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);
555 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
556 dtrace_state_t *, uint64_t, uint64_t);
557 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
559 static void dtrace_buffer_drop(dtrace_buffer_t *);
560 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
561 dtrace_state_t *, dtrace_mstate_t *);
562 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
564 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
566 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
568 uint16_t dtrace_load16(uintptr_t);
569 uint32_t dtrace_load32(uintptr_t);
570 uint64_t dtrace_load64(uintptr_t);
571 uint8_t dtrace_load8(uintptr_t);
572 void dtrace_dynvar_clean(dtrace_dstate_t *);
573 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
574 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
575 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
578 * DTrace Probe Context Functions
580 * These functions are called from probe context. Because probe context is
581 * any context in which C may be called, arbitrarily locks may be held,
582 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
583 * As a result, functions called from probe context may only call other DTrace
584 * support functions -- they may not interact at all with the system at large.
585 * (Note that the ASSERT macro is made probe-context safe by redefining it in
586 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
587 * loads are to be performed from probe context, they _must_ be in terms of
588 * the safe dtrace_load*() variants.
590 * Some functions in this block are not actually called from probe context;
591 * for these functions, there will be a comment above the function reading
592 * "Note: not called from probe context."
595 dtrace_panic(const char *format, ...)
599 va_start(alist, format);
600 dtrace_vpanic(format, alist);
605 dtrace_assfail(const char *a, const char *f, int l)
607 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
610 * We just need something here that even the most clever compiler
611 * cannot optimize away.
613 return (a[(uintptr_t)f]);
617 * Atomically increment a specified error counter from probe context.
620 dtrace_error(uint32_t *counter)
623 * Most counters stored to in probe context are per-CPU counters.
624 * However, there are some error conditions that are sufficiently
625 * arcane that they don't merit per-CPU storage. If these counters
626 * are incremented concurrently on different CPUs, scalability will be
627 * adversely affected -- but we don't expect them to be white-hot in a
628 * correctly constructed enabling...
635 if ((nval = oval + 1) == 0) {
637 * If the counter would wrap, set it to 1 -- assuring
638 * that the counter is never zero when we have seen
639 * errors. (The counter must be 32-bits because we
640 * aren't guaranteed a 64-bit compare&swap operation.)
641 * To save this code both the infamy of being fingered
642 * by a priggish news story and the indignity of being
643 * the target of a neo-puritan witch trial, we're
644 * carefully avoiding any colorful description of the
645 * likelihood of this condition -- but suffice it to
646 * say that it is only slightly more likely than the
647 * overflow of predicate cache IDs, as discussed in
648 * dtrace_predicate_create().
652 } while (dtrace_cas32(counter, oval, nval) != oval);
656 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
657 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
665 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
667 if (dest < mstate->dtms_scratch_base)
670 if (dest + size < dest)
673 if (dest + size > mstate->dtms_scratch_ptr)
680 dtrace_canstore_statvar(uint64_t addr, size_t sz,
681 dtrace_statvar_t **svars, int nsvars)
685 for (i = 0; i < nsvars; i++) {
686 dtrace_statvar_t *svar = svars[i];
688 if (svar == NULL || svar->dtsv_size == 0)
691 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
699 * Check to see if the address is within a memory region to which a store may
700 * be issued. This includes the DTrace scratch areas, and any DTrace variable
701 * region. The caller of dtrace_canstore() is responsible for performing any
702 * alignment checks that are needed before stores are actually executed.
705 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
706 dtrace_vstate_t *vstate)
709 * First, check to see if the address is in scratch space...
711 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
712 mstate->dtms_scratch_size))
716 * Now check to see if it's a dynamic variable. This check will pick
717 * up both thread-local variables and any global dynamically-allocated
720 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
721 vstate->dtvs_dynvars.dtds_size)) {
722 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
723 uintptr_t base = (uintptr_t)dstate->dtds_base +
724 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
728 * Before we assume that we can store here, we need to make
729 * sure that it isn't in our metadata -- storing to our
730 * dynamic variable metadata would corrupt our state. For
731 * the range to not include any dynamic variable metadata,
734 * (1) Start above the hash table that is at the base of
735 * the dynamic variable space
737 * (2) Have a starting chunk offset that is beyond the
738 * dtrace_dynvar_t that is at the base of every chunk
740 * (3) Not span a chunk boundary
746 chunkoffs = (addr - base) % dstate->dtds_chunksize;
748 if (chunkoffs < sizeof (dtrace_dynvar_t))
751 if (chunkoffs + sz > dstate->dtds_chunksize)
758 * Finally, check the static local and global variables. These checks
759 * take the longest, so we perform them last.
761 if (dtrace_canstore_statvar(addr, sz,
762 vstate->dtvs_locals, vstate->dtvs_nlocals))
765 if (dtrace_canstore_statvar(addr, sz,
766 vstate->dtvs_globals, vstate->dtvs_nglobals))
774 * Convenience routine to check to see if the address is within a memory
775 * region in which a load may be issued given the user's privilege level;
776 * if not, it sets the appropriate error flags and loads 'addr' into the
777 * illegal value slot.
779 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
780 * appropriate memory access protection.
783 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
784 dtrace_vstate_t *vstate)
786 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
789 * If we hold the privilege to read from kernel memory, then
790 * everything is readable.
792 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
796 * You can obviously read that which you can store.
798 if (dtrace_canstore(addr, sz, mstate, vstate))
802 * We're allowed to read from our own string table.
804 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
805 mstate->dtms_difo->dtdo_strlen))
808 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
814 * Convenience routine to check to see if a given string is within a memory
815 * region in which a load may be issued given the user's privilege level;
816 * this exists so that we don't need to issue unnecessary dtrace_strlen()
817 * calls in the event that the user has all privileges.
820 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
821 dtrace_vstate_t *vstate)
826 * If we hold the privilege to read from kernel memory, then
827 * everything is readable.
829 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
832 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
833 if (dtrace_canload(addr, strsz, mstate, vstate))
840 * Convenience routine to check to see if a given variable is within a memory
841 * region in which a load may be issued given the user's privilege level.
844 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
845 dtrace_vstate_t *vstate)
848 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
851 * If we hold the privilege to read from kernel memory, then
852 * everything is readable.
854 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
857 if (type->dtdt_kind == DIF_TYPE_STRING)
858 sz = dtrace_strlen(src,
859 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
861 sz = type->dtdt_size;
863 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
867 * Compare two strings using safe loads.
870 dtrace_strncmp(char *s1, char *s2, size_t limit)
873 volatile uint16_t *flags;
875 if (s1 == s2 || limit == 0)
878 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
884 c1 = dtrace_load8((uintptr_t)s1++);
890 c2 = dtrace_load8((uintptr_t)s2++);
895 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
901 * Compute strlen(s) for a string using safe memory accesses. The additional
902 * len parameter is used to specify a maximum length to ensure completion.
905 dtrace_strlen(const char *s, size_t lim)
909 for (len = 0; len != lim; len++) {
910 if (dtrace_load8((uintptr_t)s++) == '\0')
918 * Check if an address falls within a toxic region.
921 dtrace_istoxic(uintptr_t kaddr, size_t size)
923 uintptr_t taddr, tsize;
926 for (i = 0; i < dtrace_toxranges; i++) {
927 taddr = dtrace_toxrange[i].dtt_base;
928 tsize = dtrace_toxrange[i].dtt_limit - taddr;
930 if (kaddr - taddr < tsize) {
931 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
932 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
936 if (taddr - kaddr < size) {
937 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
938 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
947 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
948 * memory specified by the DIF program. The dst is assumed to be safe memory
949 * that we can store to directly because it is managed by DTrace. As with
950 * standard bcopy, overlapping copies are handled properly.
953 dtrace_bcopy(const void *src, void *dst, size_t len)
957 const uint8_t *s2 = src;
961 *s1++ = dtrace_load8((uintptr_t)s2++);
962 } while (--len != 0);
968 *--s1 = dtrace_load8((uintptr_t)--s2);
969 } while (--len != 0);
975 * Copy src to dst using safe memory accesses, up to either the specified
976 * length, or the point that a nul byte is encountered. The src is assumed to
977 * be unsafe memory specified by the DIF program. The dst is assumed to be
978 * safe memory that we can store to directly because it is managed by DTrace.
979 * Unlike dtrace_bcopy(), overlapping regions are not handled.
982 dtrace_strcpy(const void *src, void *dst, size_t len)
985 uint8_t *s1 = dst, c;
986 const uint8_t *s2 = src;
989 *s1++ = c = dtrace_load8((uintptr_t)s2++);
990 } while (--len != 0 && c != '\0');
995 * Copy src to dst, deriving the size and type from the specified (BYREF)
996 * variable type. The src is assumed to be unsafe memory specified by the DIF
997 * program. The dst is assumed to be DTrace variable memory that is of the
998 * specified type; we assume that we can store to directly.
1001 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1003 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1005 if (type->dtdt_kind == DIF_TYPE_STRING) {
1006 dtrace_strcpy(src, dst, type->dtdt_size);
1008 dtrace_bcopy(src, dst, type->dtdt_size);
1013 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1014 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1015 * safe memory that we can access directly because it is managed by DTrace.
1018 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1020 volatile uint16_t *flags;
1022 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1027 if (s1 == NULL || s2 == NULL)
1030 if (s1 != s2 && len != 0) {
1031 const uint8_t *ps1 = s1;
1032 const uint8_t *ps2 = s2;
1035 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1037 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1043 * Zero the specified region using a simple byte-by-byte loop. Note that this
1044 * is for safe DTrace-managed memory only.
1047 dtrace_bzero(void *dst, size_t len)
1051 for (cp = dst; len != 0; len--)
1056 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1060 result[0] = addend1[0] + addend2[0];
1061 result[1] = addend1[1] + addend2[1] +
1062 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1069 * Shift the 128-bit value in a by b. If b is positive, shift left.
1070 * If b is negative, shift right.
1073 dtrace_shift_128(uint64_t *a, int b)
1083 a[0] = a[1] >> (b - 64);
1087 mask = 1LL << (64 - b);
1089 a[0] |= ((a[1] & mask) << (64 - b));
1094 a[1] = a[0] << (b - 64);
1098 mask = a[0] >> (64 - b);
1106 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1107 * use native multiplication on those, and then re-combine into the
1108 * resulting 128-bit value.
1110 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1117 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1119 uint64_t hi1, hi2, lo1, lo2;
1122 hi1 = factor1 >> 32;
1123 hi2 = factor2 >> 32;
1125 lo1 = factor1 & DT_MASK_LO;
1126 lo2 = factor2 & DT_MASK_LO;
1128 product[0] = lo1 * lo2;
1129 product[1] = hi1 * hi2;
1133 dtrace_shift_128(tmp, 32);
1134 dtrace_add_128(product, tmp, product);
1138 dtrace_shift_128(tmp, 32);
1139 dtrace_add_128(product, tmp, product);
1143 * This privilege check should be used by actions and subroutines to
1144 * verify that the user credentials of the process that enabled the
1145 * invoking ECB match the target credentials
1148 dtrace_priv_proc_common_user(dtrace_state_t *state)
1150 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1153 * We should always have a non-NULL state cred here, since if cred
1154 * is null (anonymous tracing), we fast-path bypass this routine.
1156 ASSERT(s_cr != NULL);
1158 if ((cr = CRED()) != NULL &&
1159 s_cr->cr_uid == cr->cr_uid &&
1160 s_cr->cr_uid == cr->cr_ruid &&
1161 s_cr->cr_uid == cr->cr_suid &&
1162 s_cr->cr_gid == cr->cr_gid &&
1163 s_cr->cr_gid == cr->cr_rgid &&
1164 s_cr->cr_gid == cr->cr_sgid)
1171 * This privilege check should be used by actions and subroutines to
1172 * verify that the zone of the process that enabled the invoking ECB
1173 * matches the target credentials
1176 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1179 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1182 * We should always have a non-NULL state cred here, since if cred
1183 * is null (anonymous tracing), we fast-path bypass this routine.
1185 ASSERT(s_cr != NULL);
1187 if ((cr = CRED()) != NULL &&
1188 s_cr->cr_zone == cr->cr_zone)
1198 * This privilege check should be used by actions and subroutines to
1199 * verify that the process has not setuid or changed credentials.
1202 dtrace_priv_proc_common_nocd(void)
1206 if ((proc = ttoproc(curthread)) != NULL &&
1207 !(proc->p_flag & SNOCD))
1214 dtrace_priv_proc_destructive(dtrace_state_t *state)
1216 int action = state->dts_cred.dcr_action;
1218 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1219 dtrace_priv_proc_common_zone(state) == 0)
1222 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1223 dtrace_priv_proc_common_user(state) == 0)
1226 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1227 dtrace_priv_proc_common_nocd() == 0)
1233 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1239 dtrace_priv_proc_control(dtrace_state_t *state)
1241 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1244 if (dtrace_priv_proc_common_zone(state) &&
1245 dtrace_priv_proc_common_user(state) &&
1246 dtrace_priv_proc_common_nocd())
1249 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1255 dtrace_priv_proc(dtrace_state_t *state)
1257 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1260 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1266 dtrace_priv_kernel(dtrace_state_t *state)
1268 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1271 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1277 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1279 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1282 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1288 * Note: not called from probe context. This function is called
1289 * asynchronously (and at a regular interval) from outside of probe context to
1290 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1291 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1294 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1296 dtrace_dynvar_t *dirty;
1297 dtrace_dstate_percpu_t *dcpu;
1300 for (i = 0; i < NCPU; i++) {
1301 dcpu = &dstate->dtds_percpu[i];
1303 ASSERT(dcpu->dtdsc_rinsing == NULL);
1306 * If the dirty list is NULL, there is no dirty work to do.
1308 if (dcpu->dtdsc_dirty == NULL)
1312 * If the clean list is non-NULL, then we're not going to do
1313 * any work for this CPU -- it means that there has not been
1314 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1315 * since the last time we cleaned house.
1317 if (dcpu->dtdsc_clean != NULL)
1323 * Atomically move the dirty list aside.
1326 dirty = dcpu->dtdsc_dirty;
1329 * Before we zap the dirty list, set the rinsing list.
1330 * (This allows for a potential assertion in
1331 * dtrace_dynvar(): if a free dynamic variable appears
1332 * on a hash chain, either the dirty list or the
1333 * rinsing list for some CPU must be non-NULL.)
1335 dcpu->dtdsc_rinsing = dirty;
1336 dtrace_membar_producer();
1337 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1338 dirty, NULL) != dirty);
1343 * We have no work to do; we can simply return.
1350 for (i = 0; i < NCPU; i++) {
1351 dcpu = &dstate->dtds_percpu[i];
1353 if (dcpu->dtdsc_rinsing == NULL)
1357 * We are now guaranteed that no hash chain contains a pointer
1358 * into this dirty list; we can make it clean.
1360 ASSERT(dcpu->dtdsc_clean == NULL);
1361 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1362 dcpu->dtdsc_rinsing = NULL;
1366 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1367 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1368 * This prevents a race whereby a CPU incorrectly decides that
1369 * the state should be something other than DTRACE_DSTATE_CLEAN
1370 * after dtrace_dynvar_clean() has completed.
1374 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1378 * Depending on the value of the op parameter, this function looks-up,
1379 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1380 * allocation is requested, this function will return a pointer to a
1381 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1382 * variable can be allocated. If NULL is returned, the appropriate counter
1383 * will be incremented.
1386 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1387 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1388 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1390 uint64_t hashval = DTRACE_DYNHASH_VALID;
1391 dtrace_dynhash_t *hash = dstate->dtds_hash;
1392 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1393 processorid_t me = curcpu, cpu = me;
1394 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1395 size_t bucket, ksize;
1396 size_t chunksize = dstate->dtds_chunksize;
1397 uintptr_t kdata, lock, nstate;
1403 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1404 * algorithm. For the by-value portions, we perform the algorithm in
1405 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1406 * bit, and seems to have only a minute effect on distribution. For
1407 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1408 * over each referenced byte. It's painful to do this, but it's much
1409 * better than pathological hash distribution. The efficacy of the
1410 * hashing algorithm (and a comparison with other algorithms) may be
1411 * found by running the ::dtrace_dynstat MDB dcmd.
1413 for (i = 0; i < nkeys; i++) {
1414 if (key[i].dttk_size == 0) {
1415 uint64_t val = key[i].dttk_value;
1417 hashval += (val >> 48) & 0xffff;
1418 hashval += (hashval << 10);
1419 hashval ^= (hashval >> 6);
1421 hashval += (val >> 32) & 0xffff;
1422 hashval += (hashval << 10);
1423 hashval ^= (hashval >> 6);
1425 hashval += (val >> 16) & 0xffff;
1426 hashval += (hashval << 10);
1427 hashval ^= (hashval >> 6);
1429 hashval += val & 0xffff;
1430 hashval += (hashval << 10);
1431 hashval ^= (hashval >> 6);
1434 * This is incredibly painful, but it beats the hell
1435 * out of the alternative.
1437 uint64_t j, size = key[i].dttk_size;
1438 uintptr_t base = (uintptr_t)key[i].dttk_value;
1440 if (!dtrace_canload(base, size, mstate, vstate))
1443 for (j = 0; j < size; j++) {
1444 hashval += dtrace_load8(base + j);
1445 hashval += (hashval << 10);
1446 hashval ^= (hashval >> 6);
1451 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1454 hashval += (hashval << 3);
1455 hashval ^= (hashval >> 11);
1456 hashval += (hashval << 15);
1459 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1460 * comes out to be one of our two sentinel hash values. If this
1461 * actually happens, we set the hashval to be a value known to be a
1462 * non-sentinel value.
1464 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1465 hashval = DTRACE_DYNHASH_VALID;
1468 * Yes, it's painful to do a divide here. If the cycle count becomes
1469 * important here, tricks can be pulled to reduce it. (However, it's
1470 * critical that hash collisions be kept to an absolute minimum;
1471 * they're much more painful than a divide.) It's better to have a
1472 * solution that generates few collisions and still keeps things
1473 * relatively simple.
1475 bucket = hashval % dstate->dtds_hashsize;
1477 if (op == DTRACE_DYNVAR_DEALLOC) {
1478 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1481 while ((lock = *lockp) & 1)
1484 if (dtrace_casptr((volatile void *)lockp,
1485 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1489 dtrace_membar_producer();
1494 lock = hash[bucket].dtdh_lock;
1496 dtrace_membar_consumer();
1498 start = hash[bucket].dtdh_chain;
1499 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1500 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1501 op != DTRACE_DYNVAR_DEALLOC));
1503 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1504 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1505 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1507 if (dvar->dtdv_hashval != hashval) {
1508 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1510 * We've reached the sink, and therefore the
1511 * end of the hash chain; we can kick out of
1512 * the loop knowing that we have seen a valid
1513 * snapshot of state.
1515 ASSERT(dvar->dtdv_next == NULL);
1516 ASSERT(dvar == &dtrace_dynhash_sink);
1520 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1522 * We've gone off the rails: somewhere along
1523 * the line, one of the members of this hash
1524 * chain was deleted. Note that we could also
1525 * detect this by simply letting this loop run
1526 * to completion, as we would eventually hit
1527 * the end of the dirty list. However, we
1528 * want to avoid running the length of the
1529 * dirty list unnecessarily (it might be quite
1530 * long), so we catch this as early as
1531 * possible by detecting the hash marker. In
1532 * this case, we simply set dvar to NULL and
1533 * break; the conditional after the loop will
1534 * send us back to top.
1543 if (dtuple->dtt_nkeys != nkeys)
1546 for (i = 0; i < nkeys; i++, dkey++) {
1547 if (dkey->dttk_size != key[i].dttk_size)
1548 goto next; /* size or type mismatch */
1550 if (dkey->dttk_size != 0) {
1552 (void *)(uintptr_t)key[i].dttk_value,
1553 (void *)(uintptr_t)dkey->dttk_value,
1557 if (dkey->dttk_value != key[i].dttk_value)
1562 if (op != DTRACE_DYNVAR_DEALLOC)
1565 ASSERT(dvar->dtdv_next == NULL ||
1566 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1569 ASSERT(hash[bucket].dtdh_chain != dvar);
1570 ASSERT(start != dvar);
1571 ASSERT(prev->dtdv_next == dvar);
1572 prev->dtdv_next = dvar->dtdv_next;
1574 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1575 start, dvar->dtdv_next) != start) {
1577 * We have failed to atomically swing the
1578 * hash table head pointer, presumably because
1579 * of a conflicting allocation on another CPU.
1580 * We need to reread the hash chain and try
1587 dtrace_membar_producer();
1590 * Now set the hash value to indicate that it's free.
1592 ASSERT(hash[bucket].dtdh_chain != dvar);
1593 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1595 dtrace_membar_producer();
1598 * Set the next pointer to point at the dirty list, and
1599 * atomically swing the dirty pointer to the newly freed dvar.
1602 next = dcpu->dtdsc_dirty;
1603 dvar->dtdv_next = next;
1604 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1607 * Finally, unlock this hash bucket.
1609 ASSERT(hash[bucket].dtdh_lock == lock);
1611 hash[bucket].dtdh_lock++;
1621 * If dvar is NULL, it is because we went off the rails:
1622 * one of the elements that we traversed in the hash chain
1623 * was deleted while we were traversing it. In this case,
1624 * we assert that we aren't doing a dealloc (deallocs lock
1625 * the hash bucket to prevent themselves from racing with
1626 * one another), and retry the hash chain traversal.
1628 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1632 if (op != DTRACE_DYNVAR_ALLOC) {
1634 * If we are not to allocate a new variable, we want to
1635 * return NULL now. Before we return, check that the value
1636 * of the lock word hasn't changed. If it has, we may have
1637 * seen an inconsistent snapshot.
1639 if (op == DTRACE_DYNVAR_NOALLOC) {
1640 if (hash[bucket].dtdh_lock != lock)
1643 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1644 ASSERT(hash[bucket].dtdh_lock == lock);
1646 hash[bucket].dtdh_lock++;
1653 * We need to allocate a new dynamic variable. The size we need is the
1654 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1655 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1656 * the size of any referred-to data (dsize). We then round the final
1657 * size up to the chunksize for allocation.
1659 for (ksize = 0, i = 0; i < nkeys; i++)
1660 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1663 * This should be pretty much impossible, but could happen if, say,
1664 * strange DIF specified the tuple. Ideally, this should be an
1665 * assertion and not an error condition -- but that requires that the
1666 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1667 * bullet-proof. (That is, it must not be able to be fooled by
1668 * malicious DIF.) Given the lack of backwards branches in DIF,
1669 * solving this would presumably not amount to solving the Halting
1670 * Problem -- but it still seems awfully hard.
1672 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1673 ksize + dsize > chunksize) {
1674 dcpu->dtdsc_drops++;
1678 nstate = DTRACE_DSTATE_EMPTY;
1682 free = dcpu->dtdsc_free;
1685 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1688 if (clean == NULL) {
1690 * We're out of dynamic variable space on
1691 * this CPU. Unless we have tried all CPUs,
1692 * we'll try to allocate from a different
1695 switch (dstate->dtds_state) {
1696 case DTRACE_DSTATE_CLEAN: {
1697 void *sp = &dstate->dtds_state;
1702 if (dcpu->dtdsc_dirty != NULL &&
1703 nstate == DTRACE_DSTATE_EMPTY)
1704 nstate = DTRACE_DSTATE_DIRTY;
1706 if (dcpu->dtdsc_rinsing != NULL)
1707 nstate = DTRACE_DSTATE_RINSING;
1709 dcpu = &dstate->dtds_percpu[cpu];
1714 (void) dtrace_cas32(sp,
1715 DTRACE_DSTATE_CLEAN, nstate);
1718 * To increment the correct bean
1719 * counter, take another lap.
1724 case DTRACE_DSTATE_DIRTY:
1725 dcpu->dtdsc_dirty_drops++;
1728 case DTRACE_DSTATE_RINSING:
1729 dcpu->dtdsc_rinsing_drops++;
1732 case DTRACE_DSTATE_EMPTY:
1733 dcpu->dtdsc_drops++;
1737 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1742 * The clean list appears to be non-empty. We want to
1743 * move the clean list to the free list; we start by
1744 * moving the clean pointer aside.
1746 if (dtrace_casptr(&dcpu->dtdsc_clean,
1747 clean, NULL) != clean) {
1749 * We are in one of two situations:
1751 * (a) The clean list was switched to the
1752 * free list by another CPU.
1754 * (b) The clean list was added to by the
1757 * In either of these situations, we can
1758 * just reattempt the free list allocation.
1763 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1766 * Now we'll move the clean list to the free list.
1767 * It's impossible for this to fail: the only way
1768 * the free list can be updated is through this
1769 * code path, and only one CPU can own the clean list.
1770 * Thus, it would only be possible for this to fail if
1771 * this code were racing with dtrace_dynvar_clean().
1772 * (That is, if dtrace_dynvar_clean() updated the clean
1773 * list, and we ended up racing to update the free
1774 * list.) This race is prevented by the dtrace_sync()
1775 * in dtrace_dynvar_clean() -- which flushes the
1776 * owners of the clean lists out before resetting
1779 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1780 ASSERT(rval == NULL);
1785 new_free = dvar->dtdv_next;
1786 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1789 * We have now allocated a new chunk. We copy the tuple keys into the
1790 * tuple array and copy any referenced key data into the data space
1791 * following the tuple array. As we do this, we relocate dttk_value
1792 * in the final tuple to point to the key data address in the chunk.
1794 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1795 dvar->dtdv_data = (void *)(kdata + ksize);
1796 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1798 for (i = 0; i < nkeys; i++) {
1799 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1800 size_t kesize = key[i].dttk_size;
1804 (const void *)(uintptr_t)key[i].dttk_value,
1805 (void *)kdata, kesize);
1806 dkey->dttk_value = kdata;
1807 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1809 dkey->dttk_value = key[i].dttk_value;
1812 dkey->dttk_size = kesize;
1815 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1816 dvar->dtdv_hashval = hashval;
1817 dvar->dtdv_next = start;
1819 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1823 * The cas has failed. Either another CPU is adding an element to
1824 * this hash chain, or another CPU is deleting an element from this
1825 * hash chain. The simplest way to deal with both of these cases
1826 * (though not necessarily the most efficient) is to free our
1827 * allocated block and tail-call ourselves. Note that the free is
1828 * to the dirty list and _not_ to the free list. This is to prevent
1829 * races with allocators, above.
1831 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1833 dtrace_membar_producer();
1836 free = dcpu->dtdsc_dirty;
1837 dvar->dtdv_next = free;
1838 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1840 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1845 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1847 if ((int64_t)nval < (int64_t)*oval)
1853 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1855 if ((int64_t)nval > (int64_t)*oval)
1860 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1862 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1863 int64_t val = (int64_t)nval;
1866 for (i = 0; i < zero; i++) {
1867 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1873 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1874 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1875 quanta[i - 1] += incr;
1880 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1888 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1890 uint64_t arg = *lquanta++;
1891 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1892 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1893 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1894 int32_t val = (int32_t)nval, level;
1897 ASSERT(levels != 0);
1901 * This is an underflow.
1907 level = (val - base) / step;
1909 if (level < levels) {
1910 lquanta[level + 1] += incr;
1915 * This is an overflow.
1917 lquanta[levels + 1] += incr;
1922 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
1930 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
1932 int64_t snval = (int64_t)nval;
1939 * What we want to say here is:
1941 * data[2] += nval * nval;
1943 * But given that nval is 64-bit, we could easily overflow, so
1944 * we do this as 128-bit arithmetic.
1949 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
1950 dtrace_add_128(data + 2, tmp, data + 2);
1955 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
1962 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
1968 * Aggregate given the tuple in the principal data buffer, and the aggregating
1969 * action denoted by the specified dtrace_aggregation_t. The aggregation
1970 * buffer is specified as the buf parameter. This routine does not return
1971 * failure; if there is no space in the aggregation buffer, the data will be
1972 * dropped, and a corresponding counter incremented.
1975 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
1976 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
1978 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
1979 uint32_t i, ndx, size, fsize;
1980 uint32_t align = sizeof (uint64_t) - 1;
1981 dtrace_aggbuffer_t *agb;
1982 dtrace_aggkey_t *key;
1983 uint32_t hashval = 0, limit, isstr;
1984 caddr_t tomax, data, kdata;
1985 dtrace_actkind_t action;
1986 dtrace_action_t *act;
1992 if (!agg->dtag_hasarg) {
1994 * Currently, only quantize() and lquantize() take additional
1995 * arguments, and they have the same semantics: an increment
1996 * value that defaults to 1 when not present. If additional
1997 * aggregating actions take arguments, the setting of the
1998 * default argument value will presumably have to become more
2004 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2005 size = rec->dtrd_offset - agg->dtag_base;
2006 fsize = size + rec->dtrd_size;
2008 ASSERT(dbuf->dtb_tomax != NULL);
2009 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2011 if ((tomax = buf->dtb_tomax) == NULL) {
2012 dtrace_buffer_drop(buf);
2017 * The metastructure is always at the bottom of the buffer.
2019 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2020 sizeof (dtrace_aggbuffer_t));
2022 if (buf->dtb_offset == 0) {
2024 * We just kludge up approximately 1/8th of the size to be
2025 * buckets. If this guess ends up being routinely
2026 * off-the-mark, we may need to dynamically readjust this
2027 * based on past performance.
2029 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2031 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2032 (uintptr_t)tomax || hashsize == 0) {
2034 * We've been given a ludicrously small buffer;
2035 * increment our drop count and leave.
2037 dtrace_buffer_drop(buf);
2042 * And now, a pathetic attempt to try to get a an odd (or
2043 * perchance, a prime) hash size for better hash distribution.
2045 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2046 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2048 agb->dtagb_hashsize = hashsize;
2049 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2050 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2051 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2053 for (i = 0; i < agb->dtagb_hashsize; i++)
2054 agb->dtagb_hash[i] = NULL;
2057 ASSERT(agg->dtag_first != NULL);
2058 ASSERT(agg->dtag_first->dta_intuple);
2061 * Calculate the hash value based on the key. Note that we _don't_
2062 * include the aggid in the hashing (but we will store it as part of
2063 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2064 * algorithm: a simple, quick algorithm that has no known funnels, and
2065 * gets good distribution in practice. The efficacy of the hashing
2066 * algorithm (and a comparison with other algorithms) may be found by
2067 * running the ::dtrace_aggstat MDB dcmd.
2069 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2070 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2071 limit = i + act->dta_rec.dtrd_size;
2072 ASSERT(limit <= size);
2073 isstr = DTRACEACT_ISSTRING(act);
2075 for (; i < limit; i++) {
2077 hashval += (hashval << 10);
2078 hashval ^= (hashval >> 6);
2080 if (isstr && data[i] == '\0')
2085 hashval += (hashval << 3);
2086 hashval ^= (hashval >> 11);
2087 hashval += (hashval << 15);
2090 * Yes, the divide here is expensive -- but it's generally the least
2091 * of the performance issues given the amount of data that we iterate
2092 * over to compute hash values, compare data, etc.
2094 ndx = hashval % agb->dtagb_hashsize;
2096 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2097 ASSERT((caddr_t)key >= tomax);
2098 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2100 if (hashval != key->dtak_hashval || key->dtak_size != size)
2103 kdata = key->dtak_data;
2104 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2106 for (act = agg->dtag_first; act->dta_intuple;
2107 act = act->dta_next) {
2108 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2109 limit = i + act->dta_rec.dtrd_size;
2110 ASSERT(limit <= size);
2111 isstr = DTRACEACT_ISSTRING(act);
2113 for (; i < limit; i++) {
2114 if (kdata[i] != data[i])
2117 if (isstr && data[i] == '\0')
2122 if (action != key->dtak_action) {
2124 * We are aggregating on the same value in the same
2125 * aggregation with two different aggregating actions.
2126 * (This should have been picked up in the compiler,
2127 * so we may be dealing with errant or devious DIF.)
2128 * This is an error condition; we indicate as much,
2131 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2136 * This is a hit: we need to apply the aggregator to
2137 * the value at this key.
2139 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2146 * We didn't find it. We need to allocate some zero-filled space,
2147 * link it into the hash table appropriately, and apply the aggregator
2148 * to the (zero-filled) value.
2150 offs = buf->dtb_offset;
2151 while (offs & (align - 1))
2152 offs += sizeof (uint32_t);
2155 * If we don't have enough room to both allocate a new key _and_
2156 * its associated data, increment the drop count and return.
2158 if ((uintptr_t)tomax + offs + fsize >
2159 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2160 dtrace_buffer_drop(buf);
2165 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2166 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2167 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2169 key->dtak_data = kdata = tomax + offs;
2170 buf->dtb_offset = offs + fsize;
2173 * Now copy the data across.
2175 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2177 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2181 * Because strings are not zeroed out by default, we need to iterate
2182 * looking for actions that store strings, and we need to explicitly
2183 * pad these strings out with zeroes.
2185 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2188 if (!DTRACEACT_ISSTRING(act))
2191 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2192 limit = i + act->dta_rec.dtrd_size;
2193 ASSERT(limit <= size);
2195 for (nul = 0; i < limit; i++) {
2201 if (data[i] != '\0')
2208 for (i = size; i < fsize; i++)
2211 key->dtak_hashval = hashval;
2212 key->dtak_size = size;
2213 key->dtak_action = action;
2214 key->dtak_next = agb->dtagb_hash[ndx];
2215 agb->dtagb_hash[ndx] = key;
2218 * Finally, apply the aggregator.
2220 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2221 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2225 * Given consumer state, this routine finds a speculation in the INACTIVE
2226 * state and transitions it into the ACTIVE state. If there is no speculation
2227 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2228 * incremented -- it is up to the caller to take appropriate action.
2231 dtrace_speculation(dtrace_state_t *state)
2234 dtrace_speculation_state_t current;
2235 uint32_t *stat = &state->dts_speculations_unavail, count;
2237 while (i < state->dts_nspeculations) {
2238 dtrace_speculation_t *spec = &state->dts_speculations[i];
2240 current = spec->dtsp_state;
2242 if (current != DTRACESPEC_INACTIVE) {
2243 if (current == DTRACESPEC_COMMITTINGMANY ||
2244 current == DTRACESPEC_COMMITTING ||
2245 current == DTRACESPEC_DISCARDING)
2246 stat = &state->dts_speculations_busy;
2251 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2252 current, DTRACESPEC_ACTIVE) == current)
2257 * We couldn't find a speculation. If we found as much as a single
2258 * busy speculation buffer, we'll attribute this failure as "busy"
2259 * instead of "unavail".
2263 } while (dtrace_cas32(stat, count, count + 1) != count);
2269 * This routine commits an active speculation. If the specified speculation
2270 * is not in a valid state to perform a commit(), this routine will silently do
2271 * nothing. The state of the specified speculation is transitioned according
2272 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2275 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2276 dtrace_specid_t which)
2278 dtrace_speculation_t *spec;
2279 dtrace_buffer_t *src, *dest;
2280 uintptr_t daddr, saddr, dlimit;
2281 dtrace_speculation_state_t current, new = 0;
2287 if (which > state->dts_nspeculations) {
2288 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2292 spec = &state->dts_speculations[which - 1];
2293 src = &spec->dtsp_buffer[cpu];
2294 dest = &state->dts_buffer[cpu];
2297 current = spec->dtsp_state;
2299 if (current == DTRACESPEC_COMMITTINGMANY)
2303 case DTRACESPEC_INACTIVE:
2304 case DTRACESPEC_DISCARDING:
2307 case DTRACESPEC_COMMITTING:
2309 * This is only possible if we are (a) commit()'ing
2310 * without having done a prior speculate() on this CPU
2311 * and (b) racing with another commit() on a different
2312 * CPU. There's nothing to do -- we just assert that
2315 ASSERT(src->dtb_offset == 0);
2318 case DTRACESPEC_ACTIVE:
2319 new = DTRACESPEC_COMMITTING;
2322 case DTRACESPEC_ACTIVEONE:
2324 * This speculation is active on one CPU. If our
2325 * buffer offset is non-zero, we know that the one CPU
2326 * must be us. Otherwise, we are committing on a
2327 * different CPU from the speculate(), and we must
2328 * rely on being asynchronously cleaned.
2330 if (src->dtb_offset != 0) {
2331 new = DTRACESPEC_COMMITTING;
2336 case DTRACESPEC_ACTIVEMANY:
2337 new = DTRACESPEC_COMMITTINGMANY;
2343 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2344 current, new) != current);
2347 * We have set the state to indicate that we are committing this
2348 * speculation. Now reserve the necessary space in the destination
2351 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2352 sizeof (uint64_t), state, NULL)) < 0) {
2353 dtrace_buffer_drop(dest);
2358 * We have the space; copy the buffer across. (Note that this is a
2359 * highly subobtimal bcopy(); in the unlikely event that this becomes
2360 * a serious performance issue, a high-performance DTrace-specific
2361 * bcopy() should obviously be invented.)
2363 daddr = (uintptr_t)dest->dtb_tomax + offs;
2364 dlimit = daddr + src->dtb_offset;
2365 saddr = (uintptr_t)src->dtb_tomax;
2368 * First, the aligned portion.
2370 while (dlimit - daddr >= sizeof (uint64_t)) {
2371 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2373 daddr += sizeof (uint64_t);
2374 saddr += sizeof (uint64_t);
2378 * Now any left-over bit...
2380 while (dlimit - daddr)
2381 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2384 * Finally, commit the reserved space in the destination buffer.
2386 dest->dtb_offset = offs + src->dtb_offset;
2390 * If we're lucky enough to be the only active CPU on this speculation
2391 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2393 if (current == DTRACESPEC_ACTIVE ||
2394 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2395 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2396 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2398 ASSERT(rval == DTRACESPEC_COMMITTING);
2401 src->dtb_offset = 0;
2402 src->dtb_xamot_drops += src->dtb_drops;
2407 * This routine discards an active speculation. If the specified speculation
2408 * is not in a valid state to perform a discard(), this routine will silently
2409 * do nothing. The state of the specified speculation is transitioned
2410 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2413 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2414 dtrace_specid_t which)
2416 dtrace_speculation_t *spec;
2417 dtrace_speculation_state_t current, new = 0;
2418 dtrace_buffer_t *buf;
2423 if (which > state->dts_nspeculations) {
2424 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2428 spec = &state->dts_speculations[which - 1];
2429 buf = &spec->dtsp_buffer[cpu];
2432 current = spec->dtsp_state;
2435 case DTRACESPEC_INACTIVE:
2436 case DTRACESPEC_COMMITTINGMANY:
2437 case DTRACESPEC_COMMITTING:
2438 case DTRACESPEC_DISCARDING:
2441 case DTRACESPEC_ACTIVE:
2442 case DTRACESPEC_ACTIVEMANY:
2443 new = DTRACESPEC_DISCARDING;
2446 case DTRACESPEC_ACTIVEONE:
2447 if (buf->dtb_offset != 0) {
2448 new = DTRACESPEC_INACTIVE;
2450 new = DTRACESPEC_DISCARDING;
2457 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2458 current, new) != current);
2460 buf->dtb_offset = 0;
2465 * Note: not called from probe context. This function is called
2466 * asynchronously from cross call context to clean any speculations that are
2467 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2468 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2472 dtrace_speculation_clean_here(dtrace_state_t *state)
2474 dtrace_icookie_t cookie;
2475 processorid_t cpu = curcpu;
2476 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2479 cookie = dtrace_interrupt_disable();
2481 if (dest->dtb_tomax == NULL) {
2482 dtrace_interrupt_enable(cookie);
2486 for (i = 0; i < state->dts_nspeculations; i++) {
2487 dtrace_speculation_t *spec = &state->dts_speculations[i];
2488 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2490 if (src->dtb_tomax == NULL)
2493 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2494 src->dtb_offset = 0;
2498 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2501 if (src->dtb_offset == 0)
2504 dtrace_speculation_commit(state, cpu, i + 1);
2507 dtrace_interrupt_enable(cookie);
2511 * Note: not called from probe context. This function is called
2512 * asynchronously (and at a regular interval) to clean any speculations that
2513 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2514 * is work to be done, it cross calls all CPUs to perform that work;
2515 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2516 * INACTIVE state until they have been cleaned by all CPUs.
2519 dtrace_speculation_clean(dtrace_state_t *state)
2524 for (i = 0; i < state->dts_nspeculations; i++) {
2525 dtrace_speculation_t *spec = &state->dts_speculations[i];
2527 ASSERT(!spec->dtsp_cleaning);
2529 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2530 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2534 spec->dtsp_cleaning = 1;
2540 dtrace_xcall(DTRACE_CPUALL,
2541 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2544 * We now know that all CPUs have committed or discarded their
2545 * speculation buffers, as appropriate. We can now set the state
2548 for (i = 0; i < state->dts_nspeculations; i++) {
2549 dtrace_speculation_t *spec = &state->dts_speculations[i];
2550 dtrace_speculation_state_t current, new;
2552 if (!spec->dtsp_cleaning)
2555 current = spec->dtsp_state;
2556 ASSERT(current == DTRACESPEC_DISCARDING ||
2557 current == DTRACESPEC_COMMITTINGMANY);
2559 new = DTRACESPEC_INACTIVE;
2561 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2562 ASSERT(rv == current);
2563 spec->dtsp_cleaning = 0;
2568 * Called as part of a speculate() to get the speculative buffer associated
2569 * with a given speculation. Returns NULL if the specified speculation is not
2570 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2571 * the active CPU is not the specified CPU -- the speculation will be
2572 * atomically transitioned into the ACTIVEMANY state.
