ip4;
4620 for (i = 3; i >= 0; i--) {
4626 for (; val; val /= 10) {
4627 *end-- = '0' + (val % 10);
4634 ASSERT(end + 1 >= base);
4636 } else if (af == AF_INET6) {
4637 struct in6_addr ip6;
4638 int firstzero, tryzero, numzero, v6end;
4640 const char digits[] = "0123456789abcdef";
4643 * Stringify using RFC 1884 convention 2 - 16 bit
4644 * hexadecimal values with a zero-run compression.
4645 * Lower case hexadecimal digits are used.
4646 * eg, fe80::214:4fff:fe0b:76c8.
4647 * The IPv4 embedded form is returned for inet_ntop,
4648 * just the IPv4 string is returned for inet_ntoa6.
4652 * Safely load the IPv6 address.
4655 (void *)(uintptr_t)tupregs[argi].dttk_value,
4656 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4659 * Check an IPv6 string will fit in scratch.
4661 size = INET6_ADDRSTRLEN;
4662 if (!DTRACE_INSCRATCH(mstate, size)) {
4663 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4667 base = (char *)mstate->dtms_scratch_ptr;
4668 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4672 * Find the longest run of 16 bit zero values
4673 * for the single allowed zero compression - "::".
4678 for (i = 0; i < sizeof (struct in6_addr); i++) {
4680 if (ip6._S6_un._S6_u8[i] == 0 &&
4682 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4684 tryzero == -1 && i % 2 == 0) {
4689 if (tryzero != -1 &&
4691 (ip6._S6_un._S6_u8[i] != 0 ||
4693 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4695 i == sizeof (struct in6_addr) - 1)) {
4697 if (i - tryzero <= numzero) {
4702 firstzero = tryzero;
4703 numzero = i - i % 2 - tryzero;
4707 if (ip6._S6_un._S6_u8[i] == 0 &&
4709 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4711 i == sizeof (struct in6_addr) - 1)
4715 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4718 * Check for an IPv4 embedded address.
4720 v6end = sizeof (struct in6_addr) - 2;
4721 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4722 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4723 for (i = sizeof (struct in6_addr) - 1;
4724 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4725 ASSERT(end >= base);
4728 val = ip6._S6_un._S6_u8[i];
4730 val = ip6.__u6_addr.__u6_addr8[i];
4736 for (; val; val /= 10) {
4737 *end-- = '0' + val % 10;
4741 if (i > DTRACE_V4MAPPED_OFFSET)
4745 if (subr == DIF_SUBR_INET_NTOA6)
4749 * Set v6end to skip the IPv4 address that
4750 * we have already stringified.
4756 * Build the IPv6 string by working through the
4757 * address in reverse.
4759 for (i = v6end; i >= 0; i -= 2) {
4760 ASSERT(end >= base);
4762 if (i == firstzero + numzero - 2) {
4769 if (i < 14 && i != firstzero - 2)
4773 val = (ip6._S6_un._S6_u8[i] << 8) +
4774 ip6._S6_un._S6_u8[i + 1];
4776 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4777 ip6.__u6_addr.__u6_addr8[i + 1];
4783 for (; val; val /= 16) {
4784 *end-- = digits[val % 16];
4788 ASSERT(end + 1 >= base);
4792 * The user didn't use AH_INET or AH_INET6.
4794 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4799 inetout: regs[rd] = (uintptr_t)end + 1;
4800 mstate->dtms_scratch_ptr += size;
4804 case DIF_SUBR_MEMREF: {
4805 uintptr_t size = 2 * sizeof(uintptr_t);
4806 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4807 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4809 /* address and length */
4810 memref[0] = tupregs[0].dttk_value;
4811 memref[1] = tupregs[1].dttk_value;
4813 regs[rd] = (uintptr_t) memref;
4814 mstate->dtms_scratch_ptr += scratch_size;
4818 case DIF_SUBR_TYPEREF: {
4819 uintptr_t size = 4 * sizeof(uintptr_t);
4820 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4821 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4823 /* address, num_elements, type_str, type_len */
4824 typeref[0] = tupregs[0].dttk_value;
4825 typeref[1] = tupregs[1].dttk_value;
4826 typeref[2] = tupregs[2].dttk_value;
4827 typeref[3] = tupregs[3].dttk_value;
4829 regs[rd] = (uintptr_t) typeref;
4830 mstate->dtms_scratch_ptr += scratch_size;
4837 * Emulate the execution of DTrace IR instructions specified by the given
4838 * DIF object. This function is deliberately void of assertions as all of
4839 * the necessary checks are handled by a call to dtrace_difo_validate().
4842 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4843 dtrace_vstate_t *vstate, dtrace_state_t *state)
4845 const dif_instr_t *text = difo->dtdo_buf;
4846 const uint_t textlen = difo->dtdo_len;
4847 const char *strtab = difo->dtdo_strtab;
4848 const uint64_t *inttab = difo->dtdo_inttab;
4851 dtrace_statvar_t *svar;
4852 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4854 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4855 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4857 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4858 uint64_t regs[DIF_DIR_NREGS];
4861 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4863 uint_t pc = 0, id, opc = 0;
4869 * We stash the current DIF object into the machine state: we need it
4870 * for subsequent access checking.
4872 mstate->dtms_difo = difo;
4874 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
4876 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4880 r1 = DIF_INSTR_R1(instr);
4881 r2 = DIF_INSTR_R2(instr);
4882 rd = DIF_INSTR_RD(instr);
4884 switch (DIF_INSTR_OP(instr)) {
4886 regs[rd] = regs[r1] | regs[r2];
4889 regs[rd] = regs[r1] ^ regs[r2];
4892 regs[rd] = regs[r1] & regs[r2];
4895 regs[rd] = regs[r1] << regs[r2];
4898 regs[rd] = regs[r1] >> regs[r2];
4901 regs[rd] = regs[r1] - regs[r2];
4904 regs[rd] = regs[r1] + regs[r2];
4907 regs[rd] = regs[r1] * regs[r2];
4910 if (regs[r2] == 0) {
4912 *flags |= CPU_DTRACE_DIVZERO;
4914 regs[rd] = (int64_t)regs[r1] /
4920 if (regs[r2] == 0) {
4922 *flags |= CPU_DTRACE_DIVZERO;
4924 regs[rd] = regs[r1] / regs[r2];
4929 if (regs[r2] == 0) {
4931 *flags |= CPU_DTRACE_DIVZERO;
4933 regs[rd] = (int64_t)regs[r1] %
4939 if (regs[r2] == 0) {
4941 *flags |= CPU_DTRACE_DIVZERO;
4943 regs[rd] = regs[r1] % regs[r2];
4948 regs[rd] = ~regs[r1];
4951 regs[rd] = regs[r1];
4954 cc_r = regs[r1] - regs[r2];
4958 cc_c = regs[r1] < regs[r2];
4961 cc_n = cc_v = cc_c = 0;
4962 cc_z = regs[r1] == 0;
4965 pc = DIF_INSTR_LABEL(instr);
4969 pc = DIF_INSTR_LABEL(instr);
4973 pc = DIF_INSTR_LABEL(instr);
4976 if ((cc_z | (cc_n ^ cc_v)) == 0)
4977 pc = DIF_INSTR_LABEL(instr);
4980 if ((cc_c | cc_z) == 0)
4981 pc = DIF_INSTR_LABEL(instr);
4984 if ((cc_n ^ cc_v) == 0)
4985 pc = DIF_INSTR_LABEL(instr);
4989 pc = DIF_INSTR_LABEL(instr);
4993 pc = DIF_INSTR_LABEL(instr);
4997 pc = DIF_INSTR_LABEL(instr);
5000 if (cc_z | (cc_n ^ cc_v))
5001 pc = DIF_INSTR_LABEL(instr);
5005 pc = DIF_INSTR_LABEL(instr);
5008 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5009 *flags |= CPU_DTRACE_KPRIV;
5015 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5018 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5019 *flags |= CPU_DTRACE_KPRIV;
5025 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5028 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5029 *flags |= CPU_DTRACE_KPRIV;
5035 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5038 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5039 *flags |= CPU_DTRACE_KPRIV;
5045 regs[rd] = dtrace_load8(regs[r1]);
5048 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5049 *flags |= CPU_DTRACE_KPRIV;
5055 regs[rd] = dtrace_load16(regs[r1]);
5058 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5059 *flags |= CPU_DTRACE_KPRIV;
5065 regs[rd] = dtrace_load32(regs[r1]);
5068 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5069 *flags |= CPU_DTRACE_KPRIV;
5075 regs[rd] = dtrace_load64(regs[r1]);
5079 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5082 regs[rd] = (int16_t)
5083 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5086 regs[rd] = (int32_t)
5087 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5091 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5095 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5099 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5103 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5112 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5115 regs[rd] = (uint64_t)(uintptr_t)
5116 (strtab + DIF_INSTR_STRING(instr));
5119 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5120 uintptr_t s1 = regs[r1];
5121 uintptr_t s2 = regs[r2];
5124 !dtrace_strcanload(s1, sz, mstate, vstate))
5127 !dtrace_strcanload(s2, sz, mstate, vstate))
5130 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5138 regs[rd] = dtrace_dif_variable(mstate, state,
5142 id = DIF_INSTR_VAR(instr);
5144 if (id >= DIF_VAR_OTHER_UBASE) {
5147 id -= DIF_VAR_OTHER_UBASE;
5148 svar = vstate->dtvs_globals[id];
5149 ASSERT(svar != NULL);
5150 v = &svar->dtsv_var;
5152 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5153 regs[rd] = svar->dtsv_data;
5157 a = (uintptr_t)svar->dtsv_data;
5159 if (*(uint8_t *)a == UINT8_MAX) {
5161 * If the 0th byte is set to UINT8_MAX
5162 * then this is to be treated as a
5163 * reference to a NULL variable.
5167 regs[rd] = a + sizeof (uint64_t);
5173 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5177 id = DIF_INSTR_VAR(instr);
5179 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5180 id -= DIF_VAR_OTHER_UBASE;
5182 svar = vstate->dtvs_globals[id];
5183 ASSERT(svar != NULL);
5184 v = &svar->dtsv_var;
5186 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5187 uintptr_t a = (uintptr_t)svar->dtsv_data;
5190 ASSERT(svar->dtsv_size != 0);
5192 if (regs[rd] == 0) {
5193 *(uint8_t *)a = UINT8_MAX;
5197 a += sizeof (uint64_t);
5199 if (!dtrace_vcanload(
5200 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5204 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5205 (void *)a, &v->dtdv_type);
5209 svar->dtsv_data = regs[rd];
5214 * There are no DTrace built-in thread-local arrays at
5215 * present. This opcode is saved for future work.
5217 *flags |= CPU_DTRACE_ILLOP;
5222 id = DIF_INSTR_VAR(instr);
5224 if (id < DIF_VAR_OTHER_UBASE) {
5226 * For now, this has no meaning.
5232 id -= DIF_VAR_OTHER_UBASE;
5234 ASSERT(id < vstate->dtvs_nlocals);
5235 ASSERT(vstate->dtvs_locals != NULL);
5237 svar = vstate->dtvs_locals[id];
5238 ASSERT(svar != NULL);
5239 v = &svar->dtsv_var;
5241 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5242 uintptr_t a = (uintptr_t)svar->dtsv_data;
5243 size_t sz = v->dtdv_type.dtdt_size;
5245 sz += sizeof (uint64_t);
5246 ASSERT(svar->dtsv_size == NCPU * sz);
5249 if (*(uint8_t *)a == UINT8_MAX) {
5251 * If the 0th byte is set to UINT8_MAX
5252 * then this is to be treated as a
5253 * reference to a NULL variable.
5257 regs[rd] = a + sizeof (uint64_t);
5263 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5264 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5265 regs[rd] = tmp[curcpu];
5269 id = DIF_INSTR_VAR(instr);
5271 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5272 id -= DIF_VAR_OTHER_UBASE;
5273 ASSERT(id < vstate->dtvs_nlocals);
5275 ASSERT(vstate->dtvs_locals != NULL);
5276 svar = vstate->dtvs_locals[id];
5277 ASSERT(svar != NULL);
5278 v = &svar->dtsv_var;
5280 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5281 uintptr_t a = (uintptr_t)svar->dtsv_data;
5282 size_t sz = v->dtdv_type.dtdt_size;
5284 sz += sizeof (uint64_t);
5285 ASSERT(svar->dtsv_size == NCPU * sz);
5288 if (regs[rd] == 0) {
5289 *(uint8_t *)a = UINT8_MAX;
5293 a += sizeof (uint64_t);
5296 if (!dtrace_vcanload(
5297 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5301 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5302 (void *)a, &v->dtdv_type);
5306 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5307 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5308 tmp[curcpu] = regs[rd];
5312 dtrace_dynvar_t *dvar;
5315 id = DIF_INSTR_VAR(instr);
5316 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5317 id -= DIF_VAR_OTHER_UBASE;
5318 v = &vstate->dtvs_tlocals[id];
5320 key = &tupregs[DIF_DTR_NREGS];
5321 key[0].dttk_value = (uint64_t)id;
5322 key[0].dttk_size = 0;
5323 DTRACE_TLS_THRKEY(key[1].dttk_value);
5324 key[1].dttk_size = 0;
5326 dvar = dtrace_dynvar(dstate, 2, key,
5327 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5335 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5336 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5338 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5345 dtrace_dynvar_t *dvar;
5348 id = DIF_INSTR_VAR(instr);
5349 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5350 id -= DIF_VAR_OTHER_UBASE;
5352 key = &tupregs[DIF_DTR_NREGS];
5353 key[0].dttk_value = (uint64_t)id;
5354 key[0].dttk_size = 0;
5355 DTRACE_TLS_THRKEY(key[1].dttk_value);
5356 key[1].dttk_size = 0;
5357 v = &vstate->dtvs_tlocals[id];
5359 dvar = dtrace_dynvar(dstate, 2, key,
5360 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5361 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5362 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5363 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5366 * Given that we're storing to thread-local data,
5367 * we need to flush our predicate cache.
5369 curthread->t_predcache = 0;
5374 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5375 if (!dtrace_vcanload(
5376 (void *)(uintptr_t)regs[rd],
5377 &v->dtdv_type, mstate, vstate))
5380 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5381 dvar->dtdv_data, &v->dtdv_type);
5383 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5390 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5394 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5395 regs, tupregs, ttop, mstate, state);
5399 if (ttop == DIF_DTR_NREGS) {
5400 *flags |= CPU_DTRACE_TUPOFLOW;
5404 if (r1 == DIF_TYPE_STRING) {
5406 * If this is a string type and the size is 0,
5407 * we'll use the system-wide default string
5408 * size. Note that we are _not_ looking at
5409 * the value of the DTRACEOPT_STRSIZE option;
5410 * had this been set, we would expect to have
5411 * a non-zero size value in the "pushtr".
5413 tupregs[ttop].dttk_size =
5414 dtrace_strlen((char *)(uintptr_t)regs[rd],
5415 regs[r2] ? regs[r2] :
5416 dtrace_strsize_default) + 1;
5418 tupregs[ttop].dttk_size = regs[r2];
5421 tupregs[ttop++].dttk_value = regs[rd];
5425 if (ttop == DIF_DTR_NREGS) {
5426 *flags |= CPU_DTRACE_TUPOFLOW;
5430 tupregs[ttop].dttk_value = regs[rd];
5431 tupregs[ttop++].dttk_size = 0;
5439 case DIF_OP_FLUSHTS:
5444 case DIF_OP_LDTAA: {
5445 dtrace_dynvar_t *dvar;
5446 dtrace_key_t *key = tupregs;
5447 uint_t nkeys = ttop;
5449 id = DIF_INSTR_VAR(instr);
5450 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5451 id -= DIF_VAR_OTHER_UBASE;
5453 key[nkeys].dttk_value = (uint64_t)id;
5454 key[nkeys++].dttk_size = 0;
5456 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5457 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5458 key[nkeys++].dttk_size = 0;
5459 v = &vstate->dtvs_tlocals[id];
5461 v = &vstate->dtvs_globals[id]->dtsv_var;
5464 dvar = dtrace_dynvar(dstate, nkeys, key,
5465 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5466 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5467 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5474 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5475 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5477 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5484 case DIF_OP_STTAA: {
5485 dtrace_dynvar_t *dvar;
5486 dtrace_key_t *key = tupregs;
5487 uint_t nkeys = ttop;
5489 id = DIF_INSTR_VAR(instr);
5490 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5491 id -= DIF_VAR_OTHER_UBASE;
5493 key[nkeys].dttk_value = (uint64_t)id;
5494 key[nkeys++].dttk_size = 0;
5496 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5497 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5498 key[nkeys++].dttk_size = 0;
5499 v = &vstate->dtvs_tlocals[id];
5501 v = &vstate->dtvs_globals[id]->dtsv_var;
5504 dvar = dtrace_dynvar(dstate, nkeys, key,
5505 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5506 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5507 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5508 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5513 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5514 if (!dtrace_vcanload(
5515 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5519 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5520 dvar->dtdv_data, &v->dtdv_type);
5522 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5528 case DIF_OP_ALLOCS: {
5529 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5530 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5533 * Rounding up the user allocation size could have
5534 * overflowed large, bogus allocations (like -1ULL) to
5537 if (size < regs[r1] ||
5538 !DTRACE_INSCRATCH(mstate, size)) {
5539 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5544 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5545 mstate->dtms_scratch_ptr += size;
5551 if (!dtrace_canstore(regs[rd], regs[r2],
5553 *flags |= CPU_DTRACE_BADADDR;
5558 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5561 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5562 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5566 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5567 *flags |= CPU_DTRACE_BADADDR;
5571 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5575 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5576 *flags |= CPU_DTRACE_BADADDR;
5581 *flags |= CPU_DTRACE_BADALIGN;
5585 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5589 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5590 *flags |= CPU_DTRACE_BADADDR;
5595 *flags |= CPU_DTRACE_BADALIGN;
5599 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5603 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5604 *flags |= CPU_DTRACE_BADADDR;
5609 *flags |= CPU_DTRACE_BADALIGN;
5613 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5618 if (!(*flags & CPU_DTRACE_FAULT))
5621 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5622 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5628 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5630 dtrace_probe_t *probe = ecb->dte_probe;
5631 dtrace_provider_t *prov = probe->dtpr_provider;
5632 char c[DTRACE_FULLNAMELEN + 80], *str;
5633 char *msg = "dtrace: breakpoint action at probe ";
5634 char *ecbmsg = " (ecb ";
5635 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5636 uintptr_t val = (uintptr_t)ecb;
5637 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5639 if (dtrace_destructive_disallow)
5643 * It's impossible to be taking action on the NULL probe.
5645 ASSERT(probe != NULL);
5648 * This is a poor man's (destitute man's?) sprintf(): we want to
5649 * print the provider name, module name, function name and name of
5650 * the probe, along with the hex address of the ECB with the breakpoint
5651 * action -- all of which we must place in the character buffer by
5654 while (*msg != '\0')
5657 for (str = prov->dtpv_name; *str != '\0'; str++)
5661 for (str = probe->dtpr_mod; *str != '\0'; str++)
5665 for (str = probe->dtpr_func; *str != '\0'; str++)
5669 for (str = probe->dtpr_name; *str != '\0'; str++)
5672 while (*ecbmsg != '\0')
5675 while (shift >= 0) {
5676 mask = (uintptr_t)0xf << shift;
5678 if (val >= ((uintptr_t)1 << shift))
5679 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5689 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5694 dtrace_action_panic(dtrace_ecb_t *ecb)
5696 dtrace_probe_t *probe = ecb->dte_probe;
5699 * It's impossible to be taking action on the NULL probe.
5701 ASSERT(probe != NULL);
5703 if (dtrace_destructive_disallow)
5706 if (dtrace_panicked != NULL)
5709 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5713 * We won the right to panic. (We want to be sure that only one
5714 * thread calls panic() from dtrace_probe(), and that panic() is
5715 * called exactly once.)
5717 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5718 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5719 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5723 dtrace_action_raise(uint64_t sig)
5725 if (dtrace_destructive_disallow)
5729 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5735 * raise() has a queue depth of 1 -- we ignore all subsequent
5736 * invocations of the raise() action.
5738 if (curthread->t_dtrace_sig == 0)
5739 curthread->t_dtrace_sig = (uint8_t)sig;
5741 curthread->t_sig_check = 1;
5744 struct proc *p = curproc;
5746 kern_psignal(p, sig);
5752 dtrace_action_stop(void)
5754 if (dtrace_destructive_disallow)
5758 if (!curthread->t_dtrace_stop) {
5759 curthread->t_dtrace_stop = 1;
5760 curthread->t_sig_check = 1;
5764 struct proc *p = curproc;
5766 kern_psignal(p, SIGSTOP);
5772 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5775 volatile uint16_t *flags;
5779 cpu_t *cpu = &solaris_cpu[curcpu];
5782 if (dtrace_destructive_disallow)
5785 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5787 now = dtrace_gethrtime();
5789 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5791 * We need to advance the mark to the current time.
5793 cpu->cpu_dtrace_chillmark = now;
5794 cpu->cpu_dtrace_chilled = 0;
5798 * Now check to see if the requested chill time would take us over
5799 * the maximum amount of time allowed in the chill interval. (Or
5800 * worse, if the calculation itself induces overflow.)
5802 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5803 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5804 *flags |= CPU_DTRACE_ILLOP;
5808 while (dtrace_gethrtime() - now < val)
5812 * Normally, we assure that the value of the variable "timestamp" does
5813 * not change within an ECB. The presence of chill() represents an
5814 * exception to this rule, however.
5816 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5817 cpu->cpu_dtrace_chilled += val;
5821 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5822 uint64_t *buf, uint64_t arg)
5824 int nframes = DTRACE_USTACK_NFRAMES(arg);
5825 int strsize = DTRACE_USTACK_STRSIZE(arg);
5826 uint64_t *pcs = &buf[1], *fps;
5827 char *str = (char *)&pcs[nframes];
5828 int size, offs = 0, i, j;
5829 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5830 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5834 * Should be taking a faster path if string space has not been
5837 ASSERT(strsize != 0);
5840 * We will first allocate some temporary space for the frame pointers.
5842 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5843 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5844 (nframes * sizeof (uint64_t));
5846 if (!DTRACE_INSCRATCH(mstate, size)) {
5848 * Not enough room for our frame pointers -- need to indicate
5849 * that we ran out of scratch space.
5851 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5855 mstate->dtms_scratch_ptr += size;
5856 saved = mstate->dtms_scratch_ptr;
5859 * Now get a stack with both program counters and frame pointers.
5861 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5862 dtrace_getufpstack(buf, fps, nframes + 1);
5863 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5866 * If that faulted, we're cooked.
5868 if (*flags & CPU_DTRACE_FAULT)
5872 * Now we want to walk up the stack, calling the USTACK helper. For
5873 * each iteration, we restore the scratch pointer.
5875 for (i = 0; i < nframes; i++) {
5876 mstate->dtms_scratch_ptr = saved;
5878 if (offs >= strsize)
5881 sym = (char *)(uintptr_t)dtrace_helper(
5882 DTRACE_HELPER_ACTION_USTACK,
5883 mstate, state, pcs[i], fps[i]);
5886 * If we faulted while running the helper, we're going to
5887 * clear the fault and null out the corresponding string.
5889 if (*flags & CPU_DTRACE_FAULT) {
5890 *flags &= ~CPU_DTRACE_FAULT;
5900 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5903 * Now copy in the string that the helper returned to us.
5905 for (j = 0; offs + j < strsize; j++) {
5906 if ((str[offs + j] = sym[j]) == '\0')
5910 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5915 if (offs >= strsize) {
5917 * If we didn't have room for all of the strings, we don't
5918 * abort processing -- this needn't be a fatal error -- but we
5919 * still want to increment a counter (dts_stkstroverflows) to
5920 * allow this condition to be warned about. (If this is from
5921 * a jstack() action, it is easily tuned via jstackstrsize.)
5923 dtrace_error(&state->dts_stkstroverflows);
5926 while (offs < strsize)
5930 mstate->dtms_scratch_ptr = old;
5934 * If you're looking for the epicenter of DTrace, you just found it. This
5935 * is the function called by the provider to fire a probe -- from which all
5936 * subsequent probe-context DTrace activity emanates.
5939 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5940 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5942 processorid_t cpuid;
5943 dtrace_icookie_t cookie;
5944 dtrace_probe_t *probe;
5945 dtrace_mstate_t mstate;
5947 dtrace_action_t *act;
5951 volatile uint16_t *flags;
5954 if (panicstr != NULL)
5959 * Kick out immediately if this CPU is still being born (in which case
5960 * curthread will be set to -1) or the current thread can't allow
5961 * probes in its current context.
5963 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
5967 cookie = dtrace_interrupt_disable();
5968 probe = dtrace_probes[id - 1];
5970 onintr = CPU_ON_INTR(CPU);
5972 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
5973 probe->dtpr_predcache == curthread->t_predcache) {
5975 * We have hit in the predicate cache; we know that
5976 * this predicate would evaluate to be false.
5978 dtrace_interrupt_enable(cookie);
5983 if (panic_quiesce) {
5985 if (panicstr != NULL) {
5988 * We don't trace anything if we're panicking.
5990 dtrace_interrupt_enable(cookie);
5994 now = dtrace_gethrtime();
5995 vtime = dtrace_vtime_references != 0;
5997 if (vtime && curthread->t_dtrace_start)
5998 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6000 mstate.dtms_difo = NULL;
6001 mstate.dtms_probe = probe;
6002 mstate.dtms_strtok = 0;
6003 mstate.dtms_arg[0] = arg0;
6004 mstate.dtms_arg[1] = arg1;
6005 mstate.dtms_arg[2] = arg2;
6006 mstate.dtms_arg[3] = arg3;
6007 mstate.dtms_arg[4] = arg4;
6009 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6011 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6012 dtrace_predicate_t *pred = ecb->dte_predicate;
6013 dtrace_state_t *state = ecb->dte_state;
6014 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6015 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6016 dtrace_vstate_t *vstate = &state->dts_vstate;
6017 dtrace_provider_t *prov = probe->dtpr_provider;
6022 * A little subtlety with the following (seemingly innocuous)
6023 * declaration of the automatic 'val': by looking at the
6024 * code, you might think that it could be declared in the
6025 * action processing loop, below. (That is, it's only used in
6026 * the action processing loop.) However, it must be declared
6027 * out of that scope because in the case of DIF expression
6028 * arguments to aggregating actions, one iteration of the
6029 * action loop will use the last iteration's value.
6033 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6034 *flags &= ~CPU_DTRACE_ERROR;
6036 if (prov == dtrace_provider) {
6038 * If dtrace itself is the provider of this probe,
6039 * we're only going to continue processing the ECB if
6040 * arg0 (the dtrace_state_t) is equal to the ECB's
6041 * creating state. (This prevents disjoint consumers
6042 * from seeing one another's metaprobes.)
6044 if (arg0 != (uint64_t)(uintptr_t)state)
6048 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6050 * We're not currently active. If our provider isn't
6051 * the dtrace pseudo provider, we're not interested.
6053 if (prov != dtrace_provider)
6057 * Now we must further check if we are in the BEGIN
6058 * probe. If we are, we will only continue processing
6059 * if we're still in WARMUP -- if one BEGIN enabling
6060 * has invoked the exit() action, we don't want to
6061 * evaluate subsequent BEGIN enablings.
6063 if (probe->dtpr_id == dtrace_probeid_begin &&
6064 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6065 ASSERT(state->dts_activity ==
6066 DTRACE_ACTIVITY_DRAINING);
6071 if (ecb->dte_cond) {
6073 * If the dte_cond bits indicate that this
6074 * consumer is only allowed to see user-mode firings
6075 * of this probe, call the provider's dtps_usermode()
6076 * entry point to check that the probe was fired
6077 * while in a user context. Skip this ECB if that's
6080 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6081 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6082 probe->dtpr_id, probe->dtpr_arg) == 0)
6087 * This is more subtle than it looks. We have to be
6088 * absolutely certain that CRED() isn't going to
6089 * change out from under us so it's only legit to
6090 * examine that structure if we're in constrained
6091 * situations. Currently, the only times we'll this
6092 * check is if a non-super-user has enabled the
6093 * profile or syscall providers -- providers that
6094 * allow visibility of all processes. For the
6095 * profile case, the check above will ensure that
6096 * we're examining a user context.
6098 if (ecb->dte_cond & DTRACE_COND_OWNER) {
6101 ecb->dte_state->dts_cred.dcr_cred;
6104 ASSERT(s_cr != NULL);
6106 if ((cr = CRED()) == NULL ||
6107 s_cr->cr_uid != cr->cr_uid ||
6108 s_cr->cr_uid != cr->cr_ruid ||
6109 s_cr->cr_uid != cr->cr_suid ||
6110 s_cr->cr_gid != cr->cr_gid ||
6111 s_cr->cr_gid != cr->cr_rgid ||
6112 s_cr->cr_gid != cr->cr_sgid ||
6113 (proc = ttoproc(curthread)) == NULL ||
6114 (proc->p_flag & SNOCD))
6118 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6121 ecb->dte_state->dts_cred.dcr_cred;
6123 ASSERT(s_cr != NULL);
6125 if ((cr = CRED()) == NULL ||
6126 s_cr->cr_zone->zone_id !=
6127 cr->cr_zone->zone_id)
6133 if (now - state->dts_alive > dtrace_deadman_timeout) {
6135 * We seem to be dead. Unless we (a) have kernel
6136 * destructive permissions (b) have expicitly enabled
6137 * destructive actions and (c) destructive actions have
6138 * not been disabled, we're going to transition into
6139 * the KILLED state, from which no further processing
6140 * on this state will be performed.
6142 if (!dtrace_priv_kernel_destructive(state) ||
6143 !state->dts_cred.dcr_destructive ||
6144 dtrace_destructive_disallow) {
6145 void *activity = &state->dts_activity;
6146 dtrace_activity_t current;
6149 current = state->dts_activity;
6150 } while (dtrace_cas32(activity, current,
6151 DTRACE_ACTIVITY_KILLED) != current);
6157 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6158 ecb->dte_alignment, state, &mstate)) < 0)
6161 tomax = buf->dtb_tomax;
6162 ASSERT(tomax != NULL);
6164 if (ecb->dte_size != 0)
6165 DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
6167 mstate.dtms_epid = ecb->dte_epid;
6168 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6170 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6171 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6173 mstate.dtms_access = 0;
6176 dtrace_difo_t *dp = pred->dtp_difo;
6179 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6181 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6182 dtrace_cacheid_t cid = probe->dtpr_predcache;
6184 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6186 * Update the predicate cache...
6188 ASSERT(cid == pred->dtp_cacheid);
6189 curthread->t_predcache = cid;
6196 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6197 act != NULL; act = act->dta_next) {
6200 dtrace_recdesc_t *rec = &act->dta_rec;
6202 size = rec->dtrd_size;
6203 valoffs = offs + rec->dtrd_offset;
6205 if (DTRACEACT_ISAGG(act->dta_kind)) {
6207 dtrace_aggregation_t *agg;
6209 agg = (dtrace_aggregation_t *)act;
6211 if ((dp = act->dta_difo) != NULL)
6212 v = dtrace_dif_emulate(dp,
6213 &mstate, vstate, state);
6215 if (*flags & CPU_DTRACE_ERROR)
6219 * Note that we always pass the expression
6220 * value from the previous iteration of the
6221 * action loop. This value will only be used
6222 * if there is an expression argument to the
6223 * aggregating action, denoted by the
6224 * dtag_hasarg field.
