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 * Copyright (c) 2012 by Delphix. All rights reserved
27 * Use is subject to license terms.
30 #pragma ident "%Z%%M% %I% %E% SMI"
33 * DTrace - Dynamic Tracing for Solaris
35 * This is the implementation of the Solaris Dynamic Tracing framework
36 * (DTrace). The user-visible interface to DTrace is described at length in
37 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
38 * library, the in-kernel DTrace framework, and the DTrace providers are
39 * described in the block comments in the <sys/dtrace.h> header file. The
40 * internal architecture of DTrace is described in the block comments in the
41 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
42 * implementation very much assume mastery of all of these sources; if one has
43 * an unanswered question about the implementation, one should consult them
46 * The functions here are ordered roughly as follows:
48 * - Probe context functions
49 * - Probe hashing functions
50 * - Non-probe context utility functions
51 * - Matching functions
52 * - Provider-to-Framework API functions
53 * - Probe management functions
54 * - DIF object functions
56 * - Predicate functions
59 * - Enabling functions
61 * - Anonymous enabling functions
62 * - Consumer state functions
65 * - Driver cookbook functions
67 * Each group of functions begins with a block comment labelled the "DTrace
68 * [Group] Functions", allowing one to find each block by searching forward
69 * on capital-f functions.
71 #include <sys/errno.h>
76 #include <sys/modctl.h>
78 #include <sys/systm.h>
81 #include <sys/sunddi.h>
83 #include <sys/cpuvar.h>
86 #include <sys/strsubr.h>
88 #include <sys/sysmacros.h>
89 #include <sys/dtrace_impl.h>
90 #include <sys/atomic.h>
91 #include <sys/cmn_err.h>
93 #include <sys/mutex_impl.h>
94 #include <sys/rwlock_impl.h>
96 #include <sys/ctf_api.h>
98 #include <sys/panic.h>
99 #include <sys/priv_impl.h>
101 #include <sys/policy.h>
103 #include <sys/cred_impl.h>
104 #include <sys/procfs_isa.h>
106 #include <sys/taskq.h>
108 #include <sys/mkdev.h>
111 #include <sys/zone.h>
112 #include <sys/socket.h>
113 #include <netinet/in.h>
115 /* FreeBSD includes: */
117 #include <sys/callout.h>
118 #include <sys/ctype.h>
119 #include <sys/limits.h>
121 #include <sys/kernel.h>
122 #include <sys/malloc.h>
123 #include <sys/sysctl.h>
124 #include <sys/lock.h>
125 #include <sys/mutex.h>
126 #include <sys/rwlock.h>
128 #include <sys/dtrace_bsd.h>
129 #include <netinet/in.h>
130 #include "dtrace_cddl.h"
131 #include "dtrace_debug.c"
135 * DTrace Tunable Variables
137 * The following variables may be tuned by adding a line to /etc/system that
138 * includes both the name of the DTrace module ("dtrace") and the name of the
139 * variable. For example:
141 * set dtrace:dtrace_destructive_disallow = 1
143 * In general, the only variables that one should be tuning this way are those
144 * that affect system-wide DTrace behavior, and for which the default behavior
145 * is undesirable. Most of these variables are tunable on a per-consumer
146 * basis using DTrace options, and need not be tuned on a system-wide basis.
147 * When tuning these variables, avoid pathological values; while some attempt
148 * is made to verify the integrity of these variables, they are not considered
149 * part of the supported interface to DTrace, and they are therefore not
150 * checked comprehensively. Further, these variables should not be tuned
151 * dynamically via "mdb -kw" or other means; they should only be tuned via
154 int dtrace_destructive_disallow = 0;
155 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
156 size_t dtrace_difo_maxsize = (256 * 1024);
157 dtrace_optval_t dtrace_dof_maxsize = (256 * 1024);
158 size_t dtrace_global_maxsize = (16 * 1024);
159 size_t dtrace_actions_max = (16 * 1024);
160 size_t dtrace_retain_max = 1024;
161 dtrace_optval_t dtrace_helper_actions_max = 128;
162 dtrace_optval_t dtrace_helper_providers_max = 32;
163 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
164 size_t dtrace_strsize_default = 256;
165 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
166 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
167 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
168 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
169 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
170 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
171 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
172 dtrace_optval_t dtrace_nspec_default = 1;
173 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
174 dtrace_optval_t dtrace_stackframes_default = 20;
175 dtrace_optval_t dtrace_ustackframes_default = 20;
176 dtrace_optval_t dtrace_jstackframes_default = 50;
177 dtrace_optval_t dtrace_jstackstrsize_default = 512;
178 int dtrace_msgdsize_max = 128;
179 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
180 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
181 int dtrace_devdepth_max = 32;
182 int dtrace_err_verbose;
183 hrtime_t dtrace_deadman_interval = NANOSEC;
184 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
185 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
186 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
189 * DTrace External Variables
191 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
192 * available to DTrace consumers via the backtick (`) syntax. One of these,
193 * dtrace_zero, is made deliberately so: it is provided as a source of
194 * well-known, zero-filled memory. While this variable is not documented,
195 * it is used by some translators as an implementation detail.
197 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
200 * DTrace Internal Variables
203 static dev_info_t *dtrace_devi; /* device info */
206 static vmem_t *dtrace_arena; /* probe ID arena */
207 static vmem_t *dtrace_minor; /* minor number arena */
209 static taskq_t *dtrace_taskq; /* task queue */
210 static struct unrhdr *dtrace_arena; /* Probe ID number. */
212 static dtrace_probe_t **dtrace_probes; /* array of all probes */
213 static int dtrace_nprobes; /* number of probes */
214 static dtrace_provider_t *dtrace_provider; /* provider list */
215 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
216 static int dtrace_opens; /* number of opens */
217 static int dtrace_helpers; /* number of helpers */
219 static void *dtrace_softstate; /* softstate pointer */
221 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
222 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
223 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
224 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
225 static int dtrace_toxranges; /* number of toxic ranges */
226 static int dtrace_toxranges_max; /* size of toxic range array */
227 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
228 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
229 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
230 static kthread_t *dtrace_panicked; /* panicking thread */
231 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
232 static dtrace_genid_t dtrace_probegen; /* current probe generation */
233 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
234 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
235 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
237 static struct mtx dtrace_unr_mtx;
238 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
239 int dtrace_in_probe; /* non-zero if executing a probe */
240 #if defined(__i386__) || defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
241 uintptr_t dtrace_in_probe_addr; /* Address of invop when already in probe */
247 * DTrace is protected by three (relatively coarse-grained) locks:
249 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
250 * including enabling state, probes, ECBs, consumer state, helper state,
251 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
252 * probe context is lock-free -- synchronization is handled via the
253 * dtrace_sync() cross call mechanism.