2574 static dtrace_buffer_t *
2575 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2576 dtrace_specid_t which)
2578 dtrace_speculation_t *spec;
2579 dtrace_speculation_state_t current, new = 0;
2580 dtrace_buffer_t *buf;
2585 if (which > state->dts_nspeculations) {
2586 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2590 spec = &state->dts_speculations[which - 1];
2591 buf = &spec->dtsp_buffer[cpuid];
2594 current = spec->dtsp_state;
2597 case DTRACESPEC_INACTIVE:
2598 case DTRACESPEC_COMMITTINGMANY:
2599 case DTRACESPEC_DISCARDING:
2602 case DTRACESPEC_COMMITTING:
2603 ASSERT(buf->dtb_offset == 0);
2606 case DTRACESPEC_ACTIVEONE:
2608 * This speculation is currently active on one CPU.
2609 * Check the offset in the buffer; if it's non-zero,
2610 * that CPU must be us (and we leave the state alone).
2611 * If it's zero, assume that we're starting on a new
2612 * CPU -- and change the state to indicate that the
2613 * speculation is active on more than one CPU.
2615 if (buf->dtb_offset != 0)
2618 new = DTRACESPEC_ACTIVEMANY;
2621 case DTRACESPEC_ACTIVEMANY:
2624 case DTRACESPEC_ACTIVE:
2625 new = DTRACESPEC_ACTIVEONE;
2631 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2632 current, new) != current);
2634 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2639 * Return a string. In the event that the user lacks the privilege to access
2640 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2641 * don't fail access checking.
2643 * dtrace_dif_variable() uses this routine as a helper for various
2644 * builtin values such as 'execname' and 'probefunc.'
2647 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2648 dtrace_mstate_t *mstate)
2650 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2655 * The easy case: this probe is allowed to read all of memory, so
2656 * we can just return this as a vanilla pointer.
2658 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2662 * This is the tougher case: we copy the string in question from
2663 * kernel memory into scratch memory and return it that way: this
2664 * ensures that we won't trip up when access checking tests the
2665 * BYREF return value.
2667 strsz = dtrace_strlen((char *)addr, size) + 1;
2669 if (mstate->dtms_scratch_ptr + strsz >
2670 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2671 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2675 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2677 ret = mstate->dtms_scratch_ptr;
2678 mstate->dtms_scratch_ptr += strsz;
2683 * Return a string from a memoy address which is known to have one or
2684 * more concatenated, individually zero terminated, sub-strings.
2685 * In the event that the user lacks the privilege to access
2686 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2687 * don't fail access checking.
2689 * dtrace_dif_variable() uses this routine as a helper for various
2690 * builtin values such as 'execargs'.
2693 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2694 dtrace_mstate_t *mstate)
2700 if (mstate->dtms_scratch_ptr + strsz >
2701 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2702 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2706 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2709 /* Replace sub-string termination characters with a space. */
2710 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2715 ret = mstate->dtms_scratch_ptr;
2716 mstate->dtms_scratch_ptr += strsz;
2721 * This function implements the DIF emulator's variable lookups. The emulator
2722 * passes a reserved variable identifier and optional built-in array index.
2725 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2729 * If we're accessing one of the uncached arguments, we'll turn this
2730 * into a reference in the args array.
2732 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2733 ndx = v - DIF_VAR_ARG0;
2739 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2740 if (ndx >= sizeof (mstate->dtms_arg) /
2741 sizeof (mstate->dtms_arg[0])) {
2742 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2743 dtrace_provider_t *pv;
2746 pv = mstate->dtms_probe->dtpr_provider;
2747 if (pv->dtpv_pops.dtps_getargval != NULL)
2748 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2749 mstate->dtms_probe->dtpr_id,
2750 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2752 val = dtrace_getarg(ndx, aframes);
2755 * This is regrettably required to keep the compiler
2756 * from tail-optimizing the call to dtrace_getarg().
2757 * The condition always evaluates to true, but the
2758 * compiler has no way of figuring that out a priori.
2759 * (None of this would be necessary if the compiler
2760 * could be relied upon to _always_ tail-optimize
2761 * the call to dtrace_getarg() -- but it can't.)
2763 if (mstate->dtms_probe != NULL)
2769 return (mstate->dtms_arg[ndx]);
2772 case DIF_VAR_UREGS: {
2775 if (!dtrace_priv_proc(state))
2778 if ((lwp = curthread->t_lwp) == NULL) {
2779 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2780 cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2784 return (dtrace_getreg(lwp->lwp_regs, ndx));
2789 case DIF_VAR_CURTHREAD:
2790 if (!dtrace_priv_kernel(state))
2792 return ((uint64_t)(uintptr_t)curthread);
2794 case DIF_VAR_TIMESTAMP:
2795 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2796 mstate->dtms_timestamp = dtrace_gethrtime();
2797 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2799 return (mstate->dtms_timestamp);
2801 case DIF_VAR_VTIMESTAMP:
2802 ASSERT(dtrace_vtime_references != 0);
2803 return (curthread->t_dtrace_vtime);
2805 case DIF_VAR_WALLTIMESTAMP:
2806 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2807 mstate->dtms_walltimestamp = dtrace_gethrestime();
2808 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2810 return (mstate->dtms_walltimestamp);
2814 if (!dtrace_priv_kernel(state))
2816 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2817 mstate->dtms_ipl = dtrace_getipl();
2818 mstate->dtms_present |= DTRACE_MSTATE_IPL;
2820 return (mstate->dtms_ipl);
2824 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2825 return (mstate->dtms_epid);
2828 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2829 return (mstate->dtms_probe->dtpr_id);
2831 case DIF_VAR_STACKDEPTH:
2832 if (!dtrace_priv_kernel(state))
2834 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2835 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2837 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2838 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2840 return (mstate->dtms_stackdepth);
2843 case DIF_VAR_USTACKDEPTH:
2844 if (!dtrace_priv_proc(state))
2846 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2848 * See comment in DIF_VAR_PID.
2850 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2852 mstate->dtms_ustackdepth = 0;
2854 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2855 mstate->dtms_ustackdepth =
2856 dtrace_getustackdepth();
2857 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2859 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2861 return (mstate->dtms_ustackdepth);
2864 case DIF_VAR_CALLER:
2865 if (!dtrace_priv_kernel(state))
2867 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2868 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2870 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2872 * If this is an unanchored probe, we are
2873 * required to go through the slow path:
2874 * dtrace_caller() only guarantees correct
2875 * results for anchored probes.
2877 pc_t caller[2] = {0, 0};
2879 dtrace_getpcstack(caller, 2, aframes,
2880 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2881 mstate->dtms_caller = caller[1];
2882 } else if ((mstate->dtms_caller =
2883 dtrace_caller(aframes)) == -1) {
2885 * We have failed to do this the quick way;
2886 * we must resort to the slower approach of
2887 * calling dtrace_getpcstack().
2891 dtrace_getpcstack(&caller, 1, aframes, NULL);
2892 mstate->dtms_caller = caller;
2895 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
2897 return (mstate->dtms_caller);
2900 case DIF_VAR_UCALLER:
2901 if (!dtrace_priv_proc(state))
2904 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
2908 * dtrace_getupcstack() fills in the first uint64_t
2909 * with the current PID. The second uint64_t will
2910 * be the program counter at user-level. The third
2911 * uint64_t will contain the caller, which is what
2915 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2916 dtrace_getupcstack(ustack, 3);
2917 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2918 mstate->dtms_ucaller = ustack[2];
2919 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
2922 return (mstate->dtms_ucaller);
2925 case DIF_VAR_PROBEPROV:
2926 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2927 return (dtrace_dif_varstr(
2928 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
2931 case DIF_VAR_PROBEMOD:
2932 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2933 return (dtrace_dif_varstr(
2934 (uintptr_t)mstate->dtms_probe->dtpr_mod,
2937 case DIF_VAR_PROBEFUNC:
2938 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2939 return (dtrace_dif_varstr(
2940 (uintptr_t)mstate->dtms_probe->dtpr_func,
2943 case DIF_VAR_PROBENAME:
2944 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2945 return (dtrace_dif_varstr(
2946 (uintptr_t)mstate->dtms_probe->dtpr_name,
2950 if (!dtrace_priv_proc(state))
2955 * Note that we are assuming that an unanchored probe is
2956 * always due to a high-level interrupt. (And we're assuming
2957 * that there is only a single high level interrupt.)
2959 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2960 return (pid0.pid_id);
2963 * It is always safe to dereference one's own t_procp pointer:
2964 * it always points to a valid, allocated proc structure.
2965 * Further, it is always safe to dereference the p_pidp member
2966 * of one's own proc structure. (These are truisms becuase
2967 * threads and processes don't clean up their own state --
2968 * they leave that task to whomever reaps them.)
2970 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
2972 return ((uint64_t)curproc->p_pid);
2976 if (!dtrace_priv_proc(state))
2981 * See comment in DIF_VAR_PID.
2983 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2984 return (pid0.pid_id);
2987 * It is always safe to dereference one's own t_procp pointer:
2988 * it always points to a valid, allocated proc structure.
2989 * (This is true because threads don't clean up their own
2990 * state -- they leave that task to whomever reaps them.)
2992 return ((uint64_t)curthread->t_procp->p_ppid);
2994 return ((uint64_t)curproc->p_pptr->p_pid);
3000 * See comment in DIF_VAR_PID.
3002 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3006 return ((uint64_t)curthread->t_tid);
3008 case DIF_VAR_EXECARGS: {
3009 struct pargs *p_args = curthread->td_proc->p_args;
3014 return (dtrace_dif_varstrz(
3015 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3018 case DIF_VAR_EXECNAME:
3020 if (!dtrace_priv_proc(state))
3024 * See comment in DIF_VAR_PID.
3026 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3027 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3030 * It is always safe to dereference one's own t_procp pointer:
3031 * it always points to a valid, allocated proc structure.
3032 * (This is true because threads don't clean up their own
3033 * state -- they leave that task to whomever reaps them.)
3035 return (dtrace_dif_varstr(
3036 (uintptr_t)curthread->t_procp->p_user.u_comm,
3039 return (dtrace_dif_varstr(
3040 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3043 case DIF_VAR_ZONENAME:
3045 if (!dtrace_priv_proc(state))
3049 * See comment in DIF_VAR_PID.
3051 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3052 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3055 * It is always safe to dereference one's own t_procp pointer:
3056 * it always points to a valid, allocated proc structure.
3057 * (This is true because threads don't clean up their own
3058 * state -- they leave that task to whomever reaps them.)
3060 return (dtrace_dif_varstr(
3061 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3068 if (!dtrace_priv_proc(state))
3073 * See comment in DIF_VAR_PID.
3075 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3076 return ((uint64_t)p0.p_cred->cr_uid);
3080 * It is always safe to dereference one's own t_procp pointer:
3081 * it always points to a valid, allocated proc structure.
3082 * (This is true because threads don't clean up their own
3083 * state -- they leave that task to whomever reaps them.)
3085 * Additionally, it is safe to dereference one's own process
3086 * credential, since this is never NULL after process birth.
3088 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3091 if (!dtrace_priv_proc(state))
3096 * See comment in DIF_VAR_PID.
3098 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3099 return ((uint64_t)p0.p_cred->cr_gid);
3103 * It is always safe to dereference one's own t_procp pointer:
3104 * it always points to a valid, allocated proc structure.
3105 * (This is true because threads don't clean up their own
3106 * state -- they leave that task to whomever reaps them.)
3108 * Additionally, it is safe to dereference one's own process
3109 * credential, since this is never NULL after process birth.
3111 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3113 case DIF_VAR_ERRNO: {
3116 if (!dtrace_priv_proc(state))
3120 * See comment in DIF_VAR_PID.
3122 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3126 * It is always safe to dereference one's own t_lwp pointer in
3127 * the event that this pointer is non-NULL. (This is true
3128 * because threads and lwps don't clean up their own state --
3129 * they leave that task to whomever reaps them.)
3131 if ((lwp = curthread->t_lwp) == NULL)
3134 return ((uint64_t)lwp->lwp_errno);
3136 return (curthread->td_errno);
3140 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3146 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3147 * Notice that we don't bother validating the proper number of arguments or
3148 * their types in the tuple stack. This isn't needed because all argument
3149 * interpretation is safe because of our load safety -- the worst that can
3150 * happen is that a bogus program can obtain bogus results.
3153 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3154 dtrace_key_t *tupregs, int nargs,
3155 dtrace_mstate_t *mstate, dtrace_state_t *state)
3157 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3158 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3159 dtrace_vstate_t *vstate = &state->dts_vstate;
3172 struct thread *lowner;
3174 struct lock_object *li;
3181 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3185 case DIF_SUBR_MUTEX_OWNED:
3186 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3192 m.mx = dtrace_load64(tupregs[0].dttk_value);
3193 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3194 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3196 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3199 case DIF_SUBR_MUTEX_OWNER:
3200 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3206 m.mx = dtrace_load64(tupregs[0].dttk_value);
3207 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3208 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3209 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3214 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3215 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3221 m.mx = dtrace_load64(tupregs[0].dttk_value);
3222 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3225 case DIF_SUBR_MUTEX_TYPE_SPIN:
3226 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3232 m.mx = dtrace_load64(tupregs[0].dttk_value);
3233 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3236 case DIF_SUBR_RW_READ_HELD: {
3239 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3245 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3246 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3250 case DIF_SUBR_RW_WRITE_HELD:
3251 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3257 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3258 regs[rd] = _RW_WRITE_HELD(&r.ri);
3261 case DIF_SUBR_RW_ISWRITER:
3262 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3268 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3269 regs[rd] = _RW_ISWRITER(&r.ri);
3273 case DIF_SUBR_MUTEX_OWNED:
3274 if (!dtrace_canload(tupregs[0].dttk_value,
3275 sizeof (struct lock_object), mstate, vstate)) {
3279 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3280 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3283 case DIF_SUBR_MUTEX_OWNER:
3284 if (!dtrace_canload(tupregs[0].dttk_value,
3285 sizeof (struct lock_object), mstate, vstate)) {
3289 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3290 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3291 regs[rd] = (uintptr_t)lowner;
3294 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3295 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3300 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3301 /* XXX - should be only LC_SLEEPABLE? */
3302 regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3303 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3306 case DIF_SUBR_MUTEX_TYPE_SPIN:
3307 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3312 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3313 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3316 case DIF_SUBR_RW_READ_HELD:
3317 case DIF_SUBR_SX_SHARED_HELD:
3318 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3323 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3324 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3328 case DIF_SUBR_RW_WRITE_HELD:
3329 case DIF_SUBR_SX_EXCLUSIVE_HELD:
3330 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3335 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3336 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3337 regs[rd] = (lowner == curthread);
3340 case DIF_SUBR_RW_ISWRITER:
3341 case DIF_SUBR_SX_ISEXCLUSIVE:
3342 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3347 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3348 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3351 #endif /* ! defined(sun) */
3353 case DIF_SUBR_BCOPY: {
3355 * We need to be sure that the destination is in the scratch
3356 * region -- no other region is allowed.
3358 uintptr_t src = tupregs[0].dttk_value;
3359 uintptr_t dest = tupregs[1].dttk_value;
3360 size_t size = tupregs[2].dttk_value;
3362 if (!dtrace_inscratch(dest, size, mstate)) {
3363 *flags |= CPU_DTRACE_BADADDR;
3368 if (!dtrace_canload(src, size, mstate, vstate)) {
3373 dtrace_bcopy((void *)src, (void *)dest, size);
3377 case DIF_SUBR_ALLOCA:
3378 case DIF_SUBR_COPYIN: {
3379 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3381 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3382 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3385 * This action doesn't require any credential checks since
3386 * probes will not activate in user contexts to which the
3387 * enabling user does not have permissions.
3391 * Rounding up the user allocation size could have overflowed
3392 * a large, bogus allocation (like -1ULL) to 0.
3394 if (scratch_size < size ||
3395 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3396 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3401 if (subr == DIF_SUBR_COPYIN) {
3402 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3403 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3404 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3407 mstate->dtms_scratch_ptr += scratch_size;
3412 case DIF_SUBR_COPYINTO: {
3413 uint64_t size = tupregs[1].dttk_value;
3414 uintptr_t dest = tupregs[2].dttk_value;
3417 * This action doesn't require any credential checks since
3418 * probes will not activate in user contexts to which the
3419 * enabling user does not have permissions.
3421 if (!dtrace_inscratch(dest, size, mstate)) {
3422 *flags |= CPU_DTRACE_BADADDR;
3427 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3428 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3429 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3433 case DIF_SUBR_COPYINSTR: {
3434 uintptr_t dest = mstate->dtms_scratch_ptr;
3435 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3437 if (nargs > 1 && tupregs[1].dttk_value < size)
3438 size = tupregs[1].dttk_value + 1;
3441 * This action doesn't require any credential checks since
3442 * probes will not activate in user contexts to which the
3443 * enabling user does not have permissions.
3445 if (!DTRACE_INSCRATCH(mstate, size)) {
3446 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3451 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3452 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3453 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3455 ((char *)dest)[size - 1] = '\0';
3456 mstate->dtms_scratch_ptr += size;
3462 case DIF_SUBR_MSGSIZE:
3463 case DIF_SUBR_MSGDSIZE: {
3464 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3465 uintptr_t wptr, rptr;
3469 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3471 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3477 wptr = dtrace_loadptr(baddr +
3478 offsetof(mblk_t, b_wptr));
3480 rptr = dtrace_loadptr(baddr +
3481 offsetof(mblk_t, b_rptr));
3484 *flags |= CPU_DTRACE_BADADDR;
3485 *illval = tupregs[0].dttk_value;
3489 daddr = dtrace_loadptr(baddr +
3490 offsetof(mblk_t, b_datap));
3492 baddr = dtrace_loadptr(baddr +
3493 offsetof(mblk_t, b_cont));
3496 * We want to prevent against denial-of-service here,
3497 * so we're only going to search the list for
3498 * dtrace_msgdsize_max mblks.
3500 if (cont++ > dtrace_msgdsize_max) {
3501 *flags |= CPU_DTRACE_ILLOP;
3505 if (subr == DIF_SUBR_MSGDSIZE) {
3506 if (dtrace_load8(daddr +
3507 offsetof(dblk_t, db_type)) != M_DATA)
3511 count += wptr - rptr;
3514 if (!(*flags & CPU_DTRACE_FAULT))
3521 case DIF_SUBR_PROGENYOF: {
3522 pid_t pid = tupregs[0].dttk_value;
3526 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3528 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3530 if (p->p_pidp->pid_id == pid) {
3532 if (p->p_pid == pid) {
3539 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3545 case DIF_SUBR_SPECULATION:
3546 regs[rd] = dtrace_speculation(state);
3549 case DIF_SUBR_COPYOUT: {
3550 uintptr_t kaddr = tupregs[0].dttk_value;
3551 uintptr_t uaddr = tupregs[1].dttk_value;
3552 uint64_t size = tupregs[2].dttk_value;
3554 if (!dtrace_destructive_disallow &&
3555 dtrace_priv_proc_control(state) &&
3556 !dtrace_istoxic(kaddr, size)) {
3557 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3558 dtrace_copyout(kaddr, uaddr, size, flags);
3559 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3564 case DIF_SUBR_COPYOUTSTR: {
3565 uintptr_t kaddr = tupregs[0].dttk_value;
3566 uintptr_t uaddr = tupregs[1].dttk_value;
3567 uint64_t size = tupregs[2].dttk_value;
3569 if (!dtrace_destructive_disallow &&
3570 dtrace_priv_proc_control(state) &&
3571 !dtrace_istoxic(kaddr, size)) {
3572 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3573 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3574 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3579 case DIF_SUBR_STRLEN: {
3581 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3582 sz = dtrace_strlen((char *)addr,
3583 state->dts_options[DTRACEOPT_STRSIZE]);
3585 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3595 case DIF_SUBR_STRCHR:
3596 case DIF_SUBR_STRRCHR: {
3598 * We're going to iterate over the string looking for the
3599 * specified character. We will iterate until we have reached
3600 * the string length or we have found the character. If this
3601 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3602 * of the specified character instead of the first.
3604 uintptr_t saddr = tupregs[0].dttk_value;
3605 uintptr_t addr = tupregs[0].dttk_value;
3606 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3607 char c, target = (char)tupregs[1].dttk_value;
3609 for (regs[rd] = 0; addr < limit; addr++) {
3610 if ((c = dtrace_load8(addr)) == target) {
3613 if (subr == DIF_SUBR_STRCHR)
3621 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3629 case DIF_SUBR_STRSTR:
3630 case DIF_SUBR_INDEX:
3631 case DIF_SUBR_RINDEX: {
3633 * We're going to iterate over the string looking for the
3634 * specified string. We will iterate until we have reached
3635 * the string length or we have found the string. (Yes, this
3636 * is done in the most naive way possible -- but considering
3637 * that the string we're searching for is likely to be
3638 * relatively short, the complexity of Rabin-Karp or similar
3639 * hardly seems merited.)
3641 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3642 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3643 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3644 size_t len = dtrace_strlen(addr, size);
3645 size_t sublen = dtrace_strlen(substr, size);
3646 char *limit = addr + len, *orig = addr;
3647 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3650 regs[rd] = notfound;
3652 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3657 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3664 * strstr() and index()/rindex() have similar semantics if
3665 * both strings are the empty string: strstr() returns a
3666 * pointer to the (empty) string, and index() and rindex()
3667 * both return index 0 (regardless of any position argument).
3669 if (sublen == 0 && len == 0) {
3670 if (subr == DIF_SUBR_STRSTR)
3671 regs[rd] = (uintptr_t)addr;
3677 if (subr != DIF_SUBR_STRSTR) {
3678 if (subr == DIF_SUBR_RINDEX) {
3685 * Both index() and rindex() take an optional position
3686 * argument that denotes the starting position.
3689 int64_t pos = (int64_t)tupregs[2].dttk_value;
3692 * If the position argument to index() is
3693 * negative, Perl implicitly clamps it at
3694 * zero. This semantic is a little surprising
3695 * given the special meaning of negative
3696 * positions to similar Perl functions like
3697 * substr(), but it appears to reflect a
3698 * notion that index() can start from a
3699 * negative index and increment its way up to
3700 * the string. Given this notion, Perl's
3701 * rindex() is at least self-consistent in
3702 * that it implicitly clamps positions greater
3703 * than the string length to be the string
3704 * length. Where Perl completely loses
3705 * coherence, however, is when the specified
3706 * substring is the empty string (""). In
3707 * this case, even if the position is
3708 * negative, rindex() returns 0 -- and even if
3709 * the position is greater than the length,
3710 * index() returns the string length. These
3711 * semantics violate the notion that index()
3712 * should never return a value less than the
3713 * specified position and that rindex() should
3714 * never return a value greater than the
3715 * specified position. (One assumes that
3716 * these semantics are artifacts of Perl's
3717 * implementation and not the results of
3718 * deliberate design -- it beggars belief that
3719 * even Larry Wall could desire such oddness.)
3720 * While in the abstract one would wish for
3721 * consistent position semantics across
3722 * substr(), index() and rindex() -- or at the
3723 * very least self-consistent position
3724 * semantics for index() and rindex() -- we
3725 * instead opt to keep with the extant Perl
3726 * semantics, in all their broken glory. (Do
3727 * we have more desire to maintain Perl's
3728 * semantics than Perl does? Probably.)
3730 if (subr == DIF_SUBR_RINDEX) {
3754 for (regs[rd] = notfound; addr != limit; addr += inc) {
3755 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3756 if (subr != DIF_SUBR_STRSTR) {
3758 * As D index() and rindex() are
3759 * modeled on Perl (and not on awk),
3760 * we return a zero-based (and not a
3761 * one-based) index. (For you Perl
3762 * weenies: no, we're not going to add
3763 * $[ -- and shouldn't you be at a con
3766 regs[rd] = (uintptr_t)(addr - orig);
3770 ASSERT(subr == DIF_SUBR_STRSTR);
3771 regs[rd] = (uintptr_t)addr;
3779 case DIF_SUBR_STRTOK: {
3780 uintptr_t addr = tupregs[0].dttk_value;
3781 uintptr_t tokaddr = tupregs[1].dttk_value;
3782 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3783 uintptr_t limit, toklimit = tokaddr + size;
3784 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
3785 char *dest = (char *)mstate->dtms_scratch_ptr;
3789 * Check both the token buffer and (later) the input buffer,
3790 * since both could be non-scratch addresses.
3792 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3797 if (!DTRACE_INSCRATCH(mstate, size)) {
3798 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3805 * If the address specified is NULL, we use our saved
3806 * strtok pointer from the mstate. Note that this
3807 * means that the saved strtok pointer is _only_
3808 * valid within multiple enablings of the same probe --
3809 * it behaves like an implicit clause-local variable.
3811 addr = mstate->dtms_strtok;
3814 * If the user-specified address is non-NULL we must
3815 * access check it. This is the only time we have
3816 * a chance to do so, since this address may reside
3817 * in the string table of this clause-- future calls
3818 * (when we fetch addr from mstate->dtms_strtok)
3819 * would fail this access check.
3821 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3828 * First, zero the token map, and then process the token
3829 * string -- setting a bit in the map for every character
3830 * found in the token string.
3832 for (i = 0; i < sizeof (tokmap); i++)
3835 for (; tokaddr < toklimit; tokaddr++) {
3836 if ((c = dtrace_load8(tokaddr)) == '\0')
3839 ASSERT((c >> 3) < sizeof (tokmap));
3840 tokmap[c >> 3] |= (1 << (c & 0x7));
3843 for (limit = addr + size; addr < limit; addr++) {
3845 * We're looking for a character that is _not_ contained
3846 * in the token string.
3848 if ((c = dtrace_load8(addr)) == '\0')
3851 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3857 * We reached the end of the string without finding
3858 * any character that was not in the token string.
3859 * We return NULL in this case, and we set the saved
3860 * address to NULL as well.
3863 mstate->dtms_strtok = 0;
3868 * From here on, we're copying into the destination string.
3870 for (i = 0; addr < limit && i < size - 1; addr++) {
3871 if ((c = dtrace_load8(addr)) == '\0')
3874 if (tokmap[c >> 3] & (1 << (c & 0x7)))
3883 regs[rd] = (uintptr_t)dest;
3884 mstate->dtms_scratch_ptr += size;
3885 mstate->dtms_strtok = addr;
3889 case DIF_SUBR_SUBSTR: {
3890 uintptr_t s = tupregs[0].dttk_value;
3891 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3892 char *d = (char *)mstate->dtms_scratch_ptr;
3893 int64_t index = (int64_t)tupregs[1].dttk_value;
3894 int64_t remaining = (int64_t)tupregs[2].dttk_value;
3895 size_t len = dtrace_strlen((char *)s, size);
3898 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3903 if (!DTRACE_INSCRATCH(mstate, size)) {
3904 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3910 remaining = (int64_t)size;
3915 if (index < 0 && index + remaining > 0) {
3921 if (index >= len || index < 0) {
3923 } else if (remaining < 0) {
3924 remaining += len - index;
3925 } else if (index + remaining > size) {
3926 remaining = size - index;
3929 for (i = 0; i < remaining; i++) {
3930 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
3936 mstate->dtms_scratch_ptr += size;
3937 regs[rd] = (uintptr_t)d;
3942 case DIF_SUBR_GETMAJOR:
3944 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
3946 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
3950 case DIF_SUBR_GETMINOR:
3952 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
3954 regs[rd] = tupregs[0].dttk_value & MAXMIN;
3958 case DIF_SUBR_DDI_PATHNAME: {
3960 * This one is a galactic mess. We are going to roughly
3961 * emulate ddi_pathname(), but it's made more complicated
3962 * by the fact that we (a) want to include the minor name and
3963 * (b) must proceed iteratively instead of recursively.
3965 uintptr_t dest = mstate->dtms_scratch_ptr;
3966 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3967 char *start = (char *)dest, *end = start + size - 1;
3968 uintptr_t daddr = tupregs[0].dttk_value;
3969 int64_t minor = (int64_t)tupregs[1].dttk_value;
3971 int i, len, depth = 0;
3974 * Due to all the pointer jumping we do and context we must
3975 * rely upon, we just mandate that the user must have kernel
3976 * read privileges to use this routine.
3978 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
3979 *flags |= CPU_DTRACE_KPRIV;
3984 if (!DTRACE_INSCRATCH(mstate, size)) {
3985 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3993 * We want to have a name for the minor. In order to do this,
3994 * we need to walk the minor list from the devinfo. We want
3995 * to be sure that we don't infinitely walk a circular list,
3996 * so we check for circularity by sending a scout pointer
3997 * ahead two elements for every element that we iterate over;
3998 * if the list is circular, these will ultimately point to the
3999 * same element. You may recognize this little trick as the
4000 * answer to a stupid interview question -- one that always
4001 * seems to be asked by those who had to have it laboriously
4002 * explained to them, and who can't even concisely describe
4003 * the conditions under which one would be forced to resort to
4004 * this technique. Needless to say, those conditions are
4005 * found here -- and probably only here. Is this the only use
4006 * of this infamous trick in shipping, production code? If it
4007 * isn't, it probably should be...
4010 uintptr_t maddr = dtrace_loadptr(daddr +
4011 offsetof(struct dev_info, devi_minor));
4013 uintptr_t next = offsetof(struct ddi_minor_data, next);
4014 uintptr_t name = offsetof(struct ddi_minor_data,
4015 d_minor) + offsetof(struct ddi_minor, name);
4016 uintptr_t dev = offsetof(struct ddi_minor_data,
4017 d_minor) + offsetof(struct ddi_minor, dev);
4021 scout = dtrace_loadptr(maddr + next);
4023 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4026 m = dtrace_load64(maddr + dev) & MAXMIN64;
4028 m = dtrace_load32(maddr + dev) & MAXMIN;
4031 maddr = dtrace_loadptr(maddr + next);
4036 scout = dtrace_loadptr(scout + next);
4041 scout = dtrace_loadptr(scout + next);
4046 if (scout == maddr) {
4047 *flags |= CPU_DTRACE_ILLOP;
4055 * We have the minor data. Now we need to
4056 * copy the minor's name into the end of the
4059 s = (char *)dtrace_loadptr(maddr + name);
4060 len = dtrace_strlen(s, size);
4062 if (*flags & CPU_DTRACE_FAULT)
4066 if ((end -= (len + 1)) < start)
4072 for (i = 1; i <= len; i++)
4073 end[i] = dtrace_load8((uintptr_t)s++);
4078 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4079 ddi_node_state_t devi_state;
4081 devi_state = dtrace_load32(daddr +
4082 offsetof(struct dev_info, devi_node_state));
4084 if (*flags & CPU_DTRACE_FAULT)
4087 if (devi_state >= DS_INITIALIZED) {
4088 s = (char *)dtrace_loadptr(daddr +
4089 offsetof(struct dev_info, devi_addr));
4090 len = dtrace_strlen(s, size);
4092 if (*flags & CPU_DTRACE_FAULT)
4096 if ((end -= (len + 1)) < start)
4102 for (i = 1; i <= len; i++)
4103 end[i] = dtrace_load8((uintptr_t)s++);
4107 * Now for the node name...
4109 s = (char *)dtrace_loadptr(daddr +
4110 offsetof(struct dev_info, devi_node_name));
4112 daddr = dtrace_loadptr(daddr +
4113 offsetof(struct dev_info, devi_parent));
4116 * If our parent is NULL (that is, if we're the root
4117 * node), we're going to use the special path
4123 len = dtrace_strlen(s, size);
4124 if (*flags & CPU_DTRACE_FAULT)
4127 if ((end -= (len + 1)) < start)
4130 for (i = 1; i <= len; i++)
4131 end[i] = dtrace_load8((uintptr_t)s++);
4134 if (depth++ > dtrace_devdepth_max) {
4135 *flags |= CPU_DTRACE_ILLOP;
4141 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4144 regs[rd] = (uintptr_t)end;
4145 mstate->dtms_scratch_ptr += size;
4152 case DIF_SUBR_STRJOIN: {
4153 char *d = (char *)mstate->dtms_scratch_ptr;
4154 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4155 uintptr_t s1 = tupregs[0].dttk_value;
4156 uintptr_t s2 = tupregs[1].dttk_value;
4159 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4160 !dtrace_strcanload(s2, size, mstate, vstate)) {
4165 if (!DTRACE_INSCRATCH(mstate, size)) {
4166 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4173 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4178 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4186 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4191 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4196 mstate->dtms_scratch_ptr += i;
4197 regs[rd] = (uintptr_t)d;
4203 case DIF_SUBR_LLTOSTR: {
4204 int64_t i = (int64_t)tupregs[0].dttk_value;
4205 int64_t val = i < 0 ? i * -1 : i;
4206 uint64_t size = 22; /* enough room for 2^64 in decimal */
4207 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4209 if (!DTRACE_INSCRATCH(mstate, size)) {
4210 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4215 for (*end-- = '\0'; val; val /= 10)
4216 *end-- = '0' + (val % 10);
4224 regs[rd] = (uintptr_t)end + 1;
4225 mstate->dtms_scratch_ptr += size;
4229 case DIF_SUBR_HTONS:
4230 case DIF_SUBR_NTOHS:
4231 #if BYTE_ORDER == BIG_ENDIAN
4232 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4234 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4239 case DIF_SUBR_HTONL:
4240 case DIF_SUBR_NTOHL:
4241 #if BYTE_ORDER == BIG_ENDIAN
4242 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4244 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4249 case DIF_SUBR_HTONLL:
4250 case DIF_SUBR_NTOHLL:
4251 #if BYTE_ORDER == BIG_ENDIAN
4252 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4254 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4259 case DIF_SUBR_DIRNAME:
4260 case DIF_SUBR_BASENAME: {
4261 char *dest = (char *)mstate->dtms_scratch_ptr;
4262 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4263 uintptr_t src = tupregs[0].dttk_value;
4264 int i, j, len = dtrace_strlen((char *)src, size);
4265 int lastbase = -1, firstbase = -1, lastdir = -1;
4268 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4273 if (!DTRACE_INSCRATCH(mstate, size)) {
4274 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4280 * The basename and dirname for a zero-length string is
4285 src = (uintptr_t)".";
4289 * Start from the back of the string, moving back toward the
4290 * front until we see a character that isn't a slash. That
4291 * character is the last character in the basename.
4293 for (i = len - 1; i >= 0; i--) {
4294 if (dtrace_load8(src + i) != '/')
4302 * Starting from the last character in the basename, move
4303 * towards the front until we find a slash. The character
4304 * that we processed immediately before that is the first
4305 * character in the basename.
4307 for (; i >= 0; i--) {
4308 if (dtrace_load8(src + i) == '/')
4316 * Now keep going until we find a non-slash character. That
4317 * character is the last character in the dirname.
4319 for (; i >= 0; i--) {
4320 if (dtrace_load8(src + i) != '/')
4327 ASSERT(!(lastbase == -1 && firstbase != -1));
4328 ASSERT(!(firstbase == -1 && lastdir != -1));
4330 if (lastbase == -1) {
4332 * We didn't find a non-slash character. We know that
4333 * the length is non-zero, so the whole string must be
4334 * slashes. In either the dirname or the basename
4335 * case, we return '/'.
4337 ASSERT(firstbase == -1);
4338 firstbase = lastbase = lastdir = 0;
4341 if (firstbase == -1) {
4343 * The entire string consists only of a basename
4344 * component. If we're looking for dirname, we need
4345 * to change our string to be just "."; if we're
4346 * looking for a basename, we'll just set the first
4347 * character of the basename to be 0.
4349 if (subr == DIF_SUBR_DIRNAME) {
4350 ASSERT(lastdir == -1);
4351 src = (uintptr_t)".";
4358 if (subr == DIF_SUBR_DIRNAME) {
4359 if (lastdir == -1) {
4361 * We know that we have a slash in the name --
4362 * or lastdir would be set to 0, above. And
4363 * because lastdir is -1, we know that this
4364 * slash must be the first character. (That
4365 * is, the full string must be of the form
4366 * "/basename".) In this case, the last
4367 * character of the directory name is 0.
4375 ASSERT(subr == DIF_SUBR_BASENAME);
4376 ASSERT(firstbase != -1 && lastbase != -1);
4381 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4382 dest[j] = dtrace_load8(src + i);
4385 regs[rd] = (uintptr_t)dest;
4386 mstate->dtms_scratch_ptr += size;
4390 case DIF_SUBR_CLEANPATH: {
4391 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4392 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4393 uintptr_t src = tupregs[0].dttk_value;
4396 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4401 if (!DTRACE_INSCRATCH(mstate, size)) {
4402 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4408 * Move forward, loading each character.
4411 c = dtrace_load8(src + i++);
4413 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4421 c = dtrace_load8(src + i++);
4425 * We have two slashes -- we can just advance
4426 * to the next character.
4433 * This is not "." and it's not ".." -- we can
4434 * just store the "/" and this character and
4442 c = dtrace_load8(src + i++);
4446 * This is a "/./" component. We're not going
4447 * to store anything in the destination buffer;
4448 * we're just going to go to the next component.
4455 * This is not ".." -- we can just store the
4456 * "/." and this character and continue
4465 c = dtrace_load8(src + i++);
4467 if (c != '/' && c != '\0') {
4469 * This is not ".." -- it's "..[mumble]".
4470 * We'll store the "/.." and this character
4471 * and continue processing.
4481 * This is "/../" or "/..\0". We need to back up
4482 * our destination pointer until we find a "/".
4485 while (j != 0 && dest[--j] != '/')
4490 } while (c != '\0');
4493 regs[rd] = (uintptr_t)dest;
4494 mstate->dtms_scratch_ptr += size;
4498 case DIF_SUBR_INET_NTOA:
4499 case DIF_SUBR_INET_NTOA6:
4500 case DIF_SUBR_INET_NTOP: {
4505 if (subr == DIF_SUBR_INET_NTOP) {
4506 af = (int)tupregs[0].dttk_value;
4509 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4513 if (af == AF_INET) {
4518 * Safely load the IPv4 address.
4520 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4523 * Check an IPv4 string will fit in scratch.
4525 size = INET_ADDRSTRLEN;
4526 if (!DTRACE_INSCRATCH(mstate, size)) {
4527 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4531 base = (char *)mstate->dtms_scratch_ptr;
4532 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4535 * Stringify as a dotted decimal quad.
4538 ptr8 = (uint8_t *)&ip4;
4539 for (i = 3; i >= 0; i--) {
4545 for (; val; val /= 10) {
4546 *end-- = '0' + (val % 10);
4553 ASSERT(end + 1 >= base);
4555 } else if (af == AF_INET6) {
4556 struct in6_addr ip6;
4557 int firstzero, tryzero, numzero, v6end;
4559 const char digits[] = "0123456789abcdef";
4562 * Stringify using RFC 1884 convention 2 - 16 bit
4563 * hexadecimal values with a zero-run compression.
4564 * Lower case hexadecimal digits are used.
4565 * eg, fe80::214:4fff:fe0b:76c8.
4566 * The IPv4 embedded form is returned for inet_ntop,
4567 * just the IPv4 string is returned for inet_ntoa6.
4571 * Safely load the IPv6 address.
4574 (void *)(uintptr_t)tupregs[argi].dttk_value,
4575 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4578 * Check an IPv6 string will fit in scratch.
4580 size = INET6_ADDRSTRLEN;
4581 if (!DTRACE_INSCRATCH(mstate, size)) {
4582 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4586 base = (char *)mstate->dtms_scratch_ptr;
4587 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4591 * Find the longest run of 16 bit zero values
4592 * for the single allowed zero compression - "::".
4597 for (i = 0; i < sizeof (struct in6_addr); i++) {
4599 if (ip6._S6_un._S6_u8[i] == 0 &&
4601 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4603 tryzero == -1 && i % 2 == 0) {
4608 if (tryzero != -1 &&
4610 (ip6._S6_un._S6_u8[i] != 0 ||
4612 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4614 i == sizeof (struct in6_addr) - 1)) {
4616 if (i - tryzero <= numzero) {
4621 firstzero = tryzero;
4622 numzero = i - i % 2 - tryzero;
4626 if (ip6._S6_un._S6_u8[i] == 0 &&
4628 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4630 i == sizeof (struct in6_addr) - 1)
4634 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4637 * Check for an IPv4 embedded address.
4639 v6end = sizeof (struct in6_addr) - 2;
4640 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4641 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4642 for (i = sizeof (struct in6_addr) - 1;
4643 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4644 ASSERT(end >= base);
4647 val = ip6._S6_un._S6_u8[i];
4649 val = ip6.__u6_addr.__u6_addr8[i];
4655 for (; val; val /= 10) {
4656 *end-- = '0' + val % 10;
4660 if (i > DTRACE_V4MAPPED_OFFSET)
4664 if (subr == DIF_SUBR_INET_NTOA6)
4668 * Set v6end to skip the IPv4 address that
4669 * we have already stringified.
4675 * Build the IPv6 string by working through the
4676 * address in reverse.
4678 for (i = v6end; i >= 0; i -= 2) {
4679 ASSERT(end >= base);
4681 if (i == firstzero + numzero - 2) {
4688 if (i < 14 && i != firstzero - 2)
4692 val = (ip6._S6_un._S6_u8[i] << 8) +
4693 ip6._S6_un._S6_u8[i + 1];
4695 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4696 ip6.__u6_addr.__u6_addr8[i + 1];
4702 for (; val; val /= 16) {
4703 *end-- = digits[val % 16];
4707 ASSERT(end + 1 >= base);
4711 * The user didn't use AH_INET or AH_INET6.