6226 dtrace_aggregate(agg, buf,
6227 offs, aggbuf, v, val);
6231 switch (act->dta_kind) {
6232 case DTRACEACT_STOP:
6233 if (dtrace_priv_proc_destructive(state))
6234 dtrace_action_stop();
6237 case DTRACEACT_BREAKPOINT:
6238 if (dtrace_priv_kernel_destructive(state))
6239 dtrace_action_breakpoint(ecb);
6242 case DTRACEACT_PANIC:
6243 if (dtrace_priv_kernel_destructive(state))
6244 dtrace_action_panic(ecb);
6247 case DTRACEACT_STACK:
6248 if (!dtrace_priv_kernel(state))
6251 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6252 size / sizeof (pc_t), probe->dtpr_aframes,
6253 DTRACE_ANCHORED(probe) ? NULL :
6257 case DTRACEACT_JSTACK:
6258 case DTRACEACT_USTACK:
6259 if (!dtrace_priv_proc(state))
6263 * See comment in DIF_VAR_PID.
6265 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6267 int depth = DTRACE_USTACK_NFRAMES(
6270 dtrace_bzero((void *)(tomax + valoffs),
6271 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6272 + depth * sizeof (uint64_t));
6277 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6278 curproc->p_dtrace_helpers != NULL) {
6280 * This is the slow path -- we have
6281 * allocated string space, and we're
6282 * getting the stack of a process that
6283 * has helpers. Call into a separate
6284 * routine to perform this processing.
6286 dtrace_action_ustack(&mstate, state,
6287 (uint64_t *)(tomax + valoffs),
6292 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6293 dtrace_getupcstack((uint64_t *)
6295 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6296 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6306 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6308 if (*flags & CPU_DTRACE_ERROR)
6311 switch (act->dta_kind) {
6312 case DTRACEACT_SPECULATE:
6313 ASSERT(buf == &state->dts_buffer[cpuid]);
6314 buf = dtrace_speculation_buffer(state,
6318 *flags |= CPU_DTRACE_DROP;
6322 offs = dtrace_buffer_reserve(buf,
6323 ecb->dte_needed, ecb->dte_alignment,
6327 *flags |= CPU_DTRACE_DROP;
6331 tomax = buf->dtb_tomax;
6332 ASSERT(tomax != NULL);
6334 if (ecb->dte_size != 0)
6335 DTRACE_STORE(uint32_t, tomax, offs,
6339 case DTRACEACT_PRINTM: {
6340 /* The DIF returns a 'memref'. */
6341 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6343 /* Get the size from the memref. */
6347 * Check if the size exceeds the allocated
6350 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6352 *flags |= CPU_DTRACE_DROP;
6356 /* Store the size in the buffer first. */
6357 DTRACE_STORE(uintptr_t, tomax,
6361 * Offset the buffer address to the start
6364 valoffs += sizeof(uintptr_t);
6367 * Reset to the memory address rather than
6368 * the memref array, then let the BYREF
6369 * code below do the work to store the
6370 * memory data in the buffer.
6376 case DTRACEACT_PRINTT: {
6377 /* The DIF returns a 'typeref'. */
6378 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6383 * Get the type string length and round it
6384 * up so that the data that follows is
6385 * aligned for easy access.
6387 size_t typs = strlen((char *) typeref[2]) + 1;
6388 typs = roundup(typs, sizeof(uintptr_t));
6391 *Get the size from the typeref using the
6392 * number of elements and the type size.
6394 size = typeref[1] * typeref[3];
6397 * Check if the size exceeds the allocated
6400 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6402 *flags |= CPU_DTRACE_DROP;
6406 /* Store the size in the buffer first. */
6407 DTRACE_STORE(uintptr_t, tomax,
6409 valoffs += sizeof(uintptr_t);
6411 /* Store the type size in the buffer. */
6412 DTRACE_STORE(uintptr_t, tomax,
6413 valoffs, typeref[3]);
6414 valoffs += sizeof(uintptr_t);
6418 for (s = 0; s < typs; s++) {
6420 c = dtrace_load8(val++);
6422 DTRACE_STORE(uint8_t, tomax,
6427 * Reset to the memory address rather than
6428 * the typeref array, then let the BYREF
6429 * code below do the work to store the
6430 * memory data in the buffer.
6436 case DTRACEACT_CHILL:
6437 if (dtrace_priv_kernel_destructive(state))
6438 dtrace_action_chill(&mstate, val);
6441 case DTRACEACT_RAISE:
6442 if (dtrace_priv_proc_destructive(state))
6443 dtrace_action_raise(val);
6446 case DTRACEACT_COMMIT:
6450 * We need to commit our buffer state.
6453 buf->dtb_offset = offs + ecb->dte_size;
6454 buf = &state->dts_buffer[cpuid];
6455 dtrace_speculation_commit(state, cpuid, val);
6459 case DTRACEACT_DISCARD:
6460 dtrace_speculation_discard(state, cpuid, val);
6463 case DTRACEACT_DIFEXPR:
6464 case DTRACEACT_LIBACT:
6465 case DTRACEACT_PRINTF:
6466 case DTRACEACT_PRINTA:
6467 case DTRACEACT_SYSTEM:
6468 case DTRACEACT_FREOPEN:
6473 if (!dtrace_priv_kernel(state))
6477 case DTRACEACT_USYM:
6478 case DTRACEACT_UMOD:
6479 case DTRACEACT_UADDR: {
6481 struct pid *pid = curthread->t_procp->p_pidp;
6484 if (!dtrace_priv_proc(state))
6487 DTRACE_STORE(uint64_t, tomax,
6489 valoffs, (uint64_t)pid->pid_id);
6491 valoffs, (uint64_t) curproc->p_pid);
6493 DTRACE_STORE(uint64_t, tomax,
6494 valoffs + sizeof (uint64_t), val);
6499 case DTRACEACT_EXIT: {
6501 * For the exit action, we are going to attempt
6502 * to atomically set our activity to be
6503 * draining. If this fails (either because
6504 * another CPU has beat us to the exit action,
6505 * or because our current activity is something
6506 * other than ACTIVE or WARMUP), we will
6507 * continue. This assures that the exit action
6508 * can be successfully recorded at most once
6509 * when we're in the ACTIVE state. If we're
6510 * encountering the exit() action while in
6511 * COOLDOWN, however, we want to honor the new
6512 * status code. (We know that we're the only
6513 * thread in COOLDOWN, so there is no race.)
6515 void *activity = &state->dts_activity;
6516 dtrace_activity_t current = state->dts_activity;
6518 if (current == DTRACE_ACTIVITY_COOLDOWN)
6521 if (current != DTRACE_ACTIVITY_WARMUP)
6522 current = DTRACE_ACTIVITY_ACTIVE;
6524 if (dtrace_cas32(activity, current,
6525 DTRACE_ACTIVITY_DRAINING) != current) {
6526 *flags |= CPU_DTRACE_DROP;
6537 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6538 uintptr_t end = valoffs + size;
6540 if (!dtrace_vcanload((void *)(uintptr_t)val,
6541 &dp->dtdo_rtype, &mstate, vstate))
6545 * If this is a string, we're going to only
6546 * load until we find the zero byte -- after
6547 * which we'll store zero bytes.
6549 if (dp->dtdo_rtype.dtdt_kind ==
6552 int intuple = act->dta_intuple;
6555 for (s = 0; s < size; s++) {
6557 c = dtrace_load8(val++);
6559 DTRACE_STORE(uint8_t, tomax,
6562 if (c == '\0' && intuple)
6569 while (valoffs < end) {
6570 DTRACE_STORE(uint8_t, tomax, valoffs++,
6571 dtrace_load8(val++));
6581 case sizeof (uint8_t):
6582 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6584 case sizeof (uint16_t):
6585 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6587 case sizeof (uint32_t):
6588 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6590 case sizeof (uint64_t):
6591 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6595 * Any other size should have been returned by
6596 * reference, not by value.
6603 if (*flags & CPU_DTRACE_DROP)
6606 if (*flags & CPU_DTRACE_FAULT) {
6608 dtrace_action_t *err;
6612 if (probe->dtpr_id == dtrace_probeid_error) {
6614 * There's nothing we can do -- we had an
6615 * error on the error probe. We bump an
6616 * error counter to at least indicate that
6617 * this condition happened.
6619 dtrace_error(&state->dts_dblerrors);
6625 * Before recursing on dtrace_probe(), we
6626 * need to explicitly clear out our start
6627 * time to prevent it from being accumulated
6628 * into t_dtrace_vtime.
6630 curthread->t_dtrace_start = 0;
6634 * Iterate over the actions to figure out which action
6635 * we were processing when we experienced the error.
6636 * Note that act points _past_ the faulting action; if
6637 * act is ecb->dte_action, the fault was in the
6638 * predicate, if it's ecb->dte_action->dta_next it's
6639 * in action #1, and so on.
6641 for (err = ecb->dte_action, ndx = 0;
6642 err != act; err = err->dta_next, ndx++)
6645 dtrace_probe_error(state, ecb->dte_epid, ndx,
6646 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6647 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6648 cpu_core[cpuid].cpuc_dtrace_illval);
6654 buf->dtb_offset = offs + ecb->dte_size;
6658 curthread->t_dtrace_start = dtrace_gethrtime();
6660 dtrace_interrupt_enable(cookie);
6664 * DTrace Probe Hashing Functions
6666 * The functions in this section (and indeed, the functions in remaining
6667 * sections) are not _called_ from probe context. (Any exceptions to this are
6668 * marked with a "Note:".) Rather, they are called from elsewhere in the
6669 * DTrace framework to look-up probes in, add probes to and remove probes from
6670 * the DTrace probe hashes. (Each probe is hashed by each element of the
6671 * probe tuple -- allowing for fast lookups, regardless of what was
6675 dtrace_hash_str(const char *p)
6681 hval = (hval << 4) + *p++;
6682 if ((g = (hval & 0xf0000000)) != 0)
6689 static dtrace_hash_t *
6690 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6692 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6694 hash->dth_stroffs = stroffs;
6695 hash->dth_nextoffs = nextoffs;
6696 hash->dth_prevoffs = prevoffs;
6699 hash->dth_mask = hash->dth_size - 1;
6701 hash->dth_tab = kmem_zalloc(hash->dth_size *
6702 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6708 dtrace_hash_destroy(dtrace_hash_t *hash)
6713 for (i = 0; i < hash->dth_size; i++)
6714 ASSERT(hash->dth_tab[i] == NULL);
6717 kmem_free(hash->dth_tab,
6718 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6719 kmem_free(hash, sizeof (dtrace_hash_t));
6723 dtrace_hash_resize(dtrace_hash_t *hash)
6725 int size = hash->dth_size, i, ndx;
6726 int new_size = hash->dth_size << 1;
6727 int new_mask = new_size - 1;
6728 dtrace_hashbucket_t **new_tab, *bucket, *next;
6730 ASSERT((new_size & new_mask) == 0);
6732 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6734 for (i = 0; i < size; i++) {
6735 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6736 dtrace_probe_t *probe = bucket->dthb_chain;
6738 ASSERT(probe != NULL);
6739 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6741 next = bucket->dthb_next;
6742 bucket->dthb_next = new_tab[ndx];
6743 new_tab[ndx] = bucket;
6747 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6748 hash->dth_tab = new_tab;
6749 hash->dth_size = new_size;
6750 hash->dth_mask = new_mask;
6754 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6756 int hashval = DTRACE_HASHSTR(hash, new);
6757 int ndx = hashval & hash->dth_mask;
6758 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6759 dtrace_probe_t **nextp, **prevp;
6761 for (; bucket != NULL; bucket = bucket->dthb_next) {
6762 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6766 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6767 dtrace_hash_resize(hash);
6768 dtrace_hash_add(hash, new);
6772 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6773 bucket->dthb_next = hash->dth_tab[ndx];
6774 hash->dth_tab[ndx] = bucket;
6775 hash->dth_nbuckets++;
6778 nextp = DTRACE_HASHNEXT(hash, new);
6779 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6780 *nextp = bucket->dthb_chain;
6782 if (bucket->dthb_chain != NULL) {
6783 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6784 ASSERT(*prevp == NULL);
6788 bucket->dthb_chain = new;
6792 static dtrace_probe_t *
6793 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6795 int hashval = DTRACE_HASHSTR(hash, template);
6796 int ndx = hashval & hash->dth_mask;
6797 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6799 for (; bucket != NULL; bucket = bucket->dthb_next) {
6800 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6801 return (bucket->dthb_chain);
6808 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6810 int hashval = DTRACE_HASHSTR(hash, template);
6811 int ndx = hashval & hash->dth_mask;
6812 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6814 for (; bucket != NULL; bucket = bucket->dthb_next) {
6815 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6816 return (bucket->dthb_len);
6823 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6825 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6826 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6828 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6829 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6832 * Find the bucket that we're removing this probe from.
6834 for (; bucket != NULL; bucket = bucket->dthb_next) {
6835 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6839 ASSERT(bucket != NULL);
6841 if (*prevp == NULL) {
6842 if (*nextp == NULL) {
6844 * The removed probe was the only probe on this
6845 * bucket; we need to remove the bucket.
6847 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6849 ASSERT(bucket->dthb_chain == probe);
6853 hash->dth_tab[ndx] = bucket->dthb_next;
6855 while (b->dthb_next != bucket)
6857 b->dthb_next = bucket->dthb_next;
6860 ASSERT(hash->dth_nbuckets > 0);
6861 hash->dth_nbuckets--;
6862 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6866 bucket->dthb_chain = *nextp;
6868 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6872 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6876 * DTrace Utility Functions
6878 * These are random utility functions that are _not_ called from probe context.
6881 dtrace_badattr(const dtrace_attribute_t *a)
6883 return (a->dtat_name > DTRACE_STABILITY_MAX ||
6884 a->dtat_data > DTRACE_STABILITY_MAX ||
6885 a->dtat_class > DTRACE_CLASS_MAX);
6889 * Return a duplicate copy of a string. If the specified string is NULL,
6890 * this function returns a zero-length string.
6893 dtrace_strdup(const char *str)
6895 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6898 (void) strcpy(new, str);
6903 #define DTRACE_ISALPHA(c) \
6904 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6907 dtrace_badname(const char *s)
6911 if (s == NULL || (c = *s++) == '\0')
6914 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6917 while ((c = *s++) != '\0') {
6918 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6919 c != '-' && c != '_' && c != '.' && c != '`')
6927 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6932 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6934 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6936 priv = DTRACE_PRIV_ALL;
6938 *uidp = crgetuid(cr);
6939 *zoneidp = crgetzoneid(cr);
6942 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6943 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6944 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6945 priv |= DTRACE_PRIV_USER;
6946 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6947 priv |= DTRACE_PRIV_PROC;
6948 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6949 priv |= DTRACE_PRIV_OWNER;
6950 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6951 priv |= DTRACE_PRIV_ZONEOWNER;
6954 priv = DTRACE_PRIV_ALL;
6960 #ifdef DTRACE_ERRDEBUG
6962 dtrace_errdebug(const char *str)
6964 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
6967 mutex_enter(&dtrace_errlock);
6968 dtrace_errlast = str;
6969 dtrace_errthread = curthread;
6971 while (occupied++ < DTRACE_ERRHASHSZ) {
6972 if (dtrace_errhash[hval].dter_msg == str) {
6973 dtrace_errhash[hval].dter_count++;
6977 if (dtrace_errhash[hval].dter_msg != NULL) {
6978 hval = (hval + 1) % DTRACE_ERRHASHSZ;
6982 dtrace_errhash[hval].dter_msg = str;
6983 dtrace_errhash[hval].dter_count = 1;
6987 panic("dtrace: undersized error hash");
6989 mutex_exit(&dtrace_errlock);
6994 * DTrace Matching Functions
6996 * These functions are used to match groups of probes, given some elements of
6997 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7000 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7003 if (priv != DTRACE_PRIV_ALL) {
7004 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7005 uint32_t match = priv & ppriv;
7008 * No PRIV_DTRACE_* privileges...
7010 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7011 DTRACE_PRIV_KERNEL)) == 0)
7015 * No matching bits, but there were bits to match...
7017 if (match == 0 && ppriv != 0)
7021 * Need to have permissions to the process, but don't...
7023 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7024 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7029 * Need to be in the same zone unless we possess the
7030 * privilege to examine all zones.
7032 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7033 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7042 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7043 * consists of input pattern strings and an ops-vector to evaluate them.
7044 * This function returns >0 for match, 0 for no match, and <0 for error.
7047 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7048 uint32_t priv, uid_t uid, zoneid_t zoneid)
7050 dtrace_provider_t *pvp = prp->dtpr_provider;
7053 if (pvp->dtpv_defunct)
7056 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7059 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7062 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7065 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7068 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7075 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7076 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7077 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7078 * In addition, all of the recursion cases except for '*' matching have been
7079 * unwound. For '*', we still implement recursive evaluation, but a depth
7080 * counter is maintained and matching is aborted if we recurse too deep.
7081 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7084 dtrace_match_glob(const char *s, const char *p, int depth)
7090 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7094 s = ""; /* treat NULL as empty string */
7103 if ((c = *p++) == '\0')
7104 return (s1 == '\0');
7108 int ok = 0, notflag = 0;
7119 if ((c = *p++) == '\0')
7123 if (c == '-' && lc != '\0' && *p != ']') {
7124 if ((c = *p++) == '\0')
7126 if (c == '\\' && (c = *p++) == '\0')
7130 if (s1 < lc || s1 > c)
7134 } else if (lc <= s1 && s1 <= c)
7137 } else if (c == '\\' && (c = *p++) == '\0')
7140 lc = c; /* save left-hand 'c' for next iteration */
7150 if ((c = *p++) == '\0')
7162 if ((c = *p++) == '\0')
7178 p++; /* consecutive *'s are identical to a single one */
7183 for (s = olds; *s != '\0'; s++) {
7184 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7194 dtrace_match_string(const char *s, const char *p, int depth)
7196 return (s != NULL && strcmp(s, p) == 0);
7201 dtrace_match_nul(const char *s, const char *p, int depth)
7203 return (1); /* always match the empty pattern */
7208 dtrace_match_nonzero(const char *s, const char *p, int depth)
7210 return (s != NULL && s[0] != '\0');
7214 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7215 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7217 dtrace_probe_t template, *probe;
7218 dtrace_hash_t *hash = NULL;
7219 int len, best = INT_MAX, nmatched = 0;
7222 ASSERT(MUTEX_HELD(&dtrace_lock));
7225 * If the probe ID is specified in the key, just lookup by ID and
7226 * invoke the match callback once if a matching probe is found.
7228 if (pkp->dtpk_id != DTRACE_IDNONE) {
7229 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7230 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7231 (void) (*matched)(probe, arg);
7237 template.dtpr_mod = (char *)pkp->dtpk_mod;
7238 template.dtpr_func = (char *)pkp->dtpk_func;
7239 template.dtpr_name = (char *)pkp->dtpk_name;
7242 * We want to find the most distinct of the module name, function
7243 * name, and name. So for each one that is not a glob pattern or
7244 * empty string, we perform a lookup in the corresponding hash and
7245 * use the hash table with the fewest collisions to do our search.
7247 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7248 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7250 hash = dtrace_bymod;
7253 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7254 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7256 hash = dtrace_byfunc;
7259 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7260 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7262 hash = dtrace_byname;
7266 * If we did not select a hash table, iterate over every probe and
7267 * invoke our callback for each one that matches our input probe key.
7270 for (i = 0; i < dtrace_nprobes; i++) {
7271 if ((probe = dtrace_probes[i]) == NULL ||
7272 dtrace_match_probe(probe, pkp, priv, uid,
7278 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7286 * If we selected a hash table, iterate over each probe of the same key
7287 * name and invoke the callback for every probe that matches the other
7288 * attributes of our input probe key.
7290 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7291 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7293 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7298 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7306 * Return the function pointer dtrace_probecmp() should use to compare the
7307 * specified pattern with a string. For NULL or empty patterns, we select
7308 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7309 * For non-empty non-glob strings, we use dtrace_match_string().
7311 static dtrace_probekey_f *
7312 dtrace_probekey_func(const char *p)
7316 if (p == NULL || *p == '\0')
7317 return (&dtrace_match_nul);
7319 while ((c = *p++) != '\0') {
7320 if (c == '[' || c == '?' || c == '*' || c == '\\')
7321 return (&dtrace_match_glob);
7324 return (&dtrace_match_string);
7328 * Build a probe comparison key for use with dtrace_match_probe() from the
7329 * given probe description. By convention, a null key only matches anchored
7330 * probes: if each field is the empty string, reset dtpk_fmatch to
7331 * dtrace_match_nonzero().
7334 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7336 pkp->dtpk_prov = pdp->dtpd_provider;
7337 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7339 pkp->dtpk_mod = pdp->dtpd_mod;
7340 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7342 pkp->dtpk_func = pdp->dtpd_func;
7343 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7345 pkp->dtpk_name = pdp->dtpd_name;
7346 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7348 pkp->dtpk_id = pdp->dtpd_id;
7350 if (pkp->dtpk_id == DTRACE_IDNONE &&
7351 pkp->dtpk_pmatch == &dtrace_match_nul &&
7352 pkp->dtpk_mmatch == &dtrace_match_nul &&
7353 pkp->dtpk_fmatch == &dtrace_match_nul &&
7354 pkp->dtpk_nmatch == &dtrace_match_nul)
7355 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7359 * DTrace Provider-to-Framework API Functions
7361 * These functions implement much of the Provider-to-Framework API, as
7362 * described in <sys/dtrace.h>. The parts of the API not in this section are
7363 * the functions in the API for probe management (found below), and
7364 * dtrace_probe() itself (found above).
7368 * Register the calling provider with the DTrace framework. This should
7369 * generally be called by DTrace providers in their attach(9E) entry point.
7372 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7373 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7375 dtrace_provider_t *provider;
7377 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7378 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7379 "arguments", name ? name : "<NULL>");
7383 if (name[0] == '\0' || dtrace_badname(name)) {
7384 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7385 "provider name", name);
7389 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7390 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7391 pops->dtps_destroy == NULL ||
7392 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7393 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7394 "provider ops", name);
7398 if (dtrace_badattr(&pap->dtpa_provider) ||
7399 dtrace_badattr(&pap->dtpa_mod) ||
7400 dtrace_badattr(&pap->dtpa_func) ||
7401 dtrace_badattr(&pap->dtpa_name) ||
7402 dtrace_badattr(&pap->dtpa_args)) {
7403 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7404 "provider attributes", name);
7408 if (priv & ~DTRACE_PRIV_ALL) {
7409 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7410 "privilege attributes", name);
7414 if ((priv & DTRACE_PRIV_KERNEL) &&
7415 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7416 pops->dtps_usermode == NULL) {
7417 cmn_err(CE_WARN, "failed to register provider '%s': need "
7418 "dtps_usermode() op for given privilege attributes", name);
7422 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7423 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7424 (void) strcpy(provider->dtpv_name, name);
7426 provider->dtpv_attr = *pap;
7427 provider->dtpv_priv.dtpp_flags = priv;
7429 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7430 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7432 provider->dtpv_pops = *pops;
7434 if (pops->dtps_provide == NULL) {
7435 ASSERT(pops->dtps_provide_module != NULL);
7436 provider->dtpv_pops.dtps_provide =
7437 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7440 if (pops->dtps_provide_module == NULL) {
7441 ASSERT(pops->dtps_provide != NULL);
7442 provider->dtpv_pops.dtps_provide_module =
7443 (void (*)(void *, modctl_t *))dtrace_nullop;
7446 if (pops->dtps_suspend == NULL) {
7447 ASSERT(pops->dtps_resume == NULL);
7448 provider->dtpv_pops.dtps_suspend =
7449 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7450 provider->dtpv_pops.dtps_resume =
7451 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7454 provider->dtpv_arg = arg;
7455 *idp = (dtrace_provider_id_t)provider;
7457 if (pops == &dtrace_provider_ops) {
7458 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7459 ASSERT(MUTEX_HELD(&dtrace_lock));
7460 ASSERT(dtrace_anon.dta_enabling == NULL);
7463 * We make sure that the DTrace provider is at the head of
7464 * the provider chain.
7466 provider->dtpv_next = dtrace_provider;
7467 dtrace_provider = provider;
7471 mutex_enter(&dtrace_provider_lock);
7472 mutex_enter(&dtrace_lock);
7475 * If there is at least one provider registered, we'll add this
7476 * provider after the first provider.
7478 if (dtrace_provider != NULL) {
7479 provider->dtpv_next = dtrace_provider->dtpv_next;
7480 dtrace_provider->dtpv_next = provider;
7482 dtrace_provider = provider;
7485 if (dtrace_retained != NULL) {
7486 dtrace_enabling_provide(provider);
7489 * Now we need to call dtrace_enabling_matchall() -- which
7490 * will acquire cpu_lock and dtrace_lock. We therefore need
7491 * to drop all of our locks before calling into it...
7493 mutex_exit(&dtrace_lock);
7494 mutex_exit(&dtrace_provider_lock);
7495 dtrace_enabling_matchall();
7500 mutex_exit(&dtrace_lock);
7501 mutex_exit(&dtrace_provider_lock);
7507 * Unregister the specified provider from the DTrace framework. This should
7508 * generally be called by DTrace providers in their detach(9E) entry point.
7511 dtrace_unregister(dtrace_provider_id_t id)
7513 dtrace_provider_t *old = (dtrace_provider_t *)id;
7514 dtrace_provider_t *prev = NULL;
7516 dtrace_probe_t *probe, *first = NULL;
7518 if (old->dtpv_pops.dtps_enable ==
7519 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7521 * If DTrace itself is the provider, we're called with locks
7524 ASSERT(old == dtrace_provider);
7526 ASSERT(dtrace_devi != NULL);
7528 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7529 ASSERT(MUTEX_HELD(&dtrace_lock));
7532 if (dtrace_provider->dtpv_next != NULL) {
7534 * There's another provider here; return failure.
7539 mutex_enter(&dtrace_provider_lock);
7540 mutex_enter(&mod_lock);
7541 mutex_enter(&dtrace_lock);
7545 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7546 * probes, we refuse to let providers slither away, unless this
7547 * provider has already been explicitly invalidated.
7549 if (!old->dtpv_defunct &&
7550 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7551 dtrace_anon.dta_state->dts_necbs > 0))) {
7553 mutex_exit(&dtrace_lock);
7554 mutex_exit(&mod_lock);
7555 mutex_exit(&dtrace_provider_lock);
7561 * Attempt to destroy the probes associated with this provider.
7563 for (i = 0; i < dtrace_nprobes; i++) {
7564 if ((probe = dtrace_probes[i]) == NULL)
7567 if (probe->dtpr_provider != old)
7570 if (probe->dtpr_ecb == NULL)
7574 * We have at least one ECB; we can't remove this provider.
7577 mutex_exit(&dtrace_lock);
7578 mutex_exit(&mod_lock);
7579 mutex_exit(&dtrace_provider_lock);
7585 * All of the probes for this provider are disabled; we can safely
7586 * remove all of them from their hash chains and from the probe array.
7588 for (i = 0; i < dtrace_nprobes; i++) {
7589 if ((probe = dtrace_probes[i]) == NULL)
7592 if (probe->dtpr_provider != old)
7595 dtrace_probes[i] = NULL;
7597 dtrace_hash_remove(dtrace_bymod, probe);
7598 dtrace_hash_remove(dtrace_byfunc, probe);
7599 dtrace_hash_remove(dtrace_byname, probe);
7601 if (first == NULL) {
7603 probe->dtpr_nextmod = NULL;
7605 probe->dtpr_nextmod = first;
7611 * The provider's probes have been removed from the hash chains and
7612 * from the probe array. Now issue a dtrace_sync() to be sure that
7613 * everyone has cleared out from any probe array processing.
7617 for (probe = first; probe != NULL; probe = first) {
7618 first = probe->dtpr_nextmod;
7620 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7622 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7623 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7624 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7626 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7628 free_unr(dtrace_arena, probe->dtpr_id);
7630 kmem_free(probe, sizeof (dtrace_probe_t));
7633 if ((prev = dtrace_provider) == old) {
7635 ASSERT(self || dtrace_devi == NULL);
7636 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7638 dtrace_provider = old->dtpv_next;
7640 while (prev != NULL && prev->dtpv_next != old)
7641 prev = prev->dtpv_next;
7644 panic("attempt to unregister non-existent "
7645 "dtrace provider %p\n", (void *)id);
7648 prev->dtpv_next = old->dtpv_next;
7652 mutex_exit(&dtrace_lock);
7653 mutex_exit(&mod_lock);
7654 mutex_exit(&dtrace_provider_lock);
7657 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7658 kmem_free(old, sizeof (dtrace_provider_t));
7664 * Invalidate the specified provider. All subsequent probe lookups for the
7665 * specified provider will fail, but its probes will not be removed.
7668 dtrace_invalidate(dtrace_provider_id_t id)
7670 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7672 ASSERT(pvp->dtpv_pops.dtps_enable !=
7673 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7675 mutex_enter(&dtrace_provider_lock);
7676 mutex_enter(&dtrace_lock);
7678 pvp->dtpv_defunct = 1;
7680 mutex_exit(&dtrace_lock);
7681 mutex_exit(&dtrace_provider_lock);
7685 * Indicate whether or not DTrace has attached.
7688 dtrace_attached(void)
7691 * dtrace_provider will be non-NULL iff the DTrace driver has
7692 * attached. (It's non-NULL because DTrace is always itself a
7695 return (dtrace_provider != NULL);
7699 * Remove all the unenabled probes for the given provider. This function is
7700 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7701 * -- just as many of its associated probes as it can.
7704 dtrace_condense(dtrace_provider_id_t id)
7706 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7708 dtrace_probe_t *probe;
7711 * Make sure this isn't the dtrace provider itself.
7713 ASSERT(prov->dtpv_pops.dtps_enable !=
7714 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7716 mutex_enter(&dtrace_provider_lock);
7717 mutex_enter(&dtrace_lock);
7720 * Attempt to destroy the probes associated with this provider.
7722 for (i = 0; i < dtrace_nprobes; i++) {
7723 if ((probe = dtrace_probes[i]) == NULL)
7726 if (probe->dtpr_provider != prov)
7729 if (probe->dtpr_ecb != NULL)
7732 dtrace_probes[i] = NULL;
7734 dtrace_hash_remove(dtrace_bymod, probe);
7735 dtrace_hash_remove(dtrace_byfunc, probe);
7736 dtrace_hash_remove(dtrace_byname, probe);
7738 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7740 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7741 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7742 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7743 kmem_free(probe, sizeof (dtrace_probe_t));
7745 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7747 free_unr(dtrace_arena, i + 1);
7751 mutex_exit(&dtrace_lock);
7752 mutex_exit(&dtrace_provider_lock);
7758 * DTrace Probe Management Functions
7760 * The functions in this section perform the DTrace probe management,
7761 * including functions to create probes, look-up probes, and call into the
7762 * providers to request that probes be provided. Some of these functions are
7763 * in the Provider-to-Framework API; these functions can be identified by the
7764 * fact that they are not declared "static".
7768 * Create a probe with the specified module name, function name, and name.