255 * (2) dtrace_provider_lock is required when manipulating provider state, or
256 * when provider state must be held constant.
258 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
259 * when meta provider state must be held constant.
261 * The lock ordering between these three locks is dtrace_meta_lock before
262 * dtrace_provider_lock before dtrace_lock. (In particular, there are
263 * several places where dtrace_provider_lock is held by the framework as it
264 * calls into the providers -- which then call back into the framework,
265 * grabbing dtrace_lock.)
267 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
268 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
269 * role as a coarse-grained lock; it is acquired before both of these locks.
270 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
271 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
272 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
273 * acquired _between_ dtrace_provider_lock and dtrace_lock.
275 static kmutex_t dtrace_lock; /* probe state lock */
276 static kmutex_t dtrace_provider_lock; /* provider state lock */
277 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
280 /* XXX FreeBSD hacks. */
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 void dtrace_enabling_reap(void);
554 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 *);
558 static void dtrace_buffer_drop(dtrace_buffer_t *);
559 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
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 *);
565 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
566 uint16_t dtrace_load16(uintptr_t);
567 uint32_t dtrace_load32(uintptr_t);
568 uint64_t dtrace_load64(uintptr_t);
569 uint8_t dtrace_load8(uintptr_t);
570 void dtrace_dynvar_clean(dtrace_dstate_t *);
571 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
572 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
573 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
576 * DTrace Probe Context Functions
578 * These functions are called from probe context. Because probe context is
579 * any context in which C may be called, arbitrarily locks may be held,
580 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
581 * As a result, functions called from probe context may only call other DTrace
582 * support functions -- they may not interact at all with the system at large.
583 * (Note that the ASSERT macro is made probe-context safe by redefining it in
584 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
585 * loads are to be performed from probe context, they _must_ be in terms of
586 * the safe dtrace_load*() variants.
588 * Some functions in this block are not actually called from probe context;
589 * for these functions, there will be a comment above the function reading
590 * "Note: not called from probe context."
593 dtrace_panic(const char *format, ...)
597 va_start(alist, format);
598 dtrace_vpanic(format, alist);
603 dtrace_assfail(const char *a, const char *f, int l)
605 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
608 * We just need something here that even the most clever compiler
609 * cannot optimize away.
611 return (a[(uintptr_t)f]);
615 * Atomically increment a specified error counter from probe context.
618 dtrace_error(uint32_t *counter)
621 * Most counters stored to in probe context are per-CPU counters.
622 * However, there are some error conditions that are sufficiently
623 * arcane that they don't merit per-CPU storage. If these counters
624 * are incremented concurrently on different CPUs, scalability will be
625 * adversely affected -- but we don't expect them to be white-hot in a
626 * correctly constructed enabling...
633 if ((nval = oval + 1) == 0) {
635 * If the counter would wrap, set it to 1 -- assuring
636 * that the counter is never zero when we have seen
637 * errors. (The counter must be 32-bits because we
638 * aren't guaranteed a 64-bit compare&swap operation.)
639 * To save this code both the infamy of being fingered
640 * by a priggish news story and the indignity of being
641 * the target of a neo-puritan witch trial, we're
642 * carefully avoiding any colorful description of the
643 * likelihood of this condition -- but suffice it to
644 * say that it is only slightly more likely than the
645 * overflow of predicate cache IDs, as discussed in
646 * dtrace_predicate_create().
650 } while (dtrace_cas32(counter, oval, nval) != oval);
654 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
655 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
663 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
665 if (dest < mstate->dtms_scratch_base)
668 if (dest + size < dest)
671 if (dest + size > mstate->dtms_scratch_ptr)
678 dtrace_canstore_statvar(uint64_t addr, size_t sz,
679 dtrace_statvar_t **svars, int nsvars)
683 for (i = 0; i < nsvars; i++) {
684 dtrace_statvar_t *svar = svars[i];
686 if (svar == NULL || svar->dtsv_size == 0)
689 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
697 * Check to see if the address is within a memory region to which a store may
698 * be issued. This includes the DTrace scratch areas, and any DTrace variable
699 * region. The caller of dtrace_canstore() is responsible for performing any
700 * alignment checks that are needed before stores are actually executed.
703 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
704 dtrace_vstate_t *vstate)
707 * First, check to see if the address is in scratch space...
709 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
710 mstate->dtms_scratch_size))
714 * Now check to see if it's a dynamic variable. This check will pick
715 * up both thread-local variables and any global dynamically-allocated
718 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
719 vstate->dtvs_dynvars.dtds_size)) {
720 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
721 uintptr_t base = (uintptr_t)dstate->dtds_base +
722 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
726 * Before we assume that we can store here, we need to make
727 * sure that it isn't in our metadata -- storing to our
728 * dynamic variable metadata would corrupt our state. For
729 * the range to not include any dynamic variable metadata,
732 * (1) Start above the hash table that is at the base of
733 * the dynamic variable space
735 * (2) Have a starting chunk offset that is beyond the
736 * dtrace_dynvar_t that is at the base of every chunk
738 * (3) Not span a chunk boundary
744 chunkoffs = (addr - base) % dstate->dtds_chunksize;
746 if (chunkoffs < sizeof (dtrace_dynvar_t))
749 if (chunkoffs + sz > dstate->dtds_chunksize)
756 * Finally, check the static local and global variables. These checks
757 * take the longest, so we perform them last.
759 if (dtrace_canstore_statvar(addr, sz,
760 vstate->dtvs_locals, vstate->dtvs_nlocals))
763 if (dtrace_canstore_statvar(addr, sz,
764 vstate->dtvs_globals, vstate->dtvs_nglobals))
772 * Convenience routine to check to see if the address is within a memory
773 * region in which a load may be issued given the user's privilege level;
774 * if not, it sets the appropriate error flags and loads 'addr' into the
775 * illegal value slot.
777 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
778 * appropriate memory access protection.
781 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
782 dtrace_vstate_t *vstate)
784 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
787 * If we hold the privilege to read from kernel memory, then
788 * everything is readable.
790 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
794 * You can obviously read that which you can store.
796 if (dtrace_canstore(addr, sz, mstate, vstate))
800 * We're allowed to read from our own string table.
802 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
803 mstate->dtms_difo->dtdo_strlen))
806 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
812 * Convenience routine to check to see if a given string is within a memory
813 * region in which a load may be issued given the user's privilege level;
814 * this exists so that we don't need to issue unnecessary dtrace_strlen()
815 * calls in the event that the user has all privileges.
818 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
819 dtrace_vstate_t *vstate)
824 * If we hold the privilege to read from kernel memory, then
825 * everything is readable.