4713 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4718 inetout: regs[rd] = (uintptr_t)end + 1;
4719 mstate->dtms_scratch_ptr += size;
4723 case DIF_SUBR_MEMREF: {
4724 uintptr_t size = 2 * sizeof(uintptr_t);
4725 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4726 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4728 /* address and length */
4729 memref[0] = tupregs[0].dttk_value;
4730 memref[1] = tupregs[1].dttk_value;
4732 regs[rd] = (uintptr_t) memref;
4733 mstate->dtms_scratch_ptr += scratch_size;
4737 case DIF_SUBR_TYPEREF: {
4738 uintptr_t size = 4 * sizeof(uintptr_t);
4739 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4740 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4742 /* address, num_elements, type_str, type_len */
4743 typeref[0] = tupregs[0].dttk_value;
4744 typeref[1] = tupregs[1].dttk_value;
4745 typeref[2] = tupregs[2].dttk_value;
4746 typeref[3] = tupregs[3].dttk_value;
4748 regs[rd] = (uintptr_t) typeref;
4749 mstate->dtms_scratch_ptr += scratch_size;
4756 * Emulate the execution of DTrace IR instructions specified by the given
4757 * DIF object. This function is deliberately void of assertions as all of
4758 * the necessary checks are handled by a call to dtrace_difo_validate().
4761 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4762 dtrace_vstate_t *vstate, dtrace_state_t *state)
4764 const dif_instr_t *text = difo->dtdo_buf;
4765 const uint_t textlen = difo->dtdo_len;
4766 const char *strtab = difo->dtdo_strtab;
4767 const uint64_t *inttab = difo->dtdo_inttab;
4770 dtrace_statvar_t *svar;
4771 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4773 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4774 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4776 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4777 uint64_t regs[DIF_DIR_NREGS];
4780 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4782 uint_t pc = 0, id, opc = 0;
4788 * We stash the current DIF object into the machine state: we need it
4789 * for subsequent access checking.
4791 mstate->dtms_difo = difo;
4793 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
4795 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4799 r1 = DIF_INSTR_R1(instr);
4800 r2 = DIF_INSTR_R2(instr);
4801 rd = DIF_INSTR_RD(instr);
4803 switch (DIF_INSTR_OP(instr)) {
4805 regs[rd] = regs[r1] | regs[r2];
4808 regs[rd] = regs[r1] ^ regs[r2];
4811 regs[rd] = regs[r1] & regs[r2];
4814 regs[rd] = regs[r1] << regs[r2];
4817 regs[rd] = regs[r1] >> regs[r2];
4820 regs[rd] = regs[r1] - regs[r2];
4823 regs[rd] = regs[r1] + regs[r2];
4826 regs[rd] = regs[r1] * regs[r2];
4829 if (regs[r2] == 0) {
4831 *flags |= CPU_DTRACE_DIVZERO;
4833 regs[rd] = (int64_t)regs[r1] /
4839 if (regs[r2] == 0) {
4841 *flags |= CPU_DTRACE_DIVZERO;
4843 regs[rd] = regs[r1] / regs[r2];
4848 if (regs[r2] == 0) {
4850 *flags |= CPU_DTRACE_DIVZERO;
4852 regs[rd] = (int64_t)regs[r1] %
4858 if (regs[r2] == 0) {
4860 *flags |= CPU_DTRACE_DIVZERO;
4862 regs[rd] = regs[r1] % regs[r2];
4867 regs[rd] = ~regs[r1];
4870 regs[rd] = regs[r1];
4873 cc_r = regs[r1] - regs[r2];
4877 cc_c = regs[r1] < regs[r2];
4880 cc_n = cc_v = cc_c = 0;
4881 cc_z = regs[r1] == 0;
4884 pc = DIF_INSTR_LABEL(instr);
4888 pc = DIF_INSTR_LABEL(instr);
4892 pc = DIF_INSTR_LABEL(instr);
4895 if ((cc_z | (cc_n ^ cc_v)) == 0)
4896 pc = DIF_INSTR_LABEL(instr);
4899 if ((cc_c | cc_z) == 0)
4900 pc = DIF_INSTR_LABEL(instr);
4903 if ((cc_n ^ cc_v) == 0)
4904 pc = DIF_INSTR_LABEL(instr);
4908 pc = DIF_INSTR_LABEL(instr);
4912 pc = DIF_INSTR_LABEL(instr);
4916 pc = DIF_INSTR_LABEL(instr);
4919 if (cc_z | (cc_n ^ cc_v))
4920 pc = DIF_INSTR_LABEL(instr);
4924 pc = DIF_INSTR_LABEL(instr);
4927 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4928 *flags |= CPU_DTRACE_KPRIV;
4934 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
4937 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4938 *flags |= CPU_DTRACE_KPRIV;
4944 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
4947 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4948 *flags |= CPU_DTRACE_KPRIV;
4954 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
4957 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4958 *flags |= CPU_DTRACE_KPRIV;
4964 regs[rd] = dtrace_load8(regs[r1]);
4967 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4968 *flags |= CPU_DTRACE_KPRIV;
4974 regs[rd] = dtrace_load16(regs[r1]);
4977 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4978 *flags |= CPU_DTRACE_KPRIV;
4984 regs[rd] = dtrace_load32(regs[r1]);
4987 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
4988 *flags |= CPU_DTRACE_KPRIV;
4994 regs[rd] = dtrace_load64(regs[r1]);
4998 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5001 regs[rd] = (int16_t)
5002 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5005 regs[rd] = (int32_t)
5006 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5010 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5014 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5018 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5022 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5031 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5034 regs[rd] = (uint64_t)(uintptr_t)
5035 (strtab + DIF_INSTR_STRING(instr));
5038 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5039 uintptr_t s1 = regs[r1];
5040 uintptr_t s2 = regs[r2];
5043 !dtrace_strcanload(s1, sz, mstate, vstate))
5046 !dtrace_strcanload(s2, sz, mstate, vstate))
5049 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5057 regs[rd] = dtrace_dif_variable(mstate, state,
5061 id = DIF_INSTR_VAR(instr);
5063 if (id >= DIF_VAR_OTHER_UBASE) {
5066 id -= DIF_VAR_OTHER_UBASE;
5067 svar = vstate->dtvs_globals[id];
5068 ASSERT(svar != NULL);
5069 v = &svar->dtsv_var;
5071 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5072 regs[rd] = svar->dtsv_data;
5076 a = (uintptr_t)svar->dtsv_data;
5078 if (*(uint8_t *)a == UINT8_MAX) {
5080 * If the 0th byte is set to UINT8_MAX
5081 * then this is to be treated as a
5082 * reference to a NULL variable.
5086 regs[rd] = a + sizeof (uint64_t);
5092 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5096 id = DIF_INSTR_VAR(instr);
5098 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5099 id -= DIF_VAR_OTHER_UBASE;
5101 svar = vstate->dtvs_globals[id];
5102 ASSERT(svar != NULL);
5103 v = &svar->dtsv_var;
5105 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5106 uintptr_t a = (uintptr_t)svar->dtsv_data;
5109 ASSERT(svar->dtsv_size != 0);
5111 if (regs[rd] == 0) {
5112 *(uint8_t *)a = UINT8_MAX;
5116 a += sizeof (uint64_t);
5118 if (!dtrace_vcanload(
5119 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5123 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5124 (void *)a, &v->dtdv_type);
5128 svar->dtsv_data = regs[rd];
5133 * There are no DTrace built-in thread-local arrays at
5134 * present. This opcode is saved for future work.
5136 *flags |= CPU_DTRACE_ILLOP;
5141 id = DIF_INSTR_VAR(instr);
5143 if (id < DIF_VAR_OTHER_UBASE) {
5145 * For now, this has no meaning.
5151 id -= DIF_VAR_OTHER_UBASE;
5153 ASSERT(id < vstate->dtvs_nlocals);
5154 ASSERT(vstate->dtvs_locals != NULL);
5156 svar = vstate->dtvs_locals[id];
5157 ASSERT(svar != NULL);
5158 v = &svar->dtsv_var;
5160 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5161 uintptr_t a = (uintptr_t)svar->dtsv_data;
5162 size_t sz = v->dtdv_type.dtdt_size;
5164 sz += sizeof (uint64_t);
5165 ASSERT(svar->dtsv_size == NCPU * sz);
5168 if (*(uint8_t *)a == UINT8_MAX) {
5170 * If the 0th byte is set to UINT8_MAX
5171 * then this is to be treated as a
5172 * reference to a NULL variable.
5176 regs[rd] = a + sizeof (uint64_t);
5182 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5183 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5184 regs[rd] = tmp[curcpu];
5188 id = DIF_INSTR_VAR(instr);
5190 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5191 id -= DIF_VAR_OTHER_UBASE;
5192 ASSERT(id < vstate->dtvs_nlocals);
5194 ASSERT(vstate->dtvs_locals != NULL);
5195 svar = vstate->dtvs_locals[id];
5196 ASSERT(svar != NULL);
5197 v = &svar->dtsv_var;
5199 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5200 uintptr_t a = (uintptr_t)svar->dtsv_data;
5201 size_t sz = v->dtdv_type.dtdt_size;
5203 sz += sizeof (uint64_t);
5204 ASSERT(svar->dtsv_size == NCPU * sz);
5207 if (regs[rd] == 0) {
5208 *(uint8_t *)a = UINT8_MAX;
5212 a += sizeof (uint64_t);
5215 if (!dtrace_vcanload(
5216 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5220 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5221 (void *)a, &v->dtdv_type);
5225 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5226 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5227 tmp[curcpu] = regs[rd];
5231 dtrace_dynvar_t *dvar;
5234 id = DIF_INSTR_VAR(instr);
5235 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5236 id -= DIF_VAR_OTHER_UBASE;
5237 v = &vstate->dtvs_tlocals[id];
5239 key = &tupregs[DIF_DTR_NREGS];
5240 key[0].dttk_value = (uint64_t)id;
5241 key[0].dttk_size = 0;
5242 DTRACE_TLS_THRKEY(key[1].dttk_value);
5243 key[1].dttk_size = 0;
5245 dvar = dtrace_dynvar(dstate, 2, key,
5246 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5254 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5255 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5257 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5264 dtrace_dynvar_t *dvar;
5267 id = DIF_INSTR_VAR(instr);
5268 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5269 id -= DIF_VAR_OTHER_UBASE;
5271 key = &tupregs[DIF_DTR_NREGS];
5272 key[0].dttk_value = (uint64_t)id;
5273 key[0].dttk_size = 0;
5274 DTRACE_TLS_THRKEY(key[1].dttk_value);
5275 key[1].dttk_size = 0;
5276 v = &vstate->dtvs_tlocals[id];
5278 dvar = dtrace_dynvar(dstate, 2, key,
5279 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5280 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5281 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5282 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5285 * Given that we're storing to thread-local data,
5286 * we need to flush our predicate cache.
5288 curthread->t_predcache = 0;
5293 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5294 if (!dtrace_vcanload(
5295 (void *)(uintptr_t)regs[rd],
5296 &v->dtdv_type, mstate, vstate))
5299 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5300 dvar->dtdv_data, &v->dtdv_type);
5302 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5309 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5313 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5314 regs, tupregs, ttop, mstate, state);
5318 if (ttop == DIF_DTR_NREGS) {
5319 *flags |= CPU_DTRACE_TUPOFLOW;
5323 if (r1 == DIF_TYPE_STRING) {
5325 * If this is a string type and the size is 0,
5326 * we'll use the system-wide default string
5327 * size. Note that we are _not_ looking at
5328 * the value of the DTRACEOPT_STRSIZE option;
5329 * had this been set, we would expect to have
5330 * a non-zero size value in the "pushtr".
5332 tupregs[ttop].dttk_size =
5333 dtrace_strlen((char *)(uintptr_t)regs[rd],
5334 regs[r2] ? regs[r2] :
5335 dtrace_strsize_default) + 1;
5337 tupregs[ttop].dttk_size = regs[r2];
5340 tupregs[ttop++].dttk_value = regs[rd];
5344 if (ttop == DIF_DTR_NREGS) {
5345 *flags |= CPU_DTRACE_TUPOFLOW;
5349 tupregs[ttop].dttk_value = regs[rd];
5350 tupregs[ttop++].dttk_size = 0;
5358 case DIF_OP_FLUSHTS:
5363 case DIF_OP_LDTAA: {
5364 dtrace_dynvar_t *dvar;
5365 dtrace_key_t *key = tupregs;
5366 uint_t nkeys = ttop;
5368 id = DIF_INSTR_VAR(instr);
5369 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5370 id -= DIF_VAR_OTHER_UBASE;
5372 key[nkeys].dttk_value = (uint64_t)id;
5373 key[nkeys++].dttk_size = 0;
5375 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5376 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5377 key[nkeys++].dttk_size = 0;
5378 v = &vstate->dtvs_tlocals[id];
5380 v = &vstate->dtvs_globals[id]->dtsv_var;
5383 dvar = dtrace_dynvar(dstate, nkeys, key,
5384 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5385 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5386 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5393 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5394 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5396 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5403 case DIF_OP_STTAA: {
5404 dtrace_dynvar_t *dvar;
5405 dtrace_key_t *key = tupregs;
5406 uint_t nkeys = ttop;
5408 id = DIF_INSTR_VAR(instr);
5409 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5410 id -= DIF_VAR_OTHER_UBASE;
5412 key[nkeys].dttk_value = (uint64_t)id;
5413 key[nkeys++].dttk_size = 0;
5415 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5416 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5417 key[nkeys++].dttk_size = 0;
5418 v = &vstate->dtvs_tlocals[id];
5420 v = &vstate->dtvs_globals[id]->dtsv_var;
5423 dvar = dtrace_dynvar(dstate, nkeys, key,
5424 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5425 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5426 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5427 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5432 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5433 if (!dtrace_vcanload(
5434 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5438 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5439 dvar->dtdv_data, &v->dtdv_type);
5441 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5447 case DIF_OP_ALLOCS: {
5448 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5449 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5452 * Rounding up the user allocation size could have
5453 * overflowed large, bogus allocations (like -1ULL) to
5456 if (size < regs[r1] ||
5457 !DTRACE_INSCRATCH(mstate, size)) {
5458 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5463 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5464 mstate->dtms_scratch_ptr += size;
5470 if (!dtrace_canstore(regs[rd], regs[r2],
5472 *flags |= CPU_DTRACE_BADADDR;
5477 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5480 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5481 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5485 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5486 *flags |= CPU_DTRACE_BADADDR;
5490 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5494 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5495 *flags |= CPU_DTRACE_BADADDR;
5500 *flags |= CPU_DTRACE_BADALIGN;
5504 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5508 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5509 *flags |= CPU_DTRACE_BADADDR;
5514 *flags |= CPU_DTRACE_BADALIGN;
5518 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5522 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5523 *flags |= CPU_DTRACE_BADADDR;
5528 *flags |= CPU_DTRACE_BADALIGN;
5532 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5537 if (!(*flags & CPU_DTRACE_FAULT))
5540 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5541 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5547 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5549 dtrace_probe_t *probe = ecb->dte_probe;
5550 dtrace_provider_t *prov = probe->dtpr_provider;
5551 char c[DTRACE_FULLNAMELEN + 80], *str;
5552 char *msg = "dtrace: breakpoint action at probe ";
5553 char *ecbmsg = " (ecb ";
5554 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5555 uintptr_t val = (uintptr_t)ecb;
5556 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5558 if (dtrace_destructive_disallow)
5562 * It's impossible to be taking action on the NULL probe.
5564 ASSERT(probe != NULL);
5567 * This is a poor man's (destitute man's?) sprintf(): we want to
5568 * print the provider name, module name, function name and name of
5569 * the probe, along with the hex address of the ECB with the breakpoint
5570 * action -- all of which we must place in the character buffer by
5573 while (*msg != '\0')
5576 for (str = prov->dtpv_name; *str != '\0'; str++)
5580 for (str = probe->dtpr_mod; *str != '\0'; str++)
5584 for (str = probe->dtpr_func; *str != '\0'; str++)
5588 for (str = probe->dtpr_name; *str != '\0'; str++)
5591 while (*ecbmsg != '\0')
5594 while (shift >= 0) {
5595 mask = (uintptr_t)0xf << shift;
5597 if (val >= ((uintptr_t)1 << shift))
5598 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5608 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5613 dtrace_action_panic(dtrace_ecb_t *ecb)
5615 dtrace_probe_t *probe = ecb->dte_probe;
5618 * It's impossible to be taking action on the NULL probe.
5620 ASSERT(probe != NULL);
5622 if (dtrace_destructive_disallow)
5625 if (dtrace_panicked != NULL)
5628 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5632 * We won the right to panic. (We want to be sure that only one
5633 * thread calls panic() from dtrace_probe(), and that panic() is
5634 * called exactly once.)
5636 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5637 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5638 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5642 dtrace_action_raise(uint64_t sig)
5644 if (dtrace_destructive_disallow)
5648 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5654 * raise() has a queue depth of 1 -- we ignore all subsequent
5655 * invocations of the raise() action.
5657 if (curthread->t_dtrace_sig == 0)
5658 curthread->t_dtrace_sig = (uint8_t)sig;
5660 curthread->t_sig_check = 1;
5663 struct proc *p = curproc;
5671 dtrace_action_stop(void)
5673 if (dtrace_destructive_disallow)
5677 if (!curthread->t_dtrace_stop) {
5678 curthread->t_dtrace_stop = 1;
5679 curthread->t_sig_check = 1;
5683 struct proc *p = curproc;
5685 psignal(p, SIGSTOP);
5691 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5694 volatile uint16_t *flags;
5698 cpu_t *cpu = &solaris_cpu[curcpu];
5701 if (dtrace_destructive_disallow)
5704 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5706 now = dtrace_gethrtime();
5708 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5710 * We need to advance the mark to the current time.
5712 cpu->cpu_dtrace_chillmark = now;
5713 cpu->cpu_dtrace_chilled = 0;
5717 * Now check to see if the requested chill time would take us over
5718 * the maximum amount of time allowed in the chill interval. (Or
5719 * worse, if the calculation itself induces overflow.)
5721 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5722 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5723 *flags |= CPU_DTRACE_ILLOP;
5727 while (dtrace_gethrtime() - now < val)
5731 * Normally, we assure that the value of the variable "timestamp" does
5732 * not change within an ECB. The presence of chill() represents an
5733 * exception to this rule, however.
5735 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5736 cpu->cpu_dtrace_chilled += val;
5741 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5742 uint64_t *buf, uint64_t arg)
5744 int nframes = DTRACE_USTACK_NFRAMES(arg);
5745 int strsize = DTRACE_USTACK_STRSIZE(arg);
5746 uint64_t *pcs = &buf[1], *fps;
5747 char *str = (char *)&pcs[nframes];
5748 int size, offs = 0, i, j;
5749 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5750 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5754 * Should be taking a faster path if string space has not been
5757 ASSERT(strsize != 0);
5760 * We will first allocate some temporary space for the frame pointers.
5762 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5763 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5764 (nframes * sizeof (uint64_t));
5766 if (!DTRACE_INSCRATCH(mstate, size)) {
5768 * Not enough room for our frame pointers -- need to indicate
5769 * that we ran out of scratch space.
5771 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5775 mstate->dtms_scratch_ptr += size;
5776 saved = mstate->dtms_scratch_ptr;
5779 * Now get a stack with both program counters and frame pointers.
5781 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5782 dtrace_getufpstack(buf, fps, nframes + 1);
5783 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5786 * If that faulted, we're cooked.
5788 if (*flags & CPU_DTRACE_FAULT)
5792 * Now we want to walk up the stack, calling the USTACK helper. For
5793 * each iteration, we restore the scratch pointer.
5795 for (i = 0; i < nframes; i++) {
5796 mstate->dtms_scratch_ptr = saved;
5798 if (offs >= strsize)
5801 sym = (char *)(uintptr_t)dtrace_helper(
5802 DTRACE_HELPER_ACTION_USTACK,
5803 mstate, state, pcs[i], fps[i]);
5806 * If we faulted while running the helper, we're going to
5807 * clear the fault and null out the corresponding string.
5809 if (*flags & CPU_DTRACE_FAULT) {
5810 *flags &= ~CPU_DTRACE_FAULT;
5820 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5823 * Now copy in the string that the helper returned to us.
5825 for (j = 0; offs + j < strsize; j++) {
5826 if ((str[offs + j] = sym[j]) == '\0')
5830 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5835 if (offs >= strsize) {
5837 * If we didn't have room for all of the strings, we don't
5838 * abort processing -- this needn't be a fatal error -- but we
5839 * still want to increment a counter (dts_stkstroverflows) to
5840 * allow this condition to be warned about. (If this is from
5841 * a jstack() action, it is easily tuned via jstackstrsize.)
5843 dtrace_error(&state->dts_stkstroverflows);
5846 while (offs < strsize)
5850 mstate->dtms_scratch_ptr = old;
5855 * If you're looking for the epicenter of DTrace, you just found it. This
5856 * is the function called by the provider to fire a probe -- from which all
5857 * subsequent probe-context DTrace activity emanates.
5860 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5861 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5863 processorid_t cpuid;
5864 dtrace_icookie_t cookie;
5865 dtrace_probe_t *probe;
5866 dtrace_mstate_t mstate;
5868 dtrace_action_t *act;
5872 volatile uint16_t *flags;
5877 * Kick out immediately if this CPU is still being born (in which case
5878 * curthread will be set to -1) or the current thread can't allow
5879 * probes in its current context.
5881 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
5885 cookie = dtrace_interrupt_disable();
5886 probe = dtrace_probes[id - 1];
5888 onintr = CPU_ON_INTR(CPU);
5890 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
5891 probe->dtpr_predcache == curthread->t_predcache) {
5893 * We have hit in the predicate cache; we know that
5894 * this predicate would evaluate to be false.
5896 dtrace_interrupt_enable(cookie);
5901 if (panic_quiesce) {
5903 if (panicstr != NULL) {
5906 * We don't trace anything if we're panicking.
5908 dtrace_interrupt_enable(cookie);
5912 now = dtrace_gethrtime();
5913 vtime = dtrace_vtime_references != 0;
5915 if (vtime && curthread->t_dtrace_start)
5916 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
5918 mstate.dtms_difo = NULL;
5919 mstate.dtms_probe = probe;
5920 mstate.dtms_strtok = 0;
5921 mstate.dtms_arg[0] = arg0;
5922 mstate.dtms_arg[1] = arg1;
5923 mstate.dtms_arg[2] = arg2;
5924 mstate.dtms_arg[3] = arg3;
5925 mstate.dtms_arg[4] = arg4;
5927 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
5929 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
5930 dtrace_predicate_t *pred = ecb->dte_predicate;
5931 dtrace_state_t *state = ecb->dte_state;
5932 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
5933 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
5934 dtrace_vstate_t *vstate = &state->dts_vstate;
5935 dtrace_provider_t *prov = probe->dtpr_provider;
5940 * A little subtlety with the following (seemingly innocuous)
5941 * declaration of the automatic 'val': by looking at the
5942 * code, you might think that it could be declared in the
5943 * action processing loop, below. (That is, it's only used in
5944 * the action processing loop.) However, it must be declared
5945 * out of that scope because in the case of DIF expression
5946 * arguments to aggregating actions, one iteration of the
5947 * action loop will use the last iteration's value.
5951 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
5952 *flags &= ~CPU_DTRACE_ERROR;
5954 if (prov == dtrace_provider) {
5956 * If dtrace itself is the provider of this probe,
5957 * we're only going to continue processing the ECB if
5958 * arg0 (the dtrace_state_t) is equal to the ECB's
5959 * creating state. (This prevents disjoint consumers
5960 * from seeing one another's metaprobes.)
5962 if (arg0 != (uint64_t)(uintptr_t)state)
5966 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
5968 * We're not currently active. If our provider isn't
5969 * the dtrace pseudo provider, we're not interested.
5971 if (prov != dtrace_provider)
5975 * Now we must further check if we are in the BEGIN
5976 * probe. If we are, we will only continue processing
5977 * if we're still in WARMUP -- if one BEGIN enabling
5978 * has invoked the exit() action, we don't want to
5979 * evaluate subsequent BEGIN enablings.
5981 if (probe->dtpr_id == dtrace_probeid_begin &&
5982 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
5983 ASSERT(state->dts_activity ==
5984 DTRACE_ACTIVITY_DRAINING);
5989 if (ecb->dte_cond) {
5991 * If the dte_cond bits indicate that this
5992 * consumer is only allowed to see user-mode firings
5993 * of this probe, call the provider's dtps_usermode()
5994 * entry point to check that the probe was fired
5995 * while in a user context. Skip this ECB if that's
5998 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
5999 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6000 probe->dtpr_id, probe->dtpr_arg) == 0)
6005 * This is more subtle than it looks. We have to be
6006 * absolutely certain that CRED() isn't going to
6007 * change out from under us so it's only legit to
6008 * examine that structure if we're in constrained
6009 * situations. Currently, the only times we'll this
6010 * check is if a non-super-user has enabled the
6011 * profile or syscall providers -- providers that
6012 * allow visibility of all processes. For the
6013 * profile case, the check above will ensure that
6014 * we're examining a user context.
6016 if (ecb->dte_cond & DTRACE_COND_OWNER) {
6019 ecb->dte_state->dts_cred.dcr_cred;
6022 ASSERT(s_cr != NULL);
6024 if ((cr = CRED()) == NULL ||
6025 s_cr->cr_uid != cr->cr_uid ||
6026 s_cr->cr_uid != cr->cr_ruid ||
6027 s_cr->cr_uid != cr->cr_suid ||
6028 s_cr->cr_gid != cr->cr_gid ||
6029 s_cr->cr_gid != cr->cr_rgid ||
6030 s_cr->cr_gid != cr->cr_sgid ||
6031 (proc = ttoproc(curthread)) == NULL ||
6032 (proc->p_flag & SNOCD))
6036 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6039 ecb->dte_state->dts_cred.dcr_cred;
6041 ASSERT(s_cr != NULL);
6043 if ((cr = CRED()) == NULL ||
6044 s_cr->cr_zone->zone_id !=
6045 cr->cr_zone->zone_id)
6051 if (now - state->dts_alive > dtrace_deadman_timeout) {
6053 * We seem to be dead. Unless we (a) have kernel
6054 * destructive permissions (b) have expicitly enabled
6055 * destructive actions and (c) destructive actions have
6056 * not been disabled, we're going to transition into
6057 * the KILLED state, from which no further processing
6058 * on this state will be performed.
6060 if (!dtrace_priv_kernel_destructive(state) ||
6061 !state->dts_cred.dcr_destructive ||
6062 dtrace_destructive_disallow) {
6063 void *activity = &state->dts_activity;
6064 dtrace_activity_t current;
6067 current = state->dts_activity;
6068 } while (dtrace_cas32(activity, current,
6069 DTRACE_ACTIVITY_KILLED) != current);
6075 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6076 ecb->dte_alignment, state, &mstate)) < 0)
6079 tomax = buf->dtb_tomax;
6080 ASSERT(tomax != NULL);
6082 if (ecb->dte_size != 0)
6083 DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
6085 mstate.dtms_epid = ecb->dte_epid;
6086 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6088 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6089 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6091 mstate.dtms_access = 0;
6094 dtrace_difo_t *dp = pred->dtp_difo;
6097 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6099 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6100 dtrace_cacheid_t cid = probe->dtpr_predcache;
6102 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6104 * Update the predicate cache...
6106 ASSERT(cid == pred->dtp_cacheid);
6107 curthread->t_predcache = cid;
6114 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6115 act != NULL; act = act->dta_next) {
6118 dtrace_recdesc_t *rec = &act->dta_rec;
6120 size = rec->dtrd_size;
6121 valoffs = offs + rec->dtrd_offset;
6123 if (DTRACEACT_ISAGG(act->dta_kind)) {
6125 dtrace_aggregation_t *agg;
6127 agg = (dtrace_aggregation_t *)act;
6129 if ((dp = act->dta_difo) != NULL)
6130 v = dtrace_dif_emulate(dp,
6131 &mstate, vstate, state);
6133 if (*flags & CPU_DTRACE_ERROR)
6137 * Note that we always pass the expression
6138 * value from the previous iteration of the
6139 * action loop. This value will only be used
6140 * if there is an expression argument to the
6141 * aggregating action, denoted by the
6142 * dtag_hasarg field.
6144 dtrace_aggregate(agg, buf,
6145 offs, aggbuf, v, val);
6149 switch (act->dta_kind) {
6150 case DTRACEACT_STOP:
6151 if (dtrace_priv_proc_destructive(state))
6152 dtrace_action_stop();
6155 case DTRACEACT_BREAKPOINT:
6156 if (dtrace_priv_kernel_destructive(state))
6157 dtrace_action_breakpoint(ecb);
6160 case DTRACEACT_PANIC:
6161 if (dtrace_priv_kernel_destructive(state))
6162 dtrace_action_panic(ecb);
6165 case DTRACEACT_STACK:
6166 if (!dtrace_priv_kernel(state))
6169 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6170 size / sizeof (pc_t), probe->dtpr_aframes,
6171 DTRACE_ANCHORED(probe) ? NULL :
6176 case DTRACEACT_JSTACK:
6177 case DTRACEACT_USTACK:
6178 if (!dtrace_priv_proc(state))
6182 * See comment in DIF_VAR_PID.
6184 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6186 int depth = DTRACE_USTACK_NFRAMES(
6189 dtrace_bzero((void *)(tomax + valoffs),
6190 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6191 + depth * sizeof (uint64_t));
6196 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6197 curproc->p_dtrace_helpers != NULL) {
6199 * This is the slow path -- we have
6200 * allocated string space, and we're
6201 * getting the stack of a process that
6202 * has helpers. Call into a separate
6203 * routine to perform this processing.
6205 dtrace_action_ustack(&mstate, state,
6206 (uint64_t *)(tomax + valoffs),
6211 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6212 dtrace_getupcstack((uint64_t *)
6214 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6215 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6226 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6228 if (*flags & CPU_DTRACE_ERROR)
6231 switch (act->dta_kind) {
6232 case DTRACEACT_SPECULATE:
6233 ASSERT(buf == &state->dts_buffer[cpuid]);
6234 buf = dtrace_speculation_buffer(state,
6238 *flags |= CPU_DTRACE_DROP;
6242 offs = dtrace_buffer_reserve(buf,
6243 ecb->dte_needed, ecb->dte_alignment,
6247 *flags |= CPU_DTRACE_DROP;
6251 tomax = buf->dtb_tomax;
6252 ASSERT(tomax != NULL);
6254 if (ecb->dte_size != 0)
6255 DTRACE_STORE(uint32_t, tomax, offs,
6259 case DTRACEACT_PRINTM: {
6260 /* The DIF returns a 'memref'. */
6261 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6263 /* Get the size from the memref. */
6267 * Check if the size exceeds the allocated
6270 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6272 *flags |= CPU_DTRACE_DROP;
6276 /* Store the size in the buffer first. */
6277 DTRACE_STORE(uintptr_t, tomax,
6281 * Offset the buffer address to the start
6284 valoffs += sizeof(uintptr_t);
6287 * Reset to the memory address rather than
6288 * the memref array, then let the BYREF
6289 * code below do the work to store the
6290 * memory data in the buffer.
6296 case DTRACEACT_PRINTT: {
6297 /* The DIF returns a 'typeref'. */
6298 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6303 * Get the type string length and round it
6304 * up so that the data that follows is
6305 * aligned for easy access.
6307 size_t typs = strlen((char *) typeref[2]) + 1;
6308 typs = roundup(typs, sizeof(uintptr_t));
6311 *Get the size from the typeref using the
6312 * number of elements and the type size.
6314 size = typeref[1] * typeref[3];
6317 * Check if the size exceeds the allocated
6320 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6322 *flags |= CPU_DTRACE_DROP;
6326 /* Store the size in the buffer first. */
6327 DTRACE_STORE(uintptr_t, tomax,
6329 valoffs += sizeof(uintptr_t);
6331 /* Store the type size in the buffer. */
6332 DTRACE_STORE(uintptr_t, tomax,
6333 valoffs, typeref[3]);
6334 valoffs += sizeof(uintptr_t);
6338 for (s = 0; s < typs; s++) {
6340 c = dtrace_load8(val++);
6342 DTRACE_STORE(uint8_t, tomax,
6347 * Reset to the memory address rather than
6348 * the typeref array, then let the BYREF
6349 * code below do the work to store the
6350 * memory data in the buffer.
6356 case DTRACEACT_CHILL:
6357 if (dtrace_priv_kernel_destructive(state))
6358 dtrace_action_chill(&mstate, val);
6361 case DTRACEACT_RAISE:
6362 if (dtrace_priv_proc_destructive(state))
6363 dtrace_action_raise(val);
6366 case DTRACEACT_COMMIT:
6370 * We need to commit our buffer state.
6373 buf->dtb_offset = offs + ecb->dte_size;
6374 buf = &state->dts_buffer[cpuid];
6375 dtrace_speculation_commit(state, cpuid, val);
6379 case DTRACEACT_DISCARD:
6380 dtrace_speculation_discard(state, cpuid, val);
6383 case DTRACEACT_DIFEXPR:
6384 case DTRACEACT_LIBACT:
6385 case DTRACEACT_PRINTF:
6386 case DTRACEACT_PRINTA:
6387 case DTRACEACT_SYSTEM:
6388 case DTRACEACT_FREOPEN:
6393 if (!dtrace_priv_kernel(state))
6397 case DTRACEACT_USYM:
6398 case DTRACEACT_UMOD:
6399 case DTRACEACT_UADDR: {
6401 struct pid *pid = curthread->t_procp->p_pidp;
6404 if (!dtrace_priv_proc(state))
6407 DTRACE_STORE(uint64_t, tomax,
6409 valoffs, (uint64_t)pid->pid_id);
6411 valoffs, (uint64_t) curproc->p_pid);
6413 DTRACE_STORE(uint64_t, tomax,
6414 valoffs + sizeof (uint64_t), val);
6419 case DTRACEACT_EXIT: {
6421 * For the exit action, we are going to attempt
6422 * to atomically set our activity to be
6423 * draining. If this fails (either because
6424 * another CPU has beat us to the exit action,
6425 * or because our current activity is something
6426 * other than ACTIVE or WARMUP), we will
6427 * continue. This assures that the exit action
6428 * can be successfully recorded at most once
6429 * when we're in the ACTIVE state. If we're
6430 * encountering the exit() action while in
6431 * COOLDOWN, however, we want to honor the new
6432 * status code. (We know that we're the only
6433 * thread in COOLDOWN, so there is no race.)
6435 void *activity = &state->dts_activity;
6436 dtrace_activity_t current = state->dts_activity;
6438 if (current == DTRACE_ACTIVITY_COOLDOWN)
6441 if (current != DTRACE_ACTIVITY_WARMUP)
6442 current = DTRACE_ACTIVITY_ACTIVE;
6444 if (dtrace_cas32(activity, current,
6445 DTRACE_ACTIVITY_DRAINING) != current) {
6446 *flags |= CPU_DTRACE_DROP;
6457 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6458 uintptr_t end = valoffs + size;
6460 if (!dtrace_vcanload((void *)(uintptr_t)val,
6461 &dp->dtdo_rtype, &mstate, vstate))
6465 * If this is a string, we're going to only
6466 * load until we find the zero byte -- after
6467 * which we'll store zero bytes.
6469 if (dp->dtdo_rtype.dtdt_kind ==
6472 int intuple = act->dta_intuple;
6475 for (s = 0; s < size; s++) {
6477 c = dtrace_load8(val++);
6479 DTRACE_STORE(uint8_t, tomax,
6482 if (c == '\0' && intuple)
6489 while (valoffs < end) {
6490 DTRACE_STORE(uint8_t, tomax, valoffs++,
6491 dtrace_load8(val++));
6501 case sizeof (uint8_t):
6502 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6504 case sizeof (uint16_t):
6505 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6507 case sizeof (uint32_t):
6508 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6510 case sizeof (uint64_t):
6511 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6515 * Any other size should have been returned by
6516 * reference, not by value.
6523 if (*flags & CPU_DTRACE_DROP)
6526 if (*flags & CPU_DTRACE_FAULT) {
6528 dtrace_action_t *err;
6532 if (probe->dtpr_id == dtrace_probeid_error) {
6534 * There's nothing we can do -- we had an
6535 * error on the error probe. We bump an
6536 * error counter to at least indicate that
6537 * this condition happened.
6539 dtrace_error(&state->dts_dblerrors);
6545 * Before recursing on dtrace_probe(), we
6546 * need to explicitly clear out our start
6547 * time to prevent it from being accumulated
6548 * into t_dtrace_vtime.
6550 curthread->t_dtrace_start = 0;
6554 * Iterate over the actions to figure out which action
6555 * we were processing when we experienced the error.
6556 * Note that act points _past_ the faulting action; if
6557 * act is ecb->dte_action, the fault was in the
6558 * predicate, if it's ecb->dte_action->dta_next it's
6559 * in action #1, and so on.
6561 for (err = ecb->dte_action, ndx = 0;
6562 err != act; err = err->dta_next, ndx++)
6565 dtrace_probe_error(state, ecb->dte_epid, ndx,
6566 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6567 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6568 cpu_core[cpuid].cpuc_dtrace_illval);
6574 buf->dtb_offset = offs + ecb->dte_size;
6578 curthread->t_dtrace_start = dtrace_gethrtime();
6580 dtrace_interrupt_enable(cookie);
6584 * DTrace Probe Hashing Functions
6586 * The functions in this section (and indeed, the functions in remaining
6587 * sections) are not _called_ from probe context. (Any exceptions to this are
6588 * marked with a "Note:".) Rather, they are called from elsewhere in the
6589 * DTrace framework to look-up probes in, add probes to and remove probes from
6590 * the DTrace probe hashes. (Each probe is hashed by each element of the
6591 * probe tuple -- allowing for fast lookups, regardless of what was
6595 dtrace_hash_str(const char *p)
6601 hval = (hval << 4) + *p++;
6602 if ((g = (hval & 0xf0000000)) != 0)
6609 static dtrace_hash_t *
6610 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6612 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6614 hash->dth_stroffs = stroffs;
6615 hash->dth_nextoffs = nextoffs;
6616 hash->dth_prevoffs = prevoffs;
6619 hash->dth_mask = hash->dth_size - 1;
6621 hash->dth_tab = kmem_zalloc(hash->dth_size *
6622 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6628 dtrace_hash_destroy(dtrace_hash_t *hash)
6633 for (i = 0; i < hash->dth_size; i++)
6634 ASSERT(hash->dth_tab[i] == NULL);
6637 kmem_free(hash->dth_tab,
6638 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6639 kmem_free(hash, sizeof (dtrace_hash_t));
6643 dtrace_hash_resize(dtrace_hash_t *hash)
6645 int size = hash->dth_size, i, ndx;
6646 int new_size = hash->dth_size << 1;
6647 int new_mask = new_size - 1;
6648 dtrace_hashbucket_t **new_tab, *bucket, *next;
6650 ASSERT((new_size & new_mask) == 0);
6652 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6654 for (i = 0; i < size; i++) {
6655 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6656 dtrace_probe_t *probe = bucket->dthb_chain;
6658 ASSERT(probe != NULL);
6659 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6661 next = bucket->dthb_next;
6662 bucket->dthb_next = new_tab[ndx];
6663 new_tab[ndx] = bucket;
6667 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6668 hash->dth_tab = new_tab;
6669 hash->dth_size = new_size;
6670 hash->dth_mask = new_mask;
6674 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6676 int hashval = DTRACE_HASHSTR(hash, new);
6677 int ndx = hashval & hash->dth_mask;
6678 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6679 dtrace_probe_t **nextp, **prevp;
6681 for (; bucket != NULL; bucket = bucket->dthb_next) {
6682 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6686 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6687 dtrace_hash_resize(hash);
6688 dtrace_hash_add(hash, new);
6692 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6693 bucket->dthb_next = hash->dth_tab[ndx];
6694 hash->dth_tab[ndx] = bucket;
6695 hash->dth_nbuckets++;
6698 nextp = DTRACE_HASHNEXT(hash, new);
6699 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6700 *nextp = bucket->dthb_chain;
6702 if (bucket->dthb_chain != NULL) {
6703 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6704 ASSERT(*prevp == NULL);
6708 bucket->dthb_chain = new;
6712 static dtrace_probe_t *
6713 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6715 int hashval = DTRACE_HASHSTR(hash, template);
6716 int ndx = hashval & hash->dth_mask;
6717 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6719 for (; bucket != NULL; bucket = bucket->dthb_next) {
6720 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6721 return (bucket->dthb_chain);
6728 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6730 int hashval = DTRACE_HASHSTR(hash, template);
6731 int ndx = hashval & hash->dth_mask;
6732 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6734 for (; bucket != NULL; bucket = bucket->dthb_next) {
6735 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6736 return (bucket->dthb_len);
6743 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6745 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6746 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6748 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6749 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6752 * Find the bucket that we're removing this probe from.
6754 for (; bucket != NULL; bucket = bucket->dthb_next) {
6755 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6759 ASSERT(bucket != NULL);
6761 if (*prevp == NULL) {
6762 if (*nextp == NULL) {
6764 * The removed probe was the only probe on this
6765 * bucket; we need to remove the bucket.