7771 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7772 const char *func, const char *name, int aframes, void *arg)
7774 dtrace_probe_t *probe, **probes;
7775 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7778 if (provider == dtrace_provider) {
7779 ASSERT(MUTEX_HELD(&dtrace_lock));
7781 mutex_enter(&dtrace_lock);
7785 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7786 VM_BESTFIT | VM_SLEEP);
7788 id = alloc_unr(dtrace_arena);
7790 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7792 probe->dtpr_id = id;
7793 probe->dtpr_gen = dtrace_probegen++;
7794 probe->dtpr_mod = dtrace_strdup(mod);
7795 probe->dtpr_func = dtrace_strdup(func);
7796 probe->dtpr_name = dtrace_strdup(name);
7797 probe->dtpr_arg = arg;
7798 probe->dtpr_aframes = aframes;
7799 probe->dtpr_provider = provider;
7801 dtrace_hash_add(dtrace_bymod, probe);
7802 dtrace_hash_add(dtrace_byfunc, probe);
7803 dtrace_hash_add(dtrace_byname, probe);
7805 if (id - 1 >= dtrace_nprobes) {
7806 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7807 size_t nsize = osize << 1;
7811 ASSERT(dtrace_probes == NULL);
7812 nsize = sizeof (dtrace_probe_t *);
7815 probes = kmem_zalloc(nsize, KM_SLEEP);
7817 if (dtrace_probes == NULL) {
7819 dtrace_probes = probes;
7822 dtrace_probe_t **oprobes = dtrace_probes;
7824 bcopy(oprobes, probes, osize);
7825 dtrace_membar_producer();
7826 dtrace_probes = probes;
7831 * All CPUs are now seeing the new probes array; we can
7832 * safely free the old array.
7834 kmem_free(oprobes, osize);
7835 dtrace_nprobes <<= 1;
7838 ASSERT(id - 1 < dtrace_nprobes);
7841 ASSERT(dtrace_probes[id - 1] == NULL);
7842 dtrace_probes[id - 1] = probe;
7844 if (provider != dtrace_provider)
7845 mutex_exit(&dtrace_lock);
7850 static dtrace_probe_t *
7851 dtrace_probe_lookup_id(dtrace_id_t id)
7853 ASSERT(MUTEX_HELD(&dtrace_lock));
7855 if (id == 0 || id > dtrace_nprobes)
7858 return (dtrace_probes[id - 1]);
7862 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7864 *((dtrace_id_t *)arg) = probe->dtpr_id;
7866 return (DTRACE_MATCH_DONE);
7870 * Look up a probe based on provider and one or more of module name, function
7871 * name and probe name.
7874 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
7875 char *func, char *name)
7877 dtrace_probekey_t pkey;
7881 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7882 pkey.dtpk_pmatch = &dtrace_match_string;
7883 pkey.dtpk_mod = mod;
7884 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7885 pkey.dtpk_func = func;
7886 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7887 pkey.dtpk_name = name;
7888 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7889 pkey.dtpk_id = DTRACE_IDNONE;
7891 mutex_enter(&dtrace_lock);
7892 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7893 dtrace_probe_lookup_match, &id);
7894 mutex_exit(&dtrace_lock);
7896 ASSERT(match == 1 || match == 0);
7897 return (match ? id : 0);
7901 * Returns the probe argument associated with the specified probe.
7904 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7906 dtrace_probe_t *probe;
7909 mutex_enter(&dtrace_lock);
7911 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7912 probe->dtpr_provider == (dtrace_provider_t *)id)
7913 rval = probe->dtpr_arg;
7915 mutex_exit(&dtrace_lock);
7921 * Copy a probe into a probe description.
7924 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7926 bzero(pdp, sizeof (dtrace_probedesc_t));
7927 pdp->dtpd_id = prp->dtpr_id;
7929 (void) strncpy(pdp->dtpd_provider,
7930 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7932 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7933 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7934 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7939 dtrace_probe_provide_cb(linker_file_t lf, void *arg)
7941 dtrace_provider_t *prv = (dtrace_provider_t *) arg;
7943 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, lf);
7951 * Called to indicate that a probe -- or probes -- should be provided by a
7952 * specfied provider. If the specified description is NULL, the provider will
7953 * be told to provide all of its probes. (This is done whenever a new
7954 * consumer comes along, or whenever a retained enabling is to be matched.) If
7955 * the specified description is non-NULL, the provider is given the
7956 * opportunity to dynamically provide the specified probe, allowing providers
7957 * to support the creation of probes on-the-fly. (So-called _autocreated_
7958 * probes.) If the provider is NULL, the operations will be applied to all
7959 * providers; if the provider is non-NULL the operations will only be applied
7960 * to the specified provider. The dtrace_provider_lock must be held, and the
7961 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7962 * will need to grab the dtrace_lock when it reenters the framework through
7963 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7966 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7973 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7977 prv = dtrace_provider;
7982 * First, call the blanket provide operation.
7984 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7987 * Now call the per-module provide operation. We will grab
7988 * mod_lock to prevent the list from being modified. Note
7989 * that this also prevents the mod_busy bits from changing.
7990 * (mod_busy can only be changed with mod_lock held.)
7992 mutex_enter(&mod_lock);
7997 if (ctl->mod_busy || ctl->mod_mp == NULL)
8000 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8002 } while ((ctl = ctl->mod_next) != &modules);
8004 (void) linker_file_foreach(dtrace_probe_provide_cb, prv);
8007 mutex_exit(&mod_lock);
8008 } while (all && (prv = prv->dtpv_next) != NULL);
8013 * Iterate over each probe, and call the Framework-to-Provider API function
8017 dtrace_probe_foreach(uintptr_t offs)
8019 dtrace_provider_t *prov;
8020 void (*func)(void *, dtrace_id_t, void *);
8021 dtrace_probe_t *probe;
8022 dtrace_icookie_t cookie;
8026 * We disable interrupts to walk through the probe array. This is
8027 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8028 * won't see stale data.
8030 cookie = dtrace_interrupt_disable();
8032 for (i = 0; i < dtrace_nprobes; i++) {
8033 if ((probe = dtrace_probes[i]) == NULL)
8036 if (probe->dtpr_ecb == NULL) {
8038 * This probe isn't enabled -- don't call the function.
8043 prov = probe->dtpr_provider;
8044 func = *((void(**)(void *, dtrace_id_t, void *))
8045 ((uintptr_t)&prov->dtpv_pops + offs));
8047 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8050 dtrace_interrupt_enable(cookie);
8055 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8057 dtrace_probekey_t pkey;
8062 ASSERT(MUTEX_HELD(&dtrace_lock));
8063 dtrace_ecb_create_cache = NULL;
8067 * If we're passed a NULL description, we're being asked to
8068 * create an ECB with a NULL probe.
8070 (void) dtrace_ecb_create_enable(NULL, enab);
8074 dtrace_probekey(desc, &pkey);
8075 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8076 &priv, &uid, &zoneid);
8078 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8083 * DTrace Helper Provider Functions
8086 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8088 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8089 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8090 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8094 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8095 const dof_provider_t *dofprov, char *strtab)
8097 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8098 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8099 dofprov->dofpv_provattr);
8100 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8101 dofprov->dofpv_modattr);
8102 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8103 dofprov->dofpv_funcattr);
8104 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8105 dofprov->dofpv_nameattr);
8106 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8107 dofprov->dofpv_argsattr);
8111 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8113 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8114 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8115 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8116 dof_provider_t *provider;
8118 uint32_t *off, *enoff;
8122 dtrace_helper_provdesc_t dhpv;
8123 dtrace_helper_probedesc_t dhpb;
8124 dtrace_meta_t *meta = dtrace_meta_pid;
8125 dtrace_mops_t *mops = &meta->dtm_mops;
8128 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8129 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8130 provider->dofpv_strtab * dof->dofh_secsize);
8131 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8132 provider->dofpv_probes * dof->dofh_secsize);
8133 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8134 provider->dofpv_prargs * dof->dofh_secsize);
8135 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8136 provider->dofpv_proffs * dof->dofh_secsize);
8138 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8139 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8140 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8144 * See dtrace_helper_provider_validate().
8146 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8147 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8148 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8149 provider->dofpv_prenoffs * dof->dofh_secsize);
8150 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8153 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8156 * Create the provider.
8158 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8160 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8166 * Create the probes.
8168 for (i = 0; i < nprobes; i++) {
8169 probe = (dof_probe_t *)(uintptr_t)(daddr +
8170 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8172 dhpb.dthpb_mod = dhp->dofhp_mod;
8173 dhpb.dthpb_func = strtab + probe->dofpr_func;
8174 dhpb.dthpb_name = strtab + probe->dofpr_name;
8175 dhpb.dthpb_base = probe->dofpr_addr;
8176 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8177 dhpb.dthpb_noffs = probe->dofpr_noffs;
8178 if (enoff != NULL) {
8179 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8180 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8182 dhpb.dthpb_enoffs = NULL;
8183 dhpb.dthpb_nenoffs = 0;
8185 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8186 dhpb.dthpb_nargc = probe->dofpr_nargc;
8187 dhpb.dthpb_xargc = probe->dofpr_xargc;
8188 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8189 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8191 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8196 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8198 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8199 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8202 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8204 for (i = 0; i < dof->dofh_secnum; i++) {
8205 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8206 dof->dofh_secoff + i * dof->dofh_secsize);
8208 if (sec->dofs_type != DOF_SECT_PROVIDER)
8211 dtrace_helper_provide_one(dhp, sec, pid);
8215 * We may have just created probes, so we must now rematch against
8216 * any retained enablings. Note that this call will acquire both
8217 * cpu_lock and dtrace_lock; the fact that we are holding
8218 * dtrace_meta_lock now is what defines the ordering with respect to
8219 * these three locks.
8221 dtrace_enabling_matchall();
8225 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8227 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8228 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8230 dof_provider_t *provider;
8232 dtrace_helper_provdesc_t dhpv;
8233 dtrace_meta_t *meta = dtrace_meta_pid;
8234 dtrace_mops_t *mops = &meta->dtm_mops;
8236 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8237 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8238 provider->dofpv_strtab * dof->dofh_secsize);
8240 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8243 * Create the provider.
8245 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8247 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8253 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8255 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8256 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8259 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8261 for (i = 0; i < dof->dofh_secnum; i++) {
8262 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8263 dof->dofh_secoff + i * dof->dofh_secsize);
8265 if (sec->dofs_type != DOF_SECT_PROVIDER)
8268 dtrace_helper_provider_remove_one(dhp, sec, pid);
8273 * DTrace Meta Provider-to-Framework API Functions
8275 * These functions implement the Meta Provider-to-Framework API, as described
8276 * in <sys/dtrace.h>.
8279 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8280 dtrace_meta_provider_id_t *idp)
8282 dtrace_meta_t *meta;
8283 dtrace_helpers_t *help, *next;
8286 *idp = DTRACE_METAPROVNONE;
8289 * We strictly don't need the name, but we hold onto it for
8290 * debuggability. All hail error queues!
8293 cmn_err(CE_WARN, "failed to register meta-provider: "
8299 mops->dtms_create_probe == NULL ||
8300 mops->dtms_provide_pid == NULL ||
8301 mops->dtms_remove_pid == NULL) {
8302 cmn_err(CE_WARN, "failed to register meta-register %s: "
8303 "invalid ops", name);
8307 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8308 meta->dtm_mops = *mops;
8309 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8310 (void) strcpy(meta->dtm_name, name);
8311 meta->dtm_arg = arg;
8313 mutex_enter(&dtrace_meta_lock);
8314 mutex_enter(&dtrace_lock);
8316 if (dtrace_meta_pid != NULL) {
8317 mutex_exit(&dtrace_lock);
8318 mutex_exit(&dtrace_meta_lock);
8319 cmn_err(CE_WARN, "failed to register meta-register %s: "
8320 "user-land meta-provider exists", name);
8321 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8322 kmem_free(meta, sizeof (dtrace_meta_t));
8326 dtrace_meta_pid = meta;
8327 *idp = (dtrace_meta_provider_id_t)meta;
8330 * If there are providers and probes ready to go, pass them
8331 * off to the new meta provider now.
8334 help = dtrace_deferred_pid;
8335 dtrace_deferred_pid = NULL;
8337 mutex_exit(&dtrace_lock);
8339 while (help != NULL) {
8340 for (i = 0; i < help->dthps_nprovs; i++) {
8341 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8345 next = help->dthps_next;
8346 help->dthps_next = NULL;
8347 help->dthps_prev = NULL;
8348 help->dthps_deferred = 0;
8352 mutex_exit(&dtrace_meta_lock);
8358 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8360 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8362 mutex_enter(&dtrace_meta_lock);
8363 mutex_enter(&dtrace_lock);
8365 if (old == dtrace_meta_pid) {
8366 pp = &dtrace_meta_pid;
8368 panic("attempt to unregister non-existent "
8369 "dtrace meta-provider %p\n", (void *)old);
8372 if (old->dtm_count != 0) {
8373 mutex_exit(&dtrace_lock);
8374 mutex_exit(&dtrace_meta_lock);
8380 mutex_exit(&dtrace_lock);
8381 mutex_exit(&dtrace_meta_lock);
8383 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8384 kmem_free(old, sizeof (dtrace_meta_t));
8391 * DTrace DIF Object Functions
8394 dtrace_difo_err(uint_t pc, const char *format, ...)
8396 if (dtrace_err_verbose) {
8399 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8400 va_start(alist, format);
8401 (void) vuprintf(format, alist);
8405 #ifdef DTRACE_ERRDEBUG
8406 dtrace_errdebug(format);
8412 * Validate a DTrace DIF object by checking the IR instructions. The following
8413 * rules are currently enforced by dtrace_difo_validate():
8415 * 1. Each instruction must have a valid opcode
8416 * 2. Each register, string, variable, or subroutine reference must be valid
8417 * 3. No instruction can modify register %r0 (must be zero)
8418 * 4. All instruction reserved bits must be set to zero
8419 * 5. The last instruction must be a "ret" instruction
8420 * 6. All branch targets must reference a valid instruction _after_ the branch
8423 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8427 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8431 kcheckload = cr == NULL ||
8432 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8434 dp->dtdo_destructive = 0;
8436 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8437 dif_instr_t instr = dp->dtdo_buf[pc];
8439 uint_t r1 = DIF_INSTR_R1(instr);
8440 uint_t r2 = DIF_INSTR_R2(instr);
8441 uint_t rd = DIF_INSTR_RD(instr);
8442 uint_t rs = DIF_INSTR_RS(instr);
8443 uint_t label = DIF_INSTR_LABEL(instr);
8444 uint_t v = DIF_INSTR_VAR(instr);
8445 uint_t subr = DIF_INSTR_SUBR(instr);
8446 uint_t type = DIF_INSTR_TYPE(instr);
8447 uint_t op = DIF_INSTR_OP(instr);
8465 err += efunc(pc, "invalid register %u\n", r1);
8467 err += efunc(pc, "invalid register %u\n", r2);
8469 err += efunc(pc, "invalid register %u\n", rd);
8471 err += efunc(pc, "cannot write to %r0\n");
8477 err += efunc(pc, "invalid register %u\n", r1);
8479 err += efunc(pc, "non-zero reserved bits\n");
8481 err += efunc(pc, "invalid register %u\n", rd);
8483 err += efunc(pc, "cannot write to %r0\n");
8493 err += efunc(pc, "invalid register %u\n", r1);
8495 err += efunc(pc, "non-zero reserved bits\n");
8497 err += efunc(pc, "invalid register %u\n", rd);
8499 err += efunc(pc, "cannot write to %r0\n");
8501 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8502 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8512 err += efunc(pc, "invalid register %u\n", r1);
8514 err += efunc(pc, "non-zero reserved bits\n");
8516 err += efunc(pc, "invalid register %u\n", rd);
8518 err += efunc(pc, "cannot write to %r0\n");
8528 err += efunc(pc, "invalid register %u\n", r1);
8530 err += efunc(pc, "non-zero reserved bits\n");
8532 err += efunc(pc, "invalid register %u\n", rd);
8534 err += efunc(pc, "cannot write to %r0\n");
8541 err += efunc(pc, "invalid register %u\n", r1);
8543 err += efunc(pc, "non-zero reserved bits\n");
8545 err += efunc(pc, "invalid register %u\n", rd);
8547 err += efunc(pc, "cannot write to 0 address\n");
8552 err += efunc(pc, "invalid register %u\n", r1);
8554 err += efunc(pc, "invalid register %u\n", r2);
8556 err += efunc(pc, "non-zero reserved bits\n");
8560 err += efunc(pc, "invalid register %u\n", r1);
8561 if (r2 != 0 || rd != 0)
8562 err += efunc(pc, "non-zero reserved bits\n");
8575 if (label >= dp->dtdo_len) {
8576 err += efunc(pc, "invalid branch target %u\n",
8580 err += efunc(pc, "backward branch to %u\n",
8585 if (r1 != 0 || r2 != 0)
8586 err += efunc(pc, "non-zero reserved bits\n");
8588 err += efunc(pc, "invalid register %u\n", rd);
8592 case DIF_OP_FLUSHTS:
8593 if (r1 != 0 || r2 != 0 || rd != 0)
8594 err += efunc(pc, "non-zero reserved bits\n");
8597 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8598 err += efunc(pc, "invalid integer ref %u\n",
8599 DIF_INSTR_INTEGER(instr));
8602 err += efunc(pc, "invalid register %u\n", rd);
8604 err += efunc(pc, "cannot write to %r0\n");
8607 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8608 err += efunc(pc, "invalid string ref %u\n",
8609 DIF_INSTR_STRING(instr));
8612 err += efunc(pc, "invalid register %u\n", rd);
8614 err += efunc(pc, "cannot write to %r0\n");
8618 if (r1 > DIF_VAR_ARRAY_MAX)
8619 err += efunc(pc, "invalid array %u\n", r1);
8621 err += efunc(pc, "invalid register %u\n", r2);
8623 err += efunc(pc, "invalid register %u\n", rd);
8625 err += efunc(pc, "cannot write to %r0\n");
8632 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8633 err += efunc(pc, "invalid variable %u\n", v);
8635 err += efunc(pc, "invalid register %u\n", rd);
8637 err += efunc(pc, "cannot write to %r0\n");
8644 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8645 err += efunc(pc, "invalid variable %u\n", v);
8647 err += efunc(pc, "invalid register %u\n", rd);
8650 if (subr > DIF_SUBR_MAX)
8651 err += efunc(pc, "invalid subr %u\n", subr);
8653 err += efunc(pc, "invalid register %u\n", rd);
8655 err += efunc(pc, "cannot write to %r0\n");
8657 if (subr == DIF_SUBR_COPYOUT ||
8658 subr == DIF_SUBR_COPYOUTSTR) {
8659 dp->dtdo_destructive = 1;
8663 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8664 err += efunc(pc, "invalid ref type %u\n", type);
8666 err += efunc(pc, "invalid register %u\n", r2);
8668 err += efunc(pc, "invalid register %u\n", rs);
8671 if (type != DIF_TYPE_CTF)
8672 err += efunc(pc, "invalid val type %u\n", type);
8674 err += efunc(pc, "invalid register %u\n", r2);
8676 err += efunc(pc, "invalid register %u\n", rs);
8679 err += efunc(pc, "invalid opcode %u\n",
8680 DIF_INSTR_OP(instr));
8684 if (dp->dtdo_len != 0 &&
8685 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8686 err += efunc(dp->dtdo_len - 1,
8687 "expected 'ret' as last DIF instruction\n");
8690 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8692 * If we're not returning by reference, the size must be either
8693 * 0 or the size of one of the base types.
8695 switch (dp->dtdo_rtype.dtdt_size) {
8697 case sizeof (uint8_t):
8698 case sizeof (uint16_t):
8699 case sizeof (uint32_t):
8700 case sizeof (uint64_t):
8704 err += efunc(dp->dtdo_len - 1, "bad return size");
8708 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8709 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8710 dtrace_diftype_t *vt, *et;
8713 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8714 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8715 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8716 err += efunc(i, "unrecognized variable scope %d\n",
8721 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8722 v->dtdv_kind != DIFV_KIND_SCALAR) {
8723 err += efunc(i, "unrecognized variable type %d\n",
8728 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8729 err += efunc(i, "%d exceeds variable id limit\n", id);
8733 if (id < DIF_VAR_OTHER_UBASE)
8737 * For user-defined variables, we need to check that this
8738 * definition is identical to any previous definition that we
8741 ndx = id - DIF_VAR_OTHER_UBASE;
8743 switch (v->dtdv_scope) {
8744 case DIFV_SCOPE_GLOBAL:
8745 if (ndx < vstate->dtvs_nglobals) {
8746 dtrace_statvar_t *svar;
8748 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8749 existing = &svar->dtsv_var;
8754 case DIFV_SCOPE_THREAD:
8755 if (ndx < vstate->dtvs_ntlocals)
8756 existing = &vstate->dtvs_tlocals[ndx];
8759 case DIFV_SCOPE_LOCAL:
8760 if (ndx < vstate->dtvs_nlocals) {
8761 dtrace_statvar_t *svar;
8763 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8764 existing = &svar->dtsv_var;
8772 if (vt->dtdt_flags & DIF_TF_BYREF) {
8773 if (vt->dtdt_size == 0) {
8774 err += efunc(i, "zero-sized variable\n");
8778 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8779 vt->dtdt_size > dtrace_global_maxsize) {
8780 err += efunc(i, "oversized by-ref global\n");
8785 if (existing == NULL || existing->dtdv_id == 0)
8788 ASSERT(existing->dtdv_id == v->dtdv_id);
8789 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8791 if (existing->dtdv_kind != v->dtdv_kind)
8792 err += efunc(i, "%d changed variable kind\n", id);
8794 et = &existing->dtdv_type;
8796 if (vt->dtdt_flags != et->dtdt_flags) {
8797 err += efunc(i, "%d changed variable type flags\n", id);
8801 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8802 err += efunc(i, "%d changed variable type size\n", id);
8811 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
8812 * are much more constrained than normal DIFOs. Specifically, they may
8815 * 1. Make calls to subroutines other than copyin(), copyinstr() or
8816 * miscellaneous string routines
8817 * 2. Access DTrace variables other than the args[] array, and the
8818 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8819 * 3. Have thread-local variables.
8820 * 4. Have dynamic variables.
8823 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8825 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8829 for (pc = 0; pc < dp->dtdo_len; pc++) {
8830 dif_instr_t instr = dp->dtdo_buf[pc];
8832 uint_t v = DIF_INSTR_VAR(instr);
8833 uint_t subr = DIF_INSTR_SUBR(instr);
8834 uint_t op = DIF_INSTR_OP(instr);
8889 case DIF_OP_FLUSHTS:
8901 if (v >= DIF_VAR_OTHER_UBASE)
8904 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8907 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8908 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8909 v == DIF_VAR_EXECARGS ||
8910 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8911 v == DIF_VAR_UID || v == DIF_VAR_GID)
8914 err += efunc(pc, "illegal variable %u\n", v);
8921 err += efunc(pc, "illegal dynamic variable load\n");
8927 err += efunc(pc, "illegal dynamic variable store\n");
8931 if (subr == DIF_SUBR_ALLOCA ||
8932 subr == DIF_SUBR_BCOPY ||
8933 subr == DIF_SUBR_COPYIN ||
8934 subr == DIF_SUBR_COPYINTO ||
8935 subr == DIF_SUBR_COPYINSTR ||
8936 subr == DIF_SUBR_INDEX ||
8937 subr == DIF_SUBR_INET_NTOA ||
8938 subr == DIF_SUBR_INET_NTOA6 ||
8939 subr == DIF_SUBR_INET_NTOP ||
8940 subr == DIF_SUBR_LLTOSTR ||
8941 subr == DIF_SUBR_RINDEX ||
8942 subr == DIF_SUBR_STRCHR ||
8943 subr == DIF_SUBR_STRJOIN ||
8944 subr == DIF_SUBR_STRRCHR ||
8945 subr == DIF_SUBR_STRSTR ||
8946 subr == DIF_SUBR_HTONS ||
8947 subr == DIF_SUBR_HTONL ||
8948 subr == DIF_SUBR_HTONLL ||
8949 subr == DIF_SUBR_NTOHS ||
8950 subr == DIF_SUBR_NTOHL ||
8951 subr == DIF_SUBR_NTOHLL ||
8952 subr == DIF_SUBR_MEMREF ||
8953 subr == DIF_SUBR_TYPEREF)
8956 err += efunc(pc, "invalid subr %u\n", subr);
8960 err += efunc(pc, "invalid opcode %u\n",
8961 DIF_INSTR_OP(instr));
8969 * Returns 1 if the expression in the DIF object can be cached on a per-thread
8973 dtrace_difo_cacheable(dtrace_difo_t *dp)
8980 for (i = 0; i < dp->dtdo_varlen; i++) {
8981 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8983 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8986 switch (v->dtdv_id) {
8987 case DIF_VAR_CURTHREAD:
8990 case DIF_VAR_EXECARGS:
8991 case DIF_VAR_EXECNAME:
8992 case DIF_VAR_ZONENAME:
9001 * This DIF object may be cacheable. Now we need to look for any
9002 * array loading instructions, any memory loading instructions, or
9003 * any stores to thread-local variables.
9005 for (i = 0; i < dp->dtdo_len; i++) {
9006 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9008 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9009 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9010 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9011 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9019 dtrace_difo_hold(dtrace_difo_t *dp)
9023 ASSERT(MUTEX_HELD(&dtrace_lock));
9026 ASSERT(dp->dtdo_refcnt != 0);
9029 * We need to check this DIF object for references to the variable
9030 * DIF_VAR_VTIMESTAMP.
9032 for (i = 0; i < dp->dtdo_varlen; i++) {
9033 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9035 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9038 if (dtrace_vtime_references++ == 0)
9039 dtrace_vtime_enable();
9044 * This routine calculates the dynamic variable chunksize for a given DIF
9045 * object. The calculation is not fool-proof, and can probably be tricked by
9046 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9047 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9048 * if a dynamic variable size exceeds the chunksize.
9051 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9054 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9055 const dif_instr_t *text = dp->dtdo_buf;
9061 for (pc = 0; pc < dp->dtdo_len; pc++) {
9062 dif_instr_t instr = text[pc];
9063 uint_t op = DIF_INSTR_OP(instr);
9064 uint_t rd = DIF_INSTR_RD(instr);
9065 uint_t r1 = DIF_INSTR_R1(instr);
9069 dtrace_key_t *key = tupregs;
9073 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9078 key = &tupregs[DIF_DTR_NREGS];
9079 key[0].dttk_size = 0;
9080 key[1].dttk_size = 0;
9082 scope = DIFV_SCOPE_THREAD;
9089 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9090 key[nkeys++].dttk_size = 0;
9092 key[nkeys++].dttk_size = 0;
9094 if (op == DIF_OP_STTAA) {
9095 scope = DIFV_SCOPE_THREAD;
9097 scope = DIFV_SCOPE_GLOBAL;
9103 if (ttop == DIF_DTR_NREGS)
9106 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9108 * If the register for the size of the "pushtr"
9109 * is %r0 (or the value is 0) and the type is
9110 * a string, we'll use the system-wide default
9113 tupregs[ttop++].dttk_size =
9114 dtrace_strsize_default;
9119 tupregs[ttop++].dttk_size = sval;
9125 if (ttop == DIF_DTR_NREGS)
9128 tupregs[ttop++].dttk_size = 0;
9131 case DIF_OP_FLUSHTS:
9148 * We have a dynamic variable allocation; calculate its size.
9150 for (ksize = 0, i = 0; i < nkeys; i++)
9151 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9153 size = sizeof (dtrace_dynvar_t);
9154 size += sizeof (dtrace_key_t) * (nkeys - 1);
9158 * Now we need to determine the size of the stored data.
9160 id = DIF_INSTR_VAR(instr);
9162 for (i = 0; i < dp->dtdo_varlen; i++) {
9163 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9165 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9166 size += v->dtdv_type.dtdt_size;
9171 if (i == dp->dtdo_varlen)
9175 * We have the size. If this is larger than the chunk size
9176 * for our dynamic variable state, reset the chunk size.
9178 size = P2ROUNDUP(size, sizeof (uint64_t));
9180 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9181 vstate->dtvs_dynvars.dtds_chunksize = size;
9186 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9188 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9191 ASSERT(MUTEX_HELD(&dtrace_lock));
9192 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9194 for (i = 0; i < dp->dtdo_varlen; i++) {
9195 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9196 dtrace_statvar_t *svar, ***svarp = NULL;
9198 uint8_t scope = v->dtdv_scope;
9201 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9204 id -= DIF_VAR_OTHER_UBASE;
9207 case DIFV_SCOPE_THREAD:
9208 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9209 dtrace_difv_t *tlocals;
9211 if ((ntlocals = (otlocals << 1)) == 0)
9214 osz = otlocals * sizeof (dtrace_difv_t);
9215 nsz = ntlocals * sizeof (dtrace_difv_t);
9217 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9220 bcopy(vstate->dtvs_tlocals,
9222 kmem_free(vstate->dtvs_tlocals, osz);
9225 vstate->dtvs_tlocals = tlocals;
9226 vstate->dtvs_ntlocals = ntlocals;
9229 vstate->dtvs_tlocals[id] = *v;
9232 case DIFV_SCOPE_LOCAL:
9233 np = &vstate->dtvs_nlocals;
9234 svarp = &vstate->dtvs_locals;
9236 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9237 dsize = NCPU * (v->dtdv_type.dtdt_size +
9240 dsize = NCPU * sizeof (uint64_t);
9244 case DIFV_SCOPE_GLOBAL:
9245 np = &vstate->dtvs_nglobals;
9246 svarp = &vstate->dtvs_globals;
9248 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9249 dsize = v->dtdv_type.dtdt_size +
9258 while (id >= (oldsvars = *np)) {
9259 dtrace_statvar_t **statics;
9260 int newsvars, oldsize, newsize;
9262 if ((newsvars = (oldsvars << 1)) == 0)
9265 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9266 newsize = newsvars * sizeof (dtrace_statvar_t *);
9268 statics = kmem_zalloc(newsize, KM_SLEEP);
9271 bcopy(*svarp, statics, oldsize);
9272 kmem_free(*svarp, oldsize);
9279 if ((svar = (*svarp)[id]) == NULL) {
9280 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9281 svar->dtsv_var = *v;
9283 if ((svar->dtsv_size = dsize) != 0) {
9284 svar->dtsv_data = (uint64_t)(uintptr_t)
9285 kmem_zalloc(dsize, KM_SLEEP);
9288 (*svarp)[id] = svar;
9291 svar->dtsv_refcnt++;
9294 dtrace_difo_chunksize(dp, vstate);
9295 dtrace_difo_hold(dp);
9298 static dtrace_difo_t *
9299 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9304 ASSERT(dp->dtdo_buf != NULL);
9305 ASSERT(dp->dtdo_refcnt != 0);
9307 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9309 ASSERT(dp->dtdo_buf != NULL);
9310 sz = dp->dtdo_len * sizeof (dif_instr_t);
9311 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9312 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9313 new->dtdo_len = dp->dtdo_len;
9315 if (dp->dtdo_strtab != NULL) {
9316 ASSERT(dp->dtdo_strlen != 0);
9317 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9318 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9319 new->dtdo_strlen = dp->dtdo_strlen;
9322 if (dp->dtdo_inttab != NULL) {
9323 ASSERT(dp->dtdo_intlen != 0);
9324 sz = dp->dtdo_intlen * sizeof (uint64_t);
9325 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9326 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9327 new->dtdo_intlen = dp->dtdo_intlen;
9330 if (dp->dtdo_vartab != NULL) {
9331 ASSERT(dp->dtdo_varlen != 0);
9332 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9333 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9334 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9335 new->dtdo_varlen = dp->dtdo_varlen;
9338 dtrace_difo_init(new, vstate);
9343 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9347 ASSERT(dp->dtdo_refcnt == 0);
9349 for (i = 0; i < dp->dtdo_varlen; i++) {
9350 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9351 dtrace_statvar_t *svar, **svarp = NULL;
9353 uint8_t scope = v->dtdv_scope;
9357 case DIFV_SCOPE_THREAD:
9360 case DIFV_SCOPE_LOCAL:
9361 np = &vstate->dtvs_nlocals;
9362 svarp = vstate->dtvs_locals;
9365 case DIFV_SCOPE_GLOBAL:
9366 np = &vstate->dtvs_nglobals;
9367 svarp = vstate->dtvs_globals;
9374 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9377 id -= DIF_VAR_OTHER_UBASE;
9381 ASSERT(svar != NULL);
9382 ASSERT(svar->dtsv_refcnt > 0);
9384 if (--svar->dtsv_refcnt > 0)
9387 if (svar->dtsv_size != 0) {
9388 ASSERT(svar->dtsv_data != 0);
9389 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9393 kmem_free(svar, sizeof (dtrace_statvar_t));
9397 if (dp->dtdo_buf != NULL)
9398 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9399 if (dp->dtdo_inttab != NULL)
9400 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9401 if (dp->dtdo_strtab != NULL)
9402 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9403 if (dp->dtdo_vartab != NULL)
9404 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9406 kmem_free(dp, sizeof (dtrace_difo_t));
9410 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9414 ASSERT(MUTEX_HELD(&dtrace_lock));
9415 ASSERT(dp->dtdo_refcnt != 0);
9417 for (i = 0; i < dp->dtdo_varlen; i++) {
9418 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9420 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9423 ASSERT(dtrace_vtime_references > 0);
9424 if (--dtrace_vtime_references == 0)
9425 dtrace_vtime_disable();
9428 if (--dp->dtdo_refcnt == 0)
9429 dtrace_difo_destroy(dp, vstate);
9433 * DTrace Format Functions
9436 dtrace_format_add(dtrace_state_t *state, char *str)
9439 uint16_t ndx, len = strlen(str) + 1;
9441 fmt = kmem_zalloc(len, KM_SLEEP);
9442 bcopy(str, fmt, len);
9444 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9445 if (state->dts_formats[ndx] == NULL) {
9446 state->dts_formats[ndx] = fmt;
9451 if (state->dts_nformats == USHRT_MAX) {
9453 * This is only likely if a denial-of-service attack is being
9454 * attempted. As such, it's okay to fail silently here.