827 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
830 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
831 if (dtrace_canload(addr, strsz, mstate, vstate))
838 * Convenience routine to check to see if a given variable is within a memory
839 * region in which a load may be issued given the user's privilege level.
842 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
843 dtrace_vstate_t *vstate)
846 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
849 * If we hold the privilege to read from kernel memory, then
850 * everything is readable.
852 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
855 if (type->dtdt_kind == DIF_TYPE_STRING)
856 sz = dtrace_strlen(src,
857 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
859 sz = type->dtdt_size;
861 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
865 * Compare two strings using safe loads.
868 dtrace_strncmp(char *s1, char *s2, size_t limit)
871 volatile uint16_t *flags;
873 if (s1 == s2 || limit == 0)
876 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
882 c1 = dtrace_load8((uintptr_t)s1++);
888 c2 = dtrace_load8((uintptr_t)s2++);
893 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
899 * Compute strlen(s) for a string using safe memory accesses. The additional
900 * len parameter is used to specify a maximum length to ensure completion.
903 dtrace_strlen(const char *s, size_t lim)
907 for (len = 0; len != lim; len++) {
908 if (dtrace_load8((uintptr_t)s++) == '\0')
916 * Check if an address falls within a toxic region.
919 dtrace_istoxic(uintptr_t kaddr, size_t size)
921 uintptr_t taddr, tsize;
924 for (i = 0; i < dtrace_toxranges; i++) {
925 taddr = dtrace_toxrange[i].dtt_base;
926 tsize = dtrace_toxrange[i].dtt_limit - taddr;
928 if (kaddr - taddr < tsize) {
929 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
930 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
934 if (taddr - kaddr < size) {
935 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
936 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
945 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
946 * memory specified by the DIF program. The dst is assumed to be safe memory
947 * that we can store to directly because it is managed by DTrace. As with
948 * standard bcopy, overlapping copies are handled properly.
951 dtrace_bcopy(const void *src, void *dst, size_t len)
955 const uint8_t *s2 = src;
959 *s1++ = dtrace_load8((uintptr_t)s2++);
960 } while (--len != 0);
966 *--s1 = dtrace_load8((uintptr_t)--s2);
967 } while (--len != 0);
973 * Copy src to dst using safe memory accesses, up to either the specified
974 * length, or the point that a nul byte is encountered. The src is assumed to
975 * be unsafe memory specified by the DIF program. The dst is assumed to be
976 * safe memory that we can store to directly because it is managed by DTrace.
977 * Unlike dtrace_bcopy(), overlapping regions are not handled.
980 dtrace_strcpy(const void *src, void *dst, size_t len)
983 uint8_t *s1 = dst, c;
984 const uint8_t *s2 = src;
987 *s1++ = c = dtrace_load8((uintptr_t)s2++);
988 } while (--len != 0 && c != '\0');
993 * Copy src to dst, deriving the size and type from the specified (BYREF)
994 * variable type. The src is assumed to be unsafe memory specified by the DIF
995 * program. The dst is assumed to be DTrace variable memory that is of the
996 * specified type; we assume that we can store to directly.
999 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1001 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1003 if (type->dtdt_kind == DIF_TYPE_STRING) {
1004 dtrace_strcpy(src, dst, type->dtdt_size);
1006 dtrace_bcopy(src, dst, type->dtdt_size);
1011 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1012 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1013 * safe memory that we can access directly because it is managed by DTrace.
1016 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1018 volatile uint16_t *flags;
1020 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1025 if (s1 == NULL || s2 == NULL)
1028 if (s1 != s2 && len != 0) {
1029 const uint8_t *ps1 = s1;
1030 const uint8_t *ps2 = s2;
1033 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1035 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1041 * Zero the specified region using a simple byte-by-byte loop. Note that this
1042 * is for safe DTrace-managed memory only.
1045 dtrace_bzero(void *dst, size_t len)
1049 for (cp = dst; len != 0; len--)
1054 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1058 result[0] = addend1[0] + addend2[0];
1059 result[1] = addend1[1] + addend2[1] +
1060 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1067 * Shift the 128-bit value in a by b. If b is positive, shift left.
1068 * If b is negative, shift right.
1071 dtrace_shift_128(uint64_t *a, int b)
1081 a[0] = a[1] >> (b - 64);
1085 mask = 1LL << (64 - b);
1087 a[0] |= ((a[1] & mask) << (64 - b));
1092 a[1] = a[0] << (b - 64);
1096 mask = a[0] >> (64 - b);
1104 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1105 * use native multiplication on those, and then re-combine into the
1106 * resulting 128-bit value.
1108 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1115 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1117 uint64_t hi1, hi2, lo1, lo2;
1120 hi1 = factor1 >> 32;
1121 hi2 = factor2 >> 32;
1123 lo1 = factor1 & DT_MASK_LO;
1124 lo2 = factor2 & DT_MASK_LO;
1126 product[0] = lo1 * lo2;
1127 product[1] = hi1 * hi2;
1131 dtrace_shift_128(tmp, 32);
1132 dtrace_add_128(product, tmp, product);
1136 dtrace_shift_128(tmp, 32);
1137 dtrace_add_128(product, tmp, product);
1141 * This privilege check should be used by actions and subroutines to
1142 * verify that the user credentials of the process that enabled the
1143 * invoking ECB match the target credentials
1146 dtrace_priv_proc_common_user(dtrace_state_t *state)
1148 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1151 * We should always have a non-NULL state cred here, since if cred
1152 * is null (anonymous tracing), we fast-path bypass this routine.
1154 ASSERT(s_cr != NULL);
1156 if ((cr = CRED()) != NULL &&
1157 s_cr->cr_uid == cr->cr_uid &&
1158 s_cr->cr_uid == cr->cr_ruid &&
1159 s_cr->cr_uid == cr->cr_suid &&
1160 s_cr->cr_gid == cr->cr_gid &&
1161 s_cr->cr_gid == cr->cr_rgid &&
1162 s_cr->cr_gid == cr->cr_sgid)
1169 * This privilege check should be used by actions and subroutines to
1170 * verify that the zone of the process that enabled the invoking ECB
1171 * matches the target credentials
1174 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1177 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1180 * We should always have a non-NULL state cred here, since if cred
1181 * is null (anonymous tracing), we fast-path bypass this routine.
1183 ASSERT(s_cr != NULL);
1185 if ((cr = CRED()) != NULL &&
1186 s_cr->cr_zone == cr->cr_zone)
1196 * This privilege check should be used by actions and subroutines to
1197 * verify that the process has not setuid or changed credentials.