6767 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6769 ASSERT(bucket->dthb_chain == probe);
6773 hash->dth_tab[ndx] = bucket->dthb_next;
6775 while (b->dthb_next != bucket)
6777 b->dthb_next = bucket->dthb_next;
6780 ASSERT(hash->dth_nbuckets > 0);
6781 hash->dth_nbuckets--;
6782 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6786 bucket->dthb_chain = *nextp;
6788 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6792 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6796 * DTrace Utility Functions
6798 * These are random utility functions that are _not_ called from probe context.
6801 dtrace_badattr(const dtrace_attribute_t *a)
6803 return (a->dtat_name > DTRACE_STABILITY_MAX ||
6804 a->dtat_data > DTRACE_STABILITY_MAX ||
6805 a->dtat_class > DTRACE_CLASS_MAX);
6809 * Return a duplicate copy of a string. If the specified string is NULL,
6810 * this function returns a zero-length string.
6813 dtrace_strdup(const char *str)
6815 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6818 (void) strcpy(new, str);
6823 #define DTRACE_ISALPHA(c) \
6824 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6827 dtrace_badname(const char *s)
6831 if (s == NULL || (c = *s++) == '\0')
6834 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6837 while ((c = *s++) != '\0') {
6838 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6839 c != '-' && c != '_' && c != '.' && c != '`')
6847 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6852 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6854 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6856 priv = DTRACE_PRIV_ALL;
6858 *uidp = crgetuid(cr);
6859 *zoneidp = crgetzoneid(cr);
6862 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6863 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6864 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6865 priv |= DTRACE_PRIV_USER;
6866 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6867 priv |= DTRACE_PRIV_PROC;
6868 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6869 priv |= DTRACE_PRIV_OWNER;
6870 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6871 priv |= DTRACE_PRIV_ZONEOWNER;
6874 priv = DTRACE_PRIV_ALL;
6880 #ifdef DTRACE_ERRDEBUG
6882 dtrace_errdebug(const char *str)
6884 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
6887 mutex_enter(&dtrace_errlock);
6888 dtrace_errlast = str;
6889 dtrace_errthread = curthread;
6891 while (occupied++ < DTRACE_ERRHASHSZ) {
6892 if (dtrace_errhash[hval].dter_msg == str) {
6893 dtrace_errhash[hval].dter_count++;
6897 if (dtrace_errhash[hval].dter_msg != NULL) {
6898 hval = (hval + 1) % DTRACE_ERRHASHSZ;
6902 dtrace_errhash[hval].dter_msg = str;
6903 dtrace_errhash[hval].dter_count = 1;
6907 panic("dtrace: undersized error hash");
6909 mutex_exit(&dtrace_errlock);
6914 * DTrace Matching Functions
6916 * These functions are used to match groups of probes, given some elements of
6917 * a probe tuple, or some globbed expressions for elements of a probe tuple.
6920 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
6923 if (priv != DTRACE_PRIV_ALL) {
6924 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
6925 uint32_t match = priv & ppriv;
6928 * No PRIV_DTRACE_* privileges...
6930 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
6931 DTRACE_PRIV_KERNEL)) == 0)
6935 * No matching bits, but there were bits to match...
6937 if (match == 0 && ppriv != 0)
6941 * Need to have permissions to the process, but don't...
6943 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
6944 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
6949 * Need to be in the same zone unless we possess the
6950 * privilege to examine all zones.
6952 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
6953 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
6962 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
6963 * consists of input pattern strings and an ops-vector to evaluate them.
6964 * This function returns >0 for match, 0 for no match, and <0 for error.
6967 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
6968 uint32_t priv, uid_t uid, zoneid_t zoneid)
6970 dtrace_provider_t *pvp = prp->dtpr_provider;
6973 if (pvp->dtpv_defunct)
6976 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
6979 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
6982 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
6985 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
6988 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
6995 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
6996 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
6997 * libc's version, the kernel version only applies to 8-bit ASCII strings.
6998 * In addition, all of the recursion cases except for '*' matching have been
6999 * unwound. For '*', we still implement recursive evaluation, but a depth
7000 * counter is maintained and matching is aborted if we recurse too deep.
7001 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7004 dtrace_match_glob(const char *s, const char *p, int depth)
7010 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7014 s = ""; /* treat NULL as empty string */
7023 if ((c = *p++) == '\0')
7024 return (s1 == '\0');
7028 int ok = 0, notflag = 0;
7039 if ((c = *p++) == '\0')
7043 if (c == '-' && lc != '\0' && *p != ']') {
7044 if ((c = *p++) == '\0')
7046 if (c == '\\' && (c = *p++) == '\0')
7050 if (s1 < lc || s1 > c)
7054 } else if (lc <= s1 && s1 <= c)
7057 } else if (c == '\\' && (c = *p++) == '\0')
7060 lc = c; /* save left-hand 'c' for next iteration */
7070 if ((c = *p++) == '\0')
7082 if ((c = *p++) == '\0')
7098 p++; /* consecutive *'s are identical to a single one */
7103 for (s = olds; *s != '\0'; s++) {
7104 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7114 dtrace_match_string(const char *s, const char *p, int depth)
7116 return (s != NULL && strcmp(s, p) == 0);
7121 dtrace_match_nul(const char *s, const char *p, int depth)
7123 return (1); /* always match the empty pattern */
7128 dtrace_match_nonzero(const char *s, const char *p, int depth)
7130 return (s != NULL && s[0] != '\0');
7134 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7135 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7137 dtrace_probe_t template, *probe;
7138 dtrace_hash_t *hash = NULL;
7139 int len, best = INT_MAX, nmatched = 0;
7142 ASSERT(MUTEX_HELD(&dtrace_lock));
7145 * If the probe ID is specified in the key, just lookup by ID and
7146 * invoke the match callback once if a matching probe is found.
7148 if (pkp->dtpk_id != DTRACE_IDNONE) {
7149 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7150 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7151 (void) (*matched)(probe, arg);
7157 template.dtpr_mod = (char *)pkp->dtpk_mod;
7158 template.dtpr_func = (char *)pkp->dtpk_func;
7159 template.dtpr_name = (char *)pkp->dtpk_name;
7162 * We want to find the most distinct of the module name, function
7163 * name, and name. So for each one that is not a glob pattern or
7164 * empty string, we perform a lookup in the corresponding hash and
7165 * use the hash table with the fewest collisions to do our search.
7167 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7168 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7170 hash = dtrace_bymod;
7173 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7174 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7176 hash = dtrace_byfunc;
7179 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7180 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7182 hash = dtrace_byname;
7186 * If we did not select a hash table, iterate over every probe and
7187 * invoke our callback for each one that matches our input probe key.
7190 for (i = 0; i < dtrace_nprobes; i++) {
7191 if ((probe = dtrace_probes[i]) == NULL ||
7192 dtrace_match_probe(probe, pkp, priv, uid,
7198 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7206 * If we selected a hash table, iterate over each probe of the same key
7207 * name and invoke the callback for every probe that matches the other
7208 * attributes of our input probe key.
7210 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7211 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7213 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7218 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7226 * Return the function pointer dtrace_probecmp() should use to compare the
7227 * specified pattern with a string. For NULL or empty patterns, we select
7228 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7229 * For non-empty non-glob strings, we use dtrace_match_string().
7231 static dtrace_probekey_f *
7232 dtrace_probekey_func(const char *p)
7236 if (p == NULL || *p == '\0')
7237 return (&dtrace_match_nul);
7239 while ((c = *p++) != '\0') {
7240 if (c == '[' || c == '?' || c == '*' || c == '\\')
7241 return (&dtrace_match_glob);
7244 return (&dtrace_match_string);
7248 * Build a probe comparison key for use with dtrace_match_probe() from the
7249 * given probe description. By convention, a null key only matches anchored
7250 * probes: if each field is the empty string, reset dtpk_fmatch to
7251 * dtrace_match_nonzero().
7254 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7256 pkp->dtpk_prov = pdp->dtpd_provider;
7257 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7259 pkp->dtpk_mod = pdp->dtpd_mod;
7260 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7262 pkp->dtpk_func = pdp->dtpd_func;
7263 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7265 pkp->dtpk_name = pdp->dtpd_name;
7266 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7268 pkp->dtpk_id = pdp->dtpd_id;
7270 if (pkp->dtpk_id == DTRACE_IDNONE &&
7271 pkp->dtpk_pmatch == &dtrace_match_nul &&
7272 pkp->dtpk_mmatch == &dtrace_match_nul &&
7273 pkp->dtpk_fmatch == &dtrace_match_nul &&
7274 pkp->dtpk_nmatch == &dtrace_match_nul)
7275 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7279 * DTrace Provider-to-Framework API Functions
7281 * These functions implement much of the Provider-to-Framework API, as
7282 * described in <sys/dtrace.h>. The parts of the API not in this section are
7283 * the functions in the API for probe management (found below), and
7284 * dtrace_probe() itself (found above).
7288 * Register the calling provider with the DTrace framework. This should
7289 * generally be called by DTrace providers in their attach(9E) entry point.
7292 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7293 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7295 dtrace_provider_t *provider;
7297 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7298 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7299 "arguments", name ? name : "<NULL>");
7303 if (name[0] == '\0' || dtrace_badname(name)) {
7304 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7305 "provider name", name);
7309 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7310 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7311 pops->dtps_destroy == NULL ||
7312 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7313 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7314 "provider ops", name);
7318 if (dtrace_badattr(&pap->dtpa_provider) ||
7319 dtrace_badattr(&pap->dtpa_mod) ||
7320 dtrace_badattr(&pap->dtpa_func) ||
7321 dtrace_badattr(&pap->dtpa_name) ||
7322 dtrace_badattr(&pap->dtpa_args)) {
7323 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7324 "provider attributes", name);
7328 if (priv & ~DTRACE_PRIV_ALL) {
7329 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7330 "privilege attributes", name);
7334 if ((priv & DTRACE_PRIV_KERNEL) &&
7335 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7336 pops->dtps_usermode == NULL) {
7337 cmn_err(CE_WARN, "failed to register provider '%s': need "
7338 "dtps_usermode() op for given privilege attributes", name);
7342 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7343 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7344 (void) strcpy(provider->dtpv_name, name);
7346 provider->dtpv_attr = *pap;
7347 provider->dtpv_priv.dtpp_flags = priv;
7349 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7350 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7352 provider->dtpv_pops = *pops;
7354 if (pops->dtps_provide == NULL) {
7355 ASSERT(pops->dtps_provide_module != NULL);
7356 provider->dtpv_pops.dtps_provide =
7357 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7360 if (pops->dtps_provide_module == NULL) {
7361 ASSERT(pops->dtps_provide != NULL);
7362 provider->dtpv_pops.dtps_provide_module =
7363 (void (*)(void *, modctl_t *))dtrace_nullop;
7366 if (pops->dtps_suspend == NULL) {
7367 ASSERT(pops->dtps_resume == NULL);
7368 provider->dtpv_pops.dtps_suspend =
7369 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7370 provider->dtpv_pops.dtps_resume =
7371 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7374 provider->dtpv_arg = arg;
7375 *idp = (dtrace_provider_id_t)provider;
7377 if (pops == &dtrace_provider_ops) {
7378 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7379 ASSERT(MUTEX_HELD(&dtrace_lock));
7380 ASSERT(dtrace_anon.dta_enabling == NULL);
7383 * We make sure that the DTrace provider is at the head of
7384 * the provider chain.
7386 provider->dtpv_next = dtrace_provider;
7387 dtrace_provider = provider;
7391 mutex_enter(&dtrace_provider_lock);
7392 mutex_enter(&dtrace_lock);
7395 * If there is at least one provider registered, we'll add this
7396 * provider after the first provider.
7398 if (dtrace_provider != NULL) {
7399 provider->dtpv_next = dtrace_provider->dtpv_next;
7400 dtrace_provider->dtpv_next = provider;
7402 dtrace_provider = provider;
7405 if (dtrace_retained != NULL) {
7406 dtrace_enabling_provide(provider);
7409 * Now we need to call dtrace_enabling_matchall() -- which
7410 * will acquire cpu_lock and dtrace_lock. We therefore need
7411 * to drop all of our locks before calling into it...
7413 mutex_exit(&dtrace_lock);
7414 mutex_exit(&dtrace_provider_lock);
7415 dtrace_enabling_matchall();
7420 mutex_exit(&dtrace_lock);
7421 mutex_exit(&dtrace_provider_lock);
7427 * Unregister the specified provider from the DTrace framework. This should
7428 * generally be called by DTrace providers in their detach(9E) entry point.
7431 dtrace_unregister(dtrace_provider_id_t id)
7433 dtrace_provider_t *old = (dtrace_provider_t *)id;
7434 dtrace_provider_t *prev = NULL;
7436 dtrace_probe_t *probe, *first = NULL;
7438 if (old->dtpv_pops.dtps_enable ==
7439 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7441 * If DTrace itself is the provider, we're called with locks
7444 ASSERT(old == dtrace_provider);
7446 ASSERT(dtrace_devi != NULL);
7448 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7449 ASSERT(MUTEX_HELD(&dtrace_lock));
7452 if (dtrace_provider->dtpv_next != NULL) {
7454 * There's another provider here; return failure.
7459 mutex_enter(&dtrace_provider_lock);
7460 mutex_enter(&mod_lock);
7461 mutex_enter(&dtrace_lock);
7465 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7466 * probes, we refuse to let providers slither away, unless this
7467 * provider has already been explicitly invalidated.
7469 if (!old->dtpv_defunct &&
7470 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7471 dtrace_anon.dta_state->dts_necbs > 0))) {
7473 mutex_exit(&dtrace_lock);
7474 mutex_exit(&mod_lock);
7475 mutex_exit(&dtrace_provider_lock);
7481 * Attempt to destroy the probes associated with this provider.
7483 for (i = 0; i < dtrace_nprobes; i++) {
7484 if ((probe = dtrace_probes[i]) == NULL)
7487 if (probe->dtpr_provider != old)
7490 if (probe->dtpr_ecb == NULL)
7494 * We have at least one ECB; we can't remove this provider.
7497 mutex_exit(&dtrace_lock);
7498 mutex_exit(&mod_lock);
7499 mutex_exit(&dtrace_provider_lock);
7505 * All of the probes for this provider are disabled; we can safely
7506 * remove all of them from their hash chains and from the probe array.
7508 for (i = 0; i < dtrace_nprobes; i++) {
7509 if ((probe = dtrace_probes[i]) == NULL)
7512 if (probe->dtpr_provider != old)
7515 dtrace_probes[i] = NULL;
7517 dtrace_hash_remove(dtrace_bymod, probe);
7518 dtrace_hash_remove(dtrace_byfunc, probe);
7519 dtrace_hash_remove(dtrace_byname, probe);
7521 if (first == NULL) {
7523 probe->dtpr_nextmod = NULL;
7525 probe->dtpr_nextmod = first;
7531 * The provider's probes have been removed from the hash chains and
7532 * from the probe array. Now issue a dtrace_sync() to be sure that
7533 * everyone has cleared out from any probe array processing.
7537 for (probe = first; probe != NULL; probe = first) {
7538 first = probe->dtpr_nextmod;
7540 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7542 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7543 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7544 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7546 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7548 free_unr(dtrace_arena, probe->dtpr_id);
7550 kmem_free(probe, sizeof (dtrace_probe_t));
7553 if ((prev = dtrace_provider) == old) {
7555 ASSERT(self || dtrace_devi == NULL);
7556 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7558 dtrace_provider = old->dtpv_next;
7560 while (prev != NULL && prev->dtpv_next != old)
7561 prev = prev->dtpv_next;
7564 panic("attempt to unregister non-existent "
7565 "dtrace provider %p\n", (void *)id);
7568 prev->dtpv_next = old->dtpv_next;
7572 mutex_exit(&dtrace_lock);
7573 mutex_exit(&mod_lock);
7574 mutex_exit(&dtrace_provider_lock);
7577 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7578 kmem_free(old, sizeof (dtrace_provider_t));
7584 * Invalidate the specified provider. All subsequent probe lookups for the
7585 * specified provider will fail, but its probes will not be removed.
7588 dtrace_invalidate(dtrace_provider_id_t id)
7590 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7592 ASSERT(pvp->dtpv_pops.dtps_enable !=
7593 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7595 mutex_enter(&dtrace_provider_lock);
7596 mutex_enter(&dtrace_lock);
7598 pvp->dtpv_defunct = 1;
7600 mutex_exit(&dtrace_lock);
7601 mutex_exit(&dtrace_provider_lock);
7605 * Indicate whether or not DTrace has attached.
7608 dtrace_attached(void)
7611 * dtrace_provider will be non-NULL iff the DTrace driver has
7612 * attached. (It's non-NULL because DTrace is always itself a
7615 return (dtrace_provider != NULL);
7619 * Remove all the unenabled probes for the given provider. This function is
7620 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7621 * -- just as many of its associated probes as it can.
7624 dtrace_condense(dtrace_provider_id_t id)
7626 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7628 dtrace_probe_t *probe;
7631 * Make sure this isn't the dtrace provider itself.
7633 ASSERT(prov->dtpv_pops.dtps_enable !=
7634 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7636 mutex_enter(&dtrace_provider_lock);
7637 mutex_enter(&dtrace_lock);
7640 * Attempt to destroy the probes associated with this provider.
7642 for (i = 0; i < dtrace_nprobes; i++) {
7643 if ((probe = dtrace_probes[i]) == NULL)
7646 if (probe->dtpr_provider != prov)
7649 if (probe->dtpr_ecb != NULL)
7652 dtrace_probes[i] = NULL;
7654 dtrace_hash_remove(dtrace_bymod, probe);
7655 dtrace_hash_remove(dtrace_byfunc, probe);
7656 dtrace_hash_remove(dtrace_byname, probe);
7658 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7660 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7661 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7662 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7663 kmem_free(probe, sizeof (dtrace_probe_t));
7665 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7667 free_unr(dtrace_arena, i + 1);
7671 mutex_exit(&dtrace_lock);
7672 mutex_exit(&dtrace_provider_lock);
7678 * DTrace Probe Management Functions
7680 * The functions in this section perform the DTrace probe management,
7681 * including functions to create probes, look-up probes, and call into the
7682 * providers to request that probes be provided. Some of these functions are
7683 * in the Provider-to-Framework API; these functions can be identified by the
7684 * fact that they are not declared "static".
7688 * Create a probe with the specified module name, function name, and name.
7691 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7692 const char *func, const char *name, int aframes, void *arg)
7694 dtrace_probe_t *probe, **probes;
7695 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7698 if (provider == dtrace_provider) {
7699 ASSERT(MUTEX_HELD(&dtrace_lock));
7701 mutex_enter(&dtrace_lock);
7705 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7706 VM_BESTFIT | VM_SLEEP);
7708 id = alloc_unr(dtrace_arena);
7710 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7712 probe->dtpr_id = id;
7713 probe->dtpr_gen = dtrace_probegen++;
7714 probe->dtpr_mod = dtrace_strdup(mod);
7715 probe->dtpr_func = dtrace_strdup(func);
7716 probe->dtpr_name = dtrace_strdup(name);
7717 probe->dtpr_arg = arg;
7718 probe->dtpr_aframes = aframes;
7719 probe->dtpr_provider = provider;
7721 dtrace_hash_add(dtrace_bymod, probe);
7722 dtrace_hash_add(dtrace_byfunc, probe);
7723 dtrace_hash_add(dtrace_byname, probe);
7725 if (id - 1 >= dtrace_nprobes) {
7726 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7727 size_t nsize = osize << 1;
7731 ASSERT(dtrace_probes == NULL);
7732 nsize = sizeof (dtrace_probe_t *);
7735 probes = kmem_zalloc(nsize, KM_SLEEP);
7737 if (dtrace_probes == NULL) {
7739 dtrace_probes = probes;
7742 dtrace_probe_t **oprobes = dtrace_probes;
7744 bcopy(oprobes, probes, osize);
7745 dtrace_membar_producer();
7746 dtrace_probes = probes;
7751 * All CPUs are now seeing the new probes array; we can
7752 * safely free the old array.
7754 kmem_free(oprobes, osize);
7755 dtrace_nprobes <<= 1;
7758 ASSERT(id - 1 < dtrace_nprobes);
7761 ASSERT(dtrace_probes[id - 1] == NULL);
7762 dtrace_probes[id - 1] = probe;
7764 if (provider != dtrace_provider)
7765 mutex_exit(&dtrace_lock);
7770 static dtrace_probe_t *
7771 dtrace_probe_lookup_id(dtrace_id_t id)
7773 ASSERT(MUTEX_HELD(&dtrace_lock));
7775 if (id == 0 || id > dtrace_nprobes)
7778 return (dtrace_probes[id - 1]);
7782 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7784 *((dtrace_id_t *)arg) = probe->dtpr_id;
7786 return (DTRACE_MATCH_DONE);
7790 * Look up a probe based on provider and one or more of module name, function
7791 * name and probe name.
7794 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
7795 char *func, char *name)
7797 dtrace_probekey_t pkey;
7801 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7802 pkey.dtpk_pmatch = &dtrace_match_string;
7803 pkey.dtpk_mod = mod;
7804 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7805 pkey.dtpk_func = func;
7806 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7807 pkey.dtpk_name = name;
7808 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7809 pkey.dtpk_id = DTRACE_IDNONE;
7811 mutex_enter(&dtrace_lock);
7812 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7813 dtrace_probe_lookup_match, &id);
7814 mutex_exit(&dtrace_lock);
7816 ASSERT(match == 1 || match == 0);
7817 return (match ? id : 0);
7821 * Returns the probe argument associated with the specified probe.
7824 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7826 dtrace_probe_t *probe;
7829 mutex_enter(&dtrace_lock);
7831 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7832 probe->dtpr_provider == (dtrace_provider_t *)id)
7833 rval = probe->dtpr_arg;
7835 mutex_exit(&dtrace_lock);
7841 * Copy a probe into a probe description.
7844 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7846 bzero(pdp, sizeof (dtrace_probedesc_t));
7847 pdp->dtpd_id = prp->dtpr_id;
7849 (void) strncpy(pdp->dtpd_provider,
7850 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7852 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7853 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7854 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7859 dtrace_probe_provide_cb(linker_file_t lf, void *arg)
7861 dtrace_provider_t *prv = (dtrace_provider_t *) arg;
7863 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, lf);
7871 * Called to indicate that a probe -- or probes -- should be provided by a
7872 * specfied provider. If the specified description is NULL, the provider will
7873 * be told to provide all of its probes. (This is done whenever a new
7874 * consumer comes along, or whenever a retained enabling is to be matched.) If
7875 * the specified description is non-NULL, the provider is given the
7876 * opportunity to dynamically provide the specified probe, allowing providers
7877 * to support the creation of probes on-the-fly. (So-called _autocreated_
7878 * probes.) If the provider is NULL, the operations will be applied to all
7879 * providers; if the provider is non-NULL the operations will only be applied
7880 * to the specified provider. The dtrace_provider_lock must be held, and the
7881 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7882 * will need to grab the dtrace_lock when it reenters the framework through
7883 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7886 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7893 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7897 prv = dtrace_provider;
7902 * First, call the blanket provide operation.
7904 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7907 * Now call the per-module provide operation. We will grab
7908 * mod_lock to prevent the list from being modified. Note
7909 * that this also prevents the mod_busy bits from changing.
7910 * (mod_busy can only be changed with mod_lock held.)
7912 mutex_enter(&mod_lock);
7917 if (ctl->mod_busy || ctl->mod_mp == NULL)
7920 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
7922 } while ((ctl = ctl->mod_next) != &modules);
7924 (void) linker_file_foreach(dtrace_probe_provide_cb, prv);
7927 mutex_exit(&mod_lock);
7928 } while (all && (prv = prv->dtpv_next) != NULL);
7933 * Iterate over each probe, and call the Framework-to-Provider API function
7937 dtrace_probe_foreach(uintptr_t offs)
7939 dtrace_provider_t *prov;
7940 void (*func)(void *, dtrace_id_t, void *);
7941 dtrace_probe_t *probe;
7942 dtrace_icookie_t cookie;
7946 * We disable interrupts to walk through the probe array. This is
7947 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
7948 * won't see stale data.
7950 cookie = dtrace_interrupt_disable();
7952 for (i = 0; i < dtrace_nprobes; i++) {
7953 if ((probe = dtrace_probes[i]) == NULL)
7956 if (probe->dtpr_ecb == NULL) {
7958 * This probe isn't enabled -- don't call the function.
7963 prov = probe->dtpr_provider;
7964 func = *((void(**)(void *, dtrace_id_t, void *))
7965 ((uintptr_t)&prov->dtpv_pops + offs));
7967 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
7970 dtrace_interrupt_enable(cookie);
7975 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
7977 dtrace_probekey_t pkey;
7982 ASSERT(MUTEX_HELD(&dtrace_lock));
7983 dtrace_ecb_create_cache = NULL;
7987 * If we're passed a NULL description, we're being asked to
7988 * create an ECB with a NULL probe.
7990 (void) dtrace_ecb_create_enable(NULL, enab);
7994 dtrace_probekey(desc, &pkey);
7995 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
7996 &priv, &uid, &zoneid);
7998 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8003 * DTrace Helper Provider Functions
8006 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8008 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8009 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8010 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8014 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8015 const dof_provider_t *dofprov, char *strtab)
8017 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8018 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8019 dofprov->dofpv_provattr);
8020 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8021 dofprov->dofpv_modattr);
8022 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8023 dofprov->dofpv_funcattr);
8024 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8025 dofprov->dofpv_nameattr);
8026 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8027 dofprov->dofpv_argsattr);
8031 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8033 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8034 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8035 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8036 dof_provider_t *provider;
8038 uint32_t *off, *enoff;
8042 dtrace_helper_provdesc_t dhpv;
8043 dtrace_helper_probedesc_t dhpb;
8044 dtrace_meta_t *meta = dtrace_meta_pid;
8045 dtrace_mops_t *mops = &meta->dtm_mops;
8048 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8049 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8050 provider->dofpv_strtab * dof->dofh_secsize);
8051 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8052 provider->dofpv_probes * dof->dofh_secsize);
8053 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8054 provider->dofpv_prargs * dof->dofh_secsize);
8055 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8056 provider->dofpv_proffs * dof->dofh_secsize);
8058 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8059 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8060 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8064 * See dtrace_helper_provider_validate().
8066 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8067 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8068 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8069 provider->dofpv_prenoffs * dof->dofh_secsize);
8070 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8073 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8076 * Create the provider.
8078 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8080 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8086 * Create the probes.
8088 for (i = 0; i < nprobes; i++) {
8089 probe = (dof_probe_t *)(uintptr_t)(daddr +
8090 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8092 dhpb.dthpb_mod = dhp->dofhp_mod;
8093 dhpb.dthpb_func = strtab + probe->dofpr_func;
8094 dhpb.dthpb_name = strtab + probe->dofpr_name;
8095 dhpb.dthpb_base = probe->dofpr_addr;
8096 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8097 dhpb.dthpb_noffs = probe->dofpr_noffs;
8098 if (enoff != NULL) {
8099 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8100 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8102 dhpb.dthpb_enoffs = NULL;
8103 dhpb.dthpb_nenoffs = 0;
8105 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8106 dhpb.dthpb_nargc = probe->dofpr_nargc;
8107 dhpb.dthpb_xargc = probe->dofpr_xargc;
8108 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8109 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8111 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8116 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8118 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8119 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8122 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8124 for (i = 0; i < dof->dofh_secnum; i++) {
8125 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8126 dof->dofh_secoff + i * dof->dofh_secsize);
8128 if (sec->dofs_type != DOF_SECT_PROVIDER)
8131 dtrace_helper_provide_one(dhp, sec, pid);
8135 * We may have just created probes, so we must now rematch against
8136 * any retained enablings. Note that this call will acquire both
8137 * cpu_lock and dtrace_lock; the fact that we are holding
8138 * dtrace_meta_lock now is what defines the ordering with respect to
8139 * these three locks.
8141 dtrace_enabling_matchall();
8146 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8148 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8149 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8151 dof_provider_t *provider;
8153 dtrace_helper_provdesc_t dhpv;
8154 dtrace_meta_t *meta = dtrace_meta_pid;
8155 dtrace_mops_t *mops = &meta->dtm_mops;
8157 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8158 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8159 provider->dofpv_strtab * dof->dofh_secsize);
8161 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8164 * Create the provider.
8166 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8168 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8174 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8176 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8177 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8180 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8182 for (i = 0; i < dof->dofh_secnum; i++) {
8183 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8184 dof->dofh_secoff + i * dof->dofh_secsize);
8186 if (sec->dofs_type != DOF_SECT_PROVIDER)
8189 dtrace_helper_provider_remove_one(dhp, sec, pid);
8195 * DTrace Meta Provider-to-Framework API Functions
8197 * These functions implement the Meta Provider-to-Framework API, as described
8198 * in <sys/dtrace.h>.
8201 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8202 dtrace_meta_provider_id_t *idp)
8204 dtrace_meta_t *meta;
8205 dtrace_helpers_t *help, *next;
8208 *idp = DTRACE_METAPROVNONE;
8211 * We strictly don't need the name, but we hold onto it for
8212 * debuggability. All hail error queues!
8215 cmn_err(CE_WARN, "failed to register meta-provider: "
8221 mops->dtms_create_probe == NULL ||
8222 mops->dtms_provide_pid == NULL ||
8223 mops->dtms_remove_pid == NULL) {
8224 cmn_err(CE_WARN, "failed to register meta-register %s: "
8225 "invalid ops", name);
8229 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8230 meta->dtm_mops = *mops;
8231 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8232 (void) strcpy(meta->dtm_name, name);
8233 meta->dtm_arg = arg;
8235 mutex_enter(&dtrace_meta_lock);
8236 mutex_enter(&dtrace_lock);
8238 if (dtrace_meta_pid != NULL) {
8239 mutex_exit(&dtrace_lock);
8240 mutex_exit(&dtrace_meta_lock);
8241 cmn_err(CE_WARN, "failed to register meta-register %s: "
8242 "user-land meta-provider exists", name);
8243 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8244 kmem_free(meta, sizeof (dtrace_meta_t));
8248 dtrace_meta_pid = meta;
8249 *idp = (dtrace_meta_provider_id_t)meta;
8252 * If there are providers and probes ready to go, pass them
8253 * off to the new meta provider now.
8256 help = dtrace_deferred_pid;
8257 dtrace_deferred_pid = NULL;
8259 mutex_exit(&dtrace_lock);
8261 while (help != NULL) {
8262 for (i = 0; i < help->dthps_nprovs; i++) {
8263 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8267 next = help->dthps_next;
8268 help->dthps_next = NULL;
8269 help->dthps_prev = NULL;
8270 help->dthps_deferred = 0;
8274 mutex_exit(&dtrace_meta_lock);
8280 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8282 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8284 mutex_enter(&dtrace_meta_lock);
8285 mutex_enter(&dtrace_lock);
8287 if (old == dtrace_meta_pid) {
8288 pp = &dtrace_meta_pid;
8290 panic("attempt to unregister non-existent "
8291 "dtrace meta-provider %p\n", (void *)old);
8294 if (old->dtm_count != 0) {
8295 mutex_exit(&dtrace_lock);
8296 mutex_exit(&dtrace_meta_lock);
8302 mutex_exit(&dtrace_lock);
8303 mutex_exit(&dtrace_meta_lock);
8305 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8306 kmem_free(old, sizeof (dtrace_meta_t));
8313 * DTrace DIF Object Functions
8316 dtrace_difo_err(uint_t pc, const char *format, ...)
8318 if (dtrace_err_verbose) {
8321 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8322 va_start(alist, format);
8323 (void) vuprintf(format, alist);
8327 #ifdef DTRACE_ERRDEBUG
8328 dtrace_errdebug(format);
8334 * Validate a DTrace DIF object by checking the IR instructions. The following
8335 * rules are currently enforced by dtrace_difo_validate():
8337 * 1. Each instruction must have a valid opcode
8338 * 2. Each register, string, variable, or subroutine reference must be valid
8339 * 3. No instruction can modify register %r0 (must be zero)
8340 * 4. All instruction reserved bits must be set to zero
8341 * 5. The last instruction must be a "ret" instruction
8342 * 6. All branch targets must reference a valid instruction _after_ the branch
8345 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8349 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8353 kcheckload = cr == NULL ||
8354 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8356 dp->dtdo_destructive = 0;
8358 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8359 dif_instr_t instr = dp->dtdo_buf[pc];
8361 uint_t r1 = DIF_INSTR_R1(instr);
8362 uint_t r2 = DIF_INSTR_R2(instr);
8363 uint_t rd = DIF_INSTR_RD(instr);
8364 uint_t rs = DIF_INSTR_RS(instr);
8365 uint_t label = DIF_INSTR_LABEL(instr);
8366 uint_t v = DIF_INSTR_VAR(instr);
8367 uint_t subr = DIF_INSTR_SUBR(instr);
8368 uint_t type = DIF_INSTR_TYPE(instr);
8369 uint_t op = DIF_INSTR_OP(instr);
8387 err += efunc(pc, "invalid register %u\n", r1);
8389 err += efunc(pc, "invalid register %u\n", r2);
8391 err += efunc(pc, "invalid register %u\n", rd);
8393 err += efunc(pc, "cannot write to %r0\n");
8399 err += efunc(pc, "invalid register %u\n", r1);
8401 err += efunc(pc, "non-zero reserved bits\n");
8403 err += efunc(pc, "invalid register %u\n", rd);
8405 err += efunc(pc, "cannot write to %r0\n");
8415 err += efunc(pc, "invalid register %u\n", r1);
8417 err += efunc(pc, "non-zero reserved bits\n");
8419 err += efunc(pc, "invalid register %u\n", rd);
8421 err += efunc(pc, "cannot write to %r0\n");
8423 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8424 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8434 err += efunc(pc, "invalid register %u\n", r1);
8436 err += efunc(pc, "non-zero reserved bits\n");
8438 err += efunc(pc, "invalid register %u\n", rd);
8440 err += efunc(pc, "cannot write to %r0\n");
8450 err += efunc(pc, "invalid register %u\n", r1);
8452 err += efunc(pc, "non-zero reserved bits\n");
8454 err += efunc(pc, "invalid register %u\n", rd);
8456 err += efunc(pc, "cannot write to %r0\n");
8463 err += efunc(pc, "invalid register %u\n", r1);
8465 err += efunc(pc, "non-zero reserved bits\n");
8467 err += efunc(pc, "invalid register %u\n", rd);
8469 err += efunc(pc, "cannot write to 0 address\n");
8474 err += efunc(pc, "invalid register %u\n", r1);
8476 err += efunc(pc, "invalid register %u\n", r2);
8478 err += efunc(pc, "non-zero reserved bits\n");
8482 err += efunc(pc, "invalid register %u\n", r1);
8483 if (r2 != 0 || rd != 0)
8484 err += efunc(pc, "non-zero reserved bits\n");
8497 if (label >= dp->dtdo_len) {
8498 err += efunc(pc, "invalid branch target %u\n",
8502 err += efunc(pc, "backward branch to %u\n",
8507 if (r1 != 0 || r2 != 0)
8508 err += efunc(pc, "non-zero reserved bits\n");
8510 err += efunc(pc, "invalid register %u\n", rd);
8514 case DIF_OP_FLUSHTS:
8515 if (r1 != 0 || r2 != 0 || rd != 0)
8516 err += efunc(pc, "non-zero reserved bits\n");
8519 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8520 err += efunc(pc, "invalid integer ref %u\n",
8521 DIF_INSTR_INTEGER(instr));
8524 err += efunc(pc, "invalid register %u\n", rd);
8526 err += efunc(pc, "cannot write to %r0\n");
8529 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8530 err += efunc(pc, "invalid string ref %u\n",
8531 DIF_INSTR_STRING(instr));
8534 err += efunc(pc, "invalid register %u\n", rd);
8536 err += efunc(pc, "cannot write to %r0\n");
8540 if (r1 > DIF_VAR_ARRAY_MAX)
8541 err += efunc(pc, "invalid array %u\n", r1);
8543 err += efunc(pc, "invalid register %u\n", r2);
8545 err += efunc(pc, "invalid register %u\n", rd);
8547 err += efunc(pc, "cannot write to %r0\n");
8554 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8555 err += efunc(pc, "invalid variable %u\n", v);
8557 err += efunc(pc, "invalid register %u\n", rd);
8559 err += efunc(pc, "cannot write to %r0\n");
8566 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8567 err += efunc(pc, "invalid variable %u\n", v);
8569 err += efunc(pc, "invalid register %u\n", rd);
8572 if (subr > DIF_SUBR_MAX)
8573 err += efunc(pc, "invalid subr %u\n", subr);
8575 err += efunc(pc, "invalid register %u\n", rd);
8577 err += efunc(pc, "cannot write to %r0\n");
8579 if (subr == DIF_SUBR_COPYOUT ||
8580 subr == DIF_SUBR_COPYOUTSTR) {
8581 dp->dtdo_destructive = 1;
8585 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8586 err += efunc(pc, "invalid ref type %u\n", type);
8588 err += efunc(pc, "invalid register %u\n", r2);
8590 err += efunc(pc, "invalid register %u\n", rs);
8593 if (type != DIF_TYPE_CTF)
8594 err += efunc(pc, "invalid val type %u\n", type);
8596 err += efunc(pc, "invalid register %u\n", r2);
8598 err += efunc(pc, "invalid register %u\n", rs);
8601 err += efunc(pc, "invalid opcode %u\n",
8602 DIF_INSTR_OP(instr));
8606 if (dp->dtdo_len != 0 &&
8607 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8608 err += efunc(dp->dtdo_len - 1,
8609 "expected 'ret' as last DIF instruction\n");
8612 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8614 * If we're not returning by reference, the size must be either
8615 * 0 or the size of one of the base types.
8617 switch (dp->dtdo_rtype.dtdt_size) {
8619 case sizeof (uint8_t):
8620 case sizeof (uint16_t):
8621 case sizeof (uint32_t):
8622 case sizeof (uint64_t):
8626 err += efunc(dp->dtdo_len - 1, "bad return size");
8630 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8631 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8632 dtrace_diftype_t *vt, *et;
8635 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8636 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8637 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8638 err += efunc(i, "unrecognized variable scope %d\n",
8643 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8644 v->dtdv_kind != DIFV_KIND_SCALAR) {
8645 err += efunc(i, "unrecognized variable type %d\n",
8650 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8651 err += efunc(i, "%d exceeds variable id limit\n", id);
8655 if (id < DIF_VAR_OTHER_UBASE)
8659 * For user-defined variables, we need to check that this
8660 * definition is identical to any previous definition that we
8663 ndx = id - DIF_VAR_OTHER_UBASE;
8665 switch (v->dtdv_scope) {
8666 case DIFV_SCOPE_GLOBAL:
8667 if (ndx < vstate->dtvs_nglobals) {
8668 dtrace_statvar_t *svar;
8670 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8671 existing = &svar->dtsv_var;
8676 case DIFV_SCOPE_THREAD:
8677 if (ndx < vstate->dtvs_ntlocals)
8678 existing = &vstate->dtvs_tlocals[ndx];
8681 case DIFV_SCOPE_LOCAL:
8682 if (ndx < vstate->dtvs_nlocals) {
8683 dtrace_statvar_t *svar;
8685 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8686 existing = &svar->dtsv_var;
8694 if (vt->dtdt_flags & DIF_TF_BYREF) {
8695 if (vt->dtdt_size == 0) {
8696 err += efunc(i, "zero-sized variable\n");
8700 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8701 vt->dtdt_size > dtrace_global_maxsize) {
8702 err += efunc(i, "oversized by-ref global\n");
8707 if (existing == NULL || existing->dtdv_id == 0)
8710 ASSERT(existing->dtdv_id == v->dtdv_id);
8711 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8713 if (existing->dtdv_kind != v->dtdv_kind)
8714 err += efunc(i, "%d changed variable kind\n", id);
8716 et = &existing->dtdv_type;
8718 if (vt->dtdt_flags != et->dtdt_flags) {
8719 err += efunc(i, "%d changed variable type flags\n", id);
8723 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8724 err += efunc(i, "%d changed variable type size\n", id);
8734 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
8735 * are much more constrained than normal DIFOs. Specifically, they may
8738 * 1. Make calls to subroutines other than copyin(), copyinstr() or
8739 * miscellaneous string routines
8740 * 2. Access DTrace variables other than the args[] array, and the
8741 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8742 * 3. Have thread-local variables.
8743 * 4. Have dynamic variables.