9456 kmem_free(fmt, len);
9461 * For simplicity, we always resize the formats array to be exactly the
9462 * number of formats.
9464 ndx = state->dts_nformats++;
9465 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9467 if (state->dts_formats != NULL) {
9469 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9470 kmem_free(state->dts_formats, ndx * sizeof (char *));
9473 state->dts_formats = new;
9474 state->dts_formats[ndx] = fmt;
9480 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9484 ASSERT(state->dts_formats != NULL);
9485 ASSERT(format <= state->dts_nformats);
9486 ASSERT(state->dts_formats[format - 1] != NULL);
9488 fmt = state->dts_formats[format - 1];
9489 kmem_free(fmt, strlen(fmt) + 1);
9490 state->dts_formats[format - 1] = NULL;
9494 dtrace_format_destroy(dtrace_state_t *state)
9498 if (state->dts_nformats == 0) {
9499 ASSERT(state->dts_formats == NULL);
9503 ASSERT(state->dts_formats != NULL);
9505 for (i = 0; i < state->dts_nformats; i++) {
9506 char *fmt = state->dts_formats[i];
9511 kmem_free(fmt, strlen(fmt) + 1);
9514 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9515 state->dts_nformats = 0;
9516 state->dts_formats = NULL;
9520 * DTrace Predicate Functions
9522 static dtrace_predicate_t *
9523 dtrace_predicate_create(dtrace_difo_t *dp)
9525 dtrace_predicate_t *pred;
9527 ASSERT(MUTEX_HELD(&dtrace_lock));
9528 ASSERT(dp->dtdo_refcnt != 0);
9530 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9531 pred->dtp_difo = dp;
9532 pred->dtp_refcnt = 1;
9534 if (!dtrace_difo_cacheable(dp))
9537 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9539 * This is only theoretically possible -- we have had 2^32
9540 * cacheable predicates on this machine. We cannot allow any
9541 * more predicates to become cacheable: as unlikely as it is,
9542 * there may be a thread caching a (now stale) predicate cache
9543 * ID. (N.B.: the temptation is being successfully resisted to
9544 * have this cmn_err() "Holy shit -- we executed this code!")
9549 pred->dtp_cacheid = dtrace_predcache_id++;
9555 dtrace_predicate_hold(dtrace_predicate_t *pred)
9557 ASSERT(MUTEX_HELD(&dtrace_lock));
9558 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9559 ASSERT(pred->dtp_refcnt > 0);
9565 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9567 dtrace_difo_t *dp = pred->dtp_difo;
9569 ASSERT(MUTEX_HELD(&dtrace_lock));
9570 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9571 ASSERT(pred->dtp_refcnt > 0);
9573 if (--pred->dtp_refcnt == 0) {
9574 dtrace_difo_release(pred->dtp_difo, vstate);
9575 kmem_free(pred, sizeof (dtrace_predicate_t));
9580 * DTrace Action Description Functions
9582 static dtrace_actdesc_t *
9583 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9584 uint64_t uarg, uint64_t arg)
9586 dtrace_actdesc_t *act;
9589 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9590 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9593 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9594 act->dtad_kind = kind;
9595 act->dtad_ntuple = ntuple;
9596 act->dtad_uarg = uarg;
9597 act->dtad_arg = arg;
9598 act->dtad_refcnt = 1;
9604 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9606 ASSERT(act->dtad_refcnt >= 1);
9611 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9613 dtrace_actkind_t kind = act->dtad_kind;
9616 ASSERT(act->dtad_refcnt >= 1);
9618 if (--act->dtad_refcnt != 0)
9621 if ((dp = act->dtad_difo) != NULL)
9622 dtrace_difo_release(dp, vstate);
9624 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9625 char *str = (char *)(uintptr_t)act->dtad_arg;
9628 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9629 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9633 kmem_free(str, strlen(str) + 1);
9636 kmem_free(act, sizeof (dtrace_actdesc_t));
9640 * DTrace ECB Functions
9642 static dtrace_ecb_t *
9643 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9648 ASSERT(MUTEX_HELD(&dtrace_lock));
9650 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9651 ecb->dte_predicate = NULL;
9652 ecb->dte_probe = probe;
9655 * The default size is the size of the default action: recording
9658 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9659 ecb->dte_alignment = sizeof (dtrace_epid_t);
9661 epid = state->dts_epid++;
9663 if (epid - 1 >= state->dts_necbs) {
9664 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9665 int necbs = state->dts_necbs << 1;
9667 ASSERT(epid == state->dts_necbs + 1);
9670 ASSERT(oecbs == NULL);
9674 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9677 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9679 dtrace_membar_producer();
9680 state->dts_ecbs = ecbs;
9682 if (oecbs != NULL) {
9684 * If this state is active, we must dtrace_sync()
9685 * before we can free the old dts_ecbs array: we're
9686 * coming in hot, and there may be active ring
9687 * buffer processing (which indexes into the dts_ecbs
9688 * array) on another CPU.
9690 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9693 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9696 dtrace_membar_producer();
9697 state->dts_necbs = necbs;
9700 ecb->dte_state = state;
9702 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9703 dtrace_membar_producer();
9704 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9710 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9712 dtrace_probe_t *probe = ecb->dte_probe;
9714 ASSERT(MUTEX_HELD(&cpu_lock));
9715 ASSERT(MUTEX_HELD(&dtrace_lock));
9716 ASSERT(ecb->dte_next == NULL);
9718 if (probe == NULL) {
9720 * This is the NULL probe -- there's nothing to do.
9725 if (probe->dtpr_ecb == NULL) {
9726 dtrace_provider_t *prov = probe->dtpr_provider;
9729 * We're the first ECB on this probe.
9731 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9733 if (ecb->dte_predicate != NULL)
9734 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9736 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9737 probe->dtpr_id, probe->dtpr_arg);
9740 * This probe is already active. Swing the last pointer to
9741 * point to the new ECB, and issue a dtrace_sync() to assure
9742 * that all CPUs have seen the change.
9744 ASSERT(probe->dtpr_ecb_last != NULL);
9745 probe->dtpr_ecb_last->dte_next = ecb;
9746 probe->dtpr_ecb_last = ecb;
9747 probe->dtpr_predcache = 0;
9754 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9756 uint32_t maxalign = sizeof (dtrace_epid_t);
9757 uint32_t align = sizeof (uint8_t), offs, diff;
9758 dtrace_action_t *act;
9760 uint32_t aggbase = UINT32_MAX;
9761 dtrace_state_t *state = ecb->dte_state;
9764 * If we record anything, we always record the epid. (And we always
9767 offs = sizeof (dtrace_epid_t);
9768 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9770 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9771 dtrace_recdesc_t *rec = &act->dta_rec;
9773 if ((align = rec->dtrd_alignment) > maxalign)
9776 if (!wastuple && act->dta_intuple) {
9778 * This is the first record in a tuple. Align the
9779 * offset to be at offset 4 in an 8-byte aligned
9782 diff = offs + sizeof (dtrace_aggid_t);
9784 if ((diff = (diff & (sizeof (uint64_t) - 1))))
9785 offs += sizeof (uint64_t) - diff;
9787 aggbase = offs - sizeof (dtrace_aggid_t);
9788 ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9792 if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9794 * The current offset is not properly aligned; align it.
9796 offs += align - diff;
9799 rec->dtrd_offset = offs;
9801 if (offs + rec->dtrd_size > ecb->dte_needed) {
9802 ecb->dte_needed = offs + rec->dtrd_size;
9804 if (ecb->dte_needed > state->dts_needed)
9805 state->dts_needed = ecb->dte_needed;
9808 if (DTRACEACT_ISAGG(act->dta_kind)) {
9809 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9810 dtrace_action_t *first = agg->dtag_first, *prev;
9812 ASSERT(rec->dtrd_size != 0 && first != NULL);
9814 ASSERT(aggbase != UINT32_MAX);
9816 agg->dtag_base = aggbase;
9818 while ((prev = first->dta_prev) != NULL &&
9819 DTRACEACT_ISAGG(prev->dta_kind)) {
9820 agg = (dtrace_aggregation_t *)prev;
9821 first = agg->dtag_first;
9825 offs = prev->dta_rec.dtrd_offset +
9826 prev->dta_rec.dtrd_size;
9828 offs = sizeof (dtrace_epid_t);
9832 if (!act->dta_intuple)
9833 ecb->dte_size = offs + rec->dtrd_size;
9835 offs += rec->dtrd_size;
9838 wastuple = act->dta_intuple;
9841 if ((act = ecb->dte_action) != NULL &&
9842 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9843 ecb->dte_size == sizeof (dtrace_epid_t)) {
9845 * If the size is still sizeof (dtrace_epid_t), then all
9846 * actions store no data; set the size to 0.
9848 ecb->dte_alignment = maxalign;
9852 * If the needed space is still sizeof (dtrace_epid_t), then
9853 * all actions need no additional space; set the needed
9856 if (ecb->dte_needed == sizeof (dtrace_epid_t))
9857 ecb->dte_needed = 0;
9863 * Set our alignment, and make sure that the dte_size and dte_needed
9864 * are aligned to the size of an EPID.
9866 ecb->dte_alignment = maxalign;
9867 ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9868 ~(sizeof (dtrace_epid_t) - 1);
9869 ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9870 ~(sizeof (dtrace_epid_t) - 1);
9871 ASSERT(ecb->dte_size <= ecb->dte_needed);
9874 static dtrace_action_t *
9875 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9877 dtrace_aggregation_t *agg;
9878 size_t size = sizeof (uint64_t);
9879 int ntuple = desc->dtad_ntuple;
9880 dtrace_action_t *act;
9881 dtrace_recdesc_t *frec;
9882 dtrace_aggid_t aggid;
9883 dtrace_state_t *state = ecb->dte_state;
9885 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9886 agg->dtag_ecb = ecb;
9888 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9890 switch (desc->dtad_kind) {
9892 agg->dtag_initial = INT64_MAX;
9893 agg->dtag_aggregate = dtrace_aggregate_min;
9897 agg->dtag_initial = INT64_MIN;
9898 agg->dtag_aggregate = dtrace_aggregate_max;
9901 case DTRACEAGG_COUNT:
9902 agg->dtag_aggregate = dtrace_aggregate_count;
9905 case DTRACEAGG_QUANTIZE:
9906 agg->dtag_aggregate = dtrace_aggregate_quantize;
9907 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9911 case DTRACEAGG_LQUANTIZE: {
9912 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9913 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9915 agg->dtag_initial = desc->dtad_arg;
9916 agg->dtag_aggregate = dtrace_aggregate_lquantize;
9918 if (step == 0 || levels == 0)
9921 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9925 case DTRACEAGG_LLQUANTIZE: {
9926 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
9927 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
9928 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
9929 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
9932 agg->dtag_initial = desc->dtad_arg;
9933 agg->dtag_aggregate = dtrace_aggregate_llquantize;
9935 if (factor < 2 || low >= high || nsteps < factor)
9939 * Now check that the number of steps evenly divides a power
9940 * of the factor. (This assures both integer bucket size and
9941 * linearity within each magnitude.)
9943 for (v = factor; v < nsteps; v *= factor)
9946 if ((v % nsteps) || (nsteps % factor))
9949 size = (dtrace_aggregate_llquantize_bucket(factor,
9950 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
9955 agg->dtag_aggregate = dtrace_aggregate_avg;
9956 size = sizeof (uint64_t) * 2;
9959 case DTRACEAGG_STDDEV:
9960 agg->dtag_aggregate = dtrace_aggregate_stddev;
9961 size = sizeof (uint64_t) * 4;
9965 agg->dtag_aggregate = dtrace_aggregate_sum;
9972 agg->dtag_action.dta_rec.dtrd_size = size;
9978 * We must make sure that we have enough actions for the n-tuple.
9980 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9981 if (DTRACEACT_ISAGG(act->dta_kind))
9984 if (--ntuple == 0) {
9986 * This is the action with which our n-tuple begins.
9988 agg->dtag_first = act;
9994 * This n-tuple is short by ntuple elements. Return failure.
9996 ASSERT(ntuple != 0);
9998 kmem_free(agg, sizeof (dtrace_aggregation_t));
10003 * If the last action in the tuple has a size of zero, it's actually
10004 * an expression argument for the aggregating action.
10006 ASSERT(ecb->dte_action_last != NULL);
10007 act = ecb->dte_action_last;
10009 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10010 ASSERT(act->dta_difo != NULL);
10012 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10013 agg->dtag_hasarg = 1;
10017 * We need to allocate an id for this aggregation.
10020 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10021 VM_BESTFIT | VM_SLEEP);
10023 aggid = alloc_unr(state->dts_aggid_arena);
10026 if (aggid - 1 >= state->dts_naggregations) {
10027 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10028 dtrace_aggregation_t **aggs;
10029 int naggs = state->dts_naggregations << 1;
10030 int onaggs = state->dts_naggregations;
10032 ASSERT(aggid == state->dts_naggregations + 1);
10035 ASSERT(oaggs == NULL);
10039 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10041 if (oaggs != NULL) {
10042 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10043 kmem_free(oaggs, onaggs * sizeof (*aggs));
10046 state->dts_aggregations = aggs;
10047 state->dts_naggregations = naggs;
10050 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10051 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10053 frec = &agg->dtag_first->dta_rec;
10054 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10055 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10057 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10058 ASSERT(!act->dta_intuple);
10059 act->dta_intuple = 1;
10062 return (&agg->dtag_action);
10066 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10068 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10069 dtrace_state_t *state = ecb->dte_state;
10070 dtrace_aggid_t aggid = agg->dtag_id;
10072 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10074 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10076 free_unr(state->dts_aggid_arena, aggid);
10079 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10080 state->dts_aggregations[aggid - 1] = NULL;
10082 kmem_free(agg, sizeof (dtrace_aggregation_t));
10086 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10088 dtrace_action_t *action, *last;
10089 dtrace_difo_t *dp = desc->dtad_difo;
10090 uint32_t size = 0, align = sizeof (uint8_t), mask;
10091 uint16_t format = 0;
10092 dtrace_recdesc_t *rec;
10093 dtrace_state_t *state = ecb->dte_state;
10094 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
10095 uint64_t arg = desc->dtad_arg;
10097 ASSERT(MUTEX_HELD(&dtrace_lock));
10098 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10100 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10102 * If this is an aggregating action, there must be neither
10103 * a speculate nor a commit on the action chain.
10105 dtrace_action_t *act;
10107 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10108 if (act->dta_kind == DTRACEACT_COMMIT)
10111 if (act->dta_kind == DTRACEACT_SPECULATE)
10115 action = dtrace_ecb_aggregation_create(ecb, desc);
10117 if (action == NULL)
10120 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10121 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10122 dp != NULL && dp->dtdo_destructive)) {
10123 state->dts_destructive = 1;
10126 switch (desc->dtad_kind) {
10127 case DTRACEACT_PRINTF:
10128 case DTRACEACT_PRINTA:
10129 case DTRACEACT_SYSTEM:
10130 case DTRACEACT_FREOPEN:
10132 * We know that our arg is a string -- turn it into a
10136 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA);
10141 ASSERT(arg > KERNELBASE);
10143 format = dtrace_format_add(state,
10144 (char *)(uintptr_t)arg);
10148 case DTRACEACT_LIBACT:
10149 case DTRACEACT_DIFEXPR:
10153 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10156 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10157 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10160 size = opt[DTRACEOPT_STRSIZE];
10165 case DTRACEACT_STACK:
10166 if ((nframes = arg) == 0) {
10167 nframes = opt[DTRACEOPT_STACKFRAMES];
10168 ASSERT(nframes > 0);
10172 size = nframes * sizeof (pc_t);
10175 case DTRACEACT_JSTACK:
10176 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10177 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10179 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10180 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10182 arg = DTRACE_USTACK_ARG(nframes, strsize);
10185 case DTRACEACT_USTACK:
10186 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10187 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10188 strsize = DTRACE_USTACK_STRSIZE(arg);
10189 nframes = opt[DTRACEOPT_USTACKFRAMES];
10190 ASSERT(nframes > 0);
10191 arg = DTRACE_USTACK_ARG(nframes, strsize);
10195 * Save a slot for the pid.
10197 size = (nframes + 1) * sizeof (uint64_t);
10198 size += DTRACE_USTACK_STRSIZE(arg);
10199 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10203 case DTRACEACT_SYM:
10204 case DTRACEACT_MOD:
10205 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10206 sizeof (uint64_t)) ||
10207 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10211 case DTRACEACT_USYM:
10212 case DTRACEACT_UMOD:
10213 case DTRACEACT_UADDR:
10215 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10216 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10220 * We have a slot for the pid, plus a slot for the
10221 * argument. To keep things simple (aligned with
10222 * bitness-neutral sizing), we store each as a 64-bit
10225 size = 2 * sizeof (uint64_t);
10228 case DTRACEACT_STOP:
10229 case DTRACEACT_BREAKPOINT:
10230 case DTRACEACT_PANIC:
10233 case DTRACEACT_CHILL:
10234 case DTRACEACT_DISCARD:
10235 case DTRACEACT_RAISE:
10240 case DTRACEACT_EXIT:
10242 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10243 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10247 case DTRACEACT_SPECULATE:
10248 if (ecb->dte_size > sizeof (dtrace_epid_t))
10254 state->dts_speculates = 1;
10257 case DTRACEACT_PRINTM:
10258 size = dp->dtdo_rtype.dtdt_size;
10261 case DTRACEACT_PRINTT:
10262 size = dp->dtdo_rtype.dtdt_size;
10265 case DTRACEACT_COMMIT: {
10266 dtrace_action_t *act = ecb->dte_action;
10268 for (; act != NULL; act = act->dta_next) {
10269 if (act->dta_kind == DTRACEACT_COMMIT)
10282 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10284 * If this is a data-storing action or a speculate,
10285 * we must be sure that there isn't a commit on the
10288 dtrace_action_t *act = ecb->dte_action;
10290 for (; act != NULL; act = act->dta_next) {
10291 if (act->dta_kind == DTRACEACT_COMMIT)
10296 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10297 action->dta_rec.dtrd_size = size;
10300 action->dta_refcnt = 1;
10301 rec = &action->dta_rec;
10302 size = rec->dtrd_size;
10304 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10305 if (!(size & mask)) {
10311 action->dta_kind = desc->dtad_kind;
10313 if ((action->dta_difo = dp) != NULL)
10314 dtrace_difo_hold(dp);
10316 rec->dtrd_action = action->dta_kind;
10317 rec->dtrd_arg = arg;
10318 rec->dtrd_uarg = desc->dtad_uarg;
10319 rec->dtrd_alignment = (uint16_t)align;
10320 rec->dtrd_format = format;
10322 if ((last = ecb->dte_action_last) != NULL) {
10323 ASSERT(ecb->dte_action != NULL);
10324 action->dta_prev = last;
10325 last->dta_next = action;
10327 ASSERT(ecb->dte_action == NULL);
10328 ecb->dte_action = action;
10331 ecb->dte_action_last = action;
10337 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10339 dtrace_action_t *act = ecb->dte_action, *next;
10340 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10344 if (act != NULL && act->dta_refcnt > 1) {
10345 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10348 for (; act != NULL; act = next) {
10349 next = act->dta_next;
10350 ASSERT(next != NULL || act == ecb->dte_action_last);
10351 ASSERT(act->dta_refcnt == 1);
10353 if ((format = act->dta_rec.dtrd_format) != 0)
10354 dtrace_format_remove(ecb->dte_state, format);
10356 if ((dp = act->dta_difo) != NULL)
10357 dtrace_difo_release(dp, vstate);
10359 if (DTRACEACT_ISAGG(act->dta_kind)) {
10360 dtrace_ecb_aggregation_destroy(ecb, act);
10362 kmem_free(act, sizeof (dtrace_action_t));
10367 ecb->dte_action = NULL;
10368 ecb->dte_action_last = NULL;
10369 ecb->dte_size = sizeof (dtrace_epid_t);
10373 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10376 * We disable the ECB by removing it from its probe.
10378 dtrace_ecb_t *pecb, *prev = NULL;
10379 dtrace_probe_t *probe = ecb->dte_probe;
10381 ASSERT(MUTEX_HELD(&dtrace_lock));
10383 if (probe == NULL) {
10385 * This is the NULL probe; there is nothing to disable.
10390 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10396 ASSERT(pecb != NULL);
10398 if (prev == NULL) {
10399 probe->dtpr_ecb = ecb->dte_next;
10401 prev->dte_next = ecb->dte_next;
10404 if (ecb == probe->dtpr_ecb_last) {
10405 ASSERT(ecb->dte_next == NULL);
10406 probe->dtpr_ecb_last = prev;
10410 * The ECB has been disconnected from the probe; now sync to assure
10411 * that all CPUs have seen the change before returning.
10415 if (probe->dtpr_ecb == NULL) {
10417 * That was the last ECB on the probe; clear the predicate
10418 * cache ID for the probe, disable it and sync one more time
10419 * to assure that we'll never hit it again.
10421 dtrace_provider_t *prov = probe->dtpr_provider;
10423 ASSERT(ecb->dte_next == NULL);
10424 ASSERT(probe->dtpr_ecb_last == NULL);
10425 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10426 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10427 probe->dtpr_id, probe->dtpr_arg);
10431 * There is at least one ECB remaining on the probe. If there
10432 * is _exactly_ one, set the probe's predicate cache ID to be
10433 * the predicate cache ID of the remaining ECB.
10435 ASSERT(probe->dtpr_ecb_last != NULL);
10436 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10438 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10439 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10441 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10444 probe->dtpr_predcache = p->dtp_cacheid;
10447 ecb->dte_next = NULL;
10452 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10454 dtrace_state_t *state = ecb->dte_state;
10455 dtrace_vstate_t *vstate = &state->dts_vstate;
10456 dtrace_predicate_t *pred;
10457 dtrace_epid_t epid = ecb->dte_epid;
10459 ASSERT(MUTEX_HELD(&dtrace_lock));
10460 ASSERT(ecb->dte_next == NULL);
10461 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10463 if ((pred = ecb->dte_predicate) != NULL)
10464 dtrace_predicate_release(pred, vstate);
10466 dtrace_ecb_action_remove(ecb);
10468 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10469 state->dts_ecbs[epid - 1] = NULL;
10471 kmem_free(ecb, sizeof (dtrace_ecb_t));
10474 static dtrace_ecb_t *
10475 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10476 dtrace_enabling_t *enab)
10479 dtrace_predicate_t *pred;
10480 dtrace_actdesc_t *act;
10481 dtrace_provider_t *prov;
10482 dtrace_ecbdesc_t *desc = enab->dten_current;
10484 ASSERT(MUTEX_HELD(&dtrace_lock));
10485 ASSERT(state != NULL);
10487 ecb = dtrace_ecb_add(state, probe);
10488 ecb->dte_uarg = desc->dted_uarg;
10490 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10491 dtrace_predicate_hold(pred);
10492 ecb->dte_predicate = pred;
10495 if (probe != NULL) {
10497 * If the provider shows more leg than the consumer is old
10498 * enough to see, we need to enable the appropriate implicit
10499 * predicate bits to prevent the ecb from activating at
10502 * Providers specifying DTRACE_PRIV_USER at register time
10503 * are stating that they need the /proc-style privilege
10504 * model to be enforced, and this is what DTRACE_COND_OWNER
10505 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10507 prov = probe->dtpr_provider;
10508 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10509 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10510 ecb->dte_cond |= DTRACE_COND_OWNER;
10512 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10513 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10514 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10517 * If the provider shows us kernel innards and the user
10518 * is lacking sufficient privilege, enable the
10519 * DTRACE_COND_USERMODE implicit predicate.
10521 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10522 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10523 ecb->dte_cond |= DTRACE_COND_USERMODE;
10526 if (dtrace_ecb_create_cache != NULL) {
10528 * If we have a cached ecb, we'll use its action list instead
10529 * of creating our own (saving both time and space).
10531 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10532 dtrace_action_t *act = cached->dte_action;
10535 ASSERT(act->dta_refcnt > 0);
10537 ecb->dte_action = act;
10538 ecb->dte_action_last = cached->dte_action_last;
10539 ecb->dte_needed = cached->dte_needed;
10540 ecb->dte_size = cached->dte_size;
10541 ecb->dte_alignment = cached->dte_alignment;
10547 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10548 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10549 dtrace_ecb_destroy(ecb);
10554 dtrace_ecb_resize(ecb);
10556 return (dtrace_ecb_create_cache = ecb);
10560 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10563 dtrace_enabling_t *enab = arg;
10564 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10566 ASSERT(state != NULL);
10568 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10570 * This probe was created in a generation for which this
10571 * enabling has previously created ECBs; we don't want to
10572 * enable it again, so just kick out.
10574 return (DTRACE_MATCH_NEXT);
10577 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10578 return (DTRACE_MATCH_DONE);
10580 dtrace_ecb_enable(ecb);
10581 return (DTRACE_MATCH_NEXT);
10584 static dtrace_ecb_t *
10585 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10589 ASSERT(MUTEX_HELD(&dtrace_lock));
10591 if (id == 0 || id > state->dts_necbs)
10594 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10595 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10597 return (state->dts_ecbs[id - 1]);
10600 static dtrace_aggregation_t *
10601 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10603 dtrace_aggregation_t *agg;
10605 ASSERT(MUTEX_HELD(&dtrace_lock));
10607 if (id == 0 || id > state->dts_naggregations)
10610 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10611 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10612 agg->dtag_id == id);
10614 return (state->dts_aggregations[id - 1]);
10618 * DTrace Buffer Functions
10620 * The following functions manipulate DTrace buffers. Most of these functions
10621 * are called in the context of establishing or processing consumer state;
10622 * exceptions are explicitly noted.
10626 * Note: called from cross call context. This function switches the two
10627 * buffers on a given CPU. The atomicity of this operation is assured by
10628 * disabling interrupts while the actual switch takes place; the disabling of
10629 * interrupts serializes the execution with any execution of dtrace_probe() on
10633 dtrace_buffer_switch(dtrace_buffer_t *buf)
10635 caddr_t tomax = buf->dtb_tomax;
10636 caddr_t xamot = buf->dtb_xamot;
10637 dtrace_icookie_t cookie;
10639 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10640 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10642 cookie = dtrace_interrupt_disable();
10643 buf->dtb_tomax = xamot;
10644 buf->dtb_xamot = tomax;
10645 buf->dtb_xamot_drops = buf->dtb_drops;
10646 buf->dtb_xamot_offset = buf->dtb_offset;
10647 buf->dtb_xamot_errors = buf->dtb_errors;
10648 buf->dtb_xamot_flags = buf->dtb_flags;
10649 buf->dtb_offset = 0;
10650 buf->dtb_drops = 0;
10651 buf->dtb_errors = 0;
10652 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10653 dtrace_interrupt_enable(cookie);
10657 * Note: called from cross call context. This function activates a buffer
10658 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10659 * is guaranteed by the disabling of interrupts.
10662 dtrace_buffer_activate(dtrace_state_t *state)
10664 dtrace_buffer_t *buf;
10665 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10667 buf = &state->dts_buffer[curcpu];
10669 if (buf->dtb_tomax != NULL) {
10671 * We might like to assert that the buffer is marked inactive,
10672 * but this isn't necessarily true: the buffer for the CPU
10673 * that processes the BEGIN probe has its buffer activated
10674 * manually. In this case, we take the (harmless) action
10675 * re-clearing the bit INACTIVE bit.
10677 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10680 dtrace_interrupt_enable(cookie);
10684 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10690 dtrace_buffer_t *buf;
10693 ASSERT(MUTEX_HELD(&cpu_lock));
10694 ASSERT(MUTEX_HELD(&dtrace_lock));
10696 if (size > dtrace_nonroot_maxsize &&
10697 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10703 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10706 buf = &bufs[cp->cpu_id];
10709 * If there is already a buffer allocated for this CPU, it
10710 * is only possible that this is a DR event. In this case,
10712 if (buf->dtb_tomax != NULL) {
10713 ASSERT(buf->dtb_size == size);
10717 ASSERT(buf->dtb_xamot == NULL);
10719 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10722 buf->dtb_size = size;
10723 buf->dtb_flags = flags;
10724 buf->dtb_offset = 0;
10725 buf->dtb_drops = 0;
10727 if (flags & DTRACEBUF_NOSWITCH)
10730 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10732 } while ((cp = cp->cpu_next) != cpu_list);
10740 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10743 buf = &bufs[cp->cpu_id];
10745 if (buf->dtb_xamot != NULL) {
10746 ASSERT(buf->dtb_tomax != NULL);
10747 ASSERT(buf->dtb_size == size);
10748 kmem_free(buf->dtb_xamot, size);
10751 if (buf->dtb_tomax != NULL) {
10752 ASSERT(buf->dtb_size == size);
10753 kmem_free(buf->dtb_tomax, size);
10756 buf->dtb_tomax = NULL;
10757 buf->dtb_xamot = NULL;
10759 } while ((cp = cp->cpu_next) != cpu_list);
10765 #if defined(__amd64__)
10767 * FreeBSD isn't good at limiting the amount of memory we
10768 * ask to malloc, so let's place a limit here before trying
10769 * to do something that might well end in tears at bedtime.