1200 dtrace_priv_proc_common_nocd(void)
1204 if ((proc = ttoproc(curthread)) != NULL &&
1205 !(proc->p_flag & SNOCD))
1212 dtrace_priv_proc_destructive(dtrace_state_t *state)
1214 int action = state->dts_cred.dcr_action;
1216 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1217 dtrace_priv_proc_common_zone(state) == 0)
1220 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1221 dtrace_priv_proc_common_user(state) == 0)
1224 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1225 dtrace_priv_proc_common_nocd() == 0)
1231 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1237 dtrace_priv_proc_control(dtrace_state_t *state)
1239 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1242 if (dtrace_priv_proc_common_zone(state) &&
1243 dtrace_priv_proc_common_user(state) &&
1244 dtrace_priv_proc_common_nocd())
1247 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1253 dtrace_priv_proc(dtrace_state_t *state)
1255 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1258 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1264 dtrace_priv_kernel(dtrace_state_t *state)
1266 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1269 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1275 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1277 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1280 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1286 * Note: not called from probe context. This function is called
1287 * asynchronously (and at a regular interval) from outside of probe context to
1288 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1289 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1292 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1294 dtrace_dynvar_t *dirty;
1295 dtrace_dstate_percpu_t *dcpu;
1298 for (i = 0; i < NCPU; i++) {
1299 dcpu = &dstate->dtds_percpu[i];
1301 ASSERT(dcpu->dtdsc_rinsing == NULL);
1304 * If the dirty list is NULL, there is no dirty work to do.
1306 if (dcpu->dtdsc_dirty == NULL)
1310 * If the clean list is non-NULL, then we're not going to do
1311 * any work for this CPU -- it means that there has not been
1312 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1313 * since the last time we cleaned house.
1315 if (dcpu->dtdsc_clean != NULL)
1321 * Atomically move the dirty list aside.
1324 dirty = dcpu->dtdsc_dirty;
1327 * Before we zap the dirty list, set the rinsing list.
1328 * (This allows for a potential assertion in
1329 * dtrace_dynvar(): if a free dynamic variable appears
1330 * on a hash chain, either the dirty list or the
1331 * rinsing list for some CPU must be non-NULL.)
1333 dcpu->dtdsc_rinsing = dirty;
1334 dtrace_membar_producer();
1335 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1336 dirty, NULL) != dirty);
1341 * We have no work to do; we can simply return.
1348 for (i = 0; i < NCPU; i++) {
1349 dcpu = &dstate->dtds_percpu[i];
1351 if (dcpu->dtdsc_rinsing == NULL)
1355 * We are now guaranteed that no hash chain contains a pointer
1356 * into this dirty list; we can make it clean.
1358 ASSERT(dcpu->dtdsc_clean == NULL);
1359 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1360 dcpu->dtdsc_rinsing = NULL;
1364 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1365 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1366 * This prevents a race whereby a CPU incorrectly decides that
1367 * the state should be something other than DTRACE_DSTATE_CLEAN
1368 * after dtrace_dynvar_clean() has completed.
1372 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1376 * Depending on the value of the op parameter, this function looks-up,
1377 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1378 * allocation is requested, this function will return a pointer to a
1379 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1380 * variable can be allocated. If NULL is returned, the appropriate counter
1381 * will be incremented.
1384 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1385 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1386 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1388 uint64_t hashval = DTRACE_DYNHASH_VALID;
1389 dtrace_dynhash_t *hash = dstate->dtds_hash;
1390 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1391 processorid_t me = curcpu, cpu = me;
1392 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1393 size_t bucket, ksize;
1394 size_t chunksize = dstate->dtds_chunksize;
1395 uintptr_t kdata, lock, nstate;
1401 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1402 * algorithm. For the by-value portions, we perform the algorithm in
1403 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1404 * bit, and seems to have only a minute effect on distribution. For
1405 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1406 * over each referenced byte. It's painful to do this, but it's much
1407 * better than pathological hash distribution. The efficacy of the
1408 * hashing algorithm (and a comparison with other algorithms) may be
1409 * found by running the ::dtrace_dynstat MDB dcmd.
1411 for (i = 0; i < nkeys; i++) {
1412 if (key[i].dttk_size == 0) {
1413 uint64_t val = key[i].dttk_value;
1415 hashval += (val >> 48) & 0xffff;
1416 hashval += (hashval << 10);
1417 hashval ^= (hashval >> 6);
1419 hashval += (val >> 32) & 0xffff;
1420 hashval += (hashval << 10);
1421 hashval ^= (hashval >> 6);
1423 hashval += (val >> 16) & 0xffff;
1424 hashval += (hashval << 10);
1425 hashval ^= (hashval >> 6);
1427 hashval += val & 0xffff;
1428 hashval += (hashval << 10);
1429 hashval ^= (hashval >> 6);
1432 * This is incredibly painful, but it beats the hell
1433 * out of the alternative.
1435 uint64_t j, size = key[i].dttk_size;
1436 uintptr_t base = (uintptr_t)key[i].dttk_value;
1438 if (!dtrace_canload(base, size, mstate, vstate))
1441 for (j = 0; j < size; j++) {
1442 hashval += dtrace_load8(base + j);
1443 hashval += (hashval << 10);
1444 hashval ^= (hashval >> 6);
1449 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1452 hashval += (hashval << 3);
1453 hashval ^= (hashval >> 11);
1454 hashval += (hashval << 15);
1457 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1458 * comes out to be one of our two sentinel hash values. If this
1459 * actually happens, we set the hashval to be a value known to be a
1460 * non-sentinel value.
1462 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1463 hashval = DTRACE_DYNHASH_VALID;
1466 * Yes, it's painful to do a divide here. If the cycle count becomes
1467 * important here, tricks can be pulled to reduce it. (However, it's
1468 * critical that hash collisions be kept to an absolute minimum;
1469 * they're much more painful than a divide.) It's better to have a
1470 * solution that generates few collisions and still keeps things
1471 * relatively simple.
1473 bucket = hashval % dstate->dtds_hashsize;
1475 if (op == DTRACE_DYNVAR_DEALLOC) {
1476 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1479 while ((lock = *lockp) & 1)
1482 if (dtrace_casptr((volatile void *)lockp,
1483 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1487 dtrace_membar_producer();
1492 lock = hash[bucket].dtdh_lock;
1494 dtrace_membar_consumer();
1496 start = hash[bucket].dtdh_chain;
1497 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1498 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1499 op != DTRACE_DYNVAR_DEALLOC));
1501 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1502 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1503 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1505 if (dvar->dtdv_hashval != hashval) {
1506 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1508 * We've reached the sink, and therefore the
1509 * end of the hash chain; we can kick out of
1510 * the loop knowing that we have seen a valid
1511 * snapshot of state.