8746 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8748 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8752 for (pc = 0; pc < dp->dtdo_len; pc++) {
8753 dif_instr_t instr = dp->dtdo_buf[pc];
8755 uint_t v = DIF_INSTR_VAR(instr);
8756 uint_t subr = DIF_INSTR_SUBR(instr);
8757 uint_t op = DIF_INSTR_OP(instr);
8812 case DIF_OP_FLUSHTS:
8824 if (v >= DIF_VAR_OTHER_UBASE)
8827 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8830 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8831 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8832 v == DIF_VAR_EXECARGS ||
8833 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8834 v == DIF_VAR_UID || v == DIF_VAR_GID)
8837 err += efunc(pc, "illegal variable %u\n", v);
8844 err += efunc(pc, "illegal dynamic variable load\n");
8850 err += efunc(pc, "illegal dynamic variable store\n");
8854 if (subr == DIF_SUBR_ALLOCA ||
8855 subr == DIF_SUBR_BCOPY ||
8856 subr == DIF_SUBR_COPYIN ||
8857 subr == DIF_SUBR_COPYINTO ||
8858 subr == DIF_SUBR_COPYINSTR ||
8859 subr == DIF_SUBR_INDEX ||
8860 subr == DIF_SUBR_INET_NTOA ||
8861 subr == DIF_SUBR_INET_NTOA6 ||
8862 subr == DIF_SUBR_INET_NTOP ||
8863 subr == DIF_SUBR_LLTOSTR ||
8864 subr == DIF_SUBR_RINDEX ||
8865 subr == DIF_SUBR_STRCHR ||
8866 subr == DIF_SUBR_STRJOIN ||
8867 subr == DIF_SUBR_STRRCHR ||
8868 subr == DIF_SUBR_STRSTR ||
8869 subr == DIF_SUBR_HTONS ||
8870 subr == DIF_SUBR_HTONL ||
8871 subr == DIF_SUBR_HTONLL ||
8872 subr == DIF_SUBR_NTOHS ||
8873 subr == DIF_SUBR_NTOHL ||
8874 subr == DIF_SUBR_NTOHLL ||
8875 subr == DIF_SUBR_MEMREF ||
8876 subr == DIF_SUBR_TYPEREF)
8879 err += efunc(pc, "invalid subr %u\n", subr);
8883 err += efunc(pc, "invalid opcode %u\n",
8884 DIF_INSTR_OP(instr));
8893 * Returns 1 if the expression in the DIF object can be cached on a per-thread
8897 dtrace_difo_cacheable(dtrace_difo_t *dp)
8904 for (i = 0; i < dp->dtdo_varlen; i++) {
8905 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8907 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8910 switch (v->dtdv_id) {
8911 case DIF_VAR_CURTHREAD:
8914 case DIF_VAR_EXECARGS:
8915 case DIF_VAR_EXECNAME:
8916 case DIF_VAR_ZONENAME:
8925 * This DIF object may be cacheable. Now we need to look for any
8926 * array loading instructions, any memory loading instructions, or
8927 * any stores to thread-local variables.
8929 for (i = 0; i < dp->dtdo_len; i++) {
8930 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
8932 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
8933 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
8934 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
8935 op == DIF_OP_LDGA || op == DIF_OP_STTS)
8943 dtrace_difo_hold(dtrace_difo_t *dp)
8947 ASSERT(MUTEX_HELD(&dtrace_lock));
8950 ASSERT(dp->dtdo_refcnt != 0);
8953 * We need to check this DIF object for references to the variable
8954 * DIF_VAR_VTIMESTAMP.
8956 for (i = 0; i < dp->dtdo_varlen; i++) {
8957 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8959 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8962 if (dtrace_vtime_references++ == 0)
8963 dtrace_vtime_enable();
8968 * This routine calculates the dynamic variable chunksize for a given DIF
8969 * object. The calculation is not fool-proof, and can probably be tricked by
8970 * malicious DIF -- but it works for all compiler-generated DIF. Because this
8971 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
8972 * if a dynamic variable size exceeds the chunksize.
8975 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8978 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
8979 const dif_instr_t *text = dp->dtdo_buf;
8985 for (pc = 0; pc < dp->dtdo_len; pc++) {
8986 dif_instr_t instr = text[pc];
8987 uint_t op = DIF_INSTR_OP(instr);
8988 uint_t rd = DIF_INSTR_RD(instr);
8989 uint_t r1 = DIF_INSTR_R1(instr);
8993 dtrace_key_t *key = tupregs;
8997 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9002 key = &tupregs[DIF_DTR_NREGS];
9003 key[0].dttk_size = 0;
9004 key[1].dttk_size = 0;
9006 scope = DIFV_SCOPE_THREAD;
9013 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9014 key[nkeys++].dttk_size = 0;
9016 key[nkeys++].dttk_size = 0;
9018 if (op == DIF_OP_STTAA) {
9019 scope = DIFV_SCOPE_THREAD;
9021 scope = DIFV_SCOPE_GLOBAL;
9027 if (ttop == DIF_DTR_NREGS)
9030 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9032 * If the register for the size of the "pushtr"
9033 * is %r0 (or the value is 0) and the type is
9034 * a string, we'll use the system-wide default
9037 tupregs[ttop++].dttk_size =
9038 dtrace_strsize_default;
9043 tupregs[ttop++].dttk_size = sval;
9049 if (ttop == DIF_DTR_NREGS)
9052 tupregs[ttop++].dttk_size = 0;
9055 case DIF_OP_FLUSHTS:
9072 * We have a dynamic variable allocation; calculate its size.
9074 for (ksize = 0, i = 0; i < nkeys; i++)
9075 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9077 size = sizeof (dtrace_dynvar_t);
9078 size += sizeof (dtrace_key_t) * (nkeys - 1);
9082 * Now we need to determine the size of the stored data.
9084 id = DIF_INSTR_VAR(instr);
9086 for (i = 0; i < dp->dtdo_varlen; i++) {
9087 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9089 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9090 size += v->dtdv_type.dtdt_size;
9095 if (i == dp->dtdo_varlen)
9099 * We have the size. If this is larger than the chunk size
9100 * for our dynamic variable state, reset the chunk size.
9102 size = P2ROUNDUP(size, sizeof (uint64_t));
9104 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9105 vstate->dtvs_dynvars.dtds_chunksize = size;
9110 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9112 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9115 ASSERT(MUTEX_HELD(&dtrace_lock));
9116 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9118 for (i = 0; i < dp->dtdo_varlen; i++) {
9119 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9120 dtrace_statvar_t *svar, ***svarp = NULL;
9122 uint8_t scope = v->dtdv_scope;
9125 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9128 id -= DIF_VAR_OTHER_UBASE;
9131 case DIFV_SCOPE_THREAD:
9132 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9133 dtrace_difv_t *tlocals;
9135 if ((ntlocals = (otlocals << 1)) == 0)
9138 osz = otlocals * sizeof (dtrace_difv_t);
9139 nsz = ntlocals * sizeof (dtrace_difv_t);
9141 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9144 bcopy(vstate->dtvs_tlocals,
9146 kmem_free(vstate->dtvs_tlocals, osz);
9149 vstate->dtvs_tlocals = tlocals;
9150 vstate->dtvs_ntlocals = ntlocals;
9153 vstate->dtvs_tlocals[id] = *v;
9156 case DIFV_SCOPE_LOCAL:
9157 np = &vstate->dtvs_nlocals;
9158 svarp = &vstate->dtvs_locals;
9160 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9161 dsize = NCPU * (v->dtdv_type.dtdt_size +
9164 dsize = NCPU * sizeof (uint64_t);
9168 case DIFV_SCOPE_GLOBAL:
9169 np = &vstate->dtvs_nglobals;
9170 svarp = &vstate->dtvs_globals;
9172 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9173 dsize = v->dtdv_type.dtdt_size +
9182 while (id >= (oldsvars = *np)) {
9183 dtrace_statvar_t **statics;
9184 int newsvars, oldsize, newsize;
9186 if ((newsvars = (oldsvars << 1)) == 0)
9189 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9190 newsize = newsvars * sizeof (dtrace_statvar_t *);
9192 statics = kmem_zalloc(newsize, KM_SLEEP);
9195 bcopy(*svarp, statics, oldsize);
9196 kmem_free(*svarp, oldsize);
9203 if ((svar = (*svarp)[id]) == NULL) {
9204 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9205 svar->dtsv_var = *v;
9207 if ((svar->dtsv_size = dsize) != 0) {
9208 svar->dtsv_data = (uint64_t)(uintptr_t)
9209 kmem_zalloc(dsize, KM_SLEEP);
9212 (*svarp)[id] = svar;
9215 svar->dtsv_refcnt++;
9218 dtrace_difo_chunksize(dp, vstate);
9219 dtrace_difo_hold(dp);
9223 static dtrace_difo_t *
9224 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9229 ASSERT(dp->dtdo_buf != NULL);
9230 ASSERT(dp->dtdo_refcnt != 0);
9232 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9234 ASSERT(dp->dtdo_buf != NULL);
9235 sz = dp->dtdo_len * sizeof (dif_instr_t);
9236 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9237 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9238 new->dtdo_len = dp->dtdo_len;
9240 if (dp->dtdo_strtab != NULL) {
9241 ASSERT(dp->dtdo_strlen != 0);
9242 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9243 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9244 new->dtdo_strlen = dp->dtdo_strlen;
9247 if (dp->dtdo_inttab != NULL) {
9248 ASSERT(dp->dtdo_intlen != 0);
9249 sz = dp->dtdo_intlen * sizeof (uint64_t);
9250 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9251 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9252 new->dtdo_intlen = dp->dtdo_intlen;
9255 if (dp->dtdo_vartab != NULL) {
9256 ASSERT(dp->dtdo_varlen != 0);
9257 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9258 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9259 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9260 new->dtdo_varlen = dp->dtdo_varlen;
9263 dtrace_difo_init(new, vstate);
9269 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9273 ASSERT(dp->dtdo_refcnt == 0);
9275 for (i = 0; i < dp->dtdo_varlen; i++) {
9276 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9277 dtrace_statvar_t *svar, **svarp = NULL;
9279 uint8_t scope = v->dtdv_scope;
9283 case DIFV_SCOPE_THREAD:
9286 case DIFV_SCOPE_LOCAL:
9287 np = &vstate->dtvs_nlocals;
9288 svarp = vstate->dtvs_locals;
9291 case DIFV_SCOPE_GLOBAL:
9292 np = &vstate->dtvs_nglobals;
9293 svarp = vstate->dtvs_globals;
9300 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9303 id -= DIF_VAR_OTHER_UBASE;
9307 ASSERT(svar != NULL);
9308 ASSERT(svar->dtsv_refcnt > 0);
9310 if (--svar->dtsv_refcnt > 0)
9313 if (svar->dtsv_size != 0) {
9314 ASSERT(svar->dtsv_data != 0);
9315 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9319 kmem_free(svar, sizeof (dtrace_statvar_t));
9323 if (dp->dtdo_buf != NULL)
9324 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9325 if (dp->dtdo_inttab != NULL)
9326 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9327 if (dp->dtdo_strtab != NULL)
9328 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9329 if (dp->dtdo_vartab != NULL)
9330 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9332 kmem_free(dp, sizeof (dtrace_difo_t));
9336 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9340 ASSERT(MUTEX_HELD(&dtrace_lock));
9341 ASSERT(dp->dtdo_refcnt != 0);
9343 for (i = 0; i < dp->dtdo_varlen; i++) {
9344 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9346 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9349 ASSERT(dtrace_vtime_references > 0);
9350 if (--dtrace_vtime_references == 0)
9351 dtrace_vtime_disable();
9354 if (--dp->dtdo_refcnt == 0)
9355 dtrace_difo_destroy(dp, vstate);
9359 * DTrace Format Functions
9362 dtrace_format_add(dtrace_state_t *state, char *str)
9365 uint16_t ndx, len = strlen(str) + 1;
9367 fmt = kmem_zalloc(len, KM_SLEEP);
9368 bcopy(str, fmt, len);
9370 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9371 if (state->dts_formats[ndx] == NULL) {
9372 state->dts_formats[ndx] = fmt;
9377 if (state->dts_nformats == USHRT_MAX) {
9379 * This is only likely if a denial-of-service attack is being
9380 * attempted. As such, it's okay to fail silently here.
9382 kmem_free(fmt, len);
9387 * For simplicity, we always resize the formats array to be exactly the
9388 * number of formats.
9390 ndx = state->dts_nformats++;
9391 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9393 if (state->dts_formats != NULL) {
9395 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9396 kmem_free(state->dts_formats, ndx * sizeof (char *));
9399 state->dts_formats = new;
9400 state->dts_formats[ndx] = fmt;
9406 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9410 ASSERT(state->dts_formats != NULL);
9411 ASSERT(format <= state->dts_nformats);
9412 ASSERT(state->dts_formats[format - 1] != NULL);
9414 fmt = state->dts_formats[format - 1];
9415 kmem_free(fmt, strlen(fmt) + 1);
9416 state->dts_formats[format - 1] = NULL;
9420 dtrace_format_destroy(dtrace_state_t *state)
9424 if (state->dts_nformats == 0) {
9425 ASSERT(state->dts_formats == NULL);
9429 ASSERT(state->dts_formats != NULL);
9431 for (i = 0; i < state->dts_nformats; i++) {
9432 char *fmt = state->dts_formats[i];
9437 kmem_free(fmt, strlen(fmt) + 1);
9440 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9441 state->dts_nformats = 0;
9442 state->dts_formats = NULL;
9446 * DTrace Predicate Functions
9448 static dtrace_predicate_t *
9449 dtrace_predicate_create(dtrace_difo_t *dp)
9451 dtrace_predicate_t *pred;
9453 ASSERT(MUTEX_HELD(&dtrace_lock));
9454 ASSERT(dp->dtdo_refcnt != 0);
9456 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9457 pred->dtp_difo = dp;
9458 pred->dtp_refcnt = 1;
9460 if (!dtrace_difo_cacheable(dp))
9463 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9465 * This is only theoretically possible -- we have had 2^32
9466 * cacheable predicates on this machine. We cannot allow any
9467 * more predicates to become cacheable: as unlikely as it is,
9468 * there may be a thread caching a (now stale) predicate cache
9469 * ID. (N.B.: the temptation is being successfully resisted to
9470 * have this cmn_err() "Holy shit -- we executed this code!")
9475 pred->dtp_cacheid = dtrace_predcache_id++;
9481 dtrace_predicate_hold(dtrace_predicate_t *pred)
9483 ASSERT(MUTEX_HELD(&dtrace_lock));
9484 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9485 ASSERT(pred->dtp_refcnt > 0);
9491 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9493 dtrace_difo_t *dp = pred->dtp_difo;
9495 ASSERT(MUTEX_HELD(&dtrace_lock));
9496 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9497 ASSERT(pred->dtp_refcnt > 0);
9499 if (--pred->dtp_refcnt == 0) {
9500 dtrace_difo_release(pred->dtp_difo, vstate);
9501 kmem_free(pred, sizeof (dtrace_predicate_t));
9506 * DTrace Action Description Functions
9508 static dtrace_actdesc_t *
9509 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9510 uint64_t uarg, uint64_t arg)
9512 dtrace_actdesc_t *act;
9515 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9516 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9519 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9520 act->dtad_kind = kind;
9521 act->dtad_ntuple = ntuple;
9522 act->dtad_uarg = uarg;
9523 act->dtad_arg = arg;
9524 act->dtad_refcnt = 1;
9530 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9532 ASSERT(act->dtad_refcnt >= 1);
9537 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9539 dtrace_actkind_t kind = act->dtad_kind;
9542 ASSERT(act->dtad_refcnt >= 1);
9544 if (--act->dtad_refcnt != 0)
9547 if ((dp = act->dtad_difo) != NULL)
9548 dtrace_difo_release(dp, vstate);
9550 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9551 char *str = (char *)(uintptr_t)act->dtad_arg;
9554 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9555 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9559 kmem_free(str, strlen(str) + 1);
9562 kmem_free(act, sizeof (dtrace_actdesc_t));
9566 * DTrace ECB Functions
9568 static dtrace_ecb_t *
9569 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9574 ASSERT(MUTEX_HELD(&dtrace_lock));
9576 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9577 ecb->dte_predicate = NULL;
9578 ecb->dte_probe = probe;
9581 * The default size is the size of the default action: recording
9584 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9585 ecb->dte_alignment = sizeof (dtrace_epid_t);
9587 epid = state->dts_epid++;
9589 if (epid - 1 >= state->dts_necbs) {
9590 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9591 int necbs = state->dts_necbs << 1;
9593 ASSERT(epid == state->dts_necbs + 1);
9596 ASSERT(oecbs == NULL);
9600 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9603 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9605 dtrace_membar_producer();
9606 state->dts_ecbs = ecbs;
9608 if (oecbs != NULL) {
9610 * If this state is active, we must dtrace_sync()
9611 * before we can free the old dts_ecbs array: we're
9612 * coming in hot, and there may be active ring
9613 * buffer processing (which indexes into the dts_ecbs
9614 * array) on another CPU.
9616 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9619 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9622 dtrace_membar_producer();
9623 state->dts_necbs = necbs;
9626 ecb->dte_state = state;
9628 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9629 dtrace_membar_producer();
9630 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9636 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9638 dtrace_probe_t *probe = ecb->dte_probe;
9640 ASSERT(MUTEX_HELD(&cpu_lock));
9641 ASSERT(MUTEX_HELD(&dtrace_lock));
9642 ASSERT(ecb->dte_next == NULL);
9644 if (probe == NULL) {
9646 * This is the NULL probe -- there's nothing to do.
9651 if (probe->dtpr_ecb == NULL) {
9652 dtrace_provider_t *prov = probe->dtpr_provider;
9655 * We're the first ECB on this probe.
9657 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9659 if (ecb->dte_predicate != NULL)
9660 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9662 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9663 probe->dtpr_id, probe->dtpr_arg);
9666 * This probe is already active. Swing the last pointer to
9667 * point to the new ECB, and issue a dtrace_sync() to assure
9668 * that all CPUs have seen the change.
9670 ASSERT(probe->dtpr_ecb_last != NULL);
9671 probe->dtpr_ecb_last->dte_next = ecb;
9672 probe->dtpr_ecb_last = ecb;
9673 probe->dtpr_predcache = 0;
9680 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9682 uint32_t maxalign = sizeof (dtrace_epid_t);
9683 uint32_t align = sizeof (uint8_t), offs, diff;
9684 dtrace_action_t *act;
9686 uint32_t aggbase = UINT32_MAX;
9687 dtrace_state_t *state = ecb->dte_state;
9690 * If we record anything, we always record the epid. (And we always
9693 offs = sizeof (dtrace_epid_t);
9694 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9696 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9697 dtrace_recdesc_t *rec = &act->dta_rec;
9699 if ((align = rec->dtrd_alignment) > maxalign)
9702 if (!wastuple && act->dta_intuple) {
9704 * This is the first record in a tuple. Align the
9705 * offset to be at offset 4 in an 8-byte aligned
9708 diff = offs + sizeof (dtrace_aggid_t);
9710 if ((diff = (diff & (sizeof (uint64_t) - 1))))
9711 offs += sizeof (uint64_t) - diff;
9713 aggbase = offs - sizeof (dtrace_aggid_t);
9714 ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9718 if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9720 * The current offset is not properly aligned; align it.
9722 offs += align - diff;
9725 rec->dtrd_offset = offs;
9727 if (offs + rec->dtrd_size > ecb->dte_needed) {
9728 ecb->dte_needed = offs + rec->dtrd_size;
9730 if (ecb->dte_needed > state->dts_needed)
9731 state->dts_needed = ecb->dte_needed;
9734 if (DTRACEACT_ISAGG(act->dta_kind)) {
9735 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9736 dtrace_action_t *first = agg->dtag_first, *prev;
9738 ASSERT(rec->dtrd_size != 0 && first != NULL);
9740 ASSERT(aggbase != UINT32_MAX);
9742 agg->dtag_base = aggbase;
9744 while ((prev = first->dta_prev) != NULL &&
9745 DTRACEACT_ISAGG(prev->dta_kind)) {
9746 agg = (dtrace_aggregation_t *)prev;
9747 first = agg->dtag_first;
9751 offs = prev->dta_rec.dtrd_offset +
9752 prev->dta_rec.dtrd_size;
9754 offs = sizeof (dtrace_epid_t);
9758 if (!act->dta_intuple)
9759 ecb->dte_size = offs + rec->dtrd_size;
9761 offs += rec->dtrd_size;
9764 wastuple = act->dta_intuple;
9767 if ((act = ecb->dte_action) != NULL &&
9768 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9769 ecb->dte_size == sizeof (dtrace_epid_t)) {
9771 * If the size is still sizeof (dtrace_epid_t), then all
9772 * actions store no data; set the size to 0.
9774 ecb->dte_alignment = maxalign;
9778 * If the needed space is still sizeof (dtrace_epid_t), then
9779 * all actions need no additional space; set the needed
9782 if (ecb->dte_needed == sizeof (dtrace_epid_t))
9783 ecb->dte_needed = 0;
9789 * Set our alignment, and make sure that the dte_size and dte_needed
9790 * are aligned to the size of an EPID.
9792 ecb->dte_alignment = maxalign;
9793 ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9794 ~(sizeof (dtrace_epid_t) - 1);
9795 ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9796 ~(sizeof (dtrace_epid_t) - 1);
9797 ASSERT(ecb->dte_size <= ecb->dte_needed);
9800 static dtrace_action_t *
9801 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9803 dtrace_aggregation_t *agg;
9804 size_t size = sizeof (uint64_t);
9805 int ntuple = desc->dtad_ntuple;
9806 dtrace_action_t *act;
9807 dtrace_recdesc_t *frec;
9808 dtrace_aggid_t aggid;
9809 dtrace_state_t *state = ecb->dte_state;
9811 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9812 agg->dtag_ecb = ecb;
9814 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9816 switch (desc->dtad_kind) {
9818 agg->dtag_initial = INT64_MAX;
9819 agg->dtag_aggregate = dtrace_aggregate_min;
9823 agg->dtag_initial = INT64_MIN;
9824 agg->dtag_aggregate = dtrace_aggregate_max;
9827 case DTRACEAGG_COUNT:
9828 agg->dtag_aggregate = dtrace_aggregate_count;
9831 case DTRACEAGG_QUANTIZE:
9832 agg->dtag_aggregate = dtrace_aggregate_quantize;
9833 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9837 case DTRACEAGG_LQUANTIZE: {
9838 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9839 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9841 agg->dtag_initial = desc->dtad_arg;
9842 agg->dtag_aggregate = dtrace_aggregate_lquantize;
9844 if (step == 0 || levels == 0)
9847 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9852 agg->dtag_aggregate = dtrace_aggregate_avg;
9853 size = sizeof (uint64_t) * 2;
9856 case DTRACEAGG_STDDEV:
9857 agg->dtag_aggregate = dtrace_aggregate_stddev;
9858 size = sizeof (uint64_t) * 4;
9862 agg->dtag_aggregate = dtrace_aggregate_sum;
9869 agg->dtag_action.dta_rec.dtrd_size = size;
9875 * We must make sure that we have enough actions for the n-tuple.
9877 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9878 if (DTRACEACT_ISAGG(act->dta_kind))
9881 if (--ntuple == 0) {
9883 * This is the action with which our n-tuple begins.
9885 agg->dtag_first = act;
9891 * This n-tuple is short by ntuple elements. Return failure.
9893 ASSERT(ntuple != 0);
9895 kmem_free(agg, sizeof (dtrace_aggregation_t));
9900 * If the last action in the tuple has a size of zero, it's actually
9901 * an expression argument for the aggregating action.
9903 ASSERT(ecb->dte_action_last != NULL);
9904 act = ecb->dte_action_last;
9906 if (act->dta_kind == DTRACEACT_DIFEXPR) {
9907 ASSERT(act->dta_difo != NULL);
9909 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
9910 agg->dtag_hasarg = 1;
9914 * We need to allocate an id for this aggregation.
9917 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
9918 VM_BESTFIT | VM_SLEEP);
9920 aggid = alloc_unr(state->dts_aggid_arena);
9923 if (aggid - 1 >= state->dts_naggregations) {
9924 dtrace_aggregation_t **oaggs = state->dts_aggregations;
9925 dtrace_aggregation_t **aggs;
9926 int naggs = state->dts_naggregations << 1;
9927 int onaggs = state->dts_naggregations;
9929 ASSERT(aggid == state->dts_naggregations + 1);
9932 ASSERT(oaggs == NULL);
9936 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
9938 if (oaggs != NULL) {
9939 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
9940 kmem_free(oaggs, onaggs * sizeof (*aggs));
9943 state->dts_aggregations = aggs;
9944 state->dts_naggregations = naggs;
9947 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
9948 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
9950 frec = &agg->dtag_first->dta_rec;
9951 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
9952 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
9954 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
9955 ASSERT(!act->dta_intuple);
9956 act->dta_intuple = 1;
9959 return (&agg->dtag_action);
9963 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
9965 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9966 dtrace_state_t *state = ecb->dte_state;
9967 dtrace_aggid_t aggid = agg->dtag_id;
9969 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
9971 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
9973 free_unr(state->dts_aggid_arena, aggid);
9976 ASSERT(state->dts_aggregations[aggid - 1] == agg);
9977 state->dts_aggregations[aggid - 1] = NULL;
9979 kmem_free(agg, sizeof (dtrace_aggregation_t));
9983 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9985 dtrace_action_t *action, *last;
9986 dtrace_difo_t *dp = desc->dtad_difo;
9987 uint32_t size = 0, align = sizeof (uint8_t), mask;
9988 uint16_t format = 0;
9989 dtrace_recdesc_t *rec;
9990 dtrace_state_t *state = ecb->dte_state;
9991 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
9992 uint64_t arg = desc->dtad_arg;
9994 ASSERT(MUTEX_HELD(&dtrace_lock));
9995 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
9997 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
9999 * If this is an aggregating action, there must be neither
10000 * a speculate nor a commit on the action chain.
10002 dtrace_action_t *act;
10004 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10005 if (act->dta_kind == DTRACEACT_COMMIT)
10008 if (act->dta_kind == DTRACEACT_SPECULATE)
10012 action = dtrace_ecb_aggregation_create(ecb, desc);
10014 if (action == NULL)
10017 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10018 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10019 dp != NULL && dp->dtdo_destructive)) {
10020 state->dts_destructive = 1;
10023 switch (desc->dtad_kind) {
10024 case DTRACEACT_PRINTF:
10025 case DTRACEACT_PRINTA:
10026 case DTRACEACT_SYSTEM:
10027 case DTRACEACT_FREOPEN:
10029 * We know that our arg is a string -- turn it into a
10033 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA);
10038 ASSERT(arg > KERNELBASE);
10040 format = dtrace_format_add(state,
10041 (char *)(uintptr_t)arg);
10045 case DTRACEACT_LIBACT:
10046 case DTRACEACT_DIFEXPR:
10050 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10053 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10054 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10057 size = opt[DTRACEOPT_STRSIZE];
10062 case DTRACEACT_STACK:
10063 if ((nframes = arg) == 0) {
10064 nframes = opt[DTRACEOPT_STACKFRAMES];
10065 ASSERT(nframes > 0);
10069 size = nframes * sizeof (pc_t);
10072 case DTRACEACT_JSTACK:
10073 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10074 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10076 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10077 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10079 arg = DTRACE_USTACK_ARG(nframes, strsize);
10082 case DTRACEACT_USTACK:
10083 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10084 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10085 strsize = DTRACE_USTACK_STRSIZE(arg);
10086 nframes = opt[DTRACEOPT_USTACKFRAMES];
10087 ASSERT(nframes > 0);
10088 arg = DTRACE_USTACK_ARG(nframes, strsize);
10092 * Save a slot for the pid.
10094 size = (nframes + 1) * sizeof (uint64_t);
10095 size += DTRACE_USTACK_STRSIZE(arg);
10096 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10100 case DTRACEACT_SYM:
10101 case DTRACEACT_MOD:
10102 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10103 sizeof (uint64_t)) ||
10104 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10108 case DTRACEACT_USYM:
10109 case DTRACEACT_UMOD:
10110 case DTRACEACT_UADDR:
10112 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10113 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10117 * We have a slot for the pid, plus a slot for the
10118 * argument. To keep things simple (aligned with
10119 * bitness-neutral sizing), we store each as a 64-bit
10122 size = 2 * sizeof (uint64_t);
10125 case DTRACEACT_STOP:
10126 case DTRACEACT_BREAKPOINT:
10127 case DTRACEACT_PANIC:
10130 case DTRACEACT_CHILL:
10131 case DTRACEACT_DISCARD:
10132 case DTRACEACT_RAISE:
10137 case DTRACEACT_EXIT:
10139 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10140 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10144 case DTRACEACT_SPECULATE:
10145 if (ecb->dte_size > sizeof (dtrace_epid_t))
10151 state->dts_speculates = 1;
10154 case DTRACEACT_PRINTM:
10155 size = dp->dtdo_rtype.dtdt_size;
10158 case DTRACEACT_PRINTT:
10159 size = dp->dtdo_rtype.dtdt_size;
10162 case DTRACEACT_COMMIT: {
10163 dtrace_action_t *act = ecb->dte_action;
10165 for (; act != NULL; act = act->dta_next) {
10166 if (act->dta_kind == DTRACEACT_COMMIT)
10179 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10181 * If this is a data-storing action or a speculate,
10182 * we must be sure that there isn't a commit on the
10185 dtrace_action_t *act = ecb->dte_action;
10187 for (; act != NULL; act = act->dta_next) {
10188 if (act->dta_kind == DTRACEACT_COMMIT)
10193 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10194 action->dta_rec.dtrd_size = size;
10197 action->dta_refcnt = 1;
10198 rec = &action->dta_rec;
10199 size = rec->dtrd_size;
10201 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10202 if (!(size & mask)) {
10208 action->dta_kind = desc->dtad_kind;
10210 if ((action->dta_difo = dp) != NULL)
10211 dtrace_difo_hold(dp);
10213 rec->dtrd_action = action->dta_kind;
10214 rec->dtrd_arg = arg;
10215 rec->dtrd_uarg = desc->dtad_uarg;
10216 rec->dtrd_alignment = (uint16_t)align;
10217 rec->dtrd_format = format;
10219 if ((last = ecb->dte_action_last) != NULL) {
10220 ASSERT(ecb->dte_action != NULL);
10221 action->dta_prev = last;
10222 last->dta_next = action;
10224 ASSERT(ecb->dte_action == NULL);
10225 ecb->dte_action = action;
10228 ecb->dte_action_last = action;
10234 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10236 dtrace_action_t *act = ecb->dte_action, *next;
10237 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10241 if (act != NULL && act->dta_refcnt > 1) {
10242 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10245 for (; act != NULL; act = next) {
10246 next = act->dta_next;
10247 ASSERT(next != NULL || act == ecb->dte_action_last);
10248 ASSERT(act->dta_refcnt == 1);
10250 if ((format = act->dta_rec.dtrd_format) != 0)
10251 dtrace_format_remove(ecb->dte_state, format);
10253 if ((dp = act->dta_difo) != NULL)
10254 dtrace_difo_release(dp, vstate);
10256 if (DTRACEACT_ISAGG(act->dta_kind)) {
10257 dtrace_ecb_aggregation_destroy(ecb, act);
10259 kmem_free(act, sizeof (dtrace_action_t));
10264 ecb->dte_action = NULL;
10265 ecb->dte_action_last = NULL;
10266 ecb->dte_size = sizeof (dtrace_epid_t);
10270 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10273 * We disable the ECB by removing it from its probe.
10275 dtrace_ecb_t *pecb, *prev = NULL;
10276 dtrace_probe_t *probe = ecb->dte_probe;
10278 ASSERT(MUTEX_HELD(&dtrace_lock));
10280 if (probe == NULL) {
10282 * This is the NULL probe; there is nothing to disable.
10287 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10293 ASSERT(pecb != NULL);
10295 if (prev == NULL) {
10296 probe->dtpr_ecb = ecb->dte_next;
10298 prev->dte_next = ecb->dte_next;
10301 if (ecb == probe->dtpr_ecb_last) {
10302 ASSERT(ecb->dte_next == NULL);
10303 probe->dtpr_ecb_last = prev;
10307 * The ECB has been disconnected from the probe; now sync to assure
10308 * that all CPUs have seen the change before returning.
10312 if (probe->dtpr_ecb == NULL) {
10314 * That was the last ECB on the probe; clear the predicate
10315 * cache ID for the probe, disable it and sync one more time
10316 * to assure that we'll never hit it again.
10318 dtrace_provider_t *prov = probe->dtpr_provider;
10320 ASSERT(ecb->dte_next == NULL);
10321 ASSERT(probe->dtpr_ecb_last == NULL);
10322 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10323 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10324 probe->dtpr_id, probe->dtpr_arg);
10328 * There is at least one ECB remaining on the probe. If there
10329 * is _exactly_ one, set the probe's predicate cache ID to be
10330 * the predicate cache ID of the remaining ECB.
10332 ASSERT(probe->dtpr_ecb_last != NULL);
10333 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10335 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10336 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10338 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10341 probe->dtpr_predcache = p->dtp_cacheid;
10344 ecb->dte_next = NULL;
10349 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10351 dtrace_state_t *state = ecb->dte_state;
10352 dtrace_vstate_t *vstate = &state->dts_vstate;
10353 dtrace_predicate_t *pred;
10354 dtrace_epid_t epid = ecb->dte_epid;
10356 ASSERT(MUTEX_HELD(&dtrace_lock));
10357 ASSERT(ecb->dte_next == NULL);
10358 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10360 if ((pred = ecb->dte_predicate) != NULL)
10361 dtrace_predicate_release(pred, vstate);
10363 dtrace_ecb_action_remove(ecb);
10365 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10366 state->dts_ecbs[epid - 1] = NULL;
10368 kmem_free(ecb, sizeof (dtrace_ecb_t));
10371 static dtrace_ecb_t *
10372 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10373 dtrace_enabling_t *enab)
10376 dtrace_predicate_t *pred;
10377 dtrace_actdesc_t *act;
10378 dtrace_provider_t *prov;
10379 dtrace_ecbdesc_t *desc = enab->dten_current;
10381 ASSERT(MUTEX_HELD(&dtrace_lock));
10382 ASSERT(state != NULL);
10384 ecb = dtrace_ecb_add(state, probe);
10385 ecb->dte_uarg = desc->dted_uarg;
10387 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10388 dtrace_predicate_hold(pred);
10389 ecb->dte_predicate = pred;
10392 if (probe != NULL) {
10394 * If the provider shows more leg than the consumer is old
10395 * enough to see, we need to enable the appropriate implicit
10396 * predicate bits to prevent the ecb from activating at
10399 * Providers specifying DTRACE_PRIV_USER at register time
10400 * are stating that they need the /proc-style privilege
10401 * model to be enforced, and this is what DTRACE_COND_OWNER
10402 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10404 prov = probe->dtpr_provider;
10405 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10406 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10407 ecb->dte_cond |= DTRACE_COND_OWNER;
10409 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10410 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10411 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10414 * If the provider shows us kernel innards and the user
10415 * is lacking sufficient privilege, enable the
10416 * DTRACE_COND_USERMODE implicit predicate.
10418 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10419 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10420 ecb->dte_cond |= DTRACE_COND_USERMODE;
10423 if (dtrace_ecb_create_cache != NULL) {
10425 * If we have a cached ecb, we'll use its action list instead
10426 * of creating our own (saving both time and space).
10428 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10429 dtrace_action_t *act = cached->dte_action;
10432 ASSERT(act->dta_refcnt > 0);
10434 ecb->dte_action = act;
10435 ecb->dte_action_last = cached->dte_action_last;
10436 ecb->dte_needed = cached->dte_needed;
10437 ecb->dte_size = cached->dte_size;
10438 ecb->dte_alignment = cached->dte_alignment;
10444 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10445 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10446 dtrace_ecb_destroy(ecb);
10451 dtrace_ecb_resize(ecb);
10453 return (dtrace_ecb_create_cache = ecb);
10457 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10460 dtrace_enabling_t *enab = arg;
10461 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10463 ASSERT(state != NULL);
10465 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10467 * This probe was created in a generation for which this
10468 * enabling has previously created ECBs; we don't want to
10469 * enable it again, so just kick out.
10471 return (DTRACE_MATCH_NEXT);
10474 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10475 return (DTRACE_MATCH_DONE);
10477 dtrace_ecb_enable(ecb);
10478 return (DTRACE_MATCH_NEXT);
10481 static dtrace_ecb_t *
10482 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10486 ASSERT(MUTEX_HELD(&dtrace_lock));
10488 if (id == 0 || id > state->dts_necbs)
10491 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10492 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10494 return (state->dts_ecbs[id - 1]);
10497 static dtrace_aggregation_t *
10498 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10500 dtrace_aggregation_t *agg;
10502 ASSERT(MUTEX_HELD(&dtrace_lock));
10504 if (id == 0 || id > state->dts_naggregations)
10507 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10508 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10509 agg->dtag_id == id);
10511 return (state->dts_aggregations[id - 1]);
10515 * DTrace Buffer Functions
10517 * The following functions manipulate DTrace buffers. Most of these functions
10518 * are called in the context of establishing or processing consumer state;
10519 * exceptions are explicitly noted.
10523 * Note: called from cross call context. This function switches the two
10524 * buffers on a given CPU. The atomicity of this operation is assured by
10525 * disabling interrupts while the actual switch takes place; the disabling of
10526 * interrupts serializes the execution with any execution of dtrace_probe() on
10530 dtrace_buffer_switch(dtrace_buffer_t *buf)
10532 caddr_t tomax = buf->dtb_tomax;
10533 caddr_t xamot = buf->dtb_xamot;
10534 dtrace_icookie_t cookie;
10536 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10537 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10539 cookie = dtrace_interrupt_disable();
10540 buf->dtb_tomax = xamot;
10541 buf->dtb_xamot = tomax;
10542 buf->dtb_xamot_drops = buf->dtb_drops;
10543 buf->dtb_xamot_offset = buf->dtb_offset;
10544 buf->dtb_xamot_errors = buf->dtb_errors;
10545 buf->dtb_xamot_flags = buf->dtb_flags;
10546 buf->dtb_offset = 0;
10547 buf->dtb_drops = 0;
10548 buf->dtb_errors = 0;
10549 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10550 dtrace_interrupt_enable(cookie);
10554 * Note: called from cross call context. This function activates a buffer
10555 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10556 * is guaranteed by the disabling of interrupts.
10559 dtrace_buffer_activate(dtrace_state_t *state)
10561 dtrace_buffer_t *buf;
10562 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10564 buf = &state->dts_buffer[curcpu];
10566 if (buf->dtb_tomax != NULL) {
10568 * We might like to assert that the buffer is marked inactive,
10569 * but this isn't necessarily true: the buffer for the CPU
10570 * that processes the BEGIN probe has its buffer activated
10571 * manually. In this case, we take the (harmless) action
10572 * re-clearing the bit INACTIVE bit.
10574 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10577 dtrace_interrupt_enable(cookie);
10581 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10589 dtrace_buffer_t *buf;
10592 ASSERT(MUTEX_HELD(&cpu_lock));
10593 ASSERT(MUTEX_HELD(&dtrace_lock));
10595 if (size > dtrace_nonroot_maxsize &&
10596 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10602 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10605 buf = &bufs[cp->cpu_id];
10608 * If there is already a buffer allocated for this CPU, it
10609 * is only possible that this is a DR event. In this case,
10611 if (buf->dtb_tomax != NULL) {
10612 ASSERT(buf->dtb_size == size);
10616 ASSERT(buf->dtb_xamot == NULL);
10618 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10621 buf->dtb_size = size;
10622 buf->dtb_flags = flags;
10623 buf->dtb_offset = 0;
10624 buf->dtb_drops = 0;
10626 if (flags & DTRACEBUF_NOSWITCH)
10629 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10631 } while ((cp = cp->cpu_next) != cpu_list);
10639 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10642 buf = &bufs[cp->cpu_id];
10644 if (buf->dtb_xamot != NULL) {
10645 ASSERT(buf->dtb_tomax != NULL);
10646 ASSERT(buf->dtb_size == size);
10647 kmem_free(buf->dtb_xamot, size);
10650 if (buf->dtb_tomax != NULL) {
10651 ASSERT(buf->dtb_size == size);
10652 kmem_free(buf->dtb_tomax, size);
10655 buf->dtb_tomax = NULL;
10656 buf->dtb_xamot = NULL;
10658 } while ((cp = cp->cpu_next) != cpu_list);
10664 #if defined(__amd64__)
10666 * FreeBSD isn't good at limiting the amount of memory we
10667 * ask to malloc, so let's place a limit here before trying
10668 * to do something that might well end in tears at bedtime.
10670 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10674 ASSERT(MUTEX_HELD(&dtrace_lock));
10675 for (i = 0; i <= mp_maxid; i++) {
10676 if ((cp = pcpu_find(i)) == NULL)
10679 if (cpu != DTRACE_CPUALL && cpu != i)
10685 * If there is already a buffer allocated for this CPU, it
10686 * is only possible that this is a DR event. In this case,
10687 * the buffer size must match our specified size.
10689 if (buf->dtb_tomax != NULL) {
10690 ASSERT(buf->dtb_size == size);
10694 ASSERT(buf->dtb_xamot == NULL);
10696 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10699 buf->dtb_size = size;
10700 buf->dtb_flags = flags;
10701 buf->dtb_offset = 0;
10702 buf->dtb_drops = 0;
10704 if (flags & DTRACEBUF_NOSWITCH)
10707 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10715 * Error allocating memory, so free the buffers that were
10716 * allocated before the failed allocation.
10718 for (i = 0; i <= mp_maxid; i++) {
10719 if ((cp = pcpu_find(i)) == NULL)
10722 if (cpu != DTRACE_CPUALL && cpu != i)
10727 if (buf->dtb_xamot != NULL) {
10728 ASSERT(buf->dtb_tomax != NULL);
10729 ASSERT(buf->dtb_size == size);
10730 kmem_free(buf->dtb_xamot, size);
10733 if (buf->dtb_tomax != NULL) {
10734 ASSERT(buf->dtb_size == size);
10735 kmem_free(buf->dtb_tomax, size);
10738 buf->dtb_tomax = NULL;
10739 buf->dtb_xamot = NULL;
10749 * Note: called from probe context. This function just increments the drop
10750 * count on a buffer. It has been made a function to allow for the
10751 * possibility of understanding the source of mysterious drop counts. (A
10752 * problem for which one may be particularly disappointed that DTrace cannot
10753 * be used to understand DTrace.)