10771 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10775 ASSERT(MUTEX_HELD(&dtrace_lock));
10777 if (cpu != DTRACE_CPUALL && cpu != i)
10783 * If there is already a buffer allocated for this CPU, it
10784 * is only possible that this is a DR event. In this case,
10785 * the buffer size must match our specified size.
10787 if (buf->dtb_tomax != NULL) {
10788 ASSERT(buf->dtb_size == size);
10792 ASSERT(buf->dtb_xamot == NULL);
10794 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10797 buf->dtb_size = size;
10798 buf->dtb_flags = flags;
10799 buf->dtb_offset = 0;
10800 buf->dtb_drops = 0;
10802 if (flags & DTRACEBUF_NOSWITCH)
10805 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10813 * Error allocating memory, so free the buffers that were
10814 * allocated before the failed allocation.
10817 if (cpu != DTRACE_CPUALL && cpu != i)
10822 if (buf->dtb_xamot != NULL) {
10823 ASSERT(buf->dtb_tomax != NULL);
10824 ASSERT(buf->dtb_size == size);
10825 kmem_free(buf->dtb_xamot, size);
10828 if (buf->dtb_tomax != NULL) {
10829 ASSERT(buf->dtb_size == size);
10830 kmem_free(buf->dtb_tomax, size);
10833 buf->dtb_tomax = NULL;
10834 buf->dtb_xamot = NULL;
10844 * Note: called from probe context. This function just increments the drop
10845 * count on a buffer. It has been made a function to allow for the
10846 * possibility of understanding the source of mysterious drop counts. (A
10847 * problem for which one may be particularly disappointed that DTrace cannot
10848 * be used to understand DTrace.)
10851 dtrace_buffer_drop(dtrace_buffer_t *buf)
10857 * Note: called from probe context. This function is called to reserve space
10858 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
10859 * mstate. Returns the new offset in the buffer, or a negative value if an
10860 * error has occurred.
10863 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10864 dtrace_state_t *state, dtrace_mstate_t *mstate)
10866 intptr_t offs = buf->dtb_offset, soffs;
10871 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10874 if ((tomax = buf->dtb_tomax) == NULL) {
10875 dtrace_buffer_drop(buf);
10879 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10880 while (offs & (align - 1)) {
10882 * Assert that our alignment is off by a number which
10883 * is itself sizeof (uint32_t) aligned.
10885 ASSERT(!((align - (offs & (align - 1))) &
10886 (sizeof (uint32_t) - 1)));
10887 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10888 offs += sizeof (uint32_t);
10891 if ((soffs = offs + needed) > buf->dtb_size) {
10892 dtrace_buffer_drop(buf);
10896 if (mstate == NULL)
10899 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10900 mstate->dtms_scratch_size = buf->dtb_size - soffs;
10901 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10906 if (buf->dtb_flags & DTRACEBUF_FILL) {
10907 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10908 (buf->dtb_flags & DTRACEBUF_FULL))
10913 total = needed + (offs & (align - 1));
10916 * For a ring buffer, life is quite a bit more complicated. Before
10917 * we can store any padding, we need to adjust our wrapping offset.
10918 * (If we've never before wrapped or we're not about to, no adjustment
10921 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10922 offs + total > buf->dtb_size) {
10923 woffs = buf->dtb_xamot_offset;
10925 if (offs + total > buf->dtb_size) {
10927 * We can't fit in the end of the buffer. First, a
10928 * sanity check that we can fit in the buffer at all.
10930 if (total > buf->dtb_size) {
10931 dtrace_buffer_drop(buf);
10936 * We're going to be storing at the top of the buffer,
10937 * so now we need to deal with the wrapped offset. We
10938 * only reset our wrapped offset to 0 if it is
10939 * currently greater than the current offset. If it
10940 * is less than the current offset, it is because a
10941 * previous allocation induced a wrap -- but the
10942 * allocation didn't subsequently take the space due
10943 * to an error or false predicate evaluation. In this
10944 * case, we'll just leave the wrapped offset alone: if
10945 * the wrapped offset hasn't been advanced far enough
10946 * for this allocation, it will be adjusted in the
10949 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10957 * Now we know that we're going to be storing to the
10958 * top of the buffer and that there is room for us
10959 * there. We need to clear the buffer from the current
10960 * offset to the end (there may be old gunk there).
10962 while (offs < buf->dtb_size)
10966 * We need to set our offset to zero. And because we
10967 * are wrapping, we need to set the bit indicating as
10968 * much. We can also adjust our needed space back
10969 * down to the space required by the ECB -- we know
10970 * that the top of the buffer is aligned.
10974 buf->dtb_flags |= DTRACEBUF_WRAPPED;
10977 * There is room for us in the buffer, so we simply
10978 * need to check the wrapped offset.
10980 if (woffs < offs) {
10982 * The wrapped offset is less than the offset.
10983 * This can happen if we allocated buffer space
10984 * that induced a wrap, but then we didn't
10985 * subsequently take the space due to an error
10986 * or false predicate evaluation. This is
10987 * okay; we know that _this_ allocation isn't
10988 * going to induce a wrap. We still can't
10989 * reset the wrapped offset to be zero,
10990 * however: the space may have been trashed in
10991 * the previous failed probe attempt. But at
10992 * least the wrapped offset doesn't need to
10993 * be adjusted at all...
10999 while (offs + total > woffs) {
11000 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11003 if (epid == DTRACE_EPIDNONE) {
11004 size = sizeof (uint32_t);
11006 ASSERT(epid <= state->dts_necbs);
11007 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11009 size = state->dts_ecbs[epid - 1]->dte_size;
11012 ASSERT(woffs + size <= buf->dtb_size);
11015 if (woffs + size == buf->dtb_size) {
11017 * We've reached the end of the buffer; we want
11018 * to set the wrapped offset to 0 and break
11019 * out. However, if the offs is 0, then we're
11020 * in a strange edge-condition: the amount of
11021 * space that we want to reserve plus the size
11022 * of the record that we're overwriting is
11023 * greater than the size of the buffer. This
11024 * is problematic because if we reserve the
11025 * space but subsequently don't consume it (due
11026 * to a failed predicate or error) the wrapped
11027 * offset will be 0 -- yet the EPID at offset 0
11028 * will not be committed. This situation is
11029 * relatively easy to deal with: if we're in
11030 * this case, the buffer is indistinguishable
11031 * from one that hasn't wrapped; we need only
11032 * finish the job by clearing the wrapped bit,
11033 * explicitly setting the offset to be 0, and
11034 * zero'ing out the old data in the buffer.
11037 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11038 buf->dtb_offset = 0;
11041 while (woffs < buf->dtb_size)
11042 tomax[woffs++] = 0;
11053 * We have a wrapped offset. It may be that the wrapped offset
11054 * has become zero -- that's okay.
11056 buf->dtb_xamot_offset = woffs;
11061 * Now we can plow the buffer with any necessary padding.
11063 while (offs & (align - 1)) {
11065 * Assert that our alignment is off by a number which
11066 * is itself sizeof (uint32_t) aligned.
11068 ASSERT(!((align - (offs & (align - 1))) &
11069 (sizeof (uint32_t) - 1)));
11070 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11071 offs += sizeof (uint32_t);
11074 if (buf->dtb_flags & DTRACEBUF_FILL) {
11075 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11076 buf->dtb_flags |= DTRACEBUF_FULL;
11081 if (mstate == NULL)
11085 * For ring buffers and fill buffers, the scratch space is always
11086 * the inactive buffer.
11088 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11089 mstate->dtms_scratch_size = buf->dtb_size;
11090 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11096 dtrace_buffer_polish(dtrace_buffer_t *buf)
11098 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11099 ASSERT(MUTEX_HELD(&dtrace_lock));
11101 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11105 * We need to polish the ring buffer. There are three cases:
11107 * - The first (and presumably most common) is that there is no gap
11108 * between the buffer offset and the wrapped offset. In this case,
11109 * there is nothing in the buffer that isn't valid data; we can
11110 * mark the buffer as polished and return.
11112 * - The second (less common than the first but still more common
11113 * than the third) is that there is a gap between the buffer offset
11114 * and the wrapped offset, and the wrapped offset is larger than the
11115 * buffer offset. This can happen because of an alignment issue, or
11116 * can happen because of a call to dtrace_buffer_reserve() that
11117 * didn't subsequently consume the buffer space. In this case,
11118 * we need to zero the data from the buffer offset to the wrapped
11121 * - The third (and least common) is that there is a gap between the
11122 * buffer offset and the wrapped offset, but the wrapped offset is
11123 * _less_ than the buffer offset. This can only happen because a
11124 * call to dtrace_buffer_reserve() induced a wrap, but the space
11125 * was not subsequently consumed. In this case, we need to zero the
11126 * space from the offset to the end of the buffer _and_ from the
11127 * top of the buffer to the wrapped offset.
11129 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11130 bzero(buf->dtb_tomax + buf->dtb_offset,
11131 buf->dtb_xamot_offset - buf->dtb_offset);
11134 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11135 bzero(buf->dtb_tomax + buf->dtb_offset,
11136 buf->dtb_size - buf->dtb_offset);
11137 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11142 dtrace_buffer_free(dtrace_buffer_t *bufs)
11146 for (i = 0; i < NCPU; i++) {
11147 dtrace_buffer_t *buf = &bufs[i];
11149 if (buf->dtb_tomax == NULL) {
11150 ASSERT(buf->dtb_xamot == NULL);
11151 ASSERT(buf->dtb_size == 0);
11155 if (buf->dtb_xamot != NULL) {
11156 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11157 kmem_free(buf->dtb_xamot, buf->dtb_size);
11160 kmem_free(buf->dtb_tomax, buf->dtb_size);
11162 buf->dtb_tomax = NULL;
11163 buf->dtb_xamot = NULL;
11168 * DTrace Enabling Functions
11170 static dtrace_enabling_t *
11171 dtrace_enabling_create(dtrace_vstate_t *vstate)
11173 dtrace_enabling_t *enab;
11175 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11176 enab->dten_vstate = vstate;
11182 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11184 dtrace_ecbdesc_t **ndesc;
11185 size_t osize, nsize;
11188 * We can't add to enablings after we've enabled them, or after we've
11191 ASSERT(enab->dten_probegen == 0);
11192 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11194 if (enab->dten_ndesc < enab->dten_maxdesc) {
11195 enab->dten_desc[enab->dten_ndesc++] = ecb;
11199 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11201 if (enab->dten_maxdesc == 0) {
11202 enab->dten_maxdesc = 1;
11204 enab->dten_maxdesc <<= 1;
11207 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11209 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11210 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11211 bcopy(enab->dten_desc, ndesc, osize);
11212 if (enab->dten_desc != NULL)
11213 kmem_free(enab->dten_desc, osize);
11215 enab->dten_desc = ndesc;
11216 enab->dten_desc[enab->dten_ndesc++] = ecb;
11220 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11221 dtrace_probedesc_t *pd)
11223 dtrace_ecbdesc_t *new;
11224 dtrace_predicate_t *pred;
11225 dtrace_actdesc_t *act;
11228 * We're going to create a new ECB description that matches the
11229 * specified ECB in every way, but has the specified probe description.
11231 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11233 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11234 dtrace_predicate_hold(pred);
11236 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11237 dtrace_actdesc_hold(act);
11239 new->dted_action = ecb->dted_action;
11240 new->dted_pred = ecb->dted_pred;
11241 new->dted_probe = *pd;
11242 new->dted_uarg = ecb->dted_uarg;
11244 dtrace_enabling_add(enab, new);
11248 dtrace_enabling_dump(dtrace_enabling_t *enab)
11252 for (i = 0; i < enab->dten_ndesc; i++) {
11253 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11255 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11256 desc->dtpd_provider, desc->dtpd_mod,
11257 desc->dtpd_func, desc->dtpd_name);
11262 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11265 dtrace_ecbdesc_t *ep;
11266 dtrace_vstate_t *vstate = enab->dten_vstate;
11268 ASSERT(MUTEX_HELD(&dtrace_lock));
11270 for (i = 0; i < enab->dten_ndesc; i++) {
11271 dtrace_actdesc_t *act, *next;
11272 dtrace_predicate_t *pred;
11274 ep = enab->dten_desc[i];
11276 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11277 dtrace_predicate_release(pred, vstate);
11279 for (act = ep->dted_action; act != NULL; act = next) {
11280 next = act->dtad_next;
11281 dtrace_actdesc_release(act, vstate);
11284 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11287 if (enab->dten_desc != NULL)
11288 kmem_free(enab->dten_desc,
11289 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11292 * If this was a retained enabling, decrement the dts_nretained count
11293 * and take it off of the dtrace_retained list.
11295 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11296 dtrace_retained == enab) {
11297 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11298 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11299 enab->dten_vstate->dtvs_state->dts_nretained--;
11302 if (enab->dten_prev == NULL) {
11303 if (dtrace_retained == enab) {
11304 dtrace_retained = enab->dten_next;
11306 if (dtrace_retained != NULL)
11307 dtrace_retained->dten_prev = NULL;
11310 ASSERT(enab != dtrace_retained);
11311 ASSERT(dtrace_retained != NULL);
11312 enab->dten_prev->dten_next = enab->dten_next;
11315 if (enab->dten_next != NULL) {
11316 ASSERT(dtrace_retained != NULL);
11317 enab->dten_next->dten_prev = enab->dten_prev;
11320 kmem_free(enab, sizeof (dtrace_enabling_t));
11324 dtrace_enabling_retain(dtrace_enabling_t *enab)
11326 dtrace_state_t *state;
11328 ASSERT(MUTEX_HELD(&dtrace_lock));
11329 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11330 ASSERT(enab->dten_vstate != NULL);
11332 state = enab->dten_vstate->dtvs_state;
11333 ASSERT(state != NULL);
11336 * We only allow each state to retain dtrace_retain_max enablings.
11338 if (state->dts_nretained >= dtrace_retain_max)
11341 state->dts_nretained++;
11343 if (dtrace_retained == NULL) {
11344 dtrace_retained = enab;
11348 enab->dten_next = dtrace_retained;
11349 dtrace_retained->dten_prev = enab;
11350 dtrace_retained = enab;
11356 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11357 dtrace_probedesc_t *create)
11359 dtrace_enabling_t *new, *enab;
11360 int found = 0, err = ENOENT;
11362 ASSERT(MUTEX_HELD(&dtrace_lock));
11363 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11364 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11365 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11366 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11368 new = dtrace_enabling_create(&state->dts_vstate);
11371 * Iterate over all retained enablings, looking for enablings that
11372 * match the specified state.
11374 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11378 * dtvs_state can only be NULL for helper enablings -- and
11379 * helper enablings can't be retained.
11381 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11383 if (enab->dten_vstate->dtvs_state != state)
11387 * Now iterate over each probe description; we're looking for
11388 * an exact match to the specified probe description.
11390 for (i = 0; i < enab->dten_ndesc; i++) {
11391 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11392 dtrace_probedesc_t *pd = &ep->dted_probe;
11394 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11397 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11400 if (strcmp(pd->dtpd_func, match->dtpd_func))
11403 if (strcmp(pd->dtpd_name, match->dtpd_name))
11407 * We have a winning probe! Add it to our growing
11411 dtrace_enabling_addlike(new, ep, create);
11415 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11416 dtrace_enabling_destroy(new);
11424 dtrace_enabling_retract(dtrace_state_t *state)
11426 dtrace_enabling_t *enab, *next;
11428 ASSERT(MUTEX_HELD(&dtrace_lock));
11431 * Iterate over all retained enablings, destroy the enablings retained
11432 * for the specified state.
11434 for (enab = dtrace_retained; enab != NULL; enab = next) {
11435 next = enab->dten_next;
11438 * dtvs_state can only be NULL for helper enablings -- and
11439 * helper enablings can't be retained.
11441 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11443 if (enab->dten_vstate->dtvs_state == state) {
11444 ASSERT(state->dts_nretained > 0);
11445 dtrace_enabling_destroy(enab);
11449 ASSERT(state->dts_nretained == 0);
11453 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11458 ASSERT(MUTEX_HELD(&cpu_lock));
11459 ASSERT(MUTEX_HELD(&dtrace_lock));
11461 for (i = 0; i < enab->dten_ndesc; i++) {
11462 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11464 enab->dten_current = ep;
11465 enab->dten_error = 0;
11467 matched += dtrace_probe_enable(&ep->dted_probe, enab);
11469 if (enab->dten_error != 0) {
11471 * If we get an error half-way through enabling the
11472 * probes, we kick out -- perhaps with some number of
11473 * them enabled. Leaving enabled probes enabled may
11474 * be slightly confusing for user-level, but we expect
11475 * that no one will attempt to actually drive on in
11476 * the face of such errors. If this is an anonymous
11477 * enabling (indicated with a NULL nmatched pointer),
11478 * we cmn_err() a message. We aren't expecting to
11479 * get such an error -- such as it can exist at all,
11480 * it would be a result of corrupted DOF in the driver
11483 if (nmatched == NULL) {
11484 cmn_err(CE_WARN, "dtrace_enabling_match() "
11485 "error on %p: %d", (void *)ep,
11489 return (enab->dten_error);
11493 enab->dten_probegen = dtrace_probegen;
11494 if (nmatched != NULL)
11495 *nmatched = matched;
11501 dtrace_enabling_matchall(void)
11503 dtrace_enabling_t *enab;
11505 mutex_enter(&cpu_lock);
11506 mutex_enter(&dtrace_lock);
11509 * Iterate over all retained enablings to see if any probes match
11510 * against them. We only perform this operation on enablings for which
11511 * we have sufficient permissions by virtue of being in the global zone
11512 * or in the same zone as the DTrace client. Because we can be called
11513 * after dtrace_detach() has been called, we cannot assert that there
11514 * are retained enablings. We can safely load from dtrace_retained,
11515 * however: the taskq_destroy() at the end of dtrace_detach() will
11516 * block pending our completion.
11518 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11520 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11522 if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11524 (void) dtrace_enabling_match(enab, NULL);
11527 mutex_exit(&dtrace_lock);
11528 mutex_exit(&cpu_lock);
11532 * If an enabling is to be enabled without having matched probes (that is, if
11533 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11534 * enabling must be _primed_ by creating an ECB for every ECB description.
11535 * This must be done to assure that we know the number of speculations, the
11536 * number of aggregations, the minimum buffer size needed, etc. before we
11537 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11538 * enabling any probes, we create ECBs for every ECB decription, but with a
11539 * NULL probe -- which is exactly what this function does.
11542 dtrace_enabling_prime(dtrace_state_t *state)
11544 dtrace_enabling_t *enab;
11547 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11548 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11550 if (enab->dten_vstate->dtvs_state != state)
11554 * We don't want to prime an enabling more than once, lest
11555 * we allow a malicious user to induce resource exhaustion.
11556 * (The ECBs that result from priming an enabling aren't
11557 * leaked -- but they also aren't deallocated until the
11558 * consumer state is destroyed.)
11560 if (enab->dten_primed)
11563 for (i = 0; i < enab->dten_ndesc; i++) {
11564 enab->dten_current = enab->dten_desc[i];
11565 (void) dtrace_probe_enable(NULL, enab);
11568 enab->dten_primed = 1;
11573 * Called to indicate that probes should be provided due to retained
11574 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11575 * must take an initial lap through the enabling calling the dtps_provide()
11576 * entry point explicitly to allow for autocreated probes.
11579 dtrace_enabling_provide(dtrace_provider_t *prv)
11582 dtrace_probedesc_t desc;
11584 ASSERT(MUTEX_HELD(&dtrace_lock));
11585 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11589 prv = dtrace_provider;
11593 dtrace_enabling_t *enab = dtrace_retained;
11594 void *parg = prv->dtpv_arg;
11596 for (; enab != NULL; enab = enab->dten_next) {
11597 for (i = 0; i < enab->dten_ndesc; i++) {
11598 desc = enab->dten_desc[i]->dted_probe;
11599 mutex_exit(&dtrace_lock);
11600 prv->dtpv_pops.dtps_provide(parg, &desc);
11601 mutex_enter(&dtrace_lock);
11604 } while (all && (prv = prv->dtpv_next) != NULL);
11606 mutex_exit(&dtrace_lock);
11607 dtrace_probe_provide(NULL, all ? NULL : prv);
11608 mutex_enter(&dtrace_lock);
11612 * DTrace DOF Functions
11616 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11618 if (dtrace_err_verbose)
11619 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11621 #ifdef DTRACE_ERRDEBUG
11622 dtrace_errdebug(str);
11627 * Create DOF out of a currently enabled state. Right now, we only create
11628 * DOF containing the run-time options -- but this could be expanded to create
11629 * complete DOF representing the enabled state.
11632 dtrace_dof_create(dtrace_state_t *state)
11636 dof_optdesc_t *opt;
11637 int i, len = sizeof (dof_hdr_t) +
11638 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11639 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11641 ASSERT(MUTEX_HELD(&dtrace_lock));
11643 dof = kmem_zalloc(len, KM_SLEEP);
11644 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11645 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11646 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11647 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11649 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11650 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11651 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11652 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11653 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11654 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11656 dof->dofh_flags = 0;
11657 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11658 dof->dofh_secsize = sizeof (dof_sec_t);
11659 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11660 dof->dofh_secoff = sizeof (dof_hdr_t);
11661 dof->dofh_loadsz = len;
11662 dof->dofh_filesz = len;
11666 * Fill in the option section header...
11668 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11669 sec->dofs_type = DOF_SECT_OPTDESC;
11670 sec->dofs_align = sizeof (uint64_t);
11671 sec->dofs_flags = DOF_SECF_LOAD;
11672 sec->dofs_entsize = sizeof (dof_optdesc_t);
11674 opt = (dof_optdesc_t *)((uintptr_t)sec +
11675 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11677 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11678 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11680 for (i = 0; i < DTRACEOPT_MAX; i++) {
11681 opt[i].dofo_option = i;
11682 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11683 opt[i].dofo_value = state->dts_options[i];
11690 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11692 dof_hdr_t hdr, *dof;
11694 ASSERT(!MUTEX_HELD(&dtrace_lock));
11697 * First, we're going to copyin() the sizeof (dof_hdr_t).
11699 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11700 dtrace_dof_error(NULL, "failed to copyin DOF header");
11706 * Now we'll allocate the entire DOF and copy it in -- provided
11707 * that the length isn't outrageous.
11709 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11710 dtrace_dof_error(&hdr, "load size exceeds maximum");
11715 if (hdr.dofh_loadsz < sizeof (hdr)) {
11716 dtrace_dof_error(&hdr, "invalid load size");
11721 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11723 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
11724 kmem_free(dof, hdr.dofh_loadsz);
11733 static __inline uchar_t
11734 dtrace_dof_char(char c) {
11753 return (c - 'A' + 10);
11760 return (c - 'a' + 10);
11762 /* Should not reach here. */
11768 dtrace_dof_property(const char *name)
11772 unsigned int len, i;
11777 * Unfortunately, array of values in .conf files are always (and
11778 * only) interpreted to be integer arrays. We must read our DOF
11779 * as an integer array, and then squeeze it into a byte array.
11781 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11782 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11785 for (i = 0; i < len; i++)
11786 buf[i] = (uchar_t)(((int *)buf)[i]);
11788 if (len < sizeof (dof_hdr_t)) {
11789 ddi_prop_free(buf);
11790 dtrace_dof_error(NULL, "truncated header");
11794 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11795 ddi_prop_free(buf);
11796 dtrace_dof_error(NULL, "truncated DOF");
11800 if (loadsz >= dtrace_dof_maxsize) {
11801 ddi_prop_free(buf);
11802 dtrace_dof_error(NULL, "oversized DOF");
11806 dof = kmem_alloc(loadsz, KM_SLEEP);
11807 bcopy(buf, dof, loadsz);
11808 ddi_prop_free(buf);
11813 if ((p_env = getenv(name)) == NULL)
11816 len = strlen(p_env) / 2;
11818 buf = kmem_alloc(len, KM_SLEEP);
11820 dof = (dof_hdr_t *) buf;
11824 for (i = 0; i < len; i++) {
11825 buf[i] = (dtrace_dof_char(p[0]) << 4) |
11826 dtrace_dof_char(p[1]);
11832 if (len < sizeof (dof_hdr_t)) {
11834 dtrace_dof_error(NULL, "truncated header");
11838 if (len < (loadsz = dof->dofh_loadsz)) {
11840 dtrace_dof_error(NULL, "truncated DOF");
11844 if (loadsz >= dtrace_dof_maxsize) {
11846 dtrace_dof_error(NULL, "oversized DOF");
11855 dtrace_dof_destroy(dof_hdr_t *dof)
11857 kmem_free(dof, dof->dofh_loadsz);
11861 * Return the dof_sec_t pointer corresponding to a given section index. If the
11862 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
11863 * a type other than DOF_SECT_NONE is specified, the header is checked against
11864 * this type and NULL is returned if the types do not match.
11867 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11869 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11870 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11872 if (i >= dof->dofh_secnum) {
11873 dtrace_dof_error(dof, "referenced section index is invalid");
11877 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11878 dtrace_dof_error(dof, "referenced section is not loadable");
11882 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11883 dtrace_dof_error(dof, "referenced section is the wrong type");
11890 static dtrace_probedesc_t *
11891 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11893 dof_probedesc_t *probe;
11895 uintptr_t daddr = (uintptr_t)dof;
11899 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11900 dtrace_dof_error(dof, "invalid probe section");
11904 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11905 dtrace_dof_error(dof, "bad alignment in probe description");
11909 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11910 dtrace_dof_error(dof, "truncated probe description");
11914 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11915 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11917 if (strtab == NULL)
11920 str = daddr + strtab->dofs_offset;
11921 size = strtab->dofs_size;
11923 if (probe->dofp_provider >= strtab->dofs_size) {
11924 dtrace_dof_error(dof, "corrupt probe provider");
11928 (void) strncpy(desc->dtpd_provider,
11929 (char *)(str + probe->dofp_provider),
11930 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11932 if (probe->dofp_mod >= strtab->dofs_size) {
11933 dtrace_dof_error(dof, "corrupt probe module");
11937 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11938 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11940 if (probe->dofp_func >= strtab->dofs_size) {
11941 dtrace_dof_error(dof, "corrupt probe function");
11945 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11946 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11948 if (probe->dofp_name >= strtab->dofs_size) {
11949 dtrace_dof_error(dof, "corrupt probe name");
11953 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11954 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11959 static dtrace_difo_t *
11960 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11965 dof_difohdr_t *dofd;
11966 uintptr_t daddr = (uintptr_t)dof;
11967 size_t max = dtrace_difo_maxsize;
11970 static const struct {
11978 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11979 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11980 sizeof (dif_instr_t), "multiple DIF sections" },
11982 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11983 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11984 sizeof (uint64_t), "multiple integer tables" },
11986 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11987 offsetof(dtrace_difo_t, dtdo_strlen), 0,
11988 sizeof (char), "multiple string tables" },
11990 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11991 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11992 sizeof (uint_t), "multiple variable tables" },
11994 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
11997 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11998 dtrace_dof_error(dof, "invalid DIFO header section");
12002 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12003 dtrace_dof_error(dof, "bad alignment in DIFO header");
12007 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12008 sec->dofs_size % sizeof (dof_secidx_t)) {
12009 dtrace_dof_error(dof, "bad size in DIFO header");
12013 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12014 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12016 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12017 dp->dtdo_rtype = dofd->dofd_rtype;
12019 for (l = 0; l < n; l++) {
12024 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12025 dofd->dofd_links[l])) == NULL)
12026 goto err; /* invalid section link */
12028 if (ttl + subsec->dofs_size > max) {
12029 dtrace_dof_error(dof, "exceeds maximum size");
12033 ttl += subsec->dofs_size;
12035 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12036 if (subsec->dofs_type != difo[i].section)
12039 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12040 dtrace_dof_error(dof, "section not loaded");
12044 if (subsec->dofs_align != difo[i].align) {
12045 dtrace_dof_error(dof, "bad alignment");
12049 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12050 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12052 if (*bufp != NULL) {
12053 dtrace_dof_error(dof, difo[i].msg);
12057 if (difo[i].entsize != subsec->dofs_entsize) {
12058 dtrace_dof_error(dof, "entry size mismatch");
12062 if (subsec->dofs_entsize != 0 &&
12063 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12064 dtrace_dof_error(dof, "corrupt entry size");
12068 *lenp = subsec->dofs_size;
12069 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12070 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12071 *bufp, subsec->dofs_size);
12073 if (subsec->dofs_entsize != 0)
12074 *lenp /= subsec->dofs_entsize;
12080 * If we encounter a loadable DIFO sub-section that is not
12081 * known to us, assume this is a broken program and fail.
12083 if (difo[i].section == DOF_SECT_NONE &&
12084 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12085 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12090 if (dp->dtdo_buf == NULL) {
12092 * We can't have a DIF object without DIF text.
12094 dtrace_dof_error(dof, "missing DIF text");
12099 * Before we validate the DIF object, run through the variable table
12100 * looking for the strings -- if any of their size are under, we'll set
12101 * their size to be the system-wide default string size. Note that
12102 * this should _not_ happen if the "strsize" option has been set --
12103 * in this case, the compiler should have set the size to reflect the
12104 * setting of the option.
12106 for (i = 0; i < dp->dtdo_varlen; i++) {
12107 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12108 dtrace_diftype_t *t = &v->dtdv_type;
12110 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12113 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12114 t->dtdt_size = dtrace_strsize_default;
12117 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12120 dtrace_difo_init(dp, vstate);
12124 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12125 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12126 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12127 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12129 kmem_free(dp, sizeof (dtrace_difo_t));
12133 static dtrace_predicate_t *
12134 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12139 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12142 return (dtrace_predicate_create(dp));
12145 static dtrace_actdesc_t *
12146 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12149 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12150 dof_actdesc_t *desc;
12151 dof_sec_t *difosec;
12153 uintptr_t daddr = (uintptr_t)dof;
12155 dtrace_actkind_t kind;
12157 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12158 dtrace_dof_error(dof, "invalid action section");
12162 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12163 dtrace_dof_error(dof, "truncated action description");
12167 if (sec->dofs_align != sizeof (uint64_t)) {
12168 dtrace_dof_error(dof, "bad alignment in action description");
12172 if (sec->dofs_size < sec->dofs_entsize) {
12173 dtrace_dof_error(dof, "section entry size exceeds total size");
12177 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12178 dtrace_dof_error(dof, "bad entry size in action description");
12182 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12183 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12187 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12188 desc = (dof_actdesc_t *)(daddr +
12189 (uintptr_t)sec->dofs_offset + offs);
12190 kind = (dtrace_actkind_t)desc->dofa_kind;
12192 if (DTRACEACT_ISPRINTFLIKE(kind) &&
12193 (kind != DTRACEACT_PRINTA ||
12194 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12200 * printf()-like actions must have a format string.