1513 ASSERT(dvar->dtdv_next == NULL);
1514 ASSERT(dvar == &dtrace_dynhash_sink);
1518 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1520 * We've gone off the rails: somewhere along
1521 * the line, one of the members of this hash
1522 * chain was deleted. Note that we could also
1523 * detect this by simply letting this loop run
1524 * to completion, as we would eventually hit
1525 * the end of the dirty list. However, we
1526 * want to avoid running the length of the
1527 * dirty list unnecessarily (it might be quite
1528 * long), so we catch this as early as
1529 * possible by detecting the hash marker. In
1530 * this case, we simply set dvar to NULL and
1531 * break; the conditional after the loop will
1532 * send us back to top.
1541 if (dtuple->dtt_nkeys != nkeys)
1544 for (i = 0; i < nkeys; i++, dkey++) {
1545 if (dkey->dttk_size != key[i].dttk_size)
1546 goto next; /* size or type mismatch */
1548 if (dkey->dttk_size != 0) {
1550 (void *)(uintptr_t)key[i].dttk_value,
1551 (void *)(uintptr_t)dkey->dttk_value,
1555 if (dkey->dttk_value != key[i].dttk_value)
1560 if (op != DTRACE_DYNVAR_DEALLOC)
1563 ASSERT(dvar->dtdv_next == NULL ||
1564 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1567 ASSERT(hash[bucket].dtdh_chain != dvar);
1568 ASSERT(start != dvar);
1569 ASSERT(prev->dtdv_next == dvar);
1570 prev->dtdv_next = dvar->dtdv_next;
1572 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1573 start, dvar->dtdv_next) != start) {
1575 * We have failed to atomically swing the
1576 * hash table head pointer, presumably because
1577 * of a conflicting allocation on another CPU.
1578 * We need to reread the hash chain and try
1585 dtrace_membar_producer();
1588 * Now set the hash value to indicate that it's free.
1590 ASSERT(hash[bucket].dtdh_chain != dvar);
1591 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1593 dtrace_membar_producer();
1596 * Set the next pointer to point at the dirty list, and
1597 * atomically swing the dirty pointer to the newly freed dvar.
1600 next = dcpu->dtdsc_dirty;
1601 dvar->dtdv_next = next;
1602 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1605 * Finally, unlock this hash bucket.
1607 ASSERT(hash[bucket].dtdh_lock == lock);
1609 hash[bucket].dtdh_lock++;
1619 * If dvar is NULL, it is because we went off the rails:
1620 * one of the elements that we traversed in the hash chain
1621 * was deleted while we were traversing it. In this case,
1622 * we assert that we aren't doing a dealloc (deallocs lock
1623 * the hash bucket to prevent themselves from racing with
1624 * one another), and retry the hash chain traversal.
1626 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1630 if (op != DTRACE_DYNVAR_ALLOC) {
1632 * If we are not to allocate a new variable, we want to
1633 * return NULL now. Before we return, check that the value
1634 * of the lock word hasn't changed. If it has, we may have
1635 * seen an inconsistent snapshot.
1637 if (op == DTRACE_DYNVAR_NOALLOC) {
1638 if (hash[bucket].dtdh_lock != lock)
1641 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1642 ASSERT(hash[bucket].dtdh_lock == lock);
1644 hash[bucket].dtdh_lock++;
1651 * We need to allocate a new dynamic variable. The size we need is the
1652 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1653 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1654 * the size of any referred-to data (dsize). We then round the final
1655 * size up to the chunksize for allocation.
1657 for (ksize = 0, i = 0; i < nkeys; i++)
1658 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1661 * This should be pretty much impossible, but could happen if, say,
1662 * strange DIF specified the tuple. Ideally, this should be an
1663 * assertion and not an error condition -- but that requires that the
1664 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1665 * bullet-proof. (That is, it must not be able to be fooled by
1666 * malicious DIF.) Given the lack of backwards branches in DIF,
1667 * solving this would presumably not amount to solving the Halting
1668 * Problem -- but it still seems awfully hard.
1670 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1671 ksize + dsize > chunksize) {
1672 dcpu->dtdsc_drops++;
1676 nstate = DTRACE_DSTATE_EMPTY;
1680 free = dcpu->dtdsc_free;
1683 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1686 if (clean == NULL) {
1688 * We're out of dynamic variable space on
1689 * this CPU. Unless we have tried all CPUs,
1690 * we'll try to allocate from a different
1693 switch (dstate->dtds_state) {
1694 case DTRACE_DSTATE_CLEAN: {
1695 void *sp = &dstate->dtds_state;
1700 if (dcpu->dtdsc_dirty != NULL &&
1701 nstate == DTRACE_DSTATE_EMPTY)
1702 nstate = DTRACE_DSTATE_DIRTY;
1704 if (dcpu->dtdsc_rinsing != NULL)
1705 nstate = DTRACE_DSTATE_RINSING;
1707 dcpu = &dstate->dtds_percpu[cpu];
1712 (void) dtrace_cas32(sp,
1713 DTRACE_DSTATE_CLEAN, nstate);
1716 * To increment the correct bean
1717 * counter, take another lap.
1722 case DTRACE_DSTATE_DIRTY:
1723 dcpu->dtdsc_dirty_drops++;
1726 case DTRACE_DSTATE_RINSING:
1727 dcpu->dtdsc_rinsing_drops++;
1730 case DTRACE_DSTATE_EMPTY:
1731 dcpu->dtdsc_drops++;
1735 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1740 * The clean list appears to be non-empty. We want to
1741 * move the clean list to the free list; we start by
1742 * moving the clean pointer aside.
1744 if (dtrace_casptr(&dcpu->dtdsc_clean,
1745 clean, NULL) != clean) {
1747 * We are in one of two situations:
1749 * (a) The clean list was switched to the
1750 * free list by another CPU.
1752 * (b) The clean list was added to by the
1755 * In either of these situations, we can
1756 * just reattempt the free list allocation.
1761 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1764 * Now we'll move the clean list to the free list.
1765 * It's impossible for this to fail: the only way
1766 * the free list can be updated is through this
1767 * code path, and only one CPU can own the clean list.
1768 * Thus, it would only be possible for this to fail if
1769 * this code were racing with dtrace_dynvar_clean().
1770 * (That is, if dtrace_dynvar_clean() updated the clean
1771 * list, and we ended up racing to update the free
1772 * list.) This race is prevented by the dtrace_sync()
1773 * in dtrace_dynvar_clean() -- which flushes the
1774 * owners of the clean lists out before resetting
1777 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1778 ASSERT(rval == NULL);
1783 new_free = dvar->dtdv_next;
1784 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1787 * We have now allocated a new chunk. We copy the tuple keys into the
1788 * tuple array and copy any referenced key data into the data space
1789 * following the tuple array. As we do this, we relocate dttk_value
1790 * in the final tuple to point to the key data address in the chunk.