10756 dtrace_buffer_drop(dtrace_buffer_t *buf)
10762 * Note: called from probe context. This function is called to reserve space
10763 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
10764 * mstate. Returns the new offset in the buffer, or a negative value if an
10765 * error has occurred.
10768 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10769 dtrace_state_t *state, dtrace_mstate_t *mstate)
10771 intptr_t offs = buf->dtb_offset, soffs;
10776 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10779 if ((tomax = buf->dtb_tomax) == NULL) {
10780 dtrace_buffer_drop(buf);
10784 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10785 while (offs & (align - 1)) {
10787 * Assert that our alignment is off by a number which
10788 * is itself sizeof (uint32_t) aligned.
10790 ASSERT(!((align - (offs & (align - 1))) &
10791 (sizeof (uint32_t) - 1)));
10792 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10793 offs += sizeof (uint32_t);
10796 if ((soffs = offs + needed) > buf->dtb_size) {
10797 dtrace_buffer_drop(buf);
10801 if (mstate == NULL)
10804 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10805 mstate->dtms_scratch_size = buf->dtb_size - soffs;
10806 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10811 if (buf->dtb_flags & DTRACEBUF_FILL) {
10812 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10813 (buf->dtb_flags & DTRACEBUF_FULL))
10818 total = needed + (offs & (align - 1));
10821 * For a ring buffer, life is quite a bit more complicated. Before
10822 * we can store any padding, we need to adjust our wrapping offset.
10823 * (If we've never before wrapped or we're not about to, no adjustment
10826 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10827 offs + total > buf->dtb_size) {
10828 woffs = buf->dtb_xamot_offset;
10830 if (offs + total > buf->dtb_size) {
10832 * We can't fit in the end of the buffer. First, a
10833 * sanity check that we can fit in the buffer at all.
10835 if (total > buf->dtb_size) {
10836 dtrace_buffer_drop(buf);
10841 * We're going to be storing at the top of the buffer,
10842 * so now we need to deal with the wrapped offset. We
10843 * only reset our wrapped offset to 0 if it is
10844 * currently greater than the current offset. If it
10845 * is less than the current offset, it is because a
10846 * previous allocation induced a wrap -- but the
10847 * allocation didn't subsequently take the space due
10848 * to an error or false predicate evaluation. In this
10849 * case, we'll just leave the wrapped offset alone: if
10850 * the wrapped offset hasn't been advanced far enough
10851 * for this allocation, it will be adjusted in the
10854 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10862 * Now we know that we're going to be storing to the
10863 * top of the buffer and that there is room for us
10864 * there. We need to clear the buffer from the current
10865 * offset to the end (there may be old gunk there).
10867 while (offs < buf->dtb_size)
10871 * We need to set our offset to zero. And because we
10872 * are wrapping, we need to set the bit indicating as
10873 * much. We can also adjust our needed space back
10874 * down to the space required by the ECB -- we know
10875 * that the top of the buffer is aligned.
10879 buf->dtb_flags |= DTRACEBUF_WRAPPED;
10882 * There is room for us in the buffer, so we simply
10883 * need to check the wrapped offset.
10885 if (woffs < offs) {
10887 * The wrapped offset is less than the offset.
10888 * This can happen if we allocated buffer space
10889 * that induced a wrap, but then we didn't
10890 * subsequently take the space due to an error
10891 * or false predicate evaluation. This is
10892 * okay; we know that _this_ allocation isn't
10893 * going to induce a wrap. We still can't
10894 * reset the wrapped offset to be zero,
10895 * however: the space may have been trashed in
10896 * the previous failed probe attempt. But at
10897 * least the wrapped offset doesn't need to
10898 * be adjusted at all...
10904 while (offs + total > woffs) {
10905 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
10908 if (epid == DTRACE_EPIDNONE) {
10909 size = sizeof (uint32_t);
10911 ASSERT(epid <= state->dts_necbs);
10912 ASSERT(state->dts_ecbs[epid - 1] != NULL);
10914 size = state->dts_ecbs[epid - 1]->dte_size;
10917 ASSERT(woffs + size <= buf->dtb_size);
10920 if (woffs + size == buf->dtb_size) {
10922 * We've reached the end of the buffer; we want
10923 * to set the wrapped offset to 0 and break
10924 * out. However, if the offs is 0, then we're
10925 * in a strange edge-condition: the amount of
10926 * space that we want to reserve plus the size
10927 * of the record that we're overwriting is
10928 * greater than the size of the buffer. This
10929 * is problematic because if we reserve the
10930 * space but subsequently don't consume it (due
10931 * to a failed predicate or error) the wrapped
10932 * offset will be 0 -- yet the EPID at offset 0
10933 * will not be committed. This situation is
10934 * relatively easy to deal with: if we're in
10935 * this case, the buffer is indistinguishable
10936 * from one that hasn't wrapped; we need only
10937 * finish the job by clearing the wrapped bit,
10938 * explicitly setting the offset to be 0, and
10939 * zero'ing out the old data in the buffer.
10942 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
10943 buf->dtb_offset = 0;
10946 while (woffs < buf->dtb_size)
10947 tomax[woffs++] = 0;
10958 * We have a wrapped offset. It may be that the wrapped offset
10959 * has become zero -- that's okay.
10961 buf->dtb_xamot_offset = woffs;
10966 * Now we can plow the buffer with any necessary padding.
10968 while (offs & (align - 1)) {
10970 * Assert that our alignment is off by a number which
10971 * is itself sizeof (uint32_t) aligned.
10973 ASSERT(!((align - (offs & (align - 1))) &
10974 (sizeof (uint32_t) - 1)));
10975 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10976 offs += sizeof (uint32_t);
10979 if (buf->dtb_flags & DTRACEBUF_FILL) {
10980 if (offs + needed > buf->dtb_size - state->dts_reserve) {
10981 buf->dtb_flags |= DTRACEBUF_FULL;
10986 if (mstate == NULL)
10990 * For ring buffers and fill buffers, the scratch space is always
10991 * the inactive buffer.
10993 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
10994 mstate->dtms_scratch_size = buf->dtb_size;
10995 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11001 dtrace_buffer_polish(dtrace_buffer_t *buf)
11003 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11004 ASSERT(MUTEX_HELD(&dtrace_lock));
11006 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11010 * We need to polish the ring buffer. There are three cases:
11012 * - The first (and presumably most common) is that there is no gap
11013 * between the buffer offset and the wrapped offset. In this case,
11014 * there is nothing in the buffer that isn't valid data; we can
11015 * mark the buffer as polished and return.
11017 * - The second (less common than the first but still more common
11018 * than the third) is that there is a gap between the buffer offset
11019 * and the wrapped offset, and the wrapped offset is larger than the
11020 * buffer offset. This can happen because of an alignment issue, or
11021 * can happen because of a call to dtrace_buffer_reserve() that
11022 * didn't subsequently consume the buffer space. In this case,
11023 * we need to zero the data from the buffer offset to the wrapped
11026 * - The third (and least common) is that there is a gap between the
11027 * buffer offset and the wrapped offset, but the wrapped offset is
11028 * _less_ than the buffer offset. This can only happen because a
11029 * call to dtrace_buffer_reserve() induced a wrap, but the space
11030 * was not subsequently consumed. In this case, we need to zero the
11031 * space from the offset to the end of the buffer _and_ from the
11032 * top of the buffer to the wrapped offset.
11034 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11035 bzero(buf->dtb_tomax + buf->dtb_offset,
11036 buf->dtb_xamot_offset - buf->dtb_offset);
11039 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11040 bzero(buf->dtb_tomax + buf->dtb_offset,
11041 buf->dtb_size - buf->dtb_offset);
11042 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11047 dtrace_buffer_free(dtrace_buffer_t *bufs)
11051 for (i = 0; i < NCPU; i++) {
11052 dtrace_buffer_t *buf = &bufs[i];
11054 if (buf->dtb_tomax == NULL) {
11055 ASSERT(buf->dtb_xamot == NULL);
11056 ASSERT(buf->dtb_size == 0);
11060 if (buf->dtb_xamot != NULL) {
11061 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11062 kmem_free(buf->dtb_xamot, buf->dtb_size);
11065 kmem_free(buf->dtb_tomax, buf->dtb_size);
11067 buf->dtb_tomax = NULL;
11068 buf->dtb_xamot = NULL;
11073 * DTrace Enabling Functions
11075 static dtrace_enabling_t *
11076 dtrace_enabling_create(dtrace_vstate_t *vstate)
11078 dtrace_enabling_t *enab;
11080 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11081 enab->dten_vstate = vstate;
11087 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11089 dtrace_ecbdesc_t **ndesc;
11090 size_t osize, nsize;
11093 * We can't add to enablings after we've enabled them, or after we've
11096 ASSERT(enab->dten_probegen == 0);
11097 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11099 if (enab->dten_ndesc < enab->dten_maxdesc) {
11100 enab->dten_desc[enab->dten_ndesc++] = ecb;
11104 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11106 if (enab->dten_maxdesc == 0) {
11107 enab->dten_maxdesc = 1;
11109 enab->dten_maxdesc <<= 1;
11112 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11114 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11115 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11116 bcopy(enab->dten_desc, ndesc, osize);
11117 if (enab->dten_desc != NULL)
11118 kmem_free(enab->dten_desc, osize);
11120 enab->dten_desc = ndesc;
11121 enab->dten_desc[enab->dten_ndesc++] = ecb;
11125 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11126 dtrace_probedesc_t *pd)
11128 dtrace_ecbdesc_t *new;
11129 dtrace_predicate_t *pred;
11130 dtrace_actdesc_t *act;
11133 * We're going to create a new ECB description that matches the
11134 * specified ECB in every way, but has the specified probe description.
11136 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11138 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11139 dtrace_predicate_hold(pred);
11141 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11142 dtrace_actdesc_hold(act);
11144 new->dted_action = ecb->dted_action;
11145 new->dted_pred = ecb->dted_pred;
11146 new->dted_probe = *pd;
11147 new->dted_uarg = ecb->dted_uarg;
11149 dtrace_enabling_add(enab, new);
11153 dtrace_enabling_dump(dtrace_enabling_t *enab)
11157 for (i = 0; i < enab->dten_ndesc; i++) {
11158 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11160 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11161 desc->dtpd_provider, desc->dtpd_mod,
11162 desc->dtpd_func, desc->dtpd_name);
11167 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11170 dtrace_ecbdesc_t *ep;
11171 dtrace_vstate_t *vstate = enab->dten_vstate;
11173 ASSERT(MUTEX_HELD(&dtrace_lock));
11175 for (i = 0; i < enab->dten_ndesc; i++) {
11176 dtrace_actdesc_t *act, *next;
11177 dtrace_predicate_t *pred;
11179 ep = enab->dten_desc[i];
11181 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11182 dtrace_predicate_release(pred, vstate);
11184 for (act = ep->dted_action; act != NULL; act = next) {
11185 next = act->dtad_next;
11186 dtrace_actdesc_release(act, vstate);
11189 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11192 if (enab->dten_desc != NULL)
11193 kmem_free(enab->dten_desc,
11194 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11197 * If this was a retained enabling, decrement the dts_nretained count
11198 * and take it off of the dtrace_retained list.
11200 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11201 dtrace_retained == enab) {
11202 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11203 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11204 enab->dten_vstate->dtvs_state->dts_nretained--;
11207 if (enab->dten_prev == NULL) {
11208 if (dtrace_retained == enab) {
11209 dtrace_retained = enab->dten_next;
11211 if (dtrace_retained != NULL)
11212 dtrace_retained->dten_prev = NULL;
11215 ASSERT(enab != dtrace_retained);
11216 ASSERT(dtrace_retained != NULL);
11217 enab->dten_prev->dten_next = enab->dten_next;
11220 if (enab->dten_next != NULL) {
11221 ASSERT(dtrace_retained != NULL);
11222 enab->dten_next->dten_prev = enab->dten_prev;
11225 kmem_free(enab, sizeof (dtrace_enabling_t));
11229 dtrace_enabling_retain(dtrace_enabling_t *enab)
11231 dtrace_state_t *state;
11233 ASSERT(MUTEX_HELD(&dtrace_lock));
11234 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11235 ASSERT(enab->dten_vstate != NULL);
11237 state = enab->dten_vstate->dtvs_state;
11238 ASSERT(state != NULL);
11241 * We only allow each state to retain dtrace_retain_max enablings.
11243 if (state->dts_nretained >= dtrace_retain_max)
11246 state->dts_nretained++;
11248 if (dtrace_retained == NULL) {
11249 dtrace_retained = enab;
11253 enab->dten_next = dtrace_retained;
11254 dtrace_retained->dten_prev = enab;
11255 dtrace_retained = enab;
11261 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11262 dtrace_probedesc_t *create)
11264 dtrace_enabling_t *new, *enab;
11265 int found = 0, err = ENOENT;
11267 ASSERT(MUTEX_HELD(&dtrace_lock));
11268 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11269 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11270 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11271 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11273 new = dtrace_enabling_create(&state->dts_vstate);
11276 * Iterate over all retained enablings, looking for enablings that
11277 * match the specified state.
11279 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11283 * dtvs_state can only be NULL for helper enablings -- and
11284 * helper enablings can't be retained.
11286 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11288 if (enab->dten_vstate->dtvs_state != state)
11292 * Now iterate over each probe description; we're looking for
11293 * an exact match to the specified probe description.
11295 for (i = 0; i < enab->dten_ndesc; i++) {
11296 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11297 dtrace_probedesc_t *pd = &ep->dted_probe;
11299 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11302 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11305 if (strcmp(pd->dtpd_func, match->dtpd_func))
11308 if (strcmp(pd->dtpd_name, match->dtpd_name))
11312 * We have a winning probe! Add it to our growing
11316 dtrace_enabling_addlike(new, ep, create);
11320 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11321 dtrace_enabling_destroy(new);
11329 dtrace_enabling_retract(dtrace_state_t *state)
11331 dtrace_enabling_t *enab, *next;
11333 ASSERT(MUTEX_HELD(&dtrace_lock));
11336 * Iterate over all retained enablings, destroy the enablings retained
11337 * for the specified state.
11339 for (enab = dtrace_retained; enab != NULL; enab = next) {
11340 next = enab->dten_next;
11343 * dtvs_state can only be NULL for helper enablings -- and
11344 * helper enablings can't be retained.
11346 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11348 if (enab->dten_vstate->dtvs_state == state) {
11349 ASSERT(state->dts_nretained > 0);
11350 dtrace_enabling_destroy(enab);
11354 ASSERT(state->dts_nretained == 0);
11358 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11363 ASSERT(MUTEX_HELD(&cpu_lock));
11364 ASSERT(MUTEX_HELD(&dtrace_lock));
11366 for (i = 0; i < enab->dten_ndesc; i++) {
11367 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11369 enab->dten_current = ep;
11370 enab->dten_error = 0;
11372 matched += dtrace_probe_enable(&ep->dted_probe, enab);
11374 if (enab->dten_error != 0) {
11376 * If we get an error half-way through enabling the
11377 * probes, we kick out -- perhaps with some number of
11378 * them enabled. Leaving enabled probes enabled may
11379 * be slightly confusing for user-level, but we expect
11380 * that no one will attempt to actually drive on in
11381 * the face of such errors. If this is an anonymous
11382 * enabling (indicated with a NULL nmatched pointer),
11383 * we cmn_err() a message. We aren't expecting to
11384 * get such an error -- such as it can exist at all,
11385 * it would be a result of corrupted DOF in the driver
11388 if (nmatched == NULL) {
11389 cmn_err(CE_WARN, "dtrace_enabling_match() "
11390 "error on %p: %d", (void *)ep,
11394 return (enab->dten_error);
11398 enab->dten_probegen = dtrace_probegen;
11399 if (nmatched != NULL)
11400 *nmatched = matched;
11406 dtrace_enabling_matchall(void)
11408 dtrace_enabling_t *enab;
11410 mutex_enter(&cpu_lock);
11411 mutex_enter(&dtrace_lock);
11414 * Iterate over all retained enablings to see if any probes match
11415 * against them. We only perform this operation on enablings for which
11416 * we have sufficient permissions by virtue of being in the global zone
11417 * or in the same zone as the DTrace client. Because we can be called
11418 * after dtrace_detach() has been called, we cannot assert that there
11419 * are retained enablings. We can safely load from dtrace_retained,
11420 * however: the taskq_destroy() at the end of dtrace_detach() will
11421 * block pending our completion.
11423 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11425 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11427 if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11429 (void) dtrace_enabling_match(enab, NULL);
11432 mutex_exit(&dtrace_lock);
11433 mutex_exit(&cpu_lock);
11437 * If an enabling is to be enabled without having matched probes (that is, if
11438 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11439 * enabling must be _primed_ by creating an ECB for every ECB description.
11440 * This must be done to assure that we know the number of speculations, the
11441 * number of aggregations, the minimum buffer size needed, etc. before we
11442 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11443 * enabling any probes, we create ECBs for every ECB decription, but with a
11444 * NULL probe -- which is exactly what this function does.
11447 dtrace_enabling_prime(dtrace_state_t *state)
11449 dtrace_enabling_t *enab;
11452 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11453 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11455 if (enab->dten_vstate->dtvs_state != state)
11459 * We don't want to prime an enabling more than once, lest
11460 * we allow a malicious user to induce resource exhaustion.
11461 * (The ECBs that result from priming an enabling aren't
11462 * leaked -- but they also aren't deallocated until the
11463 * consumer state is destroyed.)
11465 if (enab->dten_primed)
11468 for (i = 0; i < enab->dten_ndesc; i++) {
11469 enab->dten_current = enab->dten_desc[i];
11470 (void) dtrace_probe_enable(NULL, enab);
11473 enab->dten_primed = 1;
11478 * Called to indicate that probes should be provided due to retained
11479 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11480 * must take an initial lap through the enabling calling the dtps_provide()
11481 * entry point explicitly to allow for autocreated probes.
11484 dtrace_enabling_provide(dtrace_provider_t *prv)
11487 dtrace_probedesc_t desc;
11489 ASSERT(MUTEX_HELD(&dtrace_lock));
11490 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11494 prv = dtrace_provider;
11498 dtrace_enabling_t *enab = dtrace_retained;
11499 void *parg = prv->dtpv_arg;
11501 for (; enab != NULL; enab = enab->dten_next) {
11502 for (i = 0; i < enab->dten_ndesc; i++) {
11503 desc = enab->dten_desc[i]->dted_probe;
11504 mutex_exit(&dtrace_lock);
11505 prv->dtpv_pops.dtps_provide(parg, &desc);
11506 mutex_enter(&dtrace_lock);
11509 } while (all && (prv = prv->dtpv_next) != NULL);
11511 mutex_exit(&dtrace_lock);
11512 dtrace_probe_provide(NULL, all ? NULL : prv);
11513 mutex_enter(&dtrace_lock);
11517 * DTrace DOF Functions
11521 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11523 if (dtrace_err_verbose)
11524 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11526 #ifdef DTRACE_ERRDEBUG
11527 dtrace_errdebug(str);
11532 * Create DOF out of a currently enabled state. Right now, we only create
11533 * DOF containing the run-time options -- but this could be expanded to create
11534 * complete DOF representing the enabled state.
11537 dtrace_dof_create(dtrace_state_t *state)
11541 dof_optdesc_t *opt;
11542 int i, len = sizeof (dof_hdr_t) +
11543 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11544 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11546 ASSERT(MUTEX_HELD(&dtrace_lock));
11548 dof = kmem_zalloc(len, KM_SLEEP);
11549 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11550 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11551 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11552 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11554 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11555 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11556 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11557 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11558 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11559 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11561 dof->dofh_flags = 0;
11562 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11563 dof->dofh_secsize = sizeof (dof_sec_t);
11564 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11565 dof->dofh_secoff = sizeof (dof_hdr_t);
11566 dof->dofh_loadsz = len;
11567 dof->dofh_filesz = len;
11571 * Fill in the option section header...
11573 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11574 sec->dofs_type = DOF_SECT_OPTDESC;
11575 sec->dofs_align = sizeof (uint64_t);
11576 sec->dofs_flags = DOF_SECF_LOAD;
11577 sec->dofs_entsize = sizeof (dof_optdesc_t);
11579 opt = (dof_optdesc_t *)((uintptr_t)sec +
11580 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11582 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11583 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11585 for (i = 0; i < DTRACEOPT_MAX; i++) {
11586 opt[i].dofo_option = i;
11587 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11588 opt[i].dofo_value = state->dts_options[i];
11595 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11597 dof_hdr_t hdr, *dof;
11599 ASSERT(!MUTEX_HELD(&dtrace_lock));
11602 * First, we're going to copyin() the sizeof (dof_hdr_t).
11604 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11605 dtrace_dof_error(NULL, "failed to copyin DOF header");
11611 * Now we'll allocate the entire DOF and copy it in -- provided
11612 * that the length isn't outrageous.
11614 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11615 dtrace_dof_error(&hdr, "load size exceeds maximum");
11620 if (hdr.dofh_loadsz < sizeof (hdr)) {
11621 dtrace_dof_error(&hdr, "invalid load size");
11626 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11628 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
11629 kmem_free(dof, hdr.dofh_loadsz);
11638 static __inline uchar_t
11639 dtrace_dof_char(char c) {
11658 return (c - 'A' + 10);
11665 return (c - 'a' + 10);
11667 /* Should not reach here. */
11673 dtrace_dof_property(const char *name)
11677 unsigned int len, i;
11682 * Unfortunately, array of values in .conf files are always (and
11683 * only) interpreted to be integer arrays. We must read our DOF
11684 * as an integer array, and then squeeze it into a byte array.
11686 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11687 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11690 for (i = 0; i < len; i++)
11691 buf[i] = (uchar_t)(((int *)buf)[i]);
11693 if (len < sizeof (dof_hdr_t)) {
11694 ddi_prop_free(buf);
11695 dtrace_dof_error(NULL, "truncated header");
11699 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11700 ddi_prop_free(buf);
11701 dtrace_dof_error(NULL, "truncated DOF");
11705 if (loadsz >= dtrace_dof_maxsize) {
11706 ddi_prop_free(buf);
11707 dtrace_dof_error(NULL, "oversized DOF");
11711 dof = kmem_alloc(loadsz, KM_SLEEP);
11712 bcopy(buf, dof, loadsz);
11713 ddi_prop_free(buf);
11718 if ((p_env = getenv(name)) == NULL)
11721 len = strlen(p_env) / 2;
11723 buf = kmem_alloc(len, KM_SLEEP);
11725 dof = (dof_hdr_t *) buf;
11729 for (i = 0; i < len; i++) {
11730 buf[i] = (dtrace_dof_char(p[0]) << 4) |
11731 dtrace_dof_char(p[1]);
11737 if (len < sizeof (dof_hdr_t)) {
11739 dtrace_dof_error(NULL, "truncated header");
11743 if (len < (loadsz = dof->dofh_loadsz)) {
11745 dtrace_dof_error(NULL, "truncated DOF");
11749 if (loadsz >= dtrace_dof_maxsize) {
11751 dtrace_dof_error(NULL, "oversized DOF");
11760 dtrace_dof_destroy(dof_hdr_t *dof)
11762 kmem_free(dof, dof->dofh_loadsz);
11766 * Return the dof_sec_t pointer corresponding to a given section index. If the
11767 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
11768 * a type other than DOF_SECT_NONE is specified, the header is checked against
11769 * this type and NULL is returned if the types do not match.
11772 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11774 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11775 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11777 if (i >= dof->dofh_secnum) {
11778 dtrace_dof_error(dof, "referenced section index is invalid");
11782 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11783 dtrace_dof_error(dof, "referenced section is not loadable");
11787 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11788 dtrace_dof_error(dof, "referenced section is the wrong type");
11795 static dtrace_probedesc_t *
11796 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11798 dof_probedesc_t *probe;
11800 uintptr_t daddr = (uintptr_t)dof;
11804 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11805 dtrace_dof_error(dof, "invalid probe section");
11809 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11810 dtrace_dof_error(dof, "bad alignment in probe description");
11814 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11815 dtrace_dof_error(dof, "truncated probe description");
11819 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11820 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11822 if (strtab == NULL)
11825 str = daddr + strtab->dofs_offset;
11826 size = strtab->dofs_size;
11828 if (probe->dofp_provider >= strtab->dofs_size) {
11829 dtrace_dof_error(dof, "corrupt probe provider");
11833 (void) strncpy(desc->dtpd_provider,
11834 (char *)(str + probe->dofp_provider),
11835 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11837 if (probe->dofp_mod >= strtab->dofs_size) {
11838 dtrace_dof_error(dof, "corrupt probe module");
11842 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11843 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11845 if (probe->dofp_func >= strtab->dofs_size) {
11846 dtrace_dof_error(dof, "corrupt probe function");
11850 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11851 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11853 if (probe->dofp_name >= strtab->dofs_size) {
11854 dtrace_dof_error(dof, "corrupt probe name");
11858 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11859 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11864 static dtrace_difo_t *
11865 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11870 dof_difohdr_t *dofd;
11871 uintptr_t daddr = (uintptr_t)dof;
11872 size_t max = dtrace_difo_maxsize;
11875 static const struct {
11883 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11884 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11885 sizeof (dif_instr_t), "multiple DIF sections" },
11887 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11888 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11889 sizeof (uint64_t), "multiple integer tables" },
11891 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11892 offsetof(dtrace_difo_t, dtdo_strlen), 0,
11893 sizeof (char), "multiple string tables" },
11895 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11896 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11897 sizeof (uint_t), "multiple variable tables" },
11899 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
11902 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11903 dtrace_dof_error(dof, "invalid DIFO header section");
11907 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11908 dtrace_dof_error(dof, "bad alignment in DIFO header");
11912 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
11913 sec->dofs_size % sizeof (dof_secidx_t)) {
11914 dtrace_dof_error(dof, "bad size in DIFO header");
11918 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11919 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
11921 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
11922 dp->dtdo_rtype = dofd->dofd_rtype;
11924 for (l = 0; l < n; l++) {
11929 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
11930 dofd->dofd_links[l])) == NULL)
11931 goto err; /* invalid section link */
11933 if (ttl + subsec->dofs_size > max) {
11934 dtrace_dof_error(dof, "exceeds maximum size");
11938 ttl += subsec->dofs_size;
11940 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
11941 if (subsec->dofs_type != difo[i].section)
11944 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
11945 dtrace_dof_error(dof, "section not loaded");
11949 if (subsec->dofs_align != difo[i].align) {
11950 dtrace_dof_error(dof, "bad alignment");
11954 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
11955 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
11957 if (*bufp != NULL) {
11958 dtrace_dof_error(dof, difo[i].msg);
11962 if (difo[i].entsize != subsec->dofs_entsize) {
11963 dtrace_dof_error(dof, "entry size mismatch");
11967 if (subsec->dofs_entsize != 0 &&
11968 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
11969 dtrace_dof_error(dof, "corrupt entry size");
11973 *lenp = subsec->dofs_size;
11974 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
11975 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
11976 *bufp, subsec->dofs_size);
11978 if (subsec->dofs_entsize != 0)
11979 *lenp /= subsec->dofs_entsize;
11985 * If we encounter a loadable DIFO sub-section that is not
11986 * known to us, assume this is a broken program and fail.
11988 if (difo[i].section == DOF_SECT_NONE &&
11989 (subsec->dofs_flags & DOF_SECF_LOAD)) {
11990 dtrace_dof_error(dof, "unrecognized DIFO subsection");
11995 if (dp->dtdo_buf == NULL) {
11997 * We can't have a DIF object without DIF text.
11999 dtrace_dof_error(dof, "missing DIF text");
12004 * Before we validate the DIF object, run through the variable table
12005 * looking for the strings -- if any of their size are under, we'll set
12006 * their size to be the system-wide default string size. Note that
12007 * this should _not_ happen if the "strsize" option has been set --
12008 * in this case, the compiler should have set the size to reflect the
12009 * setting of the option.
12011 for (i = 0; i < dp->dtdo_varlen; i++) {
12012 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12013 dtrace_diftype_t *t = &v->dtdv_type;
12015 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12018 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12019 t->dtdt_size = dtrace_strsize_default;
12022 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12025 dtrace_difo_init(dp, vstate);
12029 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12030 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12031 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12032 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12034 kmem_free(dp, sizeof (dtrace_difo_t));
12038 static dtrace_predicate_t *
12039 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12044 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12047 return (dtrace_predicate_create(dp));
12050 static dtrace_actdesc_t *
12051 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12054 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12055 dof_actdesc_t *desc;
12056 dof_sec_t *difosec;
12058 uintptr_t daddr = (uintptr_t)dof;
12060 dtrace_actkind_t kind;
12062 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12063 dtrace_dof_error(dof, "invalid action section");
12067 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12068 dtrace_dof_error(dof, "truncated action description");
12072 if (sec->dofs_align != sizeof (uint64_t)) {
12073 dtrace_dof_error(dof, "bad alignment in action description");
12077 if (sec->dofs_size < sec->dofs_entsize) {
12078 dtrace_dof_error(dof, "section entry size exceeds total size");
12082 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12083 dtrace_dof_error(dof, "bad entry size in action description");
12087 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12088 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12092 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12093 desc = (dof_actdesc_t *)(daddr +
12094 (uintptr_t)sec->dofs_offset + offs);
12095 kind = (dtrace_actkind_t)desc->dofa_kind;
12097 if (DTRACEACT_ISPRINTFLIKE(kind) &&
12098 (kind != DTRACEACT_PRINTA ||
12099 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12105 * printf()-like actions must have a format string.
12107 if ((strtab = dtrace_dof_sect(dof,
12108 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12111 str = (char *)((uintptr_t)dof +
12112 (uintptr_t)strtab->dofs_offset);
12114 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12115 if (str[i] == '\0')
12119 if (i >= strtab->dofs_size) {
12120 dtrace_dof_error(dof, "bogus format string");
12124 if (i == desc->dofa_arg) {
12125 dtrace_dof_error(dof, "empty format string");
12129 i -= desc->dofa_arg;
12130 fmt = kmem_alloc(i + 1, KM_SLEEP);
12131 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12132 arg = (uint64_t)(uintptr_t)fmt;
12134 if (kind == DTRACEACT_PRINTA) {
12135 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12138 arg = desc->dofa_arg;
12142 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12143 desc->dofa_uarg, arg);
12145 if (last != NULL) {
12146 last->dtad_next = act;
12153 if (desc->dofa_difo == DOF_SECIDX_NONE)
12156 if ((difosec = dtrace_dof_sect(dof,
12157 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12160 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12162 if (act->dtad_difo == NULL)
12166 ASSERT(first != NULL);
12170 for (act = first; act != NULL; act = next) {
12171 next = act->dtad_next;
12172 dtrace_actdesc_release(act, vstate);
12178 static dtrace_ecbdesc_t *
12179 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12182 dtrace_ecbdesc_t *ep;
12183 dof_ecbdesc_t *ecb;
12184 dtrace_probedesc_t *desc;
12185 dtrace_predicate_t *pred = NULL;
12187 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12188 dtrace_dof_error(dof, "truncated ECB description");
12192 if (sec->dofs_align != sizeof (uint64_t)) {
12193 dtrace_dof_error(dof, "bad alignment in ECB description");
12197 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12198 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12203 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12204 ep->dted_uarg = ecb->dofe_uarg;
12205 desc = &ep->dted_probe;
12207 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12210 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12211 if ((sec = dtrace_dof_sect(dof,
12212 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12215 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12218 ep->dted_pred.dtpdd_predicate = pred;
12221 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12222 if ((sec = dtrace_dof_sect(dof,
12223 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12226 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12228 if (ep->dted_action == NULL)
12236 dtrace_predicate_release(pred, vstate);
12237 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12242 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12243 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12244 * site of any user SETX relocations to account for load object base address.
12245 * In the future, if we need other relocations, this function can be extended.
12248 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12250 uintptr_t daddr = (uintptr_t)dof;
12251 dof_relohdr_t *dofr =
12252 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12253 dof_sec_t *ss, *rs, *ts;
12257 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12258 sec->dofs_align != sizeof (dof_secidx_t)) {
12259 dtrace_dof_error(dof, "invalid relocation header");
12263 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12264 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12265 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12267 if (ss == NULL || rs == NULL || ts == NULL)
12268 return (-1); /* dtrace_dof_error() has been called already */
12270 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12271 rs->dofs_align != sizeof (uint64_t)) {
12272 dtrace_dof_error(dof, "invalid relocation section");
12276 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12277 n = rs->dofs_size / rs->dofs_entsize;
12279 for (i = 0; i < n; i++) {
12280 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12282 switch (r->dofr_type) {
12283 case DOF_RELO_NONE:
12285 case DOF_RELO_SETX:
12286 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12287 sizeof (uint64_t) > ts->dofs_size) {
12288 dtrace_dof_error(dof, "bad relocation offset");
12292 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12293 dtrace_dof_error(dof, "misaligned setx relo");
12297 *(uint64_t *)taddr += ubase;
12300 dtrace_dof_error(dof, "invalid relocation type");
12304 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12311 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12312 * header: it should be at the front of a memory region that is at least
12313 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12314 * size. It need not be validated in any other way.
12317 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12318 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12320 uint64_t len = dof->dofh_loadsz, seclen;
12321 uintptr_t daddr = (uintptr_t)dof;
12322 dtrace_ecbdesc_t *ep;
12323 dtrace_enabling_t *enab;
12326 ASSERT(MUTEX_HELD(&dtrace_lock));
12327 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12330 * Check the DOF header identification bytes. In addition to checking
12331 * valid settings, we also verify that unused bits/bytes are zeroed so
12332 * we can use them later without fear of regressing existing binaries.
12334 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12335 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12336 dtrace_dof_error(dof, "DOF magic string mismatch");
12340 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12341 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12342 dtrace_dof_error(dof, "DOF has invalid data model");
12346 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12347 dtrace_dof_error(dof, "DOF encoding mismatch");
12351 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12352 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12353 dtrace_dof_error(dof, "DOF version mismatch");
12357 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12358 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12362 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12363 dtrace_dof_error(dof, "DOF uses too many integer registers");
12367 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12368 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12372 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12373 if (dof->dofh_ident[i] != 0) {
12374 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12379 if (dof->dofh_flags & ~DOF_FL_VALID) {
12380 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12384 if (dof->dofh_secsize == 0) {
12385 dtrace_dof_error(dof, "zero section header size");
12390 * Check that the section headers don't exceed the amount of DOF
12391 * data. Note that we cast the section size and number of sections
12392 * to uint64_t's to prevent possible overflow in the multiplication.
12394 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12396 if (dof->dofh_secoff > len || seclen > len ||
12397 dof->dofh_secoff + seclen > len) {
12398 dtrace_dof_error(dof, "truncated section headers");
12402 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12403 dtrace_dof_error(dof, "misaligned section headers");
12407 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12408 dtrace_dof_error(dof, "misaligned section size");
12413 * Take an initial pass through the section headers to be sure that
12414 * the headers don't have stray offsets. If the 'noprobes' flag is
12415 * set, do not permit sections relating to providers, probes, or args.
12417 for (i = 0; i < dof->dofh_secnum; i++) {
12418 dof_sec_t *sec = (dof_sec_t *)(daddr +
12419 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12422 switch (sec->dofs_type) {
12423 case DOF_SECT_PROVIDER:
12424 case DOF_SECT_PROBES:
12425 case DOF_SECT_PRARGS:
12426 case DOF_SECT_PROFFS:
12427 dtrace_dof_error(dof, "illegal sections "
12433 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12434 continue; /* just ignore non-loadable sections */
12436 if (sec->dofs_align & (sec->dofs_align - 1)) {
12437 dtrace_dof_error(dof, "bad section alignment");
12441 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12442 dtrace_dof_error(dof, "misaligned section");
12446 if (sec->dofs_offset > len || sec->dofs_size > len ||
12447 sec->dofs_offset + sec->dofs_size > len) {
12448 dtrace_dof_error(dof, "corrupt section header");
12452 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12453 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12454 dtrace_dof_error(dof, "non-terminating string table");
12460 * Take a second pass through the sections and locate and perform any
12461 * relocations that are present. We do this after the first pass to
12462 * be sure that all sections have had their headers validated.
12464 for (i = 0; i < dof->dofh_secnum; i++) {
12465 dof_sec_t *sec = (dof_sec_t *)(daddr +
12466 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12468 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12469 continue; /* skip sections that are not loadable */
12471 switch (sec->dofs_type) {
12472 case DOF_SECT_URELHDR:
12473 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12479 if ((enab = *enabp) == NULL)
12480 enab = *enabp = dtrace_enabling_create(vstate);
12482 for (i = 0; i < dof->dofh_secnum; i++) {
12483 dof_sec_t *sec = (dof_sec_t *)(daddr +
12484 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12486 if (sec->dofs_type != DOF_SECT_ECBDESC)
12489 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12490 dtrace_enabling_destroy(enab);
12495 dtrace_enabling_add(enab, ep);
12502 * Process DOF for any options. This routine assumes that the DOF has been
12503 * at least processed by dtrace_dof_slurp().
12506 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12511 dof_optdesc_t *desc;
12513 for (i = 0; i < dof->dofh_secnum; i++) {
12514 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12515 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12517 if (sec->dofs_type != DOF_SECT_OPTDESC)
12520 if (sec->dofs_align != sizeof (uint64_t)) {
12521 dtrace_dof_error(dof, "bad alignment in "
12522 "option description");
12526 if ((entsize = sec->dofs_entsize) == 0) {
12527 dtrace_dof_error(dof, "zeroed option entry size");
12531 if (entsize < sizeof (dof_optdesc_t)) {
12532 dtrace_dof_error(dof, "bad option entry size");
12536 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12537 desc = (dof_optdesc_t *)((uintptr_t)dof +
12538 (uintptr_t)sec->dofs_offset + offs);
12540 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12541 dtrace_dof_error(dof, "non-zero option string");
12545 if (desc->dofo_value == DTRACEOPT_UNSET) {
12546 dtrace_dof_error(dof, "unset option");
12550 if ((rval = dtrace_state_option(state,
12551 desc->dofo_option, desc->dofo_value)) != 0) {
12552 dtrace_dof_error(dof, "rejected option");
12562 * DTrace Consumer State Functions
12565 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12567 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12570 dtrace_dynvar_t *dvar, *next, *start;
12573 ASSERT(MUTEX_HELD(&dtrace_lock));
12574 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12576 bzero(dstate, sizeof (dtrace_dstate_t));
12578 if ((dstate->dtds_chunksize = chunksize) == 0)
12579 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12581 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12584 if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
12587 dstate->dtds_size = size;
12588 dstate->dtds_base = base;
12589 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12590 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12592 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12594 if (hashsize != 1 && (hashsize & 1))
12597 dstate->dtds_hashsize = hashsize;
12598 dstate->dtds_hash = dstate->dtds_base;
12601 * Set all of our hash buckets to point to the single sink, and (if
12602 * it hasn't already been set), set the sink's hash value to be the
12603 * sink sentinel value. The sink is needed for dynamic variable
12604 * lookups to know that they have iterated over an entire, valid hash
12607 for (i = 0; i < hashsize; i++)
12608 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12610 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12611 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12614 * Determine number of active CPUs. Divide free list evenly among
12617 start = (dtrace_dynvar_t *)
12618 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12619 limit = (uintptr_t)base + size;
12621 maxper = (limit - (uintptr_t)start) / NCPU;
12622 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12624 for (i = 0; i < NCPU; i++) {
12629 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12632 * If we don't even have enough chunks to make it once through
12633 * NCPUs, we're just going to allocate everything to the first
12634 * CPU. And if we're on the last CPU, we're going to allocate
12635 * whatever is left over. In either case, we set the limit to
12636 * be the limit of the dynamic variable space.
12638 if (maxper == 0 || i == NCPU - 1) {
12639 limit = (uintptr_t)base + size;
12642 limit = (uintptr_t)start + maxper;
12643 start = (dtrace_dynvar_t *)limit;
12646 ASSERT(limit <= (uintptr_t)base + size);
12649 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12650 dstate->dtds_chunksize);
12652 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12655 dvar->dtdv_next = next;
12667 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12669 ASSERT(MUTEX_HELD(&cpu_lock));
12671 if (dstate->dtds_base == NULL)
12674 kmem_free(dstate->dtds_base, dstate->dtds_size);
12675 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12679 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12682 * Logical XOR, where are you?
12684 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12686 if (vstate->dtvs_nglobals > 0) {
12687 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12688 sizeof (dtrace_statvar_t *));
12691 if (vstate->dtvs_ntlocals > 0) {
12692 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12693 sizeof (dtrace_difv_t));
12696 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
12698 if (vstate->dtvs_nlocals > 0) {
12699 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
12700 sizeof (dtrace_statvar_t *));
12706 dtrace_state_clean(dtrace_state_t *state)
12708 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12711 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12712 dtrace_speculation_clean(state);
12716 dtrace_state_deadman(dtrace_state_t *state)
12722 now = dtrace_gethrtime();
12724 if (state != dtrace_anon.dta_state &&
12725 now - state->dts_laststatus >= dtrace_deadman_user)
12729 * We must be sure that dts_alive never appears to be less than the
12730 * value upon entry to dtrace_state_deadman(), and because we lack a
12731 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12732 * store INT64_MAX to it, followed by a memory barrier, followed by
12733 * the new value. This assures that dts_alive never appears to be
12734 * less than its true value, regardless of the order in which the
12735 * stores to the underlying storage are issued.