12202 if ((strtab = dtrace_dof_sect(dof,
12203 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12206 str = (char *)((uintptr_t)dof +
12207 (uintptr_t)strtab->dofs_offset);
12209 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12210 if (str[i] == '\0')
12214 if (i >= strtab->dofs_size) {
12215 dtrace_dof_error(dof, "bogus format string");
12219 if (i == desc->dofa_arg) {
12220 dtrace_dof_error(dof, "empty format string");
12224 i -= desc->dofa_arg;
12225 fmt = kmem_alloc(i + 1, KM_SLEEP);
12226 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12227 arg = (uint64_t)(uintptr_t)fmt;
12229 if (kind == DTRACEACT_PRINTA) {
12230 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12233 arg = desc->dofa_arg;
12237 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12238 desc->dofa_uarg, arg);
12240 if (last != NULL) {
12241 last->dtad_next = act;
12248 if (desc->dofa_difo == DOF_SECIDX_NONE)
12251 if ((difosec = dtrace_dof_sect(dof,
12252 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12255 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12257 if (act->dtad_difo == NULL)
12261 ASSERT(first != NULL);
12265 for (act = first; act != NULL; act = next) {
12266 next = act->dtad_next;
12267 dtrace_actdesc_release(act, vstate);
12273 static dtrace_ecbdesc_t *
12274 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12277 dtrace_ecbdesc_t *ep;
12278 dof_ecbdesc_t *ecb;
12279 dtrace_probedesc_t *desc;
12280 dtrace_predicate_t *pred = NULL;
12282 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12283 dtrace_dof_error(dof, "truncated ECB description");
12287 if (sec->dofs_align != sizeof (uint64_t)) {
12288 dtrace_dof_error(dof, "bad alignment in ECB description");
12292 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12293 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12298 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12299 ep->dted_uarg = ecb->dofe_uarg;
12300 desc = &ep->dted_probe;
12302 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12305 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12306 if ((sec = dtrace_dof_sect(dof,
12307 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12310 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12313 ep->dted_pred.dtpdd_predicate = pred;
12316 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12317 if ((sec = dtrace_dof_sect(dof,
12318 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12321 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12323 if (ep->dted_action == NULL)
12331 dtrace_predicate_release(pred, vstate);
12332 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12337 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12338 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12339 * site of any user SETX relocations to account for load object base address.
12340 * In the future, if we need other relocations, this function can be extended.
12343 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12345 uintptr_t daddr = (uintptr_t)dof;
12346 dof_relohdr_t *dofr =
12347 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12348 dof_sec_t *ss, *rs, *ts;
12352 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12353 sec->dofs_align != sizeof (dof_secidx_t)) {
12354 dtrace_dof_error(dof, "invalid relocation header");
12358 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12359 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12360 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12362 if (ss == NULL || rs == NULL || ts == NULL)
12363 return (-1); /* dtrace_dof_error() has been called already */
12365 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12366 rs->dofs_align != sizeof (uint64_t)) {
12367 dtrace_dof_error(dof, "invalid relocation section");
12371 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12372 n = rs->dofs_size / rs->dofs_entsize;
12374 for (i = 0; i < n; i++) {
12375 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12377 switch (r->dofr_type) {
12378 case DOF_RELO_NONE:
12380 case DOF_RELO_SETX:
12381 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12382 sizeof (uint64_t) > ts->dofs_size) {
12383 dtrace_dof_error(dof, "bad relocation offset");
12387 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12388 dtrace_dof_error(dof, "misaligned setx relo");
12392 *(uint64_t *)taddr += ubase;
12395 dtrace_dof_error(dof, "invalid relocation type");
12399 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12406 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12407 * header: it should be at the front of a memory region that is at least
12408 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12409 * size. It need not be validated in any other way.
12412 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12413 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12415 uint64_t len = dof->dofh_loadsz, seclen;
12416 uintptr_t daddr = (uintptr_t)dof;
12417 dtrace_ecbdesc_t *ep;
12418 dtrace_enabling_t *enab;
12421 ASSERT(MUTEX_HELD(&dtrace_lock));
12422 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12425 * Check the DOF header identification bytes. In addition to checking
12426 * valid settings, we also verify that unused bits/bytes are zeroed so
12427 * we can use them later without fear of regressing existing binaries.
12429 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12430 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12431 dtrace_dof_error(dof, "DOF magic string mismatch");
12435 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12436 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12437 dtrace_dof_error(dof, "DOF has invalid data model");
12441 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12442 dtrace_dof_error(dof, "DOF encoding mismatch");
12446 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12447 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12448 dtrace_dof_error(dof, "DOF version mismatch");
12452 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12453 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12457 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12458 dtrace_dof_error(dof, "DOF uses too many integer registers");
12462 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12463 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12467 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12468 if (dof->dofh_ident[i] != 0) {
12469 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12474 if (dof->dofh_flags & ~DOF_FL_VALID) {
12475 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12479 if (dof->dofh_secsize == 0) {
12480 dtrace_dof_error(dof, "zero section header size");
12485 * Check that the section headers don't exceed the amount of DOF
12486 * data. Note that we cast the section size and number of sections
12487 * to uint64_t's to prevent possible overflow in the multiplication.
12489 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12491 if (dof->dofh_secoff > len || seclen > len ||
12492 dof->dofh_secoff + seclen > len) {
12493 dtrace_dof_error(dof, "truncated section headers");
12497 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12498 dtrace_dof_error(dof, "misaligned section headers");
12502 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12503 dtrace_dof_error(dof, "misaligned section size");
12508 * Take an initial pass through the section headers to be sure that
12509 * the headers don't have stray offsets. If the 'noprobes' flag is
12510 * set, do not permit sections relating to providers, probes, or args.
12512 for (i = 0; i < dof->dofh_secnum; i++) {
12513 dof_sec_t *sec = (dof_sec_t *)(daddr +
12514 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12517 switch (sec->dofs_type) {
12518 case DOF_SECT_PROVIDER:
12519 case DOF_SECT_PROBES:
12520 case DOF_SECT_PRARGS:
12521 case DOF_SECT_PROFFS:
12522 dtrace_dof_error(dof, "illegal sections "
12528 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12529 continue; /* just ignore non-loadable sections */
12531 if (sec->dofs_align & (sec->dofs_align - 1)) {
12532 dtrace_dof_error(dof, "bad section alignment");
12536 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12537 dtrace_dof_error(dof, "misaligned section");
12541 if (sec->dofs_offset > len || sec->dofs_size > len ||
12542 sec->dofs_offset + sec->dofs_size > len) {
12543 dtrace_dof_error(dof, "corrupt section header");
12547 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12548 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12549 dtrace_dof_error(dof, "non-terminating string table");
12555 * Take a second pass through the sections and locate and perform any
12556 * relocations that are present. We do this after the first pass to
12557 * be sure that all sections have had their headers validated.
12559 for (i = 0; i < dof->dofh_secnum; i++) {
12560 dof_sec_t *sec = (dof_sec_t *)(daddr +
12561 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12563 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12564 continue; /* skip sections that are not loadable */
12566 switch (sec->dofs_type) {
12567 case DOF_SECT_URELHDR:
12568 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12574 if ((enab = *enabp) == NULL)
12575 enab = *enabp = dtrace_enabling_create(vstate);
12577 for (i = 0; i < dof->dofh_secnum; i++) {
12578 dof_sec_t *sec = (dof_sec_t *)(daddr +
12579 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12581 if (sec->dofs_type != DOF_SECT_ECBDESC)
12584 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12585 dtrace_enabling_destroy(enab);
12590 dtrace_enabling_add(enab, ep);
12597 * Process DOF for any options. This routine assumes that the DOF has been
12598 * at least processed by dtrace_dof_slurp().
12601 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12606 dof_optdesc_t *desc;
12608 for (i = 0; i < dof->dofh_secnum; i++) {
12609 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12610 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12612 if (sec->dofs_type != DOF_SECT_OPTDESC)
12615 if (sec->dofs_align != sizeof (uint64_t)) {
12616 dtrace_dof_error(dof, "bad alignment in "
12617 "option description");
12621 if ((entsize = sec->dofs_entsize) == 0) {
12622 dtrace_dof_error(dof, "zeroed option entry size");
12626 if (entsize < sizeof (dof_optdesc_t)) {
12627 dtrace_dof_error(dof, "bad option entry size");
12631 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12632 desc = (dof_optdesc_t *)((uintptr_t)dof +
12633 (uintptr_t)sec->dofs_offset + offs);
12635 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12636 dtrace_dof_error(dof, "non-zero option string");
12640 if (desc->dofo_value == DTRACEOPT_UNSET) {
12641 dtrace_dof_error(dof, "unset option");
12645 if ((rval = dtrace_state_option(state,
12646 desc->dofo_option, desc->dofo_value)) != 0) {
12647 dtrace_dof_error(dof, "rejected option");
12657 * DTrace Consumer State Functions
12660 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12662 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12665 dtrace_dynvar_t *dvar, *next, *start;
12668 ASSERT(MUTEX_HELD(&dtrace_lock));
12669 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12671 bzero(dstate, sizeof (dtrace_dstate_t));
12673 if ((dstate->dtds_chunksize = chunksize) == 0)
12674 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12676 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12679 if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
12682 dstate->dtds_size = size;
12683 dstate->dtds_base = base;
12684 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12685 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12687 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12689 if (hashsize != 1 && (hashsize & 1))
12692 dstate->dtds_hashsize = hashsize;
12693 dstate->dtds_hash = dstate->dtds_base;
12696 * Set all of our hash buckets to point to the single sink, and (if
12697 * it hasn't already been set), set the sink's hash value to be the
12698 * sink sentinel value. The sink is needed for dynamic variable
12699 * lookups to know that they have iterated over an entire, valid hash
12702 for (i = 0; i < hashsize; i++)
12703 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12705 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12706 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12709 * Determine number of active CPUs. Divide free list evenly among
12712 start = (dtrace_dynvar_t *)
12713 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12714 limit = (uintptr_t)base + size;
12716 maxper = (limit - (uintptr_t)start) / NCPU;
12717 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12722 for (i = 0; i < NCPU; i++) {
12724 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12727 * If we don't even have enough chunks to make it once through
12728 * NCPUs, we're just going to allocate everything to the first
12729 * CPU. And if we're on the last CPU, we're going to allocate
12730 * whatever is left over. In either case, we set the limit to
12731 * be the limit of the dynamic variable space.
12733 if (maxper == 0 || i == NCPU - 1) {
12734 limit = (uintptr_t)base + size;
12737 limit = (uintptr_t)start + maxper;
12738 start = (dtrace_dynvar_t *)limit;
12741 ASSERT(limit <= (uintptr_t)base + size);
12744 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12745 dstate->dtds_chunksize);
12747 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12750 dvar->dtdv_next = next;
12762 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12764 ASSERT(MUTEX_HELD(&cpu_lock));
12766 if (dstate->dtds_base == NULL)
12769 kmem_free(dstate->dtds_base, dstate->dtds_size);
12770 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12774 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12777 * Logical XOR, where are you?
12779 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12781 if (vstate->dtvs_nglobals > 0) {
12782 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12783 sizeof (dtrace_statvar_t *));
12786 if (vstate->dtvs_ntlocals > 0) {
12787 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12788 sizeof (dtrace_difv_t));
12791 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
12793 if (vstate->dtvs_nlocals > 0) {
12794 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
12795 sizeof (dtrace_statvar_t *));
12801 dtrace_state_clean(dtrace_state_t *state)
12803 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12806 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12807 dtrace_speculation_clean(state);
12811 dtrace_state_deadman(dtrace_state_t *state)
12817 now = dtrace_gethrtime();
12819 if (state != dtrace_anon.dta_state &&
12820 now - state->dts_laststatus >= dtrace_deadman_user)
12824 * We must be sure that dts_alive never appears to be less than the
12825 * value upon entry to dtrace_state_deadman(), and because we lack a
12826 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12827 * store INT64_MAX to it, followed by a memory barrier, followed by
12828 * the new value. This assures that dts_alive never appears to be
12829 * less than its true value, regardless of the order in which the
12830 * stores to the underlying storage are issued.
12832 state->dts_alive = INT64_MAX;
12833 dtrace_membar_producer();
12834 state->dts_alive = now;
12838 dtrace_state_clean(void *arg)
12840 dtrace_state_t *state = arg;
12841 dtrace_optval_t *opt = state->dts_options;
12843 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12846 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12847 dtrace_speculation_clean(state);
12849 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
12850 dtrace_state_clean, state);
12854 dtrace_state_deadman(void *arg)
12856 dtrace_state_t *state = arg;
12861 dtrace_debug_output();
12863 now = dtrace_gethrtime();
12865 if (state != dtrace_anon.dta_state &&
12866 now - state->dts_laststatus >= dtrace_deadman_user)
12870 * We must be sure that dts_alive never appears to be less than the
12871 * value upon entry to dtrace_state_deadman(), and because we lack a
12872 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12873 * store INT64_MAX to it, followed by a memory barrier, followed by
12874 * the new value. This assures that dts_alive never appears to be
12875 * less than its true value, regardless of the order in which the
12876 * stores to the underlying storage are issued.
12878 state->dts_alive = INT64_MAX;
12879 dtrace_membar_producer();
12880 state->dts_alive = now;
12882 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
12883 dtrace_state_deadman, state);
12887 static dtrace_state_t *
12889 dtrace_state_create(dev_t *devp, cred_t *cr)
12891 dtrace_state_create(struct cdev *dev)
12902 dtrace_state_t *state;
12903 dtrace_optval_t *opt;
12904 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12906 ASSERT(MUTEX_HELD(&dtrace_lock));
12907 ASSERT(MUTEX_HELD(&cpu_lock));
12910 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12911 VM_BESTFIT | VM_SLEEP);
12913 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12914 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12918 state = ddi_get_soft_state(dtrace_softstate, minor);
12925 /* Allocate memory for the state. */
12926 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
12929 state->dts_epid = DTRACE_EPIDNONE + 1;
12931 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
12933 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12934 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12936 if (devp != NULL) {
12937 major = getemajor(*devp);
12939 major = ddi_driver_major(dtrace_devi);
12942 state->dts_dev = makedevice(major, minor);
12945 *devp = state->dts_dev;
12947 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
12948 state->dts_dev = dev;
12952 * We allocate NCPU buffers. On the one hand, this can be quite
12953 * a bit of memory per instance (nearly 36K on a Starcat). On the
12954 * other hand, it saves an additional memory reference in the probe
12957 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12958 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12961 state->dts_cleaner = CYCLIC_NONE;
12962 state->dts_deadman = CYCLIC_NONE;
12964 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
12965 callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
12967 state->dts_vstate.dtvs_state = state;
12969 for (i = 0; i < DTRACEOPT_MAX; i++)
12970 state->dts_options[i] = DTRACEOPT_UNSET;
12973 * Set the default options.
12975 opt = state->dts_options;
12976 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12977 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12978 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12979 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12980 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12981 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12982 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12983 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12984 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12985 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12986 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12987 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12988 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12989 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12991 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12994 * Depending on the user credentials, we set flag bits which alter probe
12995 * visibility or the amount of destructiveness allowed. In the case of
12996 * actual anonymous tracing, or the possession of all privileges, all of
12997 * the normal checks are bypassed.
12999 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13000 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13001 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13004 * Set up the credentials for this instantiation. We take a
13005 * hold on the credential to prevent it from disappearing on
13006 * us; this in turn prevents the zone_t referenced by this
13007 * credential from disappearing. This means that we can
13008 * examine the credential and the zone from probe context.
13011 state->dts_cred.dcr_cred = cr;
13014 * CRA_PROC means "we have *some* privilege for dtrace" and
13015 * unlocks the use of variables like pid, zonename, etc.
13017 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13018 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13019 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13023 * dtrace_user allows use of syscall and profile providers.
13024 * If the user also has proc_owner and/or proc_zone, we
13025 * extend the scope to include additional visibility and
13026 * destructive power.
13028 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13029 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13030 state->dts_cred.dcr_visible |=
13031 DTRACE_CRV_ALLPROC;
13033 state->dts_cred.dcr_action |=
13034 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13037 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13038 state->dts_cred.dcr_visible |=
13039 DTRACE_CRV_ALLZONE;
13041 state->dts_cred.dcr_action |=
13042 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13046 * If we have all privs in whatever zone this is,
13047 * we can do destructive things to processes which
13048 * have altered credentials.
13051 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13052 cr->cr_zone->zone_privset)) {
13053 state->dts_cred.dcr_action |=
13054 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13060 * Holding the dtrace_kernel privilege also implies that
13061 * the user has the dtrace_user privilege from a visibility
13062 * perspective. But without further privileges, some
13063 * destructive actions are not available.
13065 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13067 * Make all probes in all zones visible. However,
13068 * this doesn't mean that all actions become available
13071 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13072 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13074 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13077 * Holding proc_owner means that destructive actions
13078 * for *this* zone are allowed.
13080 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13081 state->dts_cred.dcr_action |=
13082 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13085 * Holding proc_zone means that destructive actions
13086 * for this user/group ID in all zones is allowed.
13088 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13089 state->dts_cred.dcr_action |=
13090 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13094 * If we have all privs in whatever zone this is,
13095 * we can do destructive things to processes which
13096 * have altered credentials.
13098 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13099 cr->cr_zone->zone_privset)) {
13100 state->dts_cred.dcr_action |=
13101 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13107 * Holding the dtrace_proc privilege gives control over fasttrap
13108 * and pid providers. We need to grant wider destructive
13109 * privileges in the event that the user has proc_owner and/or
13112 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13113 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13114 state->dts_cred.dcr_action |=
13115 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13117 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13118 state->dts_cred.dcr_action |=
13119 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13127 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13129 dtrace_optval_t *opt = state->dts_options, size;
13130 processorid_t cpu = 0;;
13131 int flags = 0, rval;
13133 ASSERT(MUTEX_HELD(&dtrace_lock));
13134 ASSERT(MUTEX_HELD(&cpu_lock));
13135 ASSERT(which < DTRACEOPT_MAX);
13136 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13137 (state == dtrace_anon.dta_state &&
13138 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13140 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13143 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13144 cpu = opt[DTRACEOPT_CPU];
13146 if (which == DTRACEOPT_SPECSIZE)
13147 flags |= DTRACEBUF_NOSWITCH;
13149 if (which == DTRACEOPT_BUFSIZE) {
13150 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13151 flags |= DTRACEBUF_RING;
13153 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13154 flags |= DTRACEBUF_FILL;
13156 if (state != dtrace_anon.dta_state ||
13157 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13158 flags |= DTRACEBUF_INACTIVE;
13161 for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
13163 * The size must be 8-byte aligned. If the size is not 8-byte
13164 * aligned, drop it down by the difference.
13166 if (size & (sizeof (uint64_t) - 1))
13167 size -= size & (sizeof (uint64_t) - 1);
13169 if (size < state->dts_reserve) {
13171 * Buffers always must be large enough to accommodate
13172 * their prereserved space. We return E2BIG instead
13173 * of ENOMEM in this case to allow for user-level
13174 * software to differentiate the cases.
13179 rval = dtrace_buffer_alloc(buf, size, flags, cpu);
13181 if (rval != ENOMEM) {
13186 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13194 dtrace_state_buffers(dtrace_state_t *state)
13196 dtrace_speculation_t *spec = state->dts_speculations;
13199 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13200 DTRACEOPT_BUFSIZE)) != 0)
13203 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13204 DTRACEOPT_AGGSIZE)) != 0)
13207 for (i = 0; i < state->dts_nspeculations; i++) {
13208 if ((rval = dtrace_state_buffer(state,
13209 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13217 dtrace_state_prereserve(dtrace_state_t *state)
13220 dtrace_probe_t *probe;
13222 state->dts_reserve = 0;
13224 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13228 * If our buffer policy is a "fill" buffer policy, we need to set the
13229 * prereserved space to be the space required by the END probes.
13231 probe = dtrace_probes[dtrace_probeid_end - 1];
13232 ASSERT(probe != NULL);
13234 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13235 if (ecb->dte_state != state)
13238 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13243 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13245 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13246 dtrace_speculation_t *spec;
13247 dtrace_buffer_t *buf;
13249 cyc_handler_t hdlr;
13252 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13253 dtrace_icookie_t cookie;
13255 mutex_enter(&cpu_lock);
13256 mutex_enter(&dtrace_lock);
13258 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13264 * Before we can perform any checks, we must prime all of the
13265 * retained enablings that correspond to this state.
13267 dtrace_enabling_prime(state);
13269 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13274 dtrace_state_prereserve(state);
13277 * Now we want to do is try to allocate our speculations.
13278 * We do not automatically resize the number of speculations; if
13279 * this fails, we will fail the operation.
13281 nspec = opt[DTRACEOPT_NSPEC];
13282 ASSERT(nspec != DTRACEOPT_UNSET);
13284 if (nspec > INT_MAX) {
13289 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
13291 if (spec == NULL) {
13296 state->dts_speculations = spec;
13297 state->dts_nspeculations = (int)nspec;
13299 for (i = 0; i < nspec; i++) {
13300 if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
13305 spec[i].dtsp_buffer = buf;
13308 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13309 if (dtrace_anon.dta_state == NULL) {
13314 if (state->dts_necbs != 0) {
13319 state->dts_anon = dtrace_anon_grab();
13320 ASSERT(state->dts_anon != NULL);
13321 state = state->dts_anon;
13324 * We want "grabanon" to be set in the grabbed state, so we'll
13325 * copy that option value from the grabbing state into the
13328 state->dts_options[DTRACEOPT_GRABANON] =
13329 opt[DTRACEOPT_GRABANON];
13331 *cpu = dtrace_anon.dta_beganon;
13334 * If the anonymous state is active (as it almost certainly
13335 * is if the anonymous enabling ultimately matched anything),
13336 * we don't allow any further option processing -- but we
13337 * don't return failure.
13339 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13343 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13344 opt[DTRACEOPT_AGGSIZE] != 0) {
13345 if (state->dts_aggregations == NULL) {
13347 * We're not going to create an aggregation buffer
13348 * because we don't have any ECBs that contain
13349 * aggregations -- set this option to 0.
13351 opt[DTRACEOPT_AGGSIZE] = 0;
13354 * If we have an aggregation buffer, we must also have
13355 * a buffer to use as scratch.
13357 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13358 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13359 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13364 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13365 opt[DTRACEOPT_SPECSIZE] != 0) {
13366 if (!state->dts_speculates) {
13368 * We're not going to create speculation buffers
13369 * because we don't have any ECBs that actually
13370 * speculate -- set the speculation size to 0.
13372 opt[DTRACEOPT_SPECSIZE] = 0;
13377 * The bare minimum size for any buffer that we're actually going to
13378 * do anything to is sizeof (uint64_t).
13380 sz = sizeof (uint64_t);
13382 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13383 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13384 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13386 * A buffer size has been explicitly set to 0 (or to a size
13387 * that will be adjusted to 0) and we need the space -- we
13388 * need to return failure. We return ENOSPC to differentiate
13389 * it from failing to allocate a buffer due to failure to meet
13390 * the reserve (for which we return E2BIG).
13396 if ((rval = dtrace_state_buffers(state)) != 0)
13399 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13400 sz = dtrace_dstate_defsize;
13403 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13408 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13410 } while (sz >>= 1);
13412 opt[DTRACEOPT_DYNVARSIZE] = sz;
13417 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13418 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13420 if (opt[DTRACEOPT_CLEANRATE] == 0)
13421 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13423 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13424 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13426 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13427 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13429 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13431 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13432 hdlr.cyh_arg = state;
13433 hdlr.cyh_level = CY_LOW_LEVEL;
13436 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13438 state->dts_cleaner = cyclic_add(&hdlr, &when);
13440 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13441 hdlr.cyh_arg = state;
13442 hdlr.cyh_level = CY_LOW_LEVEL;
13445 when.cyt_interval = dtrace_deadman_interval;
13447 state->dts_deadman = cyclic_add(&hdlr, &when);
13449 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13450 dtrace_state_clean, state);
13451 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13452 dtrace_state_deadman, state);
13455 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13458 * Now it's time to actually fire the BEGIN probe. We need to disable
13459 * interrupts here both to record the CPU on which we fired the BEGIN
13460 * probe (the data from this CPU will be processed first at user
13461 * level) and to manually activate the buffer for this CPU.
13463 cookie = dtrace_interrupt_disable();
13465 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13466 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13468 dtrace_probe(dtrace_probeid_begin,
13469 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13470 dtrace_interrupt_enable(cookie);
13472 * We may have had an exit action from a BEGIN probe; only change our
13473 * state to ACTIVE if we're still in WARMUP.
13475 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13476 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13478 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13479 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13482 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13483 * want each CPU to transition its principal buffer out of the
13484 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13485 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13486 * atomically transition from processing none of a state's ECBs to
13487 * processing all of them.
13489 dtrace_xcall(DTRACE_CPUALL,
13490 (dtrace_xcall_t)dtrace_buffer_activate, state);
13494 dtrace_buffer_free(state->dts_buffer);
13495 dtrace_buffer_free(state->dts_aggbuffer);
13497 if ((nspec = state->dts_nspeculations) == 0) {
13498 ASSERT(state->dts_speculations == NULL);
13502 spec = state->dts_speculations;
13503 ASSERT(spec != NULL);
13505 for (i = 0; i < state->dts_nspeculations; i++) {
13506 if ((buf = spec[i].dtsp_buffer) == NULL)
13509 dtrace_buffer_free(buf);
13510 kmem_free(buf, bufsize);
13513 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13514 state->dts_nspeculations = 0;
13515 state->dts_speculations = NULL;
13518 mutex_exit(&dtrace_lock);
13519 mutex_exit(&cpu_lock);
13525 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13527 dtrace_icookie_t cookie;
13529 ASSERT(MUTEX_HELD(&dtrace_lock));
13531 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13532 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13536 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13537 * to be sure that every CPU has seen it. See below for the details
13538 * on why this is done.
13540 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13544 * By this point, it is impossible for any CPU to be still processing
13545 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13546 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13547 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13548 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13549 * iff we're in the END probe.
13551 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13553 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13556 * Finally, we can release the reserve and call the END probe. We
13557 * disable interrupts across calling the END probe to allow us to
13558 * return the CPU on which we actually called the END probe. This
13559 * allows user-land to be sure that this CPU's principal buffer is
13562 state->dts_reserve = 0;
13564 cookie = dtrace_interrupt_disable();
13566 dtrace_probe(dtrace_probeid_end,
13567 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13568 dtrace_interrupt_enable(cookie);
13570 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13577 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13578 dtrace_optval_t val)
13580 ASSERT(MUTEX_HELD(&dtrace_lock));
13582 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13585 if (option >= DTRACEOPT_MAX)
13588 if (option != DTRACEOPT_CPU && val < 0)
13592 case DTRACEOPT_DESTRUCTIVE:
13593 if (dtrace_destructive_disallow)
13596 state->dts_cred.dcr_destructive = 1;
13599 case DTRACEOPT_BUFSIZE:
13600 case DTRACEOPT_DYNVARSIZE:
13601 case DTRACEOPT_AGGSIZE:
13602 case DTRACEOPT_SPECSIZE:
13603 case DTRACEOPT_STRSIZE:
13607 if (val >= LONG_MAX) {
13609 * If this is an otherwise negative value, set it to
13610 * the highest multiple of 128m less than LONG_MAX.
13611 * Technically, we're adjusting the size without
13612 * regard to the buffer resizing policy, but in fact,
13613 * this has no effect -- if we set the buffer size to
13614 * ~LONG_MAX and the buffer policy is ultimately set to
13615 * be "manual", the buffer allocation is guaranteed to
13616 * fail, if only because the allocation requires two
13617 * buffers. (We set the the size to the highest
13618 * multiple of 128m because it ensures that the size
13619 * will remain a multiple of a megabyte when
13620 * repeatedly halved -- all the way down to 15m.)
13622 val = LONG_MAX - (1 << 27) + 1;
13626 state->dts_options[option] = val;
13632 dtrace_state_destroy(dtrace_state_t *state)
13635 dtrace_vstate_t *vstate = &state->dts_vstate;
13637 minor_t minor = getminor(state->dts_dev);
13639 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13640 dtrace_speculation_t *spec = state->dts_speculations;
13641 int nspec = state->dts_nspeculations;
13644 ASSERT(MUTEX_HELD(&dtrace_lock));
13645 ASSERT(MUTEX_HELD(&cpu_lock));
13648 * First, retract any retained enablings for this state.
13650 dtrace_enabling_retract(state);
13651 ASSERT(state->dts_nretained == 0);
13653 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13654 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13656 * We have managed to come into dtrace_state_destroy() on a
13657 * hot enabling -- almost certainly because of a disorderly
13658 * shutdown of a consumer. (That is, a consumer that is
13659 * exiting without having called dtrace_stop().) In this case,
13660 * we're going to set our activity to be KILLED, and then
13661 * issue a sync to be sure that everyone is out of probe
13662 * context before we start blowing away ECBs.
13664 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13669 * Release the credential hold we took in dtrace_state_create().
13671 if (state->dts_cred.dcr_cred != NULL)
13672 crfree(state->dts_cred.dcr_cred);
13675 * Now we can safely disable and destroy any enabled probes. Because
13676 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13677 * (especially if they're all enabled), we take two passes through the
13678 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13679 * in the second we disable whatever is left over.
13681 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13682 for (i = 0; i < state->dts_necbs; i++) {
13683 if ((ecb = state->dts_ecbs[i]) == NULL)
13686 if (match && ecb->dte_probe != NULL) {
13687 dtrace_probe_t *probe = ecb->dte_probe;
13688 dtrace_provider_t *prov = probe->dtpr_provider;
13690 if (!(prov->dtpv_priv.dtpp_flags & match))
13694 dtrace_ecb_disable(ecb);
13695 dtrace_ecb_destroy(ecb);
13703 * Before we free the buffers, perform one more sync to assure that
13704 * every CPU is out of probe context.
13708 dtrace_buffer_free(state->dts_buffer);
13709 dtrace_buffer_free(state->dts_aggbuffer);
13711 for (i = 0; i < nspec; i++)
13712 dtrace_buffer_free(spec[i].dtsp_buffer);
13715 if (state->dts_cleaner != CYCLIC_NONE)
13716 cyclic_remove(state->dts_cleaner);
13718 if (state->dts_deadman != CYCLIC_NONE)
13719 cyclic_remove(state->dts_deadman);
13721 callout_stop(&state->dts_cleaner);
13722 callout_drain(&state->dts_cleaner);
13723 callout_stop(&state->dts_deadman);
13724 callout_drain(&state->dts_deadman);
13727 dtrace_dstate_fini(&vstate->dtvs_dynvars);
13728 dtrace_vstate_fini(vstate);
13729 if (state->dts_ecbs != NULL)
13730 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13732 if (state->dts_aggregations != NULL) {
13734 for (i = 0; i < state->dts_naggregations; i++)
13735 ASSERT(state->dts_aggregations[i] == NULL);
13737 ASSERT(state->dts_naggregations > 0);
13738 kmem_free(state->dts_aggregations,
13739 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13742 kmem_free(state->dts_buffer, bufsize);
13743 kmem_free(state->dts_aggbuffer, bufsize);
13745 for (i = 0; i < nspec; i++)
13746 kmem_free(spec[i].dtsp_buffer, bufsize);
13749 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13751 dtrace_format_destroy(state);
13753 if (state->dts_aggid_arena != NULL) {
13755 vmem_destroy(state->dts_aggid_arena);
13757 delete_unrhdr(state->dts_aggid_arena);
13759 state->dts_aggid_arena = NULL;
13762 ddi_soft_state_free(dtrace_softstate, minor);
13763 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13768 * DTrace Anonymous Enabling Functions
13770 static dtrace_state_t *
13771 dtrace_anon_grab(void)
13773 dtrace_state_t *state;
13775 ASSERT(MUTEX_HELD(&dtrace_lock));
13777 if ((state = dtrace_anon.dta_state) == NULL) {
13778 ASSERT(dtrace_anon.dta_enabling == NULL);
13782 ASSERT(dtrace_anon.dta_enabling != NULL);
13783 ASSERT(dtrace_retained != NULL);
13785 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13786 dtrace_anon.dta_enabling = NULL;
13787 dtrace_anon.dta_state = NULL;
13793 dtrace_anon_property(void)
13796 dtrace_state_t *state;
13798 char c[32]; /* enough for "dof-data-" + digits */
13800 ASSERT(MUTEX_HELD(&dtrace_lock));
13801 ASSERT(MUTEX_HELD(&cpu_lock));
13803 for (i = 0; ; i++) {
13804 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
13806 dtrace_err_verbose = 1;
13808 if ((dof = dtrace_dof_property(c)) == NULL) {
13809 dtrace_err_verbose = 0;
13815 * We want to create anonymous state, so we need to transition
13816 * the kernel debugger to indicate that DTrace is active. If
13817 * this fails (e.g. because the debugger has modified text in
13818 * some way), we won't continue with the processing.