1792 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1793 dvar->dtdv_data = (void *)(kdata + ksize);
1794 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1796 for (i = 0; i < nkeys; i++) {
1797 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1798 size_t kesize = key[i].dttk_size;
1802 (const void *)(uintptr_t)key[i].dttk_value,
1803 (void *)kdata, kesize);
1804 dkey->dttk_value = kdata;
1805 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1807 dkey->dttk_value = key[i].dttk_value;
1810 dkey->dttk_size = kesize;
1813 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1814 dvar->dtdv_hashval = hashval;
1815 dvar->dtdv_next = start;
1817 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1821 * The cas has failed. Either another CPU is adding an element to
1822 * this hash chain, or another CPU is deleting an element from this
1823 * hash chain. The simplest way to deal with both of these cases
1824 * (though not necessarily the most efficient) is to free our
1825 * allocated block and tail-call ourselves. Note that the free is
1826 * to the dirty list and _not_ to the free list. This is to prevent
1827 * races with allocators, above.
1829 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1831 dtrace_membar_producer();
1834 free = dcpu->dtdsc_dirty;
1835 dvar->dtdv_next = free;
1836 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1838 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1843 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1845 if ((int64_t)nval < (int64_t)*oval)
1851 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1853 if ((int64_t)nval > (int64_t)*oval)
1858 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1860 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1861 int64_t val = (int64_t)nval;
1864 for (i = 0; i < zero; i++) {
1865 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1871 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1872 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1873 quanta[i - 1] += incr;
1878 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1886 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1888 uint64_t arg = *lquanta++;
1889 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1890 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1891 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1892 int32_t val = (int32_t)nval, level;
1895 ASSERT(levels != 0);
1899 * This is an underflow.
1905 level = (val - base) / step;
1907 if (level < levels) {
1908 lquanta[level + 1] += incr;
1913 * This is an overflow.
1915 lquanta[levels + 1] += incr;
1919 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1920 uint16_t high, uint16_t nsteps, int64_t value)
1922 int64_t this = 1, last, next;
1923 int base = 1, order;
1925 ASSERT(factor <= nsteps);
1926 ASSERT(nsteps % factor == 0);
1928 for (order = 0; order < low; order++)
1932 * If our value is less than our factor taken to the power of the
1933 * low order of magnitude, it goes into the zeroth bucket.
1935 if (value < (last = this))
1938 for (this *= factor; order <= high; order++) {
1939 int nbuckets = this > nsteps ? nsteps : this;
1941 if ((next = this * factor) < this) {
1943 * We should not generally get log/linear quantizations
1944 * with a high magnitude that allows 64-bits to
1945 * overflow, but we nonetheless protect against this
1946 * by explicitly checking for overflow, and clamping
1947 * our value accordingly.
1954 * If our value lies within this order of magnitude,
1955 * determine its position by taking the offset within
1956 * the order of magnitude, dividing by the bucket
1957 * width, and adding to our (accumulated) base.
1959 return (base + (value - last) / (this / nbuckets));
1962 base += nbuckets - (nbuckets / factor);
1968 * Our value is greater than or equal to our factor taken to the
1969 * power of one plus the high magnitude -- return the top bucket.
1975 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
1977 uint64_t arg = *llquanta++;
1978 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
1979 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
1980 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
1981 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
1983 llquanta[dtrace_aggregate_llquantize_bucket(factor,
1984 low, high, nsteps, nval)] += incr;
1989 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
1997 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
1999 int64_t snval = (int64_t)nval;
2006 * What we want to say here is:
2008 * data[2] += nval * nval;
2010 * But given that nval is 64-bit, we could easily overflow, so
2011 * we do this as 128-bit arithmetic.
2016 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2017 dtrace_add_128(data + 2, tmp, data + 2);
2022 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2029 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2035 * Aggregate given the tuple in the principal data buffer, and the aggregating
2036 * action denoted by the specified dtrace_aggregation_t. The aggregation
2037 * buffer is specified as the buf parameter. This routine does not return
2038 * failure; if there is no space in the aggregation buffer, the data will be
2039 * dropped, and a corresponding counter incremented.
2042 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2043 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2045 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2046 uint32_t i, ndx, size, fsize;
2047 uint32_t align = sizeof (uint64_t) - 1;
2048 dtrace_aggbuffer_t *agb;
2049 dtrace_aggkey_t *key;
2050 uint32_t hashval = 0, limit, isstr;
2051 caddr_t tomax, data, kdata;
2052 dtrace_actkind_t action;
2053 dtrace_action_t *act;
2059 if (!agg->dtag_hasarg) {
2061 * Currently, only quantize() and lquantize() take additional
2062 * arguments, and they have the same semantics: an increment
2063 * value that defaults to 1 when not present. If additional
2064 * aggregating actions take arguments, the setting of the
2065 * default argument value will presumably have to become more
2071 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2072 size = rec->dtrd_offset - agg->dtag_base;
2073 fsize = size + rec->dtrd_size;
2075 ASSERT(dbuf->dtb_tomax != NULL);
2076 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2078 if ((tomax = buf->dtb_tomax) == NULL) {
2079 dtrace_buffer_drop(buf);
2084 * The metastructure is always at the bottom of the buffer.
2086 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2087 sizeof (dtrace_aggbuffer_t));
2089 if (buf->dtb_offset == 0) {
2091 * We just kludge up approximately 1/8th of the size to be
2092 * buckets. If this guess ends up being routinely
2093 * off-the-mark, we may need to dynamically readjust this
2094 * based on past performance.
2096 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2098 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2099 (uintptr_t)tomax || hashsize == 0) {
2101 * We've been given a ludicrously small buffer;
2102 * increment our drop count and leave.
2104 dtrace_buffer_drop(buf);
2109 * And now, a pathetic attempt to try to get a an odd (or
2110 * perchance, a prime) hash size for better hash distribution.
2112 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2113 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2115 agb->dtagb_hashsize = hashsize;
2116 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2117 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2118 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2120 for (i = 0; i < agb->dtagb_hashsize; i++)
2121 agb->dtagb_hash[i] = NULL;
2124 ASSERT(agg->dtag_first != NULL);
2125 ASSERT(agg->dtag_first->dta_intuple);
2128 * Calculate the hash value based on the key. Note that we _don't_
2129 * include the aggid in the hashing (but we will store it as part of
2130 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2131 * algorithm: a simple, quick algorithm that has no known funnels, and
2132 * gets good distribution in practice. The efficacy of the hashing
2133 * algorithm (and a comparison with other algorithms) may be found by
2134 * running the ::dtrace_aggstat MDB dcmd.