12737 state->dts_alive = INT64_MAX;
12738 dtrace_membar_producer();
12739 state->dts_alive = now;
12743 dtrace_state_clean(void *arg)
12745 dtrace_state_t *state = arg;
12746 dtrace_optval_t *opt = state->dts_options;
12748 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12751 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12752 dtrace_speculation_clean(state);
12754 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
12755 dtrace_state_clean, state);
12759 dtrace_state_deadman(void *arg)
12761 dtrace_state_t *state = arg;
12766 dtrace_debug_output();
12768 now = dtrace_gethrtime();
12770 if (state != dtrace_anon.dta_state &&
12771 now - state->dts_laststatus >= dtrace_deadman_user)
12775 * We must be sure that dts_alive never appears to be less than the
12776 * value upon entry to dtrace_state_deadman(), and because we lack a
12777 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12778 * store INT64_MAX to it, followed by a memory barrier, followed by
12779 * the new value. This assures that dts_alive never appears to be
12780 * less than its true value, regardless of the order in which the
12781 * stores to the underlying storage are issued.
12783 state->dts_alive = INT64_MAX;
12784 dtrace_membar_producer();
12785 state->dts_alive = now;
12787 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
12788 dtrace_state_deadman, state);
12792 static dtrace_state_t *
12794 dtrace_state_create(dev_t *devp, cred_t *cr)
12796 dtrace_state_create(struct cdev *dev)
12807 dtrace_state_t *state;
12808 dtrace_optval_t *opt;
12809 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12811 ASSERT(MUTEX_HELD(&dtrace_lock));
12812 ASSERT(MUTEX_HELD(&cpu_lock));
12815 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12816 VM_BESTFIT | VM_SLEEP);
12818 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12819 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12823 state = ddi_get_soft_state(dtrace_softstate, minor);
12830 /* Allocate memory for the state. */
12831 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
12834 state->dts_epid = DTRACE_EPIDNONE + 1;
12836 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
12838 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12839 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12841 if (devp != NULL) {
12842 major = getemajor(*devp);
12844 major = ddi_driver_major(dtrace_devi);
12847 state->dts_dev = makedevice(major, minor);
12850 *devp = state->dts_dev;
12852 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
12853 state->dts_dev = dev;
12857 * We allocate NCPU buffers. On the one hand, this can be quite
12858 * a bit of memory per instance (nearly 36K on a Starcat). On the
12859 * other hand, it saves an additional memory reference in the probe
12862 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12863 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12866 state->dts_cleaner = CYCLIC_NONE;
12867 state->dts_deadman = CYCLIC_NONE;
12869 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
12870 callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
12872 state->dts_vstate.dtvs_state = state;
12874 for (i = 0; i < DTRACEOPT_MAX; i++)
12875 state->dts_options[i] = DTRACEOPT_UNSET;
12878 * Set the default options.
12880 opt = state->dts_options;
12881 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12882 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12883 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12884 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12885 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12886 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12887 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12888 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12889 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12890 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12891 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12892 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12893 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12894 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12896 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12899 * Depending on the user credentials, we set flag bits which alter probe
12900 * visibility or the amount of destructiveness allowed. In the case of
12901 * actual anonymous tracing, or the possession of all privileges, all of
12902 * the normal checks are bypassed.
12904 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
12905 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
12906 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
12909 * Set up the credentials for this instantiation. We take a
12910 * hold on the credential to prevent it from disappearing on
12911 * us; this in turn prevents the zone_t referenced by this
12912 * credential from disappearing. This means that we can
12913 * examine the credential and the zone from probe context.
12916 state->dts_cred.dcr_cred = cr;
12919 * CRA_PROC means "we have *some* privilege for dtrace" and
12920 * unlocks the use of variables like pid, zonename, etc.
12922 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
12923 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12924 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
12928 * dtrace_user allows use of syscall and profile providers.
12929 * If the user also has proc_owner and/or proc_zone, we
12930 * extend the scope to include additional visibility and
12931 * destructive power.
12933 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
12934 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
12935 state->dts_cred.dcr_visible |=
12936 DTRACE_CRV_ALLPROC;
12938 state->dts_cred.dcr_action |=
12939 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12942 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
12943 state->dts_cred.dcr_visible |=
12944 DTRACE_CRV_ALLZONE;
12946 state->dts_cred.dcr_action |=
12947 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12951 * If we have all privs in whatever zone this is,
12952 * we can do destructive things to processes which
12953 * have altered credentials.
12956 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12957 cr->cr_zone->zone_privset)) {
12958 state->dts_cred.dcr_action |=
12959 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12965 * Holding the dtrace_kernel privilege also implies that
12966 * the user has the dtrace_user privilege from a visibility
12967 * perspective. But without further privileges, some
12968 * destructive actions are not available.
12970 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
12972 * Make all probes in all zones visible. However,
12973 * this doesn't mean that all actions become available
12976 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
12977 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
12979 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
12982 * Holding proc_owner means that destructive actions
12983 * for *this* zone are allowed.
12985 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12986 state->dts_cred.dcr_action |=
12987 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12990 * Holding proc_zone means that destructive actions
12991 * for this user/group ID in all zones is allowed.
12993 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12994 state->dts_cred.dcr_action |=
12995 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12999 * If we have all privs in whatever zone this is,
13000 * we can do destructive things to processes which
13001 * have altered credentials.
13003 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13004 cr->cr_zone->zone_privset)) {
13005 state->dts_cred.dcr_action |=
13006 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13012 * Holding the dtrace_proc privilege gives control over fasttrap
13013 * and pid providers. We need to grant wider destructive
13014 * privileges in the event that the user has proc_owner and/or
13017 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13018 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13019 state->dts_cred.dcr_action |=
13020 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13022 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13023 state->dts_cred.dcr_action |=
13024 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13032 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13034 dtrace_optval_t *opt = state->dts_options, size;
13035 processorid_t cpu = 0;;
13036 int flags = 0, rval;
13038 ASSERT(MUTEX_HELD(&dtrace_lock));
13039 ASSERT(MUTEX_HELD(&cpu_lock));
13040 ASSERT(which < DTRACEOPT_MAX);
13041 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13042 (state == dtrace_anon.dta_state &&
13043 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13045 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13048 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13049 cpu = opt[DTRACEOPT_CPU];
13051 if (which == DTRACEOPT_SPECSIZE)
13052 flags |= DTRACEBUF_NOSWITCH;
13054 if (which == DTRACEOPT_BUFSIZE) {
13055 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13056 flags |= DTRACEBUF_RING;
13058 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13059 flags |= DTRACEBUF_FILL;
13061 if (state != dtrace_anon.dta_state ||
13062 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13063 flags |= DTRACEBUF_INACTIVE;
13066 for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
13068 * The size must be 8-byte aligned. If the size is not 8-byte
13069 * aligned, drop it down by the difference.
13071 if (size & (sizeof (uint64_t) - 1))
13072 size -= size & (sizeof (uint64_t) - 1);
13074 if (size < state->dts_reserve) {
13076 * Buffers always must be large enough to accommodate
13077 * their prereserved space. We return E2BIG instead
13078 * of ENOMEM in this case to allow for user-level
13079 * software to differentiate the cases.
13084 rval = dtrace_buffer_alloc(buf, size, flags, cpu);
13086 if (rval != ENOMEM) {
13091 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13099 dtrace_state_buffers(dtrace_state_t *state)
13101 dtrace_speculation_t *spec = state->dts_speculations;
13104 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13105 DTRACEOPT_BUFSIZE)) != 0)
13108 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13109 DTRACEOPT_AGGSIZE)) != 0)
13112 for (i = 0; i < state->dts_nspeculations; i++) {
13113 if ((rval = dtrace_state_buffer(state,
13114 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13122 dtrace_state_prereserve(dtrace_state_t *state)
13125 dtrace_probe_t *probe;
13127 state->dts_reserve = 0;
13129 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13133 * If our buffer policy is a "fill" buffer policy, we need to set the
13134 * prereserved space to be the space required by the END probes.
13136 probe = dtrace_probes[dtrace_probeid_end - 1];
13137 ASSERT(probe != NULL);
13139 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13140 if (ecb->dte_state != state)
13143 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13148 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13150 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13151 dtrace_speculation_t *spec;
13152 dtrace_buffer_t *buf;
13154 cyc_handler_t hdlr;
13157 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13158 dtrace_icookie_t cookie;
13160 mutex_enter(&cpu_lock);
13161 mutex_enter(&dtrace_lock);
13163 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13169 * Before we can perform any checks, we must prime all of the
13170 * retained enablings that correspond to this state.
13172 dtrace_enabling_prime(state);
13174 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13179 dtrace_state_prereserve(state);
13182 * Now we want to do is try to allocate our speculations.
13183 * We do not automatically resize the number of speculations; if
13184 * this fails, we will fail the operation.
13186 nspec = opt[DTRACEOPT_NSPEC];
13187 ASSERT(nspec != DTRACEOPT_UNSET);
13189 if (nspec > INT_MAX) {
13194 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
13196 if (spec == NULL) {
13201 state->dts_speculations = spec;
13202 state->dts_nspeculations = (int)nspec;
13204 for (i = 0; i < nspec; i++) {
13205 if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
13210 spec[i].dtsp_buffer = buf;
13213 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13214 if (dtrace_anon.dta_state == NULL) {
13219 if (state->dts_necbs != 0) {
13224 state->dts_anon = dtrace_anon_grab();
13225 ASSERT(state->dts_anon != NULL);
13226 state = state->dts_anon;
13229 * We want "grabanon" to be set in the grabbed state, so we'll
13230 * copy that option value from the grabbing state into the
13233 state->dts_options[DTRACEOPT_GRABANON] =
13234 opt[DTRACEOPT_GRABANON];
13236 *cpu = dtrace_anon.dta_beganon;
13239 * If the anonymous state is active (as it almost certainly
13240 * is if the anonymous enabling ultimately matched anything),
13241 * we don't allow any further option processing -- but we
13242 * don't return failure.
13244 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13248 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13249 opt[DTRACEOPT_AGGSIZE] != 0) {
13250 if (state->dts_aggregations == NULL) {
13252 * We're not going to create an aggregation buffer
13253 * because we don't have any ECBs that contain
13254 * aggregations -- set this option to 0.
13256 opt[DTRACEOPT_AGGSIZE] = 0;
13259 * If we have an aggregation buffer, we must also have
13260 * a buffer to use as scratch.
13262 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13263 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13264 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13269 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13270 opt[DTRACEOPT_SPECSIZE] != 0) {
13271 if (!state->dts_speculates) {
13273 * We're not going to create speculation buffers
13274 * because we don't have any ECBs that actually
13275 * speculate -- set the speculation size to 0.
13277 opt[DTRACEOPT_SPECSIZE] = 0;
13282 * The bare minimum size for any buffer that we're actually going to
13283 * do anything to is sizeof (uint64_t).
13285 sz = sizeof (uint64_t);
13287 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13288 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13289 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13291 * A buffer size has been explicitly set to 0 (or to a size
13292 * that will be adjusted to 0) and we need the space -- we
13293 * need to return failure. We return ENOSPC to differentiate
13294 * it from failing to allocate a buffer due to failure to meet
13295 * the reserve (for which we return E2BIG).
13301 if ((rval = dtrace_state_buffers(state)) != 0)
13304 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13305 sz = dtrace_dstate_defsize;
13308 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13313 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13315 } while (sz >>= 1);
13317 opt[DTRACEOPT_DYNVARSIZE] = sz;
13322 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13323 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13325 if (opt[DTRACEOPT_CLEANRATE] == 0)
13326 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13328 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13329 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13331 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13332 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13334 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13336 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13337 hdlr.cyh_arg = state;
13338 hdlr.cyh_level = CY_LOW_LEVEL;
13341 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13343 state->dts_cleaner = cyclic_add(&hdlr, &when);
13345 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13346 hdlr.cyh_arg = state;
13347 hdlr.cyh_level = CY_LOW_LEVEL;
13350 when.cyt_interval = dtrace_deadman_interval;
13352 state->dts_deadman = cyclic_add(&hdlr, &when);
13354 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13355 dtrace_state_clean, state);
13356 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13357 dtrace_state_deadman, state);
13360 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13363 * Now it's time to actually fire the BEGIN probe. We need to disable
13364 * interrupts here both to record the CPU on which we fired the BEGIN
13365 * probe (the data from this CPU will be processed first at user
13366 * level) and to manually activate the buffer for this CPU.
13368 cookie = dtrace_interrupt_disable();
13370 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13371 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13373 dtrace_probe(dtrace_probeid_begin,
13374 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13375 dtrace_interrupt_enable(cookie);
13377 * We may have had an exit action from a BEGIN probe; only change our
13378 * state to ACTIVE if we're still in WARMUP.
13380 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13381 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13383 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13384 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13387 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13388 * want each CPU to transition its principal buffer out of the
13389 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13390 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13391 * atomically transition from processing none of a state's ECBs to
13392 * processing all of them.
13394 dtrace_xcall(DTRACE_CPUALL,
13395 (dtrace_xcall_t)dtrace_buffer_activate, state);
13399 dtrace_buffer_free(state->dts_buffer);
13400 dtrace_buffer_free(state->dts_aggbuffer);
13402 if ((nspec = state->dts_nspeculations) == 0) {
13403 ASSERT(state->dts_speculations == NULL);
13407 spec = state->dts_speculations;
13408 ASSERT(spec != NULL);
13410 for (i = 0; i < state->dts_nspeculations; i++) {
13411 if ((buf = spec[i].dtsp_buffer) == NULL)
13414 dtrace_buffer_free(buf);
13415 kmem_free(buf, bufsize);
13418 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13419 state->dts_nspeculations = 0;
13420 state->dts_speculations = NULL;
13423 mutex_exit(&dtrace_lock);
13424 mutex_exit(&cpu_lock);
13430 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13432 dtrace_icookie_t cookie;
13434 ASSERT(MUTEX_HELD(&dtrace_lock));
13436 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13437 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13441 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13442 * to be sure that every CPU has seen it. See below for the details
13443 * on why this is done.
13445 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13449 * By this point, it is impossible for any CPU to be still processing
13450 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13451 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13452 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13453 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13454 * iff we're in the END probe.
13456 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13458 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13461 * Finally, we can release the reserve and call the END probe. We
13462 * disable interrupts across calling the END probe to allow us to
13463 * return the CPU on which we actually called the END probe. This
13464 * allows user-land to be sure that this CPU's principal buffer is
13467 state->dts_reserve = 0;
13469 cookie = dtrace_interrupt_disable();
13471 dtrace_probe(dtrace_probeid_end,
13472 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13473 dtrace_interrupt_enable(cookie);
13475 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13482 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13483 dtrace_optval_t val)
13485 ASSERT(MUTEX_HELD(&dtrace_lock));
13487 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13490 if (option >= DTRACEOPT_MAX)
13493 if (option != DTRACEOPT_CPU && val < 0)
13497 case DTRACEOPT_DESTRUCTIVE:
13498 if (dtrace_destructive_disallow)
13501 state->dts_cred.dcr_destructive = 1;
13504 case DTRACEOPT_BUFSIZE:
13505 case DTRACEOPT_DYNVARSIZE:
13506 case DTRACEOPT_AGGSIZE:
13507 case DTRACEOPT_SPECSIZE:
13508 case DTRACEOPT_STRSIZE:
13512 if (val >= LONG_MAX) {
13514 * If this is an otherwise negative value, set it to
13515 * the highest multiple of 128m less than LONG_MAX.
13516 * Technically, we're adjusting the size without
13517 * regard to the buffer resizing policy, but in fact,
13518 * this has no effect -- if we set the buffer size to
13519 * ~LONG_MAX and the buffer policy is ultimately set to
13520 * be "manual", the buffer allocation is guaranteed to
13521 * fail, if only because the allocation requires two
13522 * buffers. (We set the the size to the highest
13523 * multiple of 128m because it ensures that the size
13524 * will remain a multiple of a megabyte when
13525 * repeatedly halved -- all the way down to 15m.)
13527 val = LONG_MAX - (1 << 27) + 1;
13531 state->dts_options[option] = val;
13537 dtrace_state_destroy(dtrace_state_t *state)
13540 dtrace_vstate_t *vstate = &state->dts_vstate;
13542 minor_t minor = getminor(state->dts_dev);
13544 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13545 dtrace_speculation_t *spec = state->dts_speculations;
13546 int nspec = state->dts_nspeculations;
13549 ASSERT(MUTEX_HELD(&dtrace_lock));
13550 ASSERT(MUTEX_HELD(&cpu_lock));
13553 * First, retract any retained enablings for this state.
13555 dtrace_enabling_retract(state);
13556 ASSERT(state->dts_nretained == 0);
13558 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13559 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13561 * We have managed to come into dtrace_state_destroy() on a
13562 * hot enabling -- almost certainly because of a disorderly
13563 * shutdown of a consumer. (That is, a consumer that is
13564 * exiting without having called dtrace_stop().) In this case,
13565 * we're going to set our activity to be KILLED, and then
13566 * issue a sync to be sure that everyone is out of probe
13567 * context before we start blowing away ECBs.
13569 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13574 * Release the credential hold we took in dtrace_state_create().
13576 if (state->dts_cred.dcr_cred != NULL)
13577 crfree(state->dts_cred.dcr_cred);
13580 * Now we can safely disable and destroy any enabled probes. Because
13581 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13582 * (especially if they're all enabled), we take two passes through the
13583 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13584 * in the second we disable whatever is left over.
13586 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13587 for (i = 0; i < state->dts_necbs; i++) {
13588 if ((ecb = state->dts_ecbs[i]) == NULL)
13591 if (match && ecb->dte_probe != NULL) {
13592 dtrace_probe_t *probe = ecb->dte_probe;
13593 dtrace_provider_t *prov = probe->dtpr_provider;
13595 if (!(prov->dtpv_priv.dtpp_flags & match))
13599 dtrace_ecb_disable(ecb);
13600 dtrace_ecb_destroy(ecb);
13608 * Before we free the buffers, perform one more sync to assure that
13609 * every CPU is out of probe context.
13613 dtrace_buffer_free(state->dts_buffer);
13614 dtrace_buffer_free(state->dts_aggbuffer);
13616 for (i = 0; i < nspec; i++)
13617 dtrace_buffer_free(spec[i].dtsp_buffer);
13620 if (state->dts_cleaner != CYCLIC_NONE)
13621 cyclic_remove(state->dts_cleaner);
13623 if (state->dts_deadman != CYCLIC_NONE)
13624 cyclic_remove(state->dts_deadman);
13626 callout_stop(&state->dts_cleaner);
13627 callout_drain(&state->dts_cleaner);
13628 callout_stop(&state->dts_deadman);
13629 callout_drain(&state->dts_deadman);
13632 dtrace_dstate_fini(&vstate->dtvs_dynvars);
13633 dtrace_vstate_fini(vstate);
13634 if (state->dts_ecbs != NULL)
13635 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13637 if (state->dts_aggregations != NULL) {
13639 for (i = 0; i < state->dts_naggregations; i++)
13640 ASSERT(state->dts_aggregations[i] == NULL);
13642 ASSERT(state->dts_naggregations > 0);
13643 kmem_free(state->dts_aggregations,
13644 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13647 kmem_free(state->dts_buffer, bufsize);
13648 kmem_free(state->dts_aggbuffer, bufsize);
13650 for (i = 0; i < nspec; i++)
13651 kmem_free(spec[i].dtsp_buffer, bufsize);
13654 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13656 dtrace_format_destroy(state);
13658 if (state->dts_aggid_arena != NULL) {
13660 vmem_destroy(state->dts_aggid_arena);
13662 delete_unrhdr(state->dts_aggid_arena);
13664 state->dts_aggid_arena = NULL;
13667 ddi_soft_state_free(dtrace_softstate, minor);
13668 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13673 * DTrace Anonymous Enabling Functions
13675 static dtrace_state_t *
13676 dtrace_anon_grab(void)
13678 dtrace_state_t *state;
13680 ASSERT(MUTEX_HELD(&dtrace_lock));
13682 if ((state = dtrace_anon.dta_state) == NULL) {
13683 ASSERT(dtrace_anon.dta_enabling == NULL);
13687 ASSERT(dtrace_anon.dta_enabling != NULL);
13688 ASSERT(dtrace_retained != NULL);
13690 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13691 dtrace_anon.dta_enabling = NULL;
13692 dtrace_anon.dta_state = NULL;
13698 dtrace_anon_property(void)
13701 dtrace_state_t *state;
13703 char c[32]; /* enough for "dof-data-" + digits */
13705 ASSERT(MUTEX_HELD(&dtrace_lock));
13706 ASSERT(MUTEX_HELD(&cpu_lock));
13708 for (i = 0; ; i++) {
13709 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
13711 dtrace_err_verbose = 1;
13713 if ((dof = dtrace_dof_property(c)) == NULL) {
13714 dtrace_err_verbose = 0;
13720 * We want to create anonymous state, so we need to transition
13721 * the kernel debugger to indicate that DTrace is active. If
13722 * this fails (e.g. because the debugger has modified text in
13723 * some way), we won't continue with the processing.
13725 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13726 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
13727 "enabling ignored.");
13728 dtrace_dof_destroy(dof);
13734 * If we haven't allocated an anonymous state, we'll do so now.
13736 if ((state = dtrace_anon.dta_state) == NULL) {
13738 state = dtrace_state_create(NULL, NULL);
13740 state = dtrace_state_create(NULL);
13742 dtrace_anon.dta_state = state;
13744 if (state == NULL) {
13746 * This basically shouldn't happen: the only
13747 * failure mode from dtrace_state_create() is a
13748 * failure of ddi_soft_state_zalloc() that
13749 * itself should never happen. Still, the
13750 * interface allows for a failure mode, and
13751 * we want to fail as gracefully as possible:
13752 * we'll emit an error message and cease
13753 * processing anonymous state in this case.
13755 cmn_err(CE_WARN, "failed to create "
13756 "anonymous state");
13757 dtrace_dof_destroy(dof);
13762 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
13763 &dtrace_anon.dta_enabling, 0, B_TRUE);
13766 rv = dtrace_dof_options(dof, state);
13768 dtrace_err_verbose = 0;
13769 dtrace_dof_destroy(dof);
13773 * This is malformed DOF; chuck any anonymous state
13776 ASSERT(dtrace_anon.dta_enabling == NULL);
13777 dtrace_state_destroy(state);
13778 dtrace_anon.dta_state = NULL;
13782 ASSERT(dtrace_anon.dta_enabling != NULL);
13785 if (dtrace_anon.dta_enabling != NULL) {
13789 * dtrace_enabling_retain() can only fail because we are
13790 * trying to retain more enablings than are allowed -- but
13791 * we only have one anonymous enabling, and we are guaranteed
13792 * to be allowed at least one retained enabling; we assert
13793 * that dtrace_enabling_retain() returns success.
13795 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
13798 dtrace_enabling_dump(dtrace_anon.dta_enabling);
13804 * DTrace Helper Functions
13807 dtrace_helper_trace(dtrace_helper_action_t *helper,
13808 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
13810 uint32_t size, next, nnext, i;
13811 dtrace_helptrace_t *ent;
13812 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
13814 if (!dtrace_helptrace_enabled)
13817 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
13820 * What would a tracing framework be without its own tracing
13821 * framework? (Well, a hell of a lot simpler, for starters...)
13823 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
13824 sizeof (uint64_t) - sizeof (uint64_t);
13827 * Iterate until we can allocate a slot in the trace buffer.
13830 next = dtrace_helptrace_next;
13832 if (next + size < dtrace_helptrace_bufsize) {
13833 nnext = next + size;
13837 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
13840 * We have our slot; fill it in.
13845 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
13846 ent->dtht_helper = helper;
13847 ent->dtht_where = where;
13848 ent->dtht_nlocals = vstate->dtvs_nlocals;
13850 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13851 mstate->dtms_fltoffs : -1;
13852 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13853 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
13855 for (i = 0; i < vstate->dtvs_nlocals; i++) {
13856 dtrace_statvar_t *svar;
13858 if ((svar = vstate->dtvs_locals[i]) == NULL)
13861 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13862 ent->dtht_locals[i] =
13863 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
13870 dtrace_helper(int which, dtrace_mstate_t *mstate,
13871 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13873 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
13874 uint64_t sarg0 = mstate->dtms_arg[0];
13875 uint64_t sarg1 = mstate->dtms_arg[1];
13877 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13878 dtrace_helper_action_t *helper;
13879 dtrace_vstate_t *vstate;
13880 dtrace_difo_t *pred;
13881 int i, trace = dtrace_helptrace_enabled;
13883 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13885 if (helpers == NULL)
13888 if ((helper = helpers->dthps_actions[which]) == NULL)
13891 vstate = &helpers->dthps_vstate;
13892 mstate->dtms_arg[0] = arg0;
13893 mstate->dtms_arg[1] = arg1;
13896 * Now iterate over each helper. If its predicate evaluates to 'true',
13897 * we'll call the corresponding actions. Note that the below calls
13898 * to dtrace_dif_emulate() may set faults in machine state. This is
13899 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
13900 * the stored DIF offset with its own (which is the desired behavior).
13901 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13902 * from machine state; this is okay, too.
13904 for (; helper != NULL; helper = helper->dtha_next) {
13905 if ((pred = helper->dtha_predicate) != NULL) {
13907 dtrace_helper_trace(helper, mstate, vstate, 0);
13909 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
13912 if (*flags & CPU_DTRACE_FAULT)
13916 for (i = 0; i < helper->dtha_nactions; i++) {
13918 dtrace_helper_trace(helper,
13919 mstate, vstate, i + 1);
13921 rval = dtrace_dif_emulate(helper->dtha_actions[i],
13922 mstate, vstate, state);
13924 if (*flags & CPU_DTRACE_FAULT)
13930 dtrace_helper_trace(helper, mstate, vstate,
13931 DTRACE_HELPTRACE_NEXT);
13935 dtrace_helper_trace(helper, mstate, vstate,
13936 DTRACE_HELPTRACE_DONE);
13939 * Restore the arg0 that we saved upon entry.
13941 mstate->dtms_arg[0] = sarg0;
13942 mstate->dtms_arg[1] = sarg1;
13948 dtrace_helper_trace(helper, mstate, vstate,
13949 DTRACE_HELPTRACE_ERR);
13952 * Restore the arg0 that we saved upon entry.
13954 mstate->dtms_arg[0] = sarg0;
13955 mstate->dtms_arg[1] = sarg1;
13961 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
13962 dtrace_vstate_t *vstate)
13966 if (helper->dtha_predicate != NULL)
13967 dtrace_difo_release(helper->dtha_predicate, vstate);
13969 for (i = 0; i < helper->dtha_nactions; i++) {
13970 ASSERT(helper->dtha_actions[i] != NULL);
13971 dtrace_difo_release(helper->dtha_actions[i], vstate);
13974 kmem_free(helper->dtha_actions,
13975 helper->dtha_nactions * sizeof (dtrace_difo_t *));
13976 kmem_free(helper, sizeof (dtrace_helper_action_t));
13980 dtrace_helper_destroygen(int gen)
13982 proc_t *p = curproc;
13983 dtrace_helpers_t *help = p->p_dtrace_helpers;
13984 dtrace_vstate_t *vstate;
13987 ASSERT(MUTEX_HELD(&dtrace_lock));
13989 if (help == NULL || gen > help->dthps_generation)
13992 vstate = &help->dthps_vstate;
13994 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13995 dtrace_helper_action_t *last = NULL, *h, *next;
13997 for (h = help->dthps_actions[i]; h != NULL; h = next) {
13998 next = h->dtha_next;
14000 if (h->dtha_generation == gen) {
14001 if (last != NULL) {
14002 last->dtha_next = next;
14004 help->dthps_actions[i] = next;
14007 dtrace_helper_action_destroy(h, vstate);
14015 * Interate until we've cleared out all helper providers with the
14016 * given generation number.
14019 dtrace_helper_provider_t *prov;
14022 * Look for a helper provider with the right generation. We
14023 * have to start back at the beginning of the list each time
14024 * because we drop dtrace_lock. It's unlikely that we'll make
14025 * more than two passes.
14027 for (i = 0; i < help->dthps_nprovs; i++) {
14028 prov = help->dthps_provs[i];
14030 if (prov->dthp_generation == gen)
14035 * If there were no matches, we're done.
14037 if (i == help->dthps_nprovs)
14041 * Move the last helper provider into this slot.
14043 help->dthps_nprovs--;
14044 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14045 help->dthps_provs[help->dthps_nprovs] = NULL;
14047 mutex_exit(&dtrace_lock);
14050 * If we have a meta provider, remove this helper provider.
14052 mutex_enter(&dtrace_meta_lock);
14053 if (dtrace_meta_pid != NULL) {
14054 ASSERT(dtrace_deferred_pid == NULL);
14055 dtrace_helper_provider_remove(&prov->dthp_prov,
14058 mutex_exit(&dtrace_meta_lock);
14060 dtrace_helper_provider_destroy(prov);
14062 mutex_enter(&dtrace_lock);
14071 dtrace_helper_validate(dtrace_helper_action_t *helper)
14076 if ((dp = helper->dtha_predicate) != NULL)
14077 err += dtrace_difo_validate_helper(dp);
14079 for (i = 0; i < helper->dtha_nactions; i++)
14080 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14088 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14090 dtrace_helpers_t *help;
14091 dtrace_helper_action_t *helper, *last;
14092 dtrace_actdesc_t *act;
14093 dtrace_vstate_t *vstate;
14094 dtrace_predicate_t *pred;
14095 int count = 0, nactions = 0, i;
14097 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14100 help = curproc->p_dtrace_helpers;
14101 last = help->dthps_actions[which];
14102 vstate = &help->dthps_vstate;
14104 for (count = 0; last != NULL; last = last->dtha_next) {
14106 if (last->dtha_next == NULL)
14111 * If we already have dtrace_helper_actions_max helper actions for this
14112 * helper action type, we'll refuse to add a new one.
14114 if (count >= dtrace_helper_actions_max)
14117 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14118 helper->dtha_generation = help->dthps_generation;
14120 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14121 ASSERT(pred->dtp_difo != NULL);
14122 dtrace_difo_hold(pred->dtp_difo);
14123 helper->dtha_predicate = pred->dtp_difo;
14126 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14127 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14130 if (act->dtad_difo == NULL)
14136 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14137 (helper->dtha_nactions = nactions), KM_SLEEP);
14139 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14140 dtrace_difo_hold(act->dtad_difo);
14141 helper->dtha_actions[i++] = act->dtad_difo;
14144 if (!dtrace_helper_validate(helper))
14147 if (last == NULL) {
14148 help->dthps_actions[which] = helper;
14150 last->dtha_next = helper;
14153 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14154 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14155 dtrace_helptrace_next = 0;
14160 dtrace_helper_action_destroy(helper, vstate);
14165 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14166 dof_helper_t *dofhp)
14168 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14170 mutex_enter(&dtrace_meta_lock);
14171 mutex_enter(&dtrace_lock);
14173 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14175 * If the dtrace module is loaded but not attached, or if
14176 * there aren't isn't a meta provider registered to deal with
14177 * these provider descriptions, we need to postpone creating
14178 * the actual providers until later.
14181 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14182 dtrace_deferred_pid != help) {
14183 help->dthps_deferred = 1;
14184 help->dthps_pid = p->p_pid;
14185 help->dthps_next = dtrace_deferred_pid;
14186 help->dthps_prev = NULL;
14187 if (dtrace_deferred_pid != NULL)
14188 dtrace_deferred_pid->dthps_prev = help;
14189 dtrace_deferred_pid = help;
14192 mutex_exit(&dtrace_lock);
14194 } else if (dofhp != NULL) {
14196 * If the dtrace module is loaded and we have a particular
14197 * helper provider description, pass that off to the
14201 mutex_exit(&dtrace_lock);
14203 dtrace_helper_provide(dofhp, p->p_pid);
14207 * Otherwise, just pass all the helper provider descriptions
14208 * off to the meta provider.
14212 mutex_exit(&dtrace_lock);
14214 for (i = 0; i < help->dthps_nprovs; i++) {
14215 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14220 mutex_exit(&dtrace_meta_lock);
14224 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14226 dtrace_helpers_t *help;
14227 dtrace_helper_provider_t *hprov, **tmp_provs;
14228 uint_t tmp_maxprovs, i;
14230 ASSERT(MUTEX_HELD(&dtrace_lock));
14232 help = curproc->p_dtrace_helpers;
14233 ASSERT(help != NULL);
14236 * If we already have dtrace_helper_providers_max helper providers,
14237 * we're refuse to add a new one.
14239 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14243 * Check to make sure this isn't a duplicate.
14245 for (i = 0; i < help->dthps_nprovs; i++) {
14246 if (dofhp->dofhp_addr ==
14247 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14251 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14252 hprov->dthp_prov = *dofhp;
14253 hprov->dthp_ref = 1;
14254 hprov->dthp_generation = gen;
14257 * Allocate a bigger table for helper providers if it's already full.
14259 if (help->dthps_maxprovs == help->dthps_nprovs) {
14260 tmp_maxprovs = help->dthps_maxprovs;
14261 tmp_provs = help->dthps_provs;
14263 if (help->dthps_maxprovs == 0)
14264 help->dthps_maxprovs = 2;
14266 help->dthps_maxprovs *= 2;
14267 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14268 help->dthps_maxprovs = dtrace_helper_providers_max;
14270 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14272 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14273 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14275 if (tmp_provs != NULL) {
14276 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14277 sizeof (dtrace_helper_provider_t *));
14278 kmem_free(tmp_provs, tmp_maxprovs *
14279 sizeof (dtrace_helper_provider_t *));
14283 help->dthps_provs[help->dthps_nprovs] = hprov;
14284 help->dthps_nprovs++;
14290 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14292 mutex_enter(&dtrace_lock);
14294 if (--hprov->dthp_ref == 0) {
14296 mutex_exit(&dtrace_lock);
14297 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14298 dtrace_dof_destroy(dof);
14299 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14301 mutex_exit(&dtrace_lock);
14306 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14308 uintptr_t daddr = (uintptr_t)dof;
14309 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14310 dof_provider_t *provider;
14311 dof_probe_t *probe;
14313 char *strtab, *typestr;
14314 dof_stridx_t typeidx;
14316 uint_t nprobes, j, k;
14318 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14320 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14321 dtrace_dof_error(dof, "misaligned section offset");
14326 * The section needs to be large enough to contain the DOF provider
14327 * structure appropriate for the given version.
14329 if (sec->dofs_size <
14330 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14331 offsetof(dof_provider_t, dofpv_prenoffs) :
14332 sizeof (dof_provider_t))) {
14333 dtrace_dof_error(dof, "provider section too small");
14337 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14338 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14339 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14340 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14341 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14343 if (str_sec == NULL || prb_sec == NULL ||
14344 arg_sec == NULL || off_sec == NULL)
14349 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14350 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14351 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14352 provider->dofpv_prenoffs)) == NULL)
14355 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14357 if (provider->dofpv_name >= str_sec->dofs_size ||
14358 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14359 dtrace_dof_error(dof, "invalid provider name");
14363 if (prb_sec->dofs_entsize == 0 ||
14364 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14365 dtrace_dof_error(dof, "invalid entry size");
14369 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14370 dtrace_dof_error(dof, "misaligned entry size");
14374 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14375 dtrace_dof_error(dof, "invalid entry size");
14379 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14380 dtrace_dof_error(dof, "misaligned section offset");
14384 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14385 dtrace_dof_error(dof, "invalid entry size");
14389 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14391 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14394 * Take a pass through the probes to check for errors.
14396 for (j = 0; j < nprobes; j++) {
14397 probe = (dof_probe_t *)(uintptr_t)(daddr +
14398 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14400 if (probe->dofpr_func >= str_sec->dofs_size) {
14401 dtrace_dof_error(dof, "invalid function name");
14405 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14406 dtrace_dof_error(dof, "function name too long");
14410 if (probe->dofpr_name >= str_sec->dofs_size ||
14411 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14412 dtrace_dof_error(dof, "invalid probe name");
14417 * The offset count must not wrap the index, and the offsets
14418 * must also not overflow the section's data.
14420 if (probe->dofpr_offidx + probe->dofpr_noffs <
14421 probe->dofpr_offidx ||
14422 (probe->dofpr_offidx + probe->dofpr_noffs) *
14423 off_sec->dofs_entsize > off_sec->dofs_size) {
14424 dtrace_dof_error(dof, "invalid probe offset");
14428 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14430 * If there's no is-enabled offset section, make sure
14431 * there aren't any is-enabled offsets. Otherwise
14432 * perform the same checks as for probe offsets
14433 * (immediately above).
14435 if (enoff_sec == NULL) {
14436 if (probe->dofpr_enoffidx != 0 ||
14437 probe->dofpr_nenoffs != 0) {
14438 dtrace_dof_error(dof, "is-enabled "
14439 "offsets with null section");
14442 } else if (probe->dofpr_enoffidx +
14443 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14444 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14445 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14446 dtrace_dof_error(dof, "invalid is-enabled "
14451 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14452 dtrace_dof_error(dof, "zero probe and "
14453 "is-enabled offsets");
14456 } else if (probe->dofpr_noffs == 0) {
14457 dtrace_dof_error(dof, "zero probe offsets");
14461 if (probe->dofpr_argidx + probe->dofpr_xargc <
14462 probe->dofpr_argidx ||
14463 (probe->dofpr_argidx + probe->dofpr_xargc) *
14464 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14465 dtrace_dof_error(dof, "invalid args");
14469 typeidx = probe->dofpr_nargv;
14470 typestr = strtab + probe->dofpr_nargv;
14471 for (k = 0; k < probe->dofpr_nargc; k++) {
14472 if (typeidx >= str_sec->dofs_size) {
14473 dtrace_dof_error(dof, "bad "
14474 "native argument type");
14478 typesz = strlen(typestr) + 1;
14479 if (typesz > DTRACE_ARGTYPELEN) {
14480 dtrace_dof_error(dof, "native "
14481 "argument type too long");
14488 typeidx = probe->dofpr_xargv;
14489 typestr = strtab + probe->dofpr_xargv;
14490 for (k = 0; k < probe->dofpr_xargc; k++) {
14491 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14492 dtrace_dof_error(dof, "bad "
14493 "native argument index");
14497 if (typeidx >= str_sec->dofs_size) {
14498 dtrace_dof_error(dof, "bad "
14499 "translated argument type");
14503 typesz = strlen(typestr) + 1;
14504 if (typesz > DTRACE_ARGTYPELEN) {
14505 dtrace_dof_error(dof, "translated argument "
14519 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14521 dtrace_helpers_t *help;
14522 dtrace_vstate_t *vstate;
14523 dtrace_enabling_t *enab = NULL;
14524 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14525 uintptr_t daddr = (uintptr_t)dof;
14527 ASSERT(MUTEX_HELD(&dtrace_lock));
14529 if ((help = curproc->p_dtrace_helpers) == NULL)
14530 help = dtrace_helpers_create(curproc);
14532 vstate = &help->dthps_vstate;
14534 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14535 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14536 dtrace_dof_destroy(dof);
14541 * Look for helper providers and validate their descriptions.
14544 for (i = 0; i < dof->dofh_secnum; i++) {
14545 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14546 dof->dofh_secoff + i * dof->dofh_secsize);
14548 if (sec->dofs_type != DOF_SECT_PROVIDER)
14551 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14552 dtrace_enabling_destroy(enab);
14553 dtrace_dof_destroy(dof);
14562 * Now we need to walk through the ECB descriptions in the enabling.
14564 for (i = 0; i < enab->dten_ndesc; i++) {
14565 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14566 dtrace_probedesc_t *desc = &ep->dted_probe;
14568 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14571 if (strcmp(desc->dtpd_mod, "helper") != 0)
14574 if (strcmp(desc->dtpd_func, "ustack") != 0)
14577 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14580 * Adding this helper action failed -- we are now going
14581 * to rip out the entire generation and return failure.
14583 (void) dtrace_helper_destroygen(help->dthps_generation);
14584 dtrace_enabling_destroy(enab);
14585 dtrace_dof_destroy(dof);
14592 if (nhelpers < enab->dten_ndesc)
14593 dtrace_dof_error(dof, "unmatched helpers");
14595 gen = help->dthps_generation++;
14596 dtrace_enabling_destroy(enab);
14598 if (dhp != NULL && nprovs > 0) {
14599 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14600 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14601 mutex_exit(&dtrace_lock);
14602 dtrace_helper_provider_register(curproc, help, dhp);
14603 mutex_enter(&dtrace_lock);
14610 dtrace_dof_destroy(dof);
14615 static dtrace_helpers_t *
14616 dtrace_helpers_create(proc_t *p)
14618 dtrace_helpers_t *help;
14620 ASSERT(MUTEX_HELD(&dtrace_lock));
14621 ASSERT(p->p_dtrace_helpers == NULL);
14623 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14624 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14625 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14627 p->p_dtrace_helpers = help;
14634 dtrace_helpers_destroy(void)
14636 dtrace_helpers_t *help;
14637 dtrace_vstate_t *vstate;
14638 proc_t *p = curproc;
14641 mutex_enter(&dtrace_lock);
14643 ASSERT(p->p_dtrace_helpers != NULL);
14644 ASSERT(dtrace_helpers > 0);
14646 help = p->p_dtrace_helpers;
14647 vstate = &help->dthps_vstate;
14650 * We're now going to lose the help from this process.
14652 p->p_dtrace_helpers = NULL;
14656 * Destory the helper actions.