13820 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13821 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
13822 "enabling ignored.");
13823 dtrace_dof_destroy(dof);
13829 * If we haven't allocated an anonymous state, we'll do so now.
13831 if ((state = dtrace_anon.dta_state) == NULL) {
13833 state = dtrace_state_create(NULL, NULL);
13835 state = dtrace_state_create(NULL);
13837 dtrace_anon.dta_state = state;
13839 if (state == NULL) {
13841 * This basically shouldn't happen: the only
13842 * failure mode from dtrace_state_create() is a
13843 * failure of ddi_soft_state_zalloc() that
13844 * itself should never happen. Still, the
13845 * interface allows for a failure mode, and
13846 * we want to fail as gracefully as possible:
13847 * we'll emit an error message and cease
13848 * processing anonymous state in this case.
13850 cmn_err(CE_WARN, "failed to create "
13851 "anonymous state");
13852 dtrace_dof_destroy(dof);
13857 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
13858 &dtrace_anon.dta_enabling, 0, B_TRUE);
13861 rv = dtrace_dof_options(dof, state);
13863 dtrace_err_verbose = 0;
13864 dtrace_dof_destroy(dof);
13868 * This is malformed DOF; chuck any anonymous state
13871 ASSERT(dtrace_anon.dta_enabling == NULL);
13872 dtrace_state_destroy(state);
13873 dtrace_anon.dta_state = NULL;
13877 ASSERT(dtrace_anon.dta_enabling != NULL);
13880 if (dtrace_anon.dta_enabling != NULL) {
13884 * dtrace_enabling_retain() can only fail because we are
13885 * trying to retain more enablings than are allowed -- but
13886 * we only have one anonymous enabling, and we are guaranteed
13887 * to be allowed at least one retained enabling; we assert
13888 * that dtrace_enabling_retain() returns success.
13890 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
13893 dtrace_enabling_dump(dtrace_anon.dta_enabling);
13898 * DTrace Helper Functions
13901 dtrace_helper_trace(dtrace_helper_action_t *helper,
13902 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
13904 uint32_t size, next, nnext, i;
13905 dtrace_helptrace_t *ent;
13906 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
13908 if (!dtrace_helptrace_enabled)
13911 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
13914 * What would a tracing framework be without its own tracing
13915 * framework? (Well, a hell of a lot simpler, for starters...)
13917 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
13918 sizeof (uint64_t) - sizeof (uint64_t);
13921 * Iterate until we can allocate a slot in the trace buffer.
13924 next = dtrace_helptrace_next;
13926 if (next + size < dtrace_helptrace_bufsize) {
13927 nnext = next + size;
13931 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
13934 * We have our slot; fill it in.
13939 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
13940 ent->dtht_helper = helper;
13941 ent->dtht_where = where;
13942 ent->dtht_nlocals = vstate->dtvs_nlocals;
13944 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13945 mstate->dtms_fltoffs : -1;
13946 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13947 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
13949 for (i = 0; i < vstate->dtvs_nlocals; i++) {
13950 dtrace_statvar_t *svar;
13952 if ((svar = vstate->dtvs_locals[i]) == NULL)
13955 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13956 ent->dtht_locals[i] =
13957 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
13962 dtrace_helper(int which, dtrace_mstate_t *mstate,
13963 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13965 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
13966 uint64_t sarg0 = mstate->dtms_arg[0];
13967 uint64_t sarg1 = mstate->dtms_arg[1];
13969 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13970 dtrace_helper_action_t *helper;
13971 dtrace_vstate_t *vstate;
13972 dtrace_difo_t *pred;
13973 int i, trace = dtrace_helptrace_enabled;
13975 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13977 if (helpers == NULL)
13980 if ((helper = helpers->dthps_actions[which]) == NULL)
13983 vstate = &helpers->dthps_vstate;
13984 mstate->dtms_arg[0] = arg0;
13985 mstate->dtms_arg[1] = arg1;
13988 * Now iterate over each helper. If its predicate evaluates to 'true',
13989 * we'll call the corresponding actions. Note that the below calls
13990 * to dtrace_dif_emulate() may set faults in machine state. This is
13991 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
13992 * the stored DIF offset with its own (which is the desired behavior).
13993 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13994 * from machine state; this is okay, too.
13996 for (; helper != NULL; helper = helper->dtha_next) {
13997 if ((pred = helper->dtha_predicate) != NULL) {
13999 dtrace_helper_trace(helper, mstate, vstate, 0);
14001 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14004 if (*flags & CPU_DTRACE_FAULT)
14008 for (i = 0; i < helper->dtha_nactions; i++) {
14010 dtrace_helper_trace(helper,
14011 mstate, vstate, i + 1);
14013 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14014 mstate, vstate, state);
14016 if (*flags & CPU_DTRACE_FAULT)
14022 dtrace_helper_trace(helper, mstate, vstate,
14023 DTRACE_HELPTRACE_NEXT);
14027 dtrace_helper_trace(helper, mstate, vstate,
14028 DTRACE_HELPTRACE_DONE);
14031 * Restore the arg0 that we saved upon entry.
14033 mstate->dtms_arg[0] = sarg0;
14034 mstate->dtms_arg[1] = sarg1;
14040 dtrace_helper_trace(helper, mstate, vstate,
14041 DTRACE_HELPTRACE_ERR);
14044 * Restore the arg0 that we saved upon entry.
14046 mstate->dtms_arg[0] = sarg0;
14047 mstate->dtms_arg[1] = sarg1;
14053 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14054 dtrace_vstate_t *vstate)
14058 if (helper->dtha_predicate != NULL)
14059 dtrace_difo_release(helper->dtha_predicate, vstate);
14061 for (i = 0; i < helper->dtha_nactions; i++) {
14062 ASSERT(helper->dtha_actions[i] != NULL);
14063 dtrace_difo_release(helper->dtha_actions[i], vstate);
14066 kmem_free(helper->dtha_actions,
14067 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14068 kmem_free(helper, sizeof (dtrace_helper_action_t));
14072 dtrace_helper_destroygen(int gen)
14074 proc_t *p = curproc;
14075 dtrace_helpers_t *help = p->p_dtrace_helpers;
14076 dtrace_vstate_t *vstate;
14079 ASSERT(MUTEX_HELD(&dtrace_lock));
14081 if (help == NULL || gen > help->dthps_generation)
14084 vstate = &help->dthps_vstate;
14086 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14087 dtrace_helper_action_t *last = NULL, *h, *next;
14089 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14090 next = h->dtha_next;
14092 if (h->dtha_generation == gen) {
14093 if (last != NULL) {
14094 last->dtha_next = next;
14096 help->dthps_actions[i] = next;
14099 dtrace_helper_action_destroy(h, vstate);
14107 * Interate until we've cleared out all helper providers with the
14108 * given generation number.
14111 dtrace_helper_provider_t *prov;
14114 * Look for a helper provider with the right generation. We
14115 * have to start back at the beginning of the list each time
14116 * because we drop dtrace_lock. It's unlikely that we'll make
14117 * more than two passes.
14119 for (i = 0; i < help->dthps_nprovs; i++) {
14120 prov = help->dthps_provs[i];
14122 if (prov->dthp_generation == gen)
14127 * If there were no matches, we're done.
14129 if (i == help->dthps_nprovs)
14133 * Move the last helper provider into this slot.
14135 help->dthps_nprovs--;
14136 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14137 help->dthps_provs[help->dthps_nprovs] = NULL;
14139 mutex_exit(&dtrace_lock);
14142 * If we have a meta provider, remove this helper provider.
14144 mutex_enter(&dtrace_meta_lock);
14145 if (dtrace_meta_pid != NULL) {
14146 ASSERT(dtrace_deferred_pid == NULL);
14147 dtrace_helper_provider_remove(&prov->dthp_prov,
14150 mutex_exit(&dtrace_meta_lock);
14152 dtrace_helper_provider_destroy(prov);
14154 mutex_enter(&dtrace_lock);
14161 dtrace_helper_validate(dtrace_helper_action_t *helper)
14166 if ((dp = helper->dtha_predicate) != NULL)
14167 err += dtrace_difo_validate_helper(dp);
14169 for (i = 0; i < helper->dtha_nactions; i++)
14170 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14176 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14178 dtrace_helpers_t *help;
14179 dtrace_helper_action_t *helper, *last;
14180 dtrace_actdesc_t *act;
14181 dtrace_vstate_t *vstate;
14182 dtrace_predicate_t *pred;
14183 int count = 0, nactions = 0, i;
14185 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14188 help = curproc->p_dtrace_helpers;
14189 last = help->dthps_actions[which];
14190 vstate = &help->dthps_vstate;
14192 for (count = 0; last != NULL; last = last->dtha_next) {
14194 if (last->dtha_next == NULL)
14199 * If we already have dtrace_helper_actions_max helper actions for this
14200 * helper action type, we'll refuse to add a new one.
14202 if (count >= dtrace_helper_actions_max)
14205 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14206 helper->dtha_generation = help->dthps_generation;
14208 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14209 ASSERT(pred->dtp_difo != NULL);
14210 dtrace_difo_hold(pred->dtp_difo);
14211 helper->dtha_predicate = pred->dtp_difo;
14214 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14215 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14218 if (act->dtad_difo == NULL)
14224 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14225 (helper->dtha_nactions = nactions), KM_SLEEP);
14227 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14228 dtrace_difo_hold(act->dtad_difo);
14229 helper->dtha_actions[i++] = act->dtad_difo;
14232 if (!dtrace_helper_validate(helper))
14235 if (last == NULL) {
14236 help->dthps_actions[which] = helper;
14238 last->dtha_next = helper;
14241 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14242 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14243 dtrace_helptrace_next = 0;
14248 dtrace_helper_action_destroy(helper, vstate);
14253 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14254 dof_helper_t *dofhp)
14256 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14258 mutex_enter(&dtrace_meta_lock);
14259 mutex_enter(&dtrace_lock);
14261 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14263 * If the dtrace module is loaded but not attached, or if
14264 * there aren't isn't a meta provider registered to deal with
14265 * these provider descriptions, we need to postpone creating
14266 * the actual providers until later.
14269 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14270 dtrace_deferred_pid != help) {
14271 help->dthps_deferred = 1;
14272 help->dthps_pid = p->p_pid;
14273 help->dthps_next = dtrace_deferred_pid;
14274 help->dthps_prev = NULL;
14275 if (dtrace_deferred_pid != NULL)
14276 dtrace_deferred_pid->dthps_prev = help;
14277 dtrace_deferred_pid = help;
14280 mutex_exit(&dtrace_lock);
14282 } else if (dofhp != NULL) {
14284 * If the dtrace module is loaded and we have a particular
14285 * helper provider description, pass that off to the
14289 mutex_exit(&dtrace_lock);
14291 dtrace_helper_provide(dofhp, p->p_pid);
14295 * Otherwise, just pass all the helper provider descriptions
14296 * off to the meta provider.
14300 mutex_exit(&dtrace_lock);
14302 for (i = 0; i < help->dthps_nprovs; i++) {
14303 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14308 mutex_exit(&dtrace_meta_lock);
14312 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14314 dtrace_helpers_t *help;
14315 dtrace_helper_provider_t *hprov, **tmp_provs;
14316 uint_t tmp_maxprovs, i;
14318 ASSERT(MUTEX_HELD(&dtrace_lock));
14320 help = curproc->p_dtrace_helpers;
14321 ASSERT(help != NULL);
14324 * If we already have dtrace_helper_providers_max helper providers,
14325 * we're refuse to add a new one.
14327 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14331 * Check to make sure this isn't a duplicate.
14333 for (i = 0; i < help->dthps_nprovs; i++) {
14334 if (dofhp->dofhp_addr ==
14335 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14339 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14340 hprov->dthp_prov = *dofhp;
14341 hprov->dthp_ref = 1;
14342 hprov->dthp_generation = gen;
14345 * Allocate a bigger table for helper providers if it's already full.
14347 if (help->dthps_maxprovs == help->dthps_nprovs) {
14348 tmp_maxprovs = help->dthps_maxprovs;
14349 tmp_provs = help->dthps_provs;
14351 if (help->dthps_maxprovs == 0)
14352 help->dthps_maxprovs = 2;
14354 help->dthps_maxprovs *= 2;
14355 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14356 help->dthps_maxprovs = dtrace_helper_providers_max;
14358 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14360 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14361 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14363 if (tmp_provs != NULL) {
14364 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14365 sizeof (dtrace_helper_provider_t *));
14366 kmem_free(tmp_provs, tmp_maxprovs *
14367 sizeof (dtrace_helper_provider_t *));
14371 help->dthps_provs[help->dthps_nprovs] = hprov;
14372 help->dthps_nprovs++;
14378 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14380 mutex_enter(&dtrace_lock);
14382 if (--hprov->dthp_ref == 0) {
14384 mutex_exit(&dtrace_lock);
14385 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14386 dtrace_dof_destroy(dof);
14387 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14389 mutex_exit(&dtrace_lock);
14394 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14396 uintptr_t daddr = (uintptr_t)dof;
14397 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14398 dof_provider_t *provider;
14399 dof_probe_t *probe;
14401 char *strtab, *typestr;
14402 dof_stridx_t typeidx;
14404 uint_t nprobes, j, k;
14406 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14408 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14409 dtrace_dof_error(dof, "misaligned section offset");
14414 * The section needs to be large enough to contain the DOF provider
14415 * structure appropriate for the given version.
14417 if (sec->dofs_size <
14418 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14419 offsetof(dof_provider_t, dofpv_prenoffs) :
14420 sizeof (dof_provider_t))) {
14421 dtrace_dof_error(dof, "provider section too small");
14425 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14426 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14427 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14428 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14429 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14431 if (str_sec == NULL || prb_sec == NULL ||
14432 arg_sec == NULL || off_sec == NULL)
14437 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14438 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14439 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14440 provider->dofpv_prenoffs)) == NULL)
14443 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14445 if (provider->dofpv_name >= str_sec->dofs_size ||
14446 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14447 dtrace_dof_error(dof, "invalid provider name");
14451 if (prb_sec->dofs_entsize == 0 ||
14452 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14453 dtrace_dof_error(dof, "invalid entry size");
14457 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14458 dtrace_dof_error(dof, "misaligned entry size");
14462 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14463 dtrace_dof_error(dof, "invalid entry size");
14467 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14468 dtrace_dof_error(dof, "misaligned section offset");
14472 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14473 dtrace_dof_error(dof, "invalid entry size");
14477 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14479 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14482 * Take a pass through the probes to check for errors.
14484 for (j = 0; j < nprobes; j++) {
14485 probe = (dof_probe_t *)(uintptr_t)(daddr +
14486 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14488 if (probe->dofpr_func >= str_sec->dofs_size) {
14489 dtrace_dof_error(dof, "invalid function name");
14493 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14494 dtrace_dof_error(dof, "function name too long");
14498 if (probe->dofpr_name >= str_sec->dofs_size ||
14499 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14500 dtrace_dof_error(dof, "invalid probe name");
14505 * The offset count must not wrap the index, and the offsets
14506 * must also not overflow the section's data.
14508 if (probe->dofpr_offidx + probe->dofpr_noffs <
14509 probe->dofpr_offidx ||
14510 (probe->dofpr_offidx + probe->dofpr_noffs) *
14511 off_sec->dofs_entsize > off_sec->dofs_size) {
14512 dtrace_dof_error(dof, "invalid probe offset");
14516 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14518 * If there's no is-enabled offset section, make sure
14519 * there aren't any is-enabled offsets. Otherwise
14520 * perform the same checks as for probe offsets
14521 * (immediately above).
14523 if (enoff_sec == NULL) {
14524 if (probe->dofpr_enoffidx != 0 ||
14525 probe->dofpr_nenoffs != 0) {
14526 dtrace_dof_error(dof, "is-enabled "
14527 "offsets with null section");
14530 } else if (probe->dofpr_enoffidx +
14531 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14532 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14533 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14534 dtrace_dof_error(dof, "invalid is-enabled "
14539 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14540 dtrace_dof_error(dof, "zero probe and "
14541 "is-enabled offsets");
14544 } else if (probe->dofpr_noffs == 0) {
14545 dtrace_dof_error(dof, "zero probe offsets");
14549 if (probe->dofpr_argidx + probe->dofpr_xargc <
14550 probe->dofpr_argidx ||
14551 (probe->dofpr_argidx + probe->dofpr_xargc) *
14552 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14553 dtrace_dof_error(dof, "invalid args");
14557 typeidx = probe->dofpr_nargv;
14558 typestr = strtab + probe->dofpr_nargv;
14559 for (k = 0; k < probe->dofpr_nargc; k++) {
14560 if (typeidx >= str_sec->dofs_size) {
14561 dtrace_dof_error(dof, "bad "
14562 "native argument type");
14566 typesz = strlen(typestr) + 1;
14567 if (typesz > DTRACE_ARGTYPELEN) {
14568 dtrace_dof_error(dof, "native "
14569 "argument type too long");
14576 typeidx = probe->dofpr_xargv;
14577 typestr = strtab + probe->dofpr_xargv;
14578 for (k = 0; k < probe->dofpr_xargc; k++) {
14579 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14580 dtrace_dof_error(dof, "bad "
14581 "native argument index");
14585 if (typeidx >= str_sec->dofs_size) {
14586 dtrace_dof_error(dof, "bad "
14587 "translated argument type");
14591 typesz = strlen(typestr) + 1;
14592 if (typesz > DTRACE_ARGTYPELEN) {
14593 dtrace_dof_error(dof, "translated argument "
14607 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14609 dtrace_helpers_t *help;
14610 dtrace_vstate_t *vstate;
14611 dtrace_enabling_t *enab = NULL;
14612 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14613 uintptr_t daddr = (uintptr_t)dof;
14615 ASSERT(MUTEX_HELD(&dtrace_lock));
14617 if ((help = curproc->p_dtrace_helpers) == NULL)
14618 help = dtrace_helpers_create(curproc);
14620 vstate = &help->dthps_vstate;
14622 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14623 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14624 dtrace_dof_destroy(dof);
14629 * Look for helper providers and validate their descriptions.
14632 for (i = 0; i < dof->dofh_secnum; i++) {
14633 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14634 dof->dofh_secoff + i * dof->dofh_secsize);
14636 if (sec->dofs_type != DOF_SECT_PROVIDER)
14639 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14640 dtrace_enabling_destroy(enab);
14641 dtrace_dof_destroy(dof);
14650 * Now we need to walk through the ECB descriptions in the enabling.
14652 for (i = 0; i < enab->dten_ndesc; i++) {
14653 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14654 dtrace_probedesc_t *desc = &ep->dted_probe;
14656 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14659 if (strcmp(desc->dtpd_mod, "helper") != 0)
14662 if (strcmp(desc->dtpd_func, "ustack") != 0)
14665 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14668 * Adding this helper action failed -- we are now going
14669 * to rip out the entire generation and return failure.
14671 (void) dtrace_helper_destroygen(help->dthps_generation);
14672 dtrace_enabling_destroy(enab);
14673 dtrace_dof_destroy(dof);
14680 if (nhelpers < enab->dten_ndesc)
14681 dtrace_dof_error(dof, "unmatched helpers");
14683 gen = help->dthps_generation++;
14684 dtrace_enabling_destroy(enab);
14686 if (dhp != NULL && nprovs > 0) {
14687 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14688 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14689 mutex_exit(&dtrace_lock);
14690 dtrace_helper_provider_register(curproc, help, dhp);
14691 mutex_enter(&dtrace_lock);
14698 dtrace_dof_destroy(dof);
14703 static dtrace_helpers_t *
14704 dtrace_helpers_create(proc_t *p)
14706 dtrace_helpers_t *help;
14708 ASSERT(MUTEX_HELD(&dtrace_lock));
14709 ASSERT(p->p_dtrace_helpers == NULL);
14711 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14712 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14713 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14715 p->p_dtrace_helpers = help;
14725 dtrace_helpers_destroy(proc_t *p)
14727 dtrace_helpers_t *help;
14728 dtrace_vstate_t *vstate;
14730 proc_t *p = curproc;
14734 mutex_enter(&dtrace_lock);
14736 ASSERT(p->p_dtrace_helpers != NULL);
14737 ASSERT(dtrace_helpers > 0);
14739 help = p->p_dtrace_helpers;
14740 vstate = &help->dthps_vstate;
14743 * We're now going to lose the help from this process.
14745 p->p_dtrace_helpers = NULL;
14749 * Destory the helper actions.
14751 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14752 dtrace_helper_action_t *h, *next;
14754 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14755 next = h->dtha_next;
14756 dtrace_helper_action_destroy(h, vstate);
14761 mutex_exit(&dtrace_lock);
14764 * Destroy the helper providers.
14766 if (help->dthps_maxprovs > 0) {
14767 mutex_enter(&dtrace_meta_lock);
14768 if (dtrace_meta_pid != NULL) {
14769 ASSERT(dtrace_deferred_pid == NULL);
14771 for (i = 0; i < help->dthps_nprovs; i++) {
14772 dtrace_helper_provider_remove(
14773 &help->dthps_provs[i]->dthp_prov, p->p_pid);
14776 mutex_enter(&dtrace_lock);
14777 ASSERT(help->dthps_deferred == 0 ||
14778 help->dthps_next != NULL ||
14779 help->dthps_prev != NULL ||
14780 help == dtrace_deferred_pid);
14783 * Remove the helper from the deferred list.
14785 if (help->dthps_next != NULL)
14786 help->dthps_next->dthps_prev = help->dthps_prev;
14787 if (help->dthps_prev != NULL)
14788 help->dthps_prev->dthps_next = help->dthps_next;
14789 if (dtrace_deferred_pid == help) {
14790 dtrace_deferred_pid = help->dthps_next;
14791 ASSERT(help->dthps_prev == NULL);
14794 mutex_exit(&dtrace_lock);
14797 mutex_exit(&dtrace_meta_lock);
14799 for (i = 0; i < help->dthps_nprovs; i++) {
14800 dtrace_helper_provider_destroy(help->dthps_provs[i]);
14803 kmem_free(help->dthps_provs, help->dthps_maxprovs *
14804 sizeof (dtrace_helper_provider_t *));
14807 mutex_enter(&dtrace_lock);
14809 dtrace_vstate_fini(&help->dthps_vstate);
14810 kmem_free(help->dthps_actions,
14811 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
14812 kmem_free(help, sizeof (dtrace_helpers_t));
14815 mutex_exit(&dtrace_lock);
14822 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
14824 dtrace_helpers_t *help, *newhelp;
14825 dtrace_helper_action_t *helper, *new, *last;
14827 dtrace_vstate_t *vstate;
14828 int i, j, sz, hasprovs = 0;
14830 mutex_enter(&dtrace_lock);
14831 ASSERT(from->p_dtrace_helpers != NULL);
14832 ASSERT(dtrace_helpers > 0);
14834 help = from->p_dtrace_helpers;
14835 newhelp = dtrace_helpers_create(to);
14836 ASSERT(to->p_dtrace_helpers != NULL);
14838 newhelp->dthps_generation = help->dthps_generation;
14839 vstate = &newhelp->dthps_vstate;
14842 * Duplicate the helper actions.
14844 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14845 if ((helper = help->dthps_actions[i]) == NULL)
14848 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
14849 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
14851 new->dtha_generation = helper->dtha_generation;
14853 if ((dp = helper->dtha_predicate) != NULL) {
14854 dp = dtrace_difo_duplicate(dp, vstate);
14855 new->dtha_predicate = dp;
14858 new->dtha_nactions = helper->dtha_nactions;
14859 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
14860 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
14862 for (j = 0; j < new->dtha_nactions; j++) {
14863 dtrace_difo_t *dp = helper->dtha_actions[j];
14865 ASSERT(dp != NULL);
14866 dp = dtrace_difo_duplicate(dp, vstate);
14867 new->dtha_actions[j] = dp;
14870 if (last != NULL) {
14871 last->dtha_next = new;
14873 newhelp->dthps_actions[i] = new;
14881 * Duplicate the helper providers and register them with the
14882 * DTrace framework.
14884 if (help->dthps_nprovs > 0) {
14885 newhelp->dthps_nprovs = help->dthps_nprovs;
14886 newhelp->dthps_maxprovs = help->dthps_nprovs;
14887 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
14888 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14889 for (i = 0; i < newhelp->dthps_nprovs; i++) {
14890 newhelp->dthps_provs[i] = help->dthps_provs[i];
14891 newhelp->dthps_provs[i]->dthp_ref++;
14897 mutex_exit(&dtrace_lock);
14900 dtrace_helper_provider_register(to, newhelp, NULL);
14905 * DTrace Hook Functions
14908 dtrace_module_loaded(modctl_t *ctl)
14910 dtrace_provider_t *prv;
14912 mutex_enter(&dtrace_provider_lock);
14913 mutex_enter(&mod_lock);
14915 ASSERT(ctl->mod_busy);
14918 * We're going to call each providers per-module provide operation
14919 * specifying only this module.
14921 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
14922 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
14924 mutex_exit(&mod_lock);
14925 mutex_exit(&dtrace_provider_lock);
14928 * If we have any retained enablings, we need to match against them.
14929 * Enabling probes requires that cpu_lock be held, and we cannot hold
14930 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
14931 * module. (In particular, this happens when loading scheduling
14932 * classes.) So if we have any retained enablings, we need to dispatch
14933 * our task queue to do the match for us.
14935 mutex_enter(&dtrace_lock);
14937 if (dtrace_retained == NULL) {
14938 mutex_exit(&dtrace_lock);
14942 (void) taskq_dispatch(dtrace_taskq,
14943 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
14945 mutex_exit(&dtrace_lock);
14948 * And now, for a little heuristic sleaze: in general, we want to
14949 * match modules as soon as they load. However, we cannot guarantee
14950 * this, because it would lead us to the lock ordering violation
14951 * outlined above. The common case, of course, is that cpu_lock is
14952 * _not_ held -- so we delay here for a clock tick, hoping that that's
14953 * long enough for the task queue to do its work. If it's not, it's
14954 * not a serious problem -- it just means that the module that we
14955 * just loaded may not be immediately instrumentable.
14961 dtrace_module_unloaded(modctl_t *ctl)
14963 dtrace_probe_t template, *probe, *first, *next;
14964 dtrace_provider_t *prov;
14966 template.dtpr_mod = ctl->mod_modname;
14968 mutex_enter(&dtrace_provider_lock);
14969 mutex_enter(&mod_lock);
14970 mutex_enter(&dtrace_lock);
14972 if (dtrace_bymod == NULL) {
14974 * The DTrace module is loaded (obviously) but not attached;
14975 * we don't have any work to do.
14977 mutex_exit(&dtrace_provider_lock);
14978 mutex_exit(&mod_lock);
14979 mutex_exit(&dtrace_lock);
14983 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14984 probe != NULL; probe = probe->dtpr_nextmod) {
14985 if (probe->dtpr_ecb != NULL) {
14986 mutex_exit(&dtrace_provider_lock);
14987 mutex_exit(&mod_lock);
14988 mutex_exit(&dtrace_lock);
14991 * This shouldn't _actually_ be possible -- we're
14992 * unloading a module that has an enabled probe in it.
14993 * (It's normally up to the provider to make sure that
14994 * this can't happen.) However, because dtps_enable()
14995 * doesn't have a failure mode, there can be an
14996 * enable/unload race. Upshot: we don't want to
14997 * assert, but we're not going to disable the
15000 if (dtrace_err_verbose) {
15001 cmn_err(CE_WARN, "unloaded module '%s' had "
15002 "enabled probes", ctl->mod_modname);
15011 for (first = NULL; probe != NULL; probe = next) {
15012 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15014 dtrace_probes[probe->dtpr_id - 1] = NULL;
15016 next = probe->dtpr_nextmod;
15017 dtrace_hash_remove(dtrace_bymod, probe);
15018 dtrace_hash_remove(dtrace_byfunc, probe);
15019 dtrace_hash_remove(dtrace_byname, probe);
15021 if (first == NULL) {
15023 probe->dtpr_nextmod = NULL;
15025 probe->dtpr_nextmod = first;
15031 * We've removed all of the module's probes from the hash chains and
15032 * from the probe array. Now issue a dtrace_sync() to be sure that
15033 * everyone has cleared out from any probe array processing.
15037 for (probe = first; probe != NULL; probe = first) {
15038 first = probe->dtpr_nextmod;
15039 prov = probe->dtpr_provider;
15040 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15042 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15043 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15044 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15045 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15046 kmem_free(probe, sizeof (dtrace_probe_t));
15049 mutex_exit(&dtrace_lock);
15050 mutex_exit(&mod_lock);
15051 mutex_exit(&dtrace_provider_lock);
15055 dtrace_suspend(void)
15057 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15061 dtrace_resume(void)
15063 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15068 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15070 ASSERT(MUTEX_HELD(&cpu_lock));
15071 mutex_enter(&dtrace_lock);
15075 dtrace_state_t *state;
15076 dtrace_optval_t *opt, rs, c;
15079 * For now, we only allocate a new buffer for anonymous state.
15081 if ((state = dtrace_anon.dta_state) == NULL)
15084 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15087 opt = state->dts_options;
15088 c = opt[DTRACEOPT_CPU];
15090 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15094 * Regardless of what the actual policy is, we're going to
15095 * temporarily set our resize policy to be manual. We're
15096 * also going to temporarily set our CPU option to denote
15097 * the newly configured CPU.
15099 rs = opt[DTRACEOPT_BUFRESIZE];
15100 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15101 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15103 (void) dtrace_state_buffers(state);
15105 opt[DTRACEOPT_BUFRESIZE] = rs;
15106 opt[DTRACEOPT_CPU] = c;
15113 * We don't free the buffer in the CPU_UNCONFIG case. (The
15114 * buffer will be freed when the consumer exits.)