2136 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2137 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2138 limit = i + act->dta_rec.dtrd_size;
2139 ASSERT(limit <= size);
2140 isstr = DTRACEACT_ISSTRING(act);
2142 for (; i < limit; i++) {
2144 hashval += (hashval << 10);
2145 hashval ^= (hashval >> 6);
2147 if (isstr && data[i] == '\0')
2152 hashval += (hashval << 3);
2153 hashval ^= (hashval >> 11);
2154 hashval += (hashval << 15);
2157 * Yes, the divide here is expensive -- but it's generally the least
2158 * of the performance issues given the amount of data that we iterate
2159 * over to compute hash values, compare data, etc.
2161 ndx = hashval % agb->dtagb_hashsize;
2163 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2164 ASSERT((caddr_t)key >= tomax);
2165 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2167 if (hashval != key->dtak_hashval || key->dtak_size != size)
2170 kdata = key->dtak_data;
2171 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2173 for (act = agg->dtag_first; act->dta_intuple;
2174 act = act->dta_next) {
2175 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2176 limit = i + act->dta_rec.dtrd_size;
2177 ASSERT(limit <= size);
2178 isstr = DTRACEACT_ISSTRING(act);
2180 for (; i < limit; i++) {
2181 if (kdata[i] != data[i])
2184 if (isstr && data[i] == '\0')
2189 if (action != key->dtak_action) {
2191 * We are aggregating on the same value in the same
2192 * aggregation with two different aggregating actions.
2193 * (This should have been picked up in the compiler,
2194 * so we may be dealing with errant or devious DIF.)
2195 * This is an error condition; we indicate as much,
2198 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2203 * This is a hit: we need to apply the aggregator to
2204 * the value at this key.
2206 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2213 * We didn't find it. We need to allocate some zero-filled space,
2214 * link it into the hash table appropriately, and apply the aggregator
2215 * to the (zero-filled) value.
2217 offs = buf->dtb_offset;
2218 while (offs & (align - 1))
2219 offs += sizeof (uint32_t);
2222 * If we don't have enough room to both allocate a new key _and_
2223 * its associated data, increment the drop count and return.
2225 if ((uintptr_t)tomax + offs + fsize >
2226 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2227 dtrace_buffer_drop(buf);
2232 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2233 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2234 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2236 key->dtak_data = kdata = tomax + offs;
2237 buf->dtb_offset = offs + fsize;
2240 * Now copy the data across.
2242 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2244 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2248 * Because strings are not zeroed out by default, we need to iterate
2249 * looking for actions that store strings, and we need to explicitly
2250 * pad these strings out with zeroes.
2252 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2255 if (!DTRACEACT_ISSTRING(act))
2258 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2259 limit = i + act->dta_rec.dtrd_size;
2260 ASSERT(limit <= size);
2262 for (nul = 0; i < limit; i++) {
2268 if (data[i] != '\0')
2275 for (i = size; i < fsize; i++)
2278 key->dtak_hashval = hashval;
2279 key->dtak_size = size;
2280 key->dtak_action = action;
2281 key->dtak_next = agb->dtagb_hash[ndx];
2282 agb->dtagb_hash[ndx] = key;
2285 * Finally, apply the aggregator.
2287 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2288 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2292 * Given consumer state, this routine finds a speculation in the INACTIVE
2293 * state and transitions it into the ACTIVE state. If there is no speculation
2294 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2295 * incremented -- it is up to the caller to take appropriate action.
2298 dtrace_speculation(dtrace_state_t *state)
2301 dtrace_speculation_state_t current;
2302 uint32_t *stat = &state->dts_speculations_unavail, count;
2304 while (i < state->dts_nspeculations) {
2305 dtrace_speculation_t *spec = &state->dts_speculations[i];
2307 current = spec->dtsp_state;
2309 if (current != DTRACESPEC_INACTIVE) {
2310 if (current == DTRACESPEC_COMMITTINGMANY ||
2311 current == DTRACESPEC_COMMITTING ||
2312 current == DTRACESPEC_DISCARDING)
2313 stat = &state->dts_speculations_busy;
2318 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2319 current, DTRACESPEC_ACTIVE) == current)
2324 * We couldn't find a speculation. If we found as much as a single
2325 * busy speculation buffer, we'll attribute this failure as "busy"
2326 * instead of "unavail".
2330 } while (dtrace_cas32(stat, count, count + 1) != count);
2336 * This routine commits an active speculation. If the specified speculation
2337 * is not in a valid state to perform a commit(), this routine will silently do
2338 * nothing. The state of the specified speculation is transitioned according
2339 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2342 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2343 dtrace_specid_t which)
2345 dtrace_speculation_t *spec;
2346 dtrace_buffer_t *src, *dest;
2347 uintptr_t daddr, saddr, dlimit, slimit;
2348 dtrace_speculation_state_t current, new = 0;
2355 if (which > state->dts_nspeculations) {
2356 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2360 spec = &state->dts_speculations[which - 1];
2361 src = &spec->dtsp_buffer[cpu];
2362 dest = &state->dts_buffer[cpu];
2365 current = spec->dtsp_state;
2367 if (current == DTRACESPEC_COMMITTINGMANY)
2371 case DTRACESPEC_INACTIVE:
2372 case DTRACESPEC_DISCARDING:
2375 case DTRACESPEC_COMMITTING:
2377 * This is only possible if we are (a) commit()'ing
2378 * without having done a prior speculate() on this CPU
2379 * and (b) racing with another commit() on a different
2380 * CPU. There's nothing to do -- we just assert that
2383 ASSERT(src->dtb_offset == 0);
2386 case DTRACESPEC_ACTIVE:
2387 new = DTRACESPEC_COMMITTING;
2390 case DTRACESPEC_ACTIVEONE:
2392 * This speculation is active on one CPU. If our
2393 * buffer offset is non-zero, we know that the one CPU
2394 * must be us. Otherwise, we are committing on a
2395 * different CPU from the speculate(), and we must
2396 * rely on being asynchronously cleaned.
2398 if (src->dtb_offset != 0) {
2399 new = DTRACESPEC_COMMITTING;
2404 case DTRACESPEC_ACTIVEMANY:
2405 new = DTRACESPEC_COMMITTINGMANY;
2411 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2412 current, new) != current);
2415 * We have set the state to indicate that we are committing this
2416 * speculation. Now reserve the necessary space in the destination
2419 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2420 sizeof (uint64_t), state, NULL)) < 0) {
2421 dtrace_buffer_drop(dest);
2426 * We have sufficient space to copy the speculative buffer into the
2427 * primary buffer. First, modify the speculative buffer, filling
2428 * in the timestamp of all entries with the current time. The data
2429 * must have the commit() time rather than the time it was traced,
2430 * so that all entries in the primary buffer are in timestamp order.