14658 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14659 dtrace_helper_action_t *h, *next;
14661 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14662 next = h->dtha_next;
14663 dtrace_helper_action_destroy(h, vstate);
14668 mutex_exit(&dtrace_lock);
14671 * Destroy the helper providers.
14673 if (help->dthps_maxprovs > 0) {
14674 mutex_enter(&dtrace_meta_lock);
14675 if (dtrace_meta_pid != NULL) {
14676 ASSERT(dtrace_deferred_pid == NULL);
14678 for (i = 0; i < help->dthps_nprovs; i++) {
14679 dtrace_helper_provider_remove(
14680 &help->dthps_provs[i]->dthp_prov, p->p_pid);
14683 mutex_enter(&dtrace_lock);
14684 ASSERT(help->dthps_deferred == 0 ||
14685 help->dthps_next != NULL ||
14686 help->dthps_prev != NULL ||
14687 help == dtrace_deferred_pid);
14690 * Remove the helper from the deferred list.
14692 if (help->dthps_next != NULL)
14693 help->dthps_next->dthps_prev = help->dthps_prev;
14694 if (help->dthps_prev != NULL)
14695 help->dthps_prev->dthps_next = help->dthps_next;
14696 if (dtrace_deferred_pid == help) {
14697 dtrace_deferred_pid = help->dthps_next;
14698 ASSERT(help->dthps_prev == NULL);
14701 mutex_exit(&dtrace_lock);
14704 mutex_exit(&dtrace_meta_lock);
14706 for (i = 0; i < help->dthps_nprovs; i++) {
14707 dtrace_helper_provider_destroy(help->dthps_provs[i]);
14710 kmem_free(help->dthps_provs, help->dthps_maxprovs *
14711 sizeof (dtrace_helper_provider_t *));
14714 mutex_enter(&dtrace_lock);
14716 dtrace_vstate_fini(&help->dthps_vstate);
14717 kmem_free(help->dthps_actions,
14718 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
14719 kmem_free(help, sizeof (dtrace_helpers_t));
14722 mutex_exit(&dtrace_lock);
14726 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
14728 dtrace_helpers_t *help, *newhelp;
14729 dtrace_helper_action_t *helper, *new, *last;
14731 dtrace_vstate_t *vstate;
14732 int i, j, sz, hasprovs = 0;
14734 mutex_enter(&dtrace_lock);
14735 ASSERT(from->p_dtrace_helpers != NULL);
14736 ASSERT(dtrace_helpers > 0);
14738 help = from->p_dtrace_helpers;
14739 newhelp = dtrace_helpers_create(to);
14740 ASSERT(to->p_dtrace_helpers != NULL);
14742 newhelp->dthps_generation = help->dthps_generation;
14743 vstate = &newhelp->dthps_vstate;
14746 * Duplicate the helper actions.
14748 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14749 if ((helper = help->dthps_actions[i]) == NULL)
14752 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
14753 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
14755 new->dtha_generation = helper->dtha_generation;
14757 if ((dp = helper->dtha_predicate) != NULL) {
14758 dp = dtrace_difo_duplicate(dp, vstate);
14759 new->dtha_predicate = dp;
14762 new->dtha_nactions = helper->dtha_nactions;
14763 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
14764 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
14766 for (j = 0; j < new->dtha_nactions; j++) {
14767 dtrace_difo_t *dp = helper->dtha_actions[j];
14769 ASSERT(dp != NULL);
14770 dp = dtrace_difo_duplicate(dp, vstate);
14771 new->dtha_actions[j] = dp;
14774 if (last != NULL) {
14775 last->dtha_next = new;
14777 newhelp->dthps_actions[i] = new;
14785 * Duplicate the helper providers and register them with the
14786 * DTrace framework.
14788 if (help->dthps_nprovs > 0) {
14789 newhelp->dthps_nprovs = help->dthps_nprovs;
14790 newhelp->dthps_maxprovs = help->dthps_nprovs;
14791 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
14792 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14793 for (i = 0; i < newhelp->dthps_nprovs; i++) {
14794 newhelp->dthps_provs[i] = help->dthps_provs[i];
14795 newhelp->dthps_provs[i]->dthp_ref++;
14801 mutex_exit(&dtrace_lock);
14804 dtrace_helper_provider_register(to, newhelp, NULL);
14810 * DTrace Hook Functions
14813 dtrace_module_loaded(modctl_t *ctl)
14815 dtrace_provider_t *prv;
14817 mutex_enter(&dtrace_provider_lock);
14818 mutex_enter(&mod_lock);
14820 ASSERT(ctl->mod_busy);
14823 * We're going to call each providers per-module provide operation
14824 * specifying only this module.
14826 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
14827 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
14829 mutex_exit(&mod_lock);
14830 mutex_exit(&dtrace_provider_lock);
14833 * If we have any retained enablings, we need to match against them.
14834 * Enabling probes requires that cpu_lock be held, and we cannot hold
14835 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
14836 * module. (In particular, this happens when loading scheduling
14837 * classes.) So if we have any retained enablings, we need to dispatch
14838 * our task queue to do the match for us.
14840 mutex_enter(&dtrace_lock);
14842 if (dtrace_retained == NULL) {
14843 mutex_exit(&dtrace_lock);
14847 (void) taskq_dispatch(dtrace_taskq,
14848 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
14850 mutex_exit(&dtrace_lock);
14853 * And now, for a little heuristic sleaze: in general, we want to
14854 * match modules as soon as they load. However, we cannot guarantee
14855 * this, because it would lead us to the lock ordering violation
14856 * outlined above. The common case, of course, is that cpu_lock is
14857 * _not_ held -- so we delay here for a clock tick, hoping that that's
14858 * long enough for the task queue to do its work. If it's not, it's
14859 * not a serious problem -- it just means that the module that we
14860 * just loaded may not be immediately instrumentable.
14866 dtrace_module_unloaded(modctl_t *ctl)
14868 dtrace_probe_t template, *probe, *first, *next;
14869 dtrace_provider_t *prov;
14871 template.dtpr_mod = ctl->mod_modname;
14873 mutex_enter(&dtrace_provider_lock);
14874 mutex_enter(&mod_lock);
14875 mutex_enter(&dtrace_lock);
14877 if (dtrace_bymod == NULL) {
14879 * The DTrace module is loaded (obviously) but not attached;
14880 * we don't have any work to do.
14882 mutex_exit(&dtrace_provider_lock);
14883 mutex_exit(&mod_lock);
14884 mutex_exit(&dtrace_lock);
14888 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14889 probe != NULL; probe = probe->dtpr_nextmod) {
14890 if (probe->dtpr_ecb != NULL) {
14891 mutex_exit(&dtrace_provider_lock);
14892 mutex_exit(&mod_lock);
14893 mutex_exit(&dtrace_lock);
14896 * This shouldn't _actually_ be possible -- we're
14897 * unloading a module that has an enabled probe in it.
14898 * (It's normally up to the provider to make sure that
14899 * this can't happen.) However, because dtps_enable()
14900 * doesn't have a failure mode, there can be an
14901 * enable/unload race. Upshot: we don't want to
14902 * assert, but we're not going to disable the
14905 if (dtrace_err_verbose) {
14906 cmn_err(CE_WARN, "unloaded module '%s' had "
14907 "enabled probes", ctl->mod_modname);
14916 for (first = NULL; probe != NULL; probe = next) {
14917 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
14919 dtrace_probes[probe->dtpr_id - 1] = NULL;
14921 next = probe->dtpr_nextmod;
14922 dtrace_hash_remove(dtrace_bymod, probe);
14923 dtrace_hash_remove(dtrace_byfunc, probe);
14924 dtrace_hash_remove(dtrace_byname, probe);
14926 if (first == NULL) {
14928 probe->dtpr_nextmod = NULL;
14930 probe->dtpr_nextmod = first;
14936 * We've removed all of the module's probes from the hash chains and
14937 * from the probe array. Now issue a dtrace_sync() to be sure that
14938 * everyone has cleared out from any probe array processing.
14942 for (probe = first; probe != NULL; probe = first) {
14943 first = probe->dtpr_nextmod;
14944 prov = probe->dtpr_provider;
14945 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
14947 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
14948 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
14949 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
14950 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
14951 kmem_free(probe, sizeof (dtrace_probe_t));
14954 mutex_exit(&dtrace_lock);
14955 mutex_exit(&mod_lock);
14956 mutex_exit(&dtrace_provider_lock);
14960 dtrace_suspend(void)
14962 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
14966 dtrace_resume(void)
14968 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
14973 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
14975 ASSERT(MUTEX_HELD(&cpu_lock));
14976 mutex_enter(&dtrace_lock);
14980 dtrace_state_t *state;
14981 dtrace_optval_t *opt, rs, c;
14984 * For now, we only allocate a new buffer for anonymous state.
14986 if ((state = dtrace_anon.dta_state) == NULL)
14989 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14992 opt = state->dts_options;
14993 c = opt[DTRACEOPT_CPU];
14995 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
14999 * Regardless of what the actual policy is, we're going to
15000 * temporarily set our resize policy to be manual. We're
15001 * also going to temporarily set our CPU option to denote
15002 * the newly configured CPU.
15004 rs = opt[DTRACEOPT_BUFRESIZE];
15005 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15006 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15008 (void) dtrace_state_buffers(state);
15010 opt[DTRACEOPT_BUFRESIZE] = rs;
15011 opt[DTRACEOPT_CPU] = c;
15018 * We don't free the buffer in the CPU_UNCONFIG case. (The
15019 * buffer will be freed when the consumer exits.)
15027 mutex_exit(&dtrace_lock);
15033 dtrace_cpu_setup_initial(processorid_t cpu)
15035 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15040 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15042 if (dtrace_toxranges >= dtrace_toxranges_max) {
15044 dtrace_toxrange_t *range;
15046 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15049 ASSERT(dtrace_toxrange == NULL);
15050 ASSERT(dtrace_toxranges_max == 0);
15051 dtrace_toxranges_max = 1;
15053 dtrace_toxranges_max <<= 1;
15056 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15057 range = kmem_zalloc(nsize, KM_SLEEP);
15059 if (dtrace_toxrange != NULL) {
15060 ASSERT(osize != 0);
15061 bcopy(dtrace_toxrange, range, osize);
15062 kmem_free(dtrace_toxrange, osize);
15065 dtrace_toxrange = range;
15068 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15069 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15071 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15072 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15073 dtrace_toxranges++;
15077 * DTrace Driver Cookbook Functions
15082 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15084 dtrace_provider_id_t id;
15085 dtrace_state_t *state = NULL;
15086 dtrace_enabling_t *enab;
15088 mutex_enter(&cpu_lock);
15089 mutex_enter(&dtrace_provider_lock);
15090 mutex_enter(&dtrace_lock);
15092 if (ddi_soft_state_init(&dtrace_softstate,
15093 sizeof (dtrace_state_t), 0) != 0) {
15094 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15095 mutex_exit(&cpu_lock);
15096 mutex_exit(&dtrace_provider_lock);
15097 mutex_exit(&dtrace_lock);
15098 return (DDI_FAILURE);
15101 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15102 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15103 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15104 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15105 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15106 ddi_remove_minor_node(devi, NULL);
15107 ddi_soft_state_fini(&dtrace_softstate);
15108 mutex_exit(&cpu_lock);
15109 mutex_exit(&dtrace_provider_lock);
15110 mutex_exit(&dtrace_lock);
15111 return (DDI_FAILURE);
15114 ddi_report_dev(devi);
15115 dtrace_devi = devi;
15117 dtrace_modload = dtrace_module_loaded;
15118 dtrace_modunload = dtrace_module_unloaded;
15119 dtrace_cpu_init = dtrace_cpu_setup_initial;
15120 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15121 dtrace_helpers_fork = dtrace_helpers_duplicate;
15122 dtrace_cpustart_init = dtrace_suspend;
15123 dtrace_cpustart_fini = dtrace_resume;
15124 dtrace_debugger_init = dtrace_suspend;
15125 dtrace_debugger_fini = dtrace_resume;
15127 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15129 ASSERT(MUTEX_HELD(&cpu_lock));
15131 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15132 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15133 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15134 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15135 VM_SLEEP | VMC_IDENTIFIER);
15136 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15139 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15140 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15141 NULL, NULL, NULL, NULL, NULL, 0);
15143 ASSERT(MUTEX_HELD(&cpu_lock));
15144 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15145 offsetof(dtrace_probe_t, dtpr_nextmod),
15146 offsetof(dtrace_probe_t, dtpr_prevmod));
15148 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15149 offsetof(dtrace_probe_t, dtpr_nextfunc),
15150 offsetof(dtrace_probe_t, dtpr_prevfunc));
15152 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15153 offsetof(dtrace_probe_t, dtpr_nextname),
15154 offsetof(dtrace_probe_t, dtpr_prevname));
15156 if (dtrace_retain_max < 1) {
15157 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15158 "setting to 1", dtrace_retain_max);
15159 dtrace_retain_max = 1;
15163 * Now discover our toxic ranges.
15165 dtrace_toxic_ranges(dtrace_toxrange_add);
15168 * Before we register ourselves as a provider to our own framework,
15169 * we would like to assert that dtrace_provider is NULL -- but that's
15170 * not true if we were loaded as a dependency of a DTrace provider.
15171 * Once we've registered, we can assert that dtrace_provider is our
15174 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15175 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15177 ASSERT(dtrace_provider != NULL);
15178 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15180 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15181 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15182 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15183 dtrace_provider, NULL, NULL, "END", 0, NULL);
15184 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15185 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15187 dtrace_anon_property();
15188 mutex_exit(&cpu_lock);
15191 * If DTrace helper tracing is enabled, we need to allocate the
15192 * trace buffer and initialize the values.
15194 if (dtrace_helptrace_enabled) {
15195 ASSERT(dtrace_helptrace_buffer == NULL);
15196 dtrace_helptrace_buffer =
15197 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15198 dtrace_helptrace_next = 0;
15202 * If there are already providers, we must ask them to provide their
15203 * probes, and then match any anonymous enabling against them. Note
15204 * that there should be no other retained enablings at this time:
15205 * the only retained enablings at this time should be the anonymous
15208 if (dtrace_anon.dta_enabling != NULL) {
15209 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15211 dtrace_enabling_provide(NULL);
15212 state = dtrace_anon.dta_state;
15215 * We couldn't hold cpu_lock across the above call to
15216 * dtrace_enabling_provide(), but we must hold it to actually
15217 * enable the probes. We have to drop all of our locks, pick
15218 * up cpu_lock, and regain our locks before matching the
15219 * retained anonymous enabling.
15221 mutex_exit(&dtrace_lock);
15222 mutex_exit(&dtrace_provider_lock);
15224 mutex_enter(&cpu_lock);
15225 mutex_enter(&dtrace_provider_lock);
15226 mutex_enter(&dtrace_lock);
15228 if ((enab = dtrace_anon.dta_enabling) != NULL)
15229 (void) dtrace_enabling_match(enab, NULL);
15231 mutex_exit(&cpu_lock);
15234 mutex_exit(&dtrace_lock);
15235 mutex_exit(&dtrace_provider_lock);
15237 if (state != NULL) {
15239 * If we created any anonymous state, set it going now.
15241 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15244 return (DDI_SUCCESS);
15249 #if __FreeBSD_version >= 800039
15251 dtrace_dtr(void *data __unused)
15260 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15262 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15265 dtrace_state_t *state;
15271 if (getminor(*devp) == DTRACEMNRN_HELPER)
15275 * If this wasn't an open with the "helper" minor, then it must be
15276 * the "dtrace" minor.
15278 ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15280 cred_t *cred_p = NULL;
15282 #if __FreeBSD_version < 800039
15284 * The first minor device is the one that is cloned so there is
15285 * nothing more to do here.
15287 if (dev2unit(dev) == 0)
15291 * Devices are cloned, so if the DTrace state has already
15292 * been allocated, that means this device belongs to a
15293 * different client. Each client should open '/dev/dtrace'
15294 * to get a cloned device.
15296 if (dev->si_drv1 != NULL)
15300 cred_p = dev->si_cred;
15304 * If no DTRACE_PRIV_* bits are set in the credential, then the
15305 * caller lacks sufficient permission to do anything with DTrace.
15307 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15308 if (priv == DTRACE_PRIV_NONE) {
15310 #if __FreeBSD_version < 800039
15311 /* Destroy the cloned device. */
15320 * Ask all providers to provide all their probes.
15322 mutex_enter(&dtrace_provider_lock);
15323 dtrace_probe_provide(NULL, NULL);
15324 mutex_exit(&dtrace_provider_lock);
15326 mutex_enter(&cpu_lock);
15327 mutex_enter(&dtrace_lock);
15329 dtrace_membar_producer();
15333 * If the kernel debugger is active (that is, if the kernel debugger
15334 * modified text in some way), we won't allow the open.
15336 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15338 mutex_exit(&cpu_lock);
15339 mutex_exit(&dtrace_lock);
15343 state = dtrace_state_create(devp, cred_p);
15345 state = dtrace_state_create(dev);
15346 #if __FreeBSD_version < 800039
15347 dev->si_drv1 = state;
15349 devfs_set_cdevpriv(state, dtrace_dtr);
15353 mutex_exit(&cpu_lock);
15355 if (state == NULL) {
15357 if (--dtrace_opens == 0)
15358 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15362 mutex_exit(&dtrace_lock);
15364 #if __FreeBSD_version < 800039
15365 /* Destroy the cloned device. */
15372 mutex_exit(&dtrace_lock);
15380 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15382 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15386 minor_t minor = getminor(dev);
15387 dtrace_state_t *state;
15389 if (minor == DTRACEMNRN_HELPER)
15392 state = ddi_get_soft_state(dtrace_softstate, minor);
15394 #if __FreeBSD_version < 800039
15395 dtrace_state_t *state = dev->si_drv1;
15397 /* Check if this is not a cloned device. */
15398 if (dev2unit(dev) == 0)
15401 dtrace_state_t *state;
15402 devfs_get_cdevpriv((void **) &state);
15407 mutex_enter(&cpu_lock);
15408 mutex_enter(&dtrace_lock);
15410 if (state != NULL) {
15411 if (state->dts_anon) {
15413 * There is anonymous state. Destroy that first.
15415 ASSERT(dtrace_anon.dta_state == NULL);
15416 dtrace_state_destroy(state->dts_anon);
15419 dtrace_state_destroy(state);
15422 kmem_free(state, 0);
15423 #if __FreeBSD_version < 800039
15424 dev->si_drv1 = NULL;
15426 devfs_clear_cdevpriv();
15431 ASSERT(dtrace_opens > 0);
15433 if (--dtrace_opens == 0)
15434 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15439 mutex_exit(&dtrace_lock);
15440 mutex_exit(&cpu_lock);
15442 #if __FreeBSD_version < 800039
15443 /* Schedule this cloned device to be destroyed. */
15444 destroy_dev_sched(dev);
15453 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15456 dof_helper_t help, *dhp = NULL;
15459 case DTRACEHIOC_ADDDOF:
15460 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15461 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15466 arg = (intptr_t)help.dofhp_dof;
15469 case DTRACEHIOC_ADD: {
15470 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15475 mutex_enter(&dtrace_lock);
15478 * dtrace_helper_slurp() takes responsibility for the dof --
15479 * it may free it now or it may save it and free it later.
15481 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15488 mutex_exit(&dtrace_lock);
15492 case DTRACEHIOC_REMOVE: {
15493 mutex_enter(&dtrace_lock);
15494 rval = dtrace_helper_destroygen(arg);
15495 mutex_exit(&dtrace_lock);
15509 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15511 minor_t minor = getminor(dev);
15512 dtrace_state_t *state;
15515 if (minor == DTRACEMNRN_HELPER)
15516 return (dtrace_ioctl_helper(cmd, arg, rv));
15518 state = ddi_get_soft_state(dtrace_softstate, minor);
15520 if (state->dts_anon) {
15521 ASSERT(dtrace_anon.dta_state == NULL);
15522 state = state->dts_anon;
15526 case DTRACEIOC_PROVIDER: {
15527 dtrace_providerdesc_t pvd;
15528 dtrace_provider_t *pvp;
15530 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15533 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15534 mutex_enter(&dtrace_provider_lock);
15536 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15537 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15541 mutex_exit(&dtrace_provider_lock);
15546 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15547 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15549 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15555 case DTRACEIOC_EPROBE: {
15556 dtrace_eprobedesc_t epdesc;
15558 dtrace_action_t *act;
15564 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15567 mutex_enter(&dtrace_lock);
15569 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15570 mutex_exit(&dtrace_lock);
15574 if (ecb->dte_probe == NULL) {
15575 mutex_exit(&dtrace_lock);
15579 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15580 epdesc.dtepd_uarg = ecb->dte_uarg;
15581 epdesc.dtepd_size = ecb->dte_size;
15583 nrecs = epdesc.dtepd_nrecs;
15584 epdesc.dtepd_nrecs = 0;
15585 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15586 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15589 epdesc.dtepd_nrecs++;
15593 * Now that we have the size, we need to allocate a temporary
15594 * buffer in which to store the complete description. We need
15595 * the temporary buffer to be able to drop dtrace_lock()
15596 * across the copyout(), below.
15598 size = sizeof (dtrace_eprobedesc_t) +
15599 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15601 buf = kmem_alloc(size, KM_SLEEP);
15602 dest = (uintptr_t)buf;
15604 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15605 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15607 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15608 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15614 bcopy(&act->dta_rec, (void *)dest,
15615 sizeof (dtrace_recdesc_t));
15616 dest += sizeof (dtrace_recdesc_t);
15619 mutex_exit(&dtrace_lock);
15621 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15622 kmem_free(buf, size);
15626 kmem_free(buf, size);
15630 case DTRACEIOC_AGGDESC: {
15631 dtrace_aggdesc_t aggdesc;
15632 dtrace_action_t *act;
15633 dtrace_aggregation_t *agg;
15636 dtrace_recdesc_t *lrec;
15641 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15644 mutex_enter(&dtrace_lock);
15646 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15647 mutex_exit(&dtrace_lock);
15651 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15653 nrecs = aggdesc.dtagd_nrecs;
15654 aggdesc.dtagd_nrecs = 0;
15656 offs = agg->dtag_base;
15657 lrec = &agg->dtag_action.dta_rec;
15658 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15660 for (act = agg->dtag_first; ; act = act->dta_next) {
15661 ASSERT(act->dta_intuple ||
15662 DTRACEACT_ISAGG(act->dta_kind));
15665 * If this action has a record size of zero, it
15666 * denotes an argument to the aggregating action.
15667 * Because the presence of this record doesn't (or
15668 * shouldn't) affect the way the data is interpreted,
15669 * we don't copy it out to save user-level the
15670 * confusion of dealing with a zero-length record.
15672 if (act->dta_rec.dtrd_size == 0) {
15673 ASSERT(agg->dtag_hasarg);
15677 aggdesc.dtagd_nrecs++;
15679 if (act == &agg->dtag_action)
15684 * Now that we have the size, we need to allocate a temporary
15685 * buffer in which to store the complete description. We need
15686 * the temporary buffer to be able to drop dtrace_lock()
15687 * across the copyout(), below.
15689 size = sizeof (dtrace_aggdesc_t) +
15690 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15692 buf = kmem_alloc(size, KM_SLEEP);
15693 dest = (uintptr_t)buf;
15695 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15696 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15698 for (act = agg->dtag_first; ; act = act->dta_next) {
15699 dtrace_recdesc_t rec = act->dta_rec;
15702 * See the comment in the above loop for why we pass
15703 * over zero-length records.
15705 if (rec.dtrd_size == 0) {
15706 ASSERT(agg->dtag_hasarg);
15713 rec.dtrd_offset -= offs;
15714 bcopy(&rec, (void *)dest, sizeof (rec));
15715 dest += sizeof (dtrace_recdesc_t);
15717 if (act == &agg->dtag_action)
15721 mutex_exit(&dtrace_lock);
15723 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15724 kmem_free(buf, size);
15728 kmem_free(buf, size);
15732 case DTRACEIOC_ENABLE: {
15734 dtrace_enabling_t *enab = NULL;
15735 dtrace_vstate_t *vstate;
15741 * If a NULL argument has been passed, we take this as our
15742 * cue to reevaluate our enablings.
15745 dtrace_enabling_matchall();
15750 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
15753 mutex_enter(&cpu_lock);
15754 mutex_enter(&dtrace_lock);
15755 vstate = &state->dts_vstate;
15757 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
15758 mutex_exit(&dtrace_lock);
15759 mutex_exit(&cpu_lock);
15760 dtrace_dof_destroy(dof);
15764 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
15765 mutex_exit(&dtrace_lock);
15766 mutex_exit(&cpu_lock);
15767 dtrace_dof_destroy(dof);
15771 if ((rval = dtrace_dof_options(dof, state)) != 0) {
15772 dtrace_enabling_destroy(enab);
15773 mutex_exit(&dtrace_lock);
15774 mutex_exit(&cpu_lock);
15775 dtrace_dof_destroy(dof);
15779 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
15780 err = dtrace_enabling_retain(enab);
15782 dtrace_enabling_destroy(enab);
15785 mutex_exit(&cpu_lock);
15786 mutex_exit(&dtrace_lock);
15787 dtrace_dof_destroy(dof);
15792 case DTRACEIOC_REPLICATE: {
15793 dtrace_repldesc_t desc;
15794 dtrace_probedesc_t *match = &desc.dtrpd_match;
15795 dtrace_probedesc_t *create = &desc.dtrpd_create;
15798 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15801 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15802 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15803 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15804 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15806 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15807 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15808 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15809 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15811 mutex_enter(&dtrace_lock);
15812 err = dtrace_enabling_replicate(state, match, create);
15813 mutex_exit(&dtrace_lock);
15818 case DTRACEIOC_PROBEMATCH:
15819 case DTRACEIOC_PROBES: {
15820 dtrace_probe_t *probe = NULL;
15821 dtrace_probedesc_t desc;
15822 dtrace_probekey_t pkey;
15829 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15832 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15833 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15834 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15835 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15838 * Before we attempt to match this probe, we want to give
15839 * all providers the opportunity to provide it.
15841 if (desc.dtpd_id == DTRACE_IDNONE) {
15842 mutex_enter(&dtrace_provider_lock);
15843 dtrace_probe_provide(&desc, NULL);
15844 mutex_exit(&dtrace_provider_lock);
15848 if (cmd == DTRACEIOC_PROBEMATCH) {
15849 dtrace_probekey(&desc, &pkey);
15850 pkey.dtpk_id = DTRACE_IDNONE;
15853 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
15855 mutex_enter(&dtrace_lock);
15857 if (cmd == DTRACEIOC_PROBEMATCH) {
15858 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15859 if ((probe = dtrace_probes[i - 1]) != NULL &&
15860 (m = dtrace_match_probe(probe, &pkey,
15861 priv, uid, zoneid)) != 0)
15866 mutex_exit(&dtrace_lock);
15871 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15872 if ((probe = dtrace_probes[i - 1]) != NULL &&
15873 dtrace_match_priv(probe, priv, uid, zoneid))
15878 if (probe == NULL) {
15879 mutex_exit(&dtrace_lock);
15883 dtrace_probe_description(probe, &desc);
15884 mutex_exit(&dtrace_lock);
15886 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15892 case DTRACEIOC_PROBEARG: {
15893 dtrace_argdesc_t desc;
15894 dtrace_probe_t *probe;
15895 dtrace_provider_t *prov;
15897 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15900 if (desc.dtargd_id == DTRACE_IDNONE)
15903 if (desc.dtargd_ndx == DTRACE_ARGNONE)
15906 mutex_enter(&dtrace_provider_lock);
15907 mutex_enter(&mod_lock);
15908 mutex_enter(&dtrace_lock);
15910 if (desc.dtargd_id > dtrace_nprobes) {
15911 mutex_exit(&dtrace_lock);
15912 mutex_exit(&mod_lock);
15913 mutex_exit(&dtrace_provider_lock);
15917 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
15918 mutex_exit(&dtrace_lock);
15919 mutex_exit(&mod_lock);
15920 mutex_exit(&dtrace_provider_lock);
15924 mutex_exit(&dtrace_lock);
15926 prov = probe->dtpr_provider;
15928 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
15930 * There isn't any typed information for this probe.
15931 * Set the argument number to DTRACE_ARGNONE.
15933 desc.dtargd_ndx = DTRACE_ARGNONE;
15935 desc.dtargd_native[0] = '\0';
15936 desc.dtargd_xlate[0] = '\0';
15937 desc.dtargd_mapping = desc.dtargd_ndx;
15939 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
15940 probe->dtpr_id, probe->dtpr_arg, &desc);
15943 mutex_exit(&mod_lock);
15944 mutex_exit(&dtrace_provider_lock);
15946 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15952 case DTRACEIOC_GO: {
15953 processorid_t cpuid;
15954 rval = dtrace_state_go(state, &cpuid);
15959 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15965 case DTRACEIOC_STOP: {
15966 processorid_t cpuid;
15968 mutex_enter(&dtrace_lock);
15969 rval = dtrace_state_stop(state, &cpuid);
15970 mutex_exit(&dtrace_lock);
15975 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15981 case DTRACEIOC_DOFGET: {
15982 dof_hdr_t hdr, *dof;
15985 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
15988 mutex_enter(&dtrace_lock);
15989 dof = dtrace_dof_create(state);
15990 mutex_exit(&dtrace_lock);
15992 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
15993 rval = copyout(dof, (void *)arg, len);
15994 dtrace_dof_destroy(dof);
15996 return (rval == 0 ? 0 : EFAULT);
15999 case DTRACEIOC_AGGSNAP:
16000 case DTRACEIOC_BUFSNAP: {
16001 dtrace_bufdesc_t desc;
16003 dtrace_buffer_t *buf;
16005 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16008 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16011 mutex_enter(&dtrace_lock);
16013 if (cmd == DTRACEIOC_BUFSNAP) {
16014 buf = &state->dts_buffer[desc.dtbd_cpu];
16016 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16019 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16020 size_t sz = buf->dtb_offset;
16022 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16023 mutex_exit(&dtrace_lock);
16028 * If this buffer has already been consumed, we're
16029 * going to indicate that there's nothing left here
16032 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16033 mutex_exit(&dtrace_lock);
16035 desc.dtbd_size = 0;
16036 desc.dtbd_drops = 0;
16037 desc.dtbd_errors = 0;
16038 desc.dtbd_oldest = 0;
16039 sz = sizeof (desc);
16041 if (copyout(&desc, (void *)arg, sz) != 0)
16048 * If this is a ring buffer that has wrapped, we want
16049 * to copy the whole thing out.
16051 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16052 dtrace_buffer_polish(buf);
16053 sz = buf->dtb_size;
16056 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16057 mutex_exit(&dtrace_lock);
16061 desc.dtbd_size = sz;
16062 desc.dtbd_drops = buf->dtb_drops;
16063 desc.dtbd_errors = buf->dtb_errors;
16064 desc.dtbd_oldest = buf->dtb_xamot_offset;
16066 mutex_exit(&dtrace_lock);
16068 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16071 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16076 if (buf->dtb_tomax == NULL) {
16077 ASSERT(buf->dtb_xamot == NULL);
16078 mutex_exit(&dtrace_lock);
16082 cached = buf->dtb_tomax;
16083 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16085 dtrace_xcall(desc.dtbd_cpu,
16086 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16088 state->dts_errors += buf->dtb_xamot_errors;
16091 * If the buffers did not actually switch, then the cross call
16092 * did not take place -- presumably because the given CPU is
16093 * not in the ready set. If this is the case, we'll return
16096 if (buf->dtb_tomax == cached) {
16097 ASSERT(buf->dtb_xamot != cached);
16098 mutex_exit(&dtrace_lock);
16102 ASSERT(cached == buf->dtb_xamot);
16105 * We have our snapshot; now copy it out.
16107 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16108 buf->dtb_xamot_offset) != 0) {
16109 mutex_exit(&dtrace_lock);
16113 desc.dtbd_size = buf->dtb_xamot_offset;
16114 desc.dtbd_drops = buf->dtb_xamot_drops;
16115 desc.dtbd_errors = buf->dtb_xamot_errors;
16116 desc.dtbd_oldest = 0;
16118 mutex_exit(&dtrace_lock);
16121 * Finally, copy out the buffer description.
16123 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16129 case DTRACEIOC_CONF: {
16130 dtrace_conf_t conf;
16132 bzero(&conf, sizeof (conf));
16133 conf.dtc_difversion = DIF_VERSION;
16134 conf.dtc_difintregs = DIF_DIR_NREGS;
16135 conf.dtc_diftupregs = DIF_DTR_NREGS;
16136 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16138 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16144 case DTRACEIOC_STATUS: {
16145 dtrace_status_t stat;
16146 dtrace_dstate_t *dstate;
16151 * See the comment in dtrace_state_deadman() for the reason
16152 * for setting dts_laststatus to INT64_MAX before setting
16153 * it to the correct value.
16155 state->dts_laststatus = INT64_MAX;
16156 dtrace_membar_producer();
16157 state->dts_laststatus = dtrace_gethrtime();
16159 bzero(&stat, sizeof (stat));
16161 mutex_enter(&dtrace_lock);
16163 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16164 mutex_exit(&dtrace_lock);
16168 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16169 stat.dtst_exiting = 1;
16171 nerrs = state->dts_errors;
16172 dstate = &state->dts_vstate.dtvs_dynvars;
16174 for (i = 0; i < NCPU; i++) {
16175 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16177 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16178 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16179 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16181 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16182 stat.dtst_filled++;
16184 nerrs += state->dts_buffer[i].dtb_errors;
16186 for (j = 0; j < state->dts_nspeculations; j++) {
16187 dtrace_speculation_t *spec;
16188 dtrace_buffer_t *buf;
16190 spec = &state->dts_speculations[j];
16191 buf = &spec->dtsp_buffer[i];
16192 stat.dtst_specdrops += buf->dtb_xamot_drops;
16196 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16197 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16198 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16199 stat.dtst_dblerrors = state->dts_dblerrors;
16201 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16202 stat.dtst_errors = nerrs;
16204 mutex_exit(&dtrace_lock);
16206 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16212 case DTRACEIOC_FORMAT: {
16213 dtrace_fmtdesc_t fmt;
16217 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16220 mutex_enter(&dtrace_lock);
16222 if (fmt.dtfd_format == 0 ||
16223 fmt.dtfd_format > state->dts_nformats) {
16224 mutex_exit(&dtrace_lock);
16229 * Format strings are allocated contiguously and they are
16230 * never freed; if a format index is less than the number
16231 * of formats, we can assert that the format map is non-NULL
16232 * and that the format for the specified index is non-NULL.
16234 ASSERT(state->dts_formats != NULL);
16235 str = state->dts_formats[fmt.dtfd_format - 1];
16236 ASSERT(str != NULL);
16238 len = strlen(str) + 1;
16240 if (len > fmt.dtfd_length) {
16241 fmt.dtfd_length = len;
16243 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16244 mutex_exit(&dtrace_lock);
16248 if (copyout(str, fmt.dtfd_string, len) != 0) {
16249 mutex_exit(&dtrace_lock);
16254 mutex_exit(&dtrace_lock);
16267 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16269 dtrace_state_t *state;
16276 return (DDI_SUCCESS);
16279 return (DDI_FAILURE);
16282 mutex_enter(&cpu_lock);
16283 mutex_enter(&dtrace_provider_lock);
16284 mutex_enter(&dtrace_lock);
16286 ASSERT(dtrace_opens == 0);
16288 if (dtrace_helpers > 0) {
16289 mutex_exit(&dtrace_provider_lock);
16290 mutex_exit(&dtrace_lock);
16291 mutex_exit(&cpu_lock);
16292 return (DDI_FAILURE);
16295 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16296 mutex_exit(&dtrace_provider_lock);
16297 mutex_exit(&dtrace_lock);
16298 mutex_exit(&cpu_lock);
16299 return (DDI_FAILURE);
16302 dtrace_provider = NULL;
16304 if ((state = dtrace_anon_grab()) != NULL) {
16306 * If there were ECBs on this state, the provider should
16307 * have not been allowed to detach; assert that there is
16310 ASSERT(state->dts_necbs == 0);
16311 dtrace_state_destroy(state);
16314 * If we're being detached with anonymous state, we need to
16315 * indicate to the kernel debugger that DTrace is now inactive.
16317 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16320 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16321 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16322 dtrace_cpu_init = NULL;
16323 dtrace_helpers_cleanup = NULL;
16324 dtrace_helpers_fork = NULL;
16325 dtrace_cpustart_init = NULL;
16326 dtrace_cpustart_fini = NULL;
16327 dtrace_debugger_init = NULL;
16328 dtrace_debugger_fini = NULL;
16329 dtrace_modload = NULL;
16330 dtrace_modunload = NULL;
16332 mutex_exit(&cpu_lock);
16334 if (dtrace_helptrace_enabled) {
16335 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16336 dtrace_helptrace_buffer = NULL;
16339 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16340 dtrace_probes = NULL;
16341 dtrace_nprobes = 0;
16343 dtrace_hash_destroy(dtrace_bymod);
16344 dtrace_hash_destroy(dtrace_byfunc);
16345 dtrace_hash_destroy(dtrace_byname);
16346 dtrace_bymod = NULL;
16347 dtrace_byfunc = NULL;
16348 dtrace_byname = NULL;
16350 kmem_cache_destroy(dtrace_state_cache);
16351 vmem_destroy(dtrace_minor);
16352 vmem_destroy(dtrace_arena);
16354 if (dtrace_toxrange != NULL) {
16355 kmem_free(dtrace_toxrange,
16356 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16357 dtrace_toxrange = NULL;
16358 dtrace_toxranges = 0;
16359 dtrace_toxranges_max = 0;
16362 ddi_remove_minor_node(dtrace_devi, NULL);
16363 dtrace_devi = NULL;
16365 ddi_soft_state_fini(&dtrace_softstate);
16367 ASSERT(dtrace_vtime_references == 0);
16368 ASSERT(dtrace_opens == 0);
16369 ASSERT(dtrace_retained == NULL);
16371 mutex_exit(&dtrace_lock);
16372 mutex_exit(&dtrace_provider_lock);
16375 * We don't destroy the task queue until after we have dropped our
16376 * locks (taskq_destroy() may block on running tasks). To prevent
16377 * attempting to do work after we have effectively detached but before
16378 * the task queue has been destroyed, all tasks dispatched via the
16379 * task queue must check that DTrace is still attached before
16380 * performing any operation.
16382 taskq_destroy(dtrace_taskq);
16383 dtrace_taskq = NULL;
16385 return (DDI_SUCCESS);
16392 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16397 case DDI_INFO_DEVT2DEVINFO:
16398 *result = (void *)dtrace_devi;
16399 error = DDI_SUCCESS;
16401 case DDI_INFO_DEVT2INSTANCE:
16402 *result = (void *)0;
16403 error = DDI_SUCCESS;
16406 error = DDI_FAILURE;
16413 static struct cb_ops dtrace_cb_ops = {
16414 dtrace_open, /* open */
16415 dtrace_close, /* close */
16416 nulldev, /* strategy */
16417 nulldev, /* print */
16421 dtrace_ioctl, /* ioctl */
16422 nodev, /* devmap */
16424 nodev, /* segmap */
16425 nochpoll, /* poll */
16426 ddi_prop_op, /* cb_prop_op */
16428 D_NEW | D_MP /* Driver compatibility flag */
16431 static struct dev_ops dtrace_ops = {
16432 DEVO_REV, /* devo_rev */
16434 dtrace_info, /* get_dev_info */
16435 nulldev, /* identify */
16436 nulldev, /* probe */
16437 dtrace_attach, /* attach */
16438 dtrace_detach, /* detach */
16440 &dtrace_cb_ops, /* driver operations */
16441 NULL, /* bus operations */
16442 nodev /* dev power */
16445 static struct modldrv modldrv = {
16446 &mod_driverops, /* module type (this is a pseudo driver) */
16447 "Dynamic Tracing", /* name of module */
16448 &dtrace_ops, /* driver ops */
16451 static struct modlinkage modlinkage = {
16460 return (mod_install(&modlinkage));
16464 _info(struct modinfo *modinfop)
16466 return (mod_info(&modlinkage, modinfop));
16472 return (mod_remove(&modlinkage));
16476 static d_ioctl_t dtrace_ioctl;
16477 static void dtrace_load(void *);
16478 static int dtrace_unload(void);
16479 #if __FreeBSD_version < 800039
16480 static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16481 static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */
16482 static eventhandler_tag eh_tag; /* Event handler tag. */
16484 static struct cdev *dtrace_dev;
16487 void dtrace_invop_init(void);
16488 void dtrace_invop_uninit(void);
16490 static struct cdevsw dtrace_cdevsw = {
16491 .d_version = D_VERSION,
16492 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16493 .d_close = dtrace_close,
16494 .d_ioctl = dtrace_ioctl,
16495 .d_open = dtrace_open,
16496 .d_name = "dtrace",
16499 #include <dtrace_anon.c>
16500 #if __FreeBSD_version < 800039
16501 #include <dtrace_clone.c>
16503 #include <dtrace_ioctl.c>
16504 #include <dtrace_load.c>
16505 #include <dtrace_modevent.c>
16506 #include <dtrace_sysctl.c>
16507 #include <dtrace_unload.c>
16508 #include <dtrace_vtime.c>
16509 #include <dtrace_hacks.c>
16510 #include <dtrace_isa.c>
16512 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16513 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16514 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16516 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16517 MODULE_VERSION(dtrace, 1);
16518 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
16519 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);