15122 mutex_exit(&dtrace_lock);
15128 dtrace_cpu_setup_initial(processorid_t cpu)
15130 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15135 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15137 if (dtrace_toxranges >= dtrace_toxranges_max) {
15139 dtrace_toxrange_t *range;
15141 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15144 ASSERT(dtrace_toxrange == NULL);
15145 ASSERT(dtrace_toxranges_max == 0);
15146 dtrace_toxranges_max = 1;
15148 dtrace_toxranges_max <<= 1;
15151 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15152 range = kmem_zalloc(nsize, KM_SLEEP);
15154 if (dtrace_toxrange != NULL) {
15155 ASSERT(osize != 0);
15156 bcopy(dtrace_toxrange, range, osize);
15157 kmem_free(dtrace_toxrange, osize);
15160 dtrace_toxrange = range;
15163 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15164 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15166 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15167 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15168 dtrace_toxranges++;
15172 * DTrace Driver Cookbook Functions
15177 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15179 dtrace_provider_id_t id;
15180 dtrace_state_t *state = NULL;
15181 dtrace_enabling_t *enab;
15183 mutex_enter(&cpu_lock);
15184 mutex_enter(&dtrace_provider_lock);
15185 mutex_enter(&dtrace_lock);
15187 if (ddi_soft_state_init(&dtrace_softstate,
15188 sizeof (dtrace_state_t), 0) != 0) {
15189 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15190 mutex_exit(&cpu_lock);
15191 mutex_exit(&dtrace_provider_lock);
15192 mutex_exit(&dtrace_lock);
15193 return (DDI_FAILURE);
15196 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15197 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15198 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15199 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15200 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15201 ddi_remove_minor_node(devi, NULL);
15202 ddi_soft_state_fini(&dtrace_softstate);
15203 mutex_exit(&cpu_lock);
15204 mutex_exit(&dtrace_provider_lock);
15205 mutex_exit(&dtrace_lock);
15206 return (DDI_FAILURE);
15209 ddi_report_dev(devi);
15210 dtrace_devi = devi;
15212 dtrace_modload = dtrace_module_loaded;
15213 dtrace_modunload = dtrace_module_unloaded;
15214 dtrace_cpu_init = dtrace_cpu_setup_initial;
15215 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15216 dtrace_helpers_fork = dtrace_helpers_duplicate;
15217 dtrace_cpustart_init = dtrace_suspend;
15218 dtrace_cpustart_fini = dtrace_resume;
15219 dtrace_debugger_init = dtrace_suspend;
15220 dtrace_debugger_fini = dtrace_resume;
15222 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15224 ASSERT(MUTEX_HELD(&cpu_lock));
15226 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15227 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15228 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15229 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15230 VM_SLEEP | VMC_IDENTIFIER);
15231 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15234 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15235 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15236 NULL, NULL, NULL, NULL, NULL, 0);
15238 ASSERT(MUTEX_HELD(&cpu_lock));
15239 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15240 offsetof(dtrace_probe_t, dtpr_nextmod),
15241 offsetof(dtrace_probe_t, dtpr_prevmod));
15243 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15244 offsetof(dtrace_probe_t, dtpr_nextfunc),
15245 offsetof(dtrace_probe_t, dtpr_prevfunc));
15247 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15248 offsetof(dtrace_probe_t, dtpr_nextname),
15249 offsetof(dtrace_probe_t, dtpr_prevname));
15251 if (dtrace_retain_max < 1) {
15252 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15253 "setting to 1", dtrace_retain_max);
15254 dtrace_retain_max = 1;
15258 * Now discover our toxic ranges.
15260 dtrace_toxic_ranges(dtrace_toxrange_add);
15263 * Before we register ourselves as a provider to our own framework,
15264 * we would like to assert that dtrace_provider is NULL -- but that's
15265 * not true if we were loaded as a dependency of a DTrace provider.
15266 * Once we've registered, we can assert that dtrace_provider is our
15269 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15270 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15272 ASSERT(dtrace_provider != NULL);
15273 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15275 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15276 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15277 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15278 dtrace_provider, NULL, NULL, "END", 0, NULL);
15279 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15280 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15282 dtrace_anon_property();
15283 mutex_exit(&cpu_lock);
15286 * If DTrace helper tracing is enabled, we need to allocate the
15287 * trace buffer and initialize the values.
15289 if (dtrace_helptrace_enabled) {
15290 ASSERT(dtrace_helptrace_buffer == NULL);
15291 dtrace_helptrace_buffer =
15292 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15293 dtrace_helptrace_next = 0;
15297 * If there are already providers, we must ask them to provide their
15298 * probes, and then match any anonymous enabling against them. Note
15299 * that there should be no other retained enablings at this time:
15300 * the only retained enablings at this time should be the anonymous
15303 if (dtrace_anon.dta_enabling != NULL) {
15304 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15306 dtrace_enabling_provide(NULL);
15307 state = dtrace_anon.dta_state;
15310 * We couldn't hold cpu_lock across the above call to
15311 * dtrace_enabling_provide(), but we must hold it to actually
15312 * enable the probes. We have to drop all of our locks, pick
15313 * up cpu_lock, and regain our locks before matching the
15314 * retained anonymous enabling.
15316 mutex_exit(&dtrace_lock);
15317 mutex_exit(&dtrace_provider_lock);
15319 mutex_enter(&cpu_lock);
15320 mutex_enter(&dtrace_provider_lock);
15321 mutex_enter(&dtrace_lock);
15323 if ((enab = dtrace_anon.dta_enabling) != NULL)
15324 (void) dtrace_enabling_match(enab, NULL);
15326 mutex_exit(&cpu_lock);
15329 mutex_exit(&dtrace_lock);
15330 mutex_exit(&dtrace_provider_lock);
15332 if (state != NULL) {
15334 * If we created any anonymous state, set it going now.
15336 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15339 return (DDI_SUCCESS);
15344 #if __FreeBSD_version >= 800039
15346 dtrace_dtr(void *data __unused)
15355 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15357 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15360 dtrace_state_t *state;
15366 if (getminor(*devp) == DTRACEMNRN_HELPER)
15370 * If this wasn't an open with the "helper" minor, then it must be
15371 * the "dtrace" minor.
15373 ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15375 cred_t *cred_p = NULL;
15377 #if __FreeBSD_version < 800039
15379 * The first minor device is the one that is cloned so there is
15380 * nothing more to do here.
15382 if (dev2unit(dev) == 0)
15386 * Devices are cloned, so if the DTrace state has already
15387 * been allocated, that means this device belongs to a
15388 * different client. Each client should open '/dev/dtrace'
15389 * to get a cloned device.
15391 if (dev->si_drv1 != NULL)
15395 cred_p = dev->si_cred;
15399 * If no DTRACE_PRIV_* bits are set in the credential, then the
15400 * caller lacks sufficient permission to do anything with DTrace.
15402 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15403 if (priv == DTRACE_PRIV_NONE) {
15405 #if __FreeBSD_version < 800039
15406 /* Destroy the cloned device. */
15415 * Ask all providers to provide all their probes.
15417 mutex_enter(&dtrace_provider_lock);
15418 dtrace_probe_provide(NULL, NULL);
15419 mutex_exit(&dtrace_provider_lock);
15421 mutex_enter(&cpu_lock);
15422 mutex_enter(&dtrace_lock);
15424 dtrace_membar_producer();
15428 * If the kernel debugger is active (that is, if the kernel debugger
15429 * modified text in some way), we won't allow the open.
15431 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15433 mutex_exit(&cpu_lock);
15434 mutex_exit(&dtrace_lock);
15438 state = dtrace_state_create(devp, cred_p);
15440 state = dtrace_state_create(dev);
15441 #if __FreeBSD_version < 800039
15442 dev->si_drv1 = state;
15444 devfs_set_cdevpriv(state, dtrace_dtr);
15448 mutex_exit(&cpu_lock);
15450 if (state == NULL) {
15452 if (--dtrace_opens == 0)
15453 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15457 mutex_exit(&dtrace_lock);
15459 #if __FreeBSD_version < 800039
15460 /* Destroy the cloned device. */
15467 mutex_exit(&dtrace_lock);
15475 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15477 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15481 minor_t minor = getminor(dev);
15482 dtrace_state_t *state;
15484 if (minor == DTRACEMNRN_HELPER)
15487 state = ddi_get_soft_state(dtrace_softstate, minor);
15489 #if __FreeBSD_version < 800039
15490 dtrace_state_t *state = dev->si_drv1;
15492 /* Check if this is not a cloned device. */
15493 if (dev2unit(dev) == 0)
15496 dtrace_state_t *state;
15497 devfs_get_cdevpriv((void **) &state);
15502 mutex_enter(&cpu_lock);
15503 mutex_enter(&dtrace_lock);
15505 if (state != NULL) {
15506 if (state->dts_anon) {
15508 * There is anonymous state. Destroy that first.
15510 ASSERT(dtrace_anon.dta_state == NULL);
15511 dtrace_state_destroy(state->dts_anon);
15514 dtrace_state_destroy(state);
15517 kmem_free(state, 0);
15518 #if __FreeBSD_version < 800039
15519 dev->si_drv1 = NULL;
15521 devfs_clear_cdevpriv();
15526 ASSERT(dtrace_opens > 0);
15528 if (--dtrace_opens == 0)
15529 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15534 mutex_exit(&dtrace_lock);
15535 mutex_exit(&cpu_lock);
15537 #if __FreeBSD_version < 800039
15538 /* Schedule this cloned device to be destroyed. */
15539 destroy_dev_sched(dev);
15548 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15551 dof_helper_t help, *dhp = NULL;
15554 case DTRACEHIOC_ADDDOF:
15555 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15556 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15561 arg = (intptr_t)help.dofhp_dof;
15564 case DTRACEHIOC_ADD: {
15565 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15570 mutex_enter(&dtrace_lock);
15573 * dtrace_helper_slurp() takes responsibility for the dof --
15574 * it may free it now or it may save it and free it later.
15576 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15583 mutex_exit(&dtrace_lock);
15587 case DTRACEHIOC_REMOVE: {
15588 mutex_enter(&dtrace_lock);
15589 rval = dtrace_helper_destroygen(arg);
15590 mutex_exit(&dtrace_lock);
15604 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15606 minor_t minor = getminor(dev);
15607 dtrace_state_t *state;
15610 if (minor == DTRACEMNRN_HELPER)
15611 return (dtrace_ioctl_helper(cmd, arg, rv));
15613 state = ddi_get_soft_state(dtrace_softstate, minor);
15615 if (state->dts_anon) {
15616 ASSERT(dtrace_anon.dta_state == NULL);
15617 state = state->dts_anon;
15621 case DTRACEIOC_PROVIDER: {
15622 dtrace_providerdesc_t pvd;
15623 dtrace_provider_t *pvp;
15625 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15628 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15629 mutex_enter(&dtrace_provider_lock);
15631 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15632 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15636 mutex_exit(&dtrace_provider_lock);
15641 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15642 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15644 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15650 case DTRACEIOC_EPROBE: {
15651 dtrace_eprobedesc_t epdesc;
15653 dtrace_action_t *act;
15659 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15662 mutex_enter(&dtrace_lock);
15664 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15665 mutex_exit(&dtrace_lock);
15669 if (ecb->dte_probe == NULL) {
15670 mutex_exit(&dtrace_lock);
15674 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15675 epdesc.dtepd_uarg = ecb->dte_uarg;
15676 epdesc.dtepd_size = ecb->dte_size;
15678 nrecs = epdesc.dtepd_nrecs;
15679 epdesc.dtepd_nrecs = 0;
15680 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15681 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15684 epdesc.dtepd_nrecs++;
15688 * Now that we have the size, we need to allocate a temporary
15689 * buffer in which to store the complete description. We need
15690 * the temporary buffer to be able to drop dtrace_lock()
15691 * across the copyout(), below.
15693 size = sizeof (dtrace_eprobedesc_t) +
15694 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15696 buf = kmem_alloc(size, KM_SLEEP);
15697 dest = (uintptr_t)buf;
15699 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15700 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15702 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15703 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15709 bcopy(&act->dta_rec, (void *)dest,
15710 sizeof (dtrace_recdesc_t));
15711 dest += sizeof (dtrace_recdesc_t);
15714 mutex_exit(&dtrace_lock);
15716 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15717 kmem_free(buf, size);
15721 kmem_free(buf, size);
15725 case DTRACEIOC_AGGDESC: {
15726 dtrace_aggdesc_t aggdesc;
15727 dtrace_action_t *act;
15728 dtrace_aggregation_t *agg;
15731 dtrace_recdesc_t *lrec;
15736 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15739 mutex_enter(&dtrace_lock);
15741 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15742 mutex_exit(&dtrace_lock);
15746 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15748 nrecs = aggdesc.dtagd_nrecs;
15749 aggdesc.dtagd_nrecs = 0;
15751 offs = agg->dtag_base;
15752 lrec = &agg->dtag_action.dta_rec;
15753 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15755 for (act = agg->dtag_first; ; act = act->dta_next) {
15756 ASSERT(act->dta_intuple ||
15757 DTRACEACT_ISAGG(act->dta_kind));
15760 * If this action has a record size of zero, it
15761 * denotes an argument to the aggregating action.
15762 * Because the presence of this record doesn't (or
15763 * shouldn't) affect the way the data is interpreted,
15764 * we don't copy it out to save user-level the
15765 * confusion of dealing with a zero-length record.
15767 if (act->dta_rec.dtrd_size == 0) {
15768 ASSERT(agg->dtag_hasarg);
15772 aggdesc.dtagd_nrecs++;
15774 if (act == &agg->dtag_action)
15779 * Now that we have the size, we need to allocate a temporary
15780 * buffer in which to store the complete description. We need
15781 * the temporary buffer to be able to drop dtrace_lock()
15782 * across the copyout(), below.
15784 size = sizeof (dtrace_aggdesc_t) +
15785 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15787 buf = kmem_alloc(size, KM_SLEEP);
15788 dest = (uintptr_t)buf;
15790 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15791 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15793 for (act = agg->dtag_first; ; act = act->dta_next) {
15794 dtrace_recdesc_t rec = act->dta_rec;
15797 * See the comment in the above loop for why we pass
15798 * over zero-length records.
15800 if (rec.dtrd_size == 0) {
15801 ASSERT(agg->dtag_hasarg);
15808 rec.dtrd_offset -= offs;
15809 bcopy(&rec, (void *)dest, sizeof (rec));
15810 dest += sizeof (dtrace_recdesc_t);
15812 if (act == &agg->dtag_action)
15816 mutex_exit(&dtrace_lock);
15818 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15819 kmem_free(buf, size);
15823 kmem_free(buf, size);
15827 case DTRACEIOC_ENABLE: {
15829 dtrace_enabling_t *enab = NULL;
15830 dtrace_vstate_t *vstate;
15836 * If a NULL argument has been passed, we take this as our
15837 * cue to reevaluate our enablings.
15840 dtrace_enabling_matchall();
15845 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
15848 mutex_enter(&cpu_lock);
15849 mutex_enter(&dtrace_lock);
15850 vstate = &state->dts_vstate;
15852 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
15853 mutex_exit(&dtrace_lock);
15854 mutex_exit(&cpu_lock);
15855 dtrace_dof_destroy(dof);
15859 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
15860 mutex_exit(&dtrace_lock);
15861 mutex_exit(&cpu_lock);
15862 dtrace_dof_destroy(dof);
15866 if ((rval = dtrace_dof_options(dof, state)) != 0) {
15867 dtrace_enabling_destroy(enab);
15868 mutex_exit(&dtrace_lock);
15869 mutex_exit(&cpu_lock);
15870 dtrace_dof_destroy(dof);
15874 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
15875 err = dtrace_enabling_retain(enab);
15877 dtrace_enabling_destroy(enab);
15880 mutex_exit(&cpu_lock);
15881 mutex_exit(&dtrace_lock);
15882 dtrace_dof_destroy(dof);
15887 case DTRACEIOC_REPLICATE: {
15888 dtrace_repldesc_t desc;
15889 dtrace_probedesc_t *match = &desc.dtrpd_match;
15890 dtrace_probedesc_t *create = &desc.dtrpd_create;
15893 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15896 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15897 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15898 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15899 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15901 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15902 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15903 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15904 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15906 mutex_enter(&dtrace_lock);
15907 err = dtrace_enabling_replicate(state, match, create);
15908 mutex_exit(&dtrace_lock);
15913 case DTRACEIOC_PROBEMATCH:
15914 case DTRACEIOC_PROBES: {
15915 dtrace_probe_t *probe = NULL;
15916 dtrace_probedesc_t desc;
15917 dtrace_probekey_t pkey;
15924 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15927 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15928 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15929 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15930 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15933 * Before we attempt to match this probe, we want to give
15934 * all providers the opportunity to provide it.
15936 if (desc.dtpd_id == DTRACE_IDNONE) {
15937 mutex_enter(&dtrace_provider_lock);
15938 dtrace_probe_provide(&desc, NULL);
15939 mutex_exit(&dtrace_provider_lock);
15943 if (cmd == DTRACEIOC_PROBEMATCH) {
15944 dtrace_probekey(&desc, &pkey);
15945 pkey.dtpk_id = DTRACE_IDNONE;
15948 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
15950 mutex_enter(&dtrace_lock);
15952 if (cmd == DTRACEIOC_PROBEMATCH) {
15953 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15954 if ((probe = dtrace_probes[i - 1]) != NULL &&
15955 (m = dtrace_match_probe(probe, &pkey,
15956 priv, uid, zoneid)) != 0)
15961 mutex_exit(&dtrace_lock);
15966 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15967 if ((probe = dtrace_probes[i - 1]) != NULL &&
15968 dtrace_match_priv(probe, priv, uid, zoneid))
15973 if (probe == NULL) {
15974 mutex_exit(&dtrace_lock);
15978 dtrace_probe_description(probe, &desc);
15979 mutex_exit(&dtrace_lock);
15981 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15987 case DTRACEIOC_PROBEARG: {
15988 dtrace_argdesc_t desc;
15989 dtrace_probe_t *probe;
15990 dtrace_provider_t *prov;
15992 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15995 if (desc.dtargd_id == DTRACE_IDNONE)
15998 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16001 mutex_enter(&dtrace_provider_lock);
16002 mutex_enter(&mod_lock);
16003 mutex_enter(&dtrace_lock);
16005 if (desc.dtargd_id > dtrace_nprobes) {
16006 mutex_exit(&dtrace_lock);
16007 mutex_exit(&mod_lock);
16008 mutex_exit(&dtrace_provider_lock);
16012 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16013 mutex_exit(&dtrace_lock);
16014 mutex_exit(&mod_lock);
16015 mutex_exit(&dtrace_provider_lock);
16019 mutex_exit(&dtrace_lock);
16021 prov = probe->dtpr_provider;
16023 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16025 * There isn't any typed information for this probe.
16026 * Set the argument number to DTRACE_ARGNONE.
16028 desc.dtargd_ndx = DTRACE_ARGNONE;
16030 desc.dtargd_native[0] = '\0';
16031 desc.dtargd_xlate[0] = '\0';
16032 desc.dtargd_mapping = desc.dtargd_ndx;
16034 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16035 probe->dtpr_id, probe->dtpr_arg, &desc);
16038 mutex_exit(&mod_lock);
16039 mutex_exit(&dtrace_provider_lock);
16041 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16047 case DTRACEIOC_GO: {
16048 processorid_t cpuid;
16049 rval = dtrace_state_go(state, &cpuid);
16054 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16060 case DTRACEIOC_STOP: {
16061 processorid_t cpuid;
16063 mutex_enter(&dtrace_lock);
16064 rval = dtrace_state_stop(state, &cpuid);
16065 mutex_exit(&dtrace_lock);
16070 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16076 case DTRACEIOC_DOFGET: {
16077 dof_hdr_t hdr, *dof;
16080 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16083 mutex_enter(&dtrace_lock);
16084 dof = dtrace_dof_create(state);
16085 mutex_exit(&dtrace_lock);
16087 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16088 rval = copyout(dof, (void *)arg, len);
16089 dtrace_dof_destroy(dof);
16091 return (rval == 0 ? 0 : EFAULT);
16094 case DTRACEIOC_AGGSNAP:
16095 case DTRACEIOC_BUFSNAP: {
16096 dtrace_bufdesc_t desc;
16098 dtrace_buffer_t *buf;
16100 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16103 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16106 mutex_enter(&dtrace_lock);
16108 if (cmd == DTRACEIOC_BUFSNAP) {
16109 buf = &state->dts_buffer[desc.dtbd_cpu];
16111 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16114 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16115 size_t sz = buf->dtb_offset;
16117 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16118 mutex_exit(&dtrace_lock);
16123 * If this buffer has already been consumed, we're
16124 * going to indicate that there's nothing left here
16127 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16128 mutex_exit(&dtrace_lock);
16130 desc.dtbd_size = 0;
16131 desc.dtbd_drops = 0;
16132 desc.dtbd_errors = 0;
16133 desc.dtbd_oldest = 0;
16134 sz = sizeof (desc);
16136 if (copyout(&desc, (void *)arg, sz) != 0)
16143 * If this is a ring buffer that has wrapped, we want
16144 * to copy the whole thing out.
16146 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16147 dtrace_buffer_polish(buf);
16148 sz = buf->dtb_size;
16151 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16152 mutex_exit(&dtrace_lock);
16156 desc.dtbd_size = sz;
16157 desc.dtbd_drops = buf->dtb_drops;
16158 desc.dtbd_errors = buf->dtb_errors;
16159 desc.dtbd_oldest = buf->dtb_xamot_offset;
16161 mutex_exit(&dtrace_lock);
16163 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16166 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16171 if (buf->dtb_tomax == NULL) {
16172 ASSERT(buf->dtb_xamot == NULL);
16173 mutex_exit(&dtrace_lock);
16177 cached = buf->dtb_tomax;
16178 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16180 dtrace_xcall(desc.dtbd_cpu,
16181 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16183 state->dts_errors += buf->dtb_xamot_errors;
16186 * If the buffers did not actually switch, then the cross call
16187 * did not take place -- presumably because the given CPU is
16188 * not in the ready set. If this is the case, we'll return
16191 if (buf->dtb_tomax == cached) {
16192 ASSERT(buf->dtb_xamot != cached);
16193 mutex_exit(&dtrace_lock);
16197 ASSERT(cached == buf->dtb_xamot);
16200 * We have our snapshot; now copy it out.
16202 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16203 buf->dtb_xamot_offset) != 0) {
16204 mutex_exit(&dtrace_lock);
16208 desc.dtbd_size = buf->dtb_xamot_offset;
16209 desc.dtbd_drops = buf->dtb_xamot_drops;
16210 desc.dtbd_errors = buf->dtb_xamot_errors;
16211 desc.dtbd_oldest = 0;
16213 mutex_exit(&dtrace_lock);
16216 * Finally, copy out the buffer description.
16218 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16224 case DTRACEIOC_CONF: {
16225 dtrace_conf_t conf;
16227 bzero(&conf, sizeof (conf));
16228 conf.dtc_difversion = DIF_VERSION;
16229 conf.dtc_difintregs = DIF_DIR_NREGS;
16230 conf.dtc_diftupregs = DIF_DTR_NREGS;
16231 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16233 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16239 case DTRACEIOC_STATUS: {
16240 dtrace_status_t stat;
16241 dtrace_dstate_t *dstate;
16246 * See the comment in dtrace_state_deadman() for the reason
16247 * for setting dts_laststatus to INT64_MAX before setting
16248 * it to the correct value.
16250 state->dts_laststatus = INT64_MAX;
16251 dtrace_membar_producer();
16252 state->dts_laststatus = dtrace_gethrtime();
16254 bzero(&stat, sizeof (stat));
16256 mutex_enter(&dtrace_lock);
16258 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16259 mutex_exit(&dtrace_lock);
16263 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16264 stat.dtst_exiting = 1;
16266 nerrs = state->dts_errors;
16267 dstate = &state->dts_vstate.dtvs_dynvars;
16269 for (i = 0; i < NCPU; i++) {
16270 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16272 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16273 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16274 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16276 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16277 stat.dtst_filled++;
16279 nerrs += state->dts_buffer[i].dtb_errors;
16281 for (j = 0; j < state->dts_nspeculations; j++) {
16282 dtrace_speculation_t *spec;
16283 dtrace_buffer_t *buf;
16285 spec = &state->dts_speculations[j];
16286 buf = &spec->dtsp_buffer[i];
16287 stat.dtst_specdrops += buf->dtb_xamot_drops;
16291 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16292 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16293 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16294 stat.dtst_dblerrors = state->dts_dblerrors;
16296 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16297 stat.dtst_errors = nerrs;
16299 mutex_exit(&dtrace_lock);
16301 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16307 case DTRACEIOC_FORMAT: {
16308 dtrace_fmtdesc_t fmt;
16312 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16315 mutex_enter(&dtrace_lock);
16317 if (fmt.dtfd_format == 0 ||
16318 fmt.dtfd_format > state->dts_nformats) {
16319 mutex_exit(&dtrace_lock);
16324 * Format strings are allocated contiguously and they are
16325 * never freed; if a format index is less than the number
16326 * of formats, we can assert that the format map is non-NULL
16327 * and that the format for the specified index is non-NULL.
16329 ASSERT(state->dts_formats != NULL);
16330 str = state->dts_formats[fmt.dtfd_format - 1];
16331 ASSERT(str != NULL);
16333 len = strlen(str) + 1;
16335 if (len > fmt.dtfd_length) {
16336 fmt.dtfd_length = len;
16338 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16339 mutex_exit(&dtrace_lock);
16343 if (copyout(str, fmt.dtfd_string, len) != 0) {
16344 mutex_exit(&dtrace_lock);
16349 mutex_exit(&dtrace_lock);
16362 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16364 dtrace_state_t *state;
16371 return (DDI_SUCCESS);
16374 return (DDI_FAILURE);
16377 mutex_enter(&cpu_lock);
16378 mutex_enter(&dtrace_provider_lock);
16379 mutex_enter(&dtrace_lock);
16381 ASSERT(dtrace_opens == 0);
16383 if (dtrace_helpers > 0) {
16384 mutex_exit(&dtrace_provider_lock);
16385 mutex_exit(&dtrace_lock);
16386 mutex_exit(&cpu_lock);
16387 return (DDI_FAILURE);
16390 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16391 mutex_exit(&dtrace_provider_lock);
16392 mutex_exit(&dtrace_lock);
16393 mutex_exit(&cpu_lock);
16394 return (DDI_FAILURE);
16397 dtrace_provider = NULL;
16399 if ((state = dtrace_anon_grab()) != NULL) {
16401 * If there were ECBs on this state, the provider should
16402 * have not been allowed to detach; assert that there is
16405 ASSERT(state->dts_necbs == 0);
16406 dtrace_state_destroy(state);
16409 * If we're being detached with anonymous state, we need to
16410 * indicate to the kernel debugger that DTrace is now inactive.
16412 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16415 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16416 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16417 dtrace_cpu_init = NULL;
16418 dtrace_helpers_cleanup = NULL;
16419 dtrace_helpers_fork = NULL;
16420 dtrace_cpustart_init = NULL;
16421 dtrace_cpustart_fini = NULL;
16422 dtrace_debugger_init = NULL;
16423 dtrace_debugger_fini = NULL;
16424 dtrace_modload = NULL;
16425 dtrace_modunload = NULL;
16427 mutex_exit(&cpu_lock);
16429 if (dtrace_helptrace_enabled) {
16430 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16431 dtrace_helptrace_buffer = NULL;
16434 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16435 dtrace_probes = NULL;
16436 dtrace_nprobes = 0;
16438 dtrace_hash_destroy(dtrace_bymod);
16439 dtrace_hash_destroy(dtrace_byfunc);
16440 dtrace_hash_destroy(dtrace_byname);
16441 dtrace_bymod = NULL;
16442 dtrace_byfunc = NULL;
16443 dtrace_byname = NULL;
16445 kmem_cache_destroy(dtrace_state_cache);
16446 vmem_destroy(dtrace_minor);
16447 vmem_destroy(dtrace_arena);
16449 if (dtrace_toxrange != NULL) {
16450 kmem_free(dtrace_toxrange,
16451 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16452 dtrace_toxrange = NULL;
16453 dtrace_toxranges = 0;
16454 dtrace_toxranges_max = 0;
16457 ddi_remove_minor_node(dtrace_devi, NULL);
16458 dtrace_devi = NULL;
16460 ddi_soft_state_fini(&dtrace_softstate);
16462 ASSERT(dtrace_vtime_references == 0);
16463 ASSERT(dtrace_opens == 0);
16464 ASSERT(dtrace_retained == NULL);
16466 mutex_exit(&dtrace_lock);
16467 mutex_exit(&dtrace_provider_lock);
16470 * We don't destroy the task queue until after we have dropped our
16471 * locks (taskq_destroy() may block on running tasks). To prevent
16472 * attempting to do work after we have effectively detached but before
16473 * the task queue has been destroyed, all tasks dispatched via the
16474 * task queue must check that DTrace is still attached before
16475 * performing any operation.
16477 taskq_destroy(dtrace_taskq);
16478 dtrace_taskq = NULL;
16480 return (DDI_SUCCESS);
16487 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16492 case DDI_INFO_DEVT2DEVINFO:
16493 *result = (void *)dtrace_devi;
16494 error = DDI_SUCCESS;
16496 case DDI_INFO_DEVT2INSTANCE:
16497 *result = (void *)0;
16498 error = DDI_SUCCESS;
16501 error = DDI_FAILURE;
16508 static struct cb_ops dtrace_cb_ops = {
16509 dtrace_open, /* open */
16510 dtrace_close, /* close */
16511 nulldev, /* strategy */
16512 nulldev, /* print */
16516 dtrace_ioctl, /* ioctl */
16517 nodev, /* devmap */
16519 nodev, /* segmap */
16520 nochpoll, /* poll */
16521 ddi_prop_op, /* cb_prop_op */
16523 D_NEW | D_MP /* Driver compatibility flag */
16526 static struct dev_ops dtrace_ops = {
16527 DEVO_REV, /* devo_rev */
16529 dtrace_info, /* get_dev_info */
16530 nulldev, /* identify */
16531 nulldev, /* probe */
16532 dtrace_attach, /* attach */
16533 dtrace_detach, /* detach */
16535 &dtrace_cb_ops, /* driver operations */
16536 NULL, /* bus operations */
16537 nodev /* dev power */
16540 static struct modldrv modldrv = {
16541 &mod_driverops, /* module type (this is a pseudo driver) */
16542 "Dynamic Tracing", /* name of module */
16543 &dtrace_ops, /* driver ops */
16546 static struct modlinkage modlinkage = {
16555 return (mod_install(&modlinkage));
16559 _info(struct modinfo *modinfop)
16561 return (mod_info(&modlinkage, modinfop));
16567 return (mod_remove(&modlinkage));
16571 static d_ioctl_t dtrace_ioctl;
16572 static d_ioctl_t dtrace_ioctl_helper;
16573 static void dtrace_load(void *);
16574 static int dtrace_unload(void);
16575 #if __FreeBSD_version < 800039
16576 static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16577 static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */
16578 static eventhandler_tag eh_tag; /* Event handler tag. */
16580 static struct cdev *dtrace_dev;
16581 static struct cdev *helper_dev;
16584 void dtrace_invop_init(void);
16585 void dtrace_invop_uninit(void);
16587 static struct cdevsw dtrace_cdevsw = {
16588 .d_version = D_VERSION,
16589 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16590 .d_close = dtrace_close,
16591 .d_ioctl = dtrace_ioctl,
16592 .d_open = dtrace_open,
16593 .d_name = "dtrace",
16596 static struct cdevsw helper_cdevsw = {
16597 .d_version = D_VERSION,
16598 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16599 .d_ioctl = dtrace_ioctl_helper,
16600 .d_name = "helper",
16603 #include <dtrace_anon.c>
16604 #if __FreeBSD_version < 800039
16605 #include <dtrace_clone.c>
16607 #include <dtrace_ioctl.c>
16608 #include <dtrace_load.c>
16609 #include <dtrace_modevent.c>
16610 #include <dtrace_sysctl.c>
16611 #include <dtrace_unload.c>
16612 #include <dtrace_vtime.c>
16613 #include <dtrace_hacks.c>
16614 #include <dtrace_isa.c>
16616 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16617 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16618 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16620 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16621 MODULE_VERSION(dtrace, 1);
16622 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
16623 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);