2432 timestamp = dtrace_gethrtime();
2433 saddr = (uintptr_t)src->dtb_tomax;
2434 slimit = saddr + src->dtb_offset;
2435 while (saddr < slimit) {
2437 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2439 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2440 saddr += sizeof (dtrace_epid_t);
2443 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2444 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2446 ASSERT3U(saddr + size, <=, slimit);
2447 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2448 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2450 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2456 * Copy the buffer across. (Note that this is a
2457 * highly subobtimal bcopy(); in the unlikely event that this becomes
2458 * a serious performance issue, a high-performance DTrace-specific
2459 * bcopy() should obviously be invented.)
2461 daddr = (uintptr_t)dest->dtb_tomax + offs;
2462 dlimit = daddr + src->dtb_offset;
2463 saddr = (uintptr_t)src->dtb_tomax;
2466 * First, the aligned portion.
2468 while (dlimit - daddr >= sizeof (uint64_t)) {
2469 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2471 daddr += sizeof (uint64_t);
2472 saddr += sizeof (uint64_t);
2476 * Now any left-over bit...
2478 while (dlimit - daddr)
2479 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2482 * Finally, commit the reserved space in the destination buffer.
2484 dest->dtb_offset = offs + src->dtb_offset;
2488 * If we're lucky enough to be the only active CPU on this speculation
2489 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2491 if (current == DTRACESPEC_ACTIVE ||
2492 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2493 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2494 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2496 ASSERT(rval == DTRACESPEC_COMMITTING);
2499 src->dtb_offset = 0;
2500 src->dtb_xamot_drops += src->dtb_drops;
2505 * This routine discards an active speculation. If the specified speculation
2506 * is not in a valid state to perform a discard(), this routine will silently
2507 * do nothing. The state of the specified speculation is transitioned
2508 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2511 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2512 dtrace_specid_t which)
2514 dtrace_speculation_t *spec;
2515 dtrace_speculation_state_t current, new = 0;
2516 dtrace_buffer_t *buf;
2521 if (which > state->dts_nspeculations) {
2522 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2526 spec = &state->dts_speculations[which - 1];
2527 buf = &spec->dtsp_buffer[cpu];
2530 current = spec->dtsp_state;
2533 case DTRACESPEC_INACTIVE:
2534 case DTRACESPEC_COMMITTINGMANY:
2535 case DTRACESPEC_COMMITTING:
2536 case DTRACESPEC_DISCARDING:
2539 case DTRACESPEC_ACTIVE:
2540 case DTRACESPEC_ACTIVEMANY:
2541 new = DTRACESPEC_DISCARDING;
2544 case DTRACESPEC_ACTIVEONE:
2545 if (buf->dtb_offset != 0) {
2546 new = DTRACESPEC_INACTIVE;
2548 new = DTRACESPEC_DISCARDING;
2555 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2556 current, new) != current);
2558 buf->dtb_offset = 0;
2563 * Note: not called from probe context. This function is called
2564 * asynchronously from cross call context to clean any speculations that are
2565 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2566 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2570 dtrace_speculation_clean_here(dtrace_state_t *state)
2572 dtrace_icookie_t cookie;
2573 processorid_t cpu = curcpu;
2574 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2577 cookie = dtrace_interrupt_disable();
2579 if (dest->dtb_tomax == NULL) {
2580 dtrace_interrupt_enable(cookie);
2584 for (i = 0; i < state->dts_nspeculations; i++) {
2585 dtrace_speculation_t *spec = &state->dts_speculations[i];
2586 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2588 if (src->dtb_tomax == NULL)
2591 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2592 src->dtb_offset = 0;
2596 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2599 if (src->dtb_offset == 0)
2602 dtrace_speculation_commit(state, cpu, i + 1);
2605 dtrace_interrupt_enable(cookie);
2609 * Note: not called from probe context. This function is called
2610 * asynchronously (and at a regular interval) to clean any speculations that
2611 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2612 * is work to be done, it cross calls all CPUs to perform that work;
2613 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2614 * INACTIVE state until they have been cleaned by all CPUs.
2617 dtrace_speculation_clean(dtrace_state_t *state)
2622 for (i = 0; i < state->dts_nspeculations; i++) {
2623 dtrace_speculation_t *spec = &state->dts_speculations[i];
2625 ASSERT(!spec->dtsp_cleaning);
2627 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2628 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2632 spec->dtsp_cleaning = 1;
2638 dtrace_xcall(DTRACE_CPUALL,
2639 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2642 * We now know that all CPUs have committed or discarded their
2643 * speculation buffers, as appropriate. We can now set the state
2646 for (i = 0; i < state->dts_nspeculations; i++) {
2647 dtrace_speculation_t *spec = &state->dts_speculations[i];
2648 dtrace_speculation_state_t current, new;
2650 if (!spec->dtsp_cleaning)
2653 current = spec->dtsp_state;
2654 ASSERT(current == DTRACESPEC_DISCARDING ||
2655 current == DTRACESPEC_COMMITTINGMANY);
2657 new = DTRACESPEC_INACTIVE;
2659 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2660 ASSERT(rv == current);
2661 spec->dtsp_cleaning = 0;
2666 * Called as part of a speculate() to get the speculative buffer associated
2667 * with a given speculation. Returns NULL if the specified speculation is not
2668 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2669 * the active CPU is not the specified CPU -- the speculation will be
2670 * atomically transitioned into the ACTIVEMANY state.
2672 static dtrace_buffer_t *
2673 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2674 dtrace_specid_t which)
2676 dtrace_speculation_t *spec;
2677 dtrace_speculation_state_t current, new = 0;
2678 dtrace_buffer_t *buf;
2683 if (which > state->dts_nspeculations) {
2684 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2688 spec = &state->dts_speculations[which - 1];
2689 buf = &spec->dtsp_buffer[cpuid];
2692 current = spec->dtsp_state;
2695 case DTRACESPEC_INACTIVE:
2696 case DTRACESPEC_COMMITTINGMANY:
2697 case DTRACESPEC_DISCARDING:
2700 case DTRACESPEC_COMMITTING:
2701 ASSERT(buf->dtb_offset == 0);
2704 case DTRACESPEC_ACTIVEONE:
2706 * This speculation is currently active on one CPU.
2707 * Check the offset in the buffer; if it's non-zero,
2708 * that CPU must be us (and we leave the state alone).
2709 * If it's zero, assume that we're starting on a new
2710 * CPU -- and change the state to indicate that the
2711 * speculation is active on more than one CPU.
2713 if (buf->dtb_offset != 0)
2716 new = DTRACESPEC_ACTIVEMANY;
2719 case DTRACESPEC_ACTIVEMANY:
2722 case DTRACESPEC_ACTIVE:
2723 new = DTRACESPEC_ACTIVEONE;
2729 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2730 current, new) != current);
2732 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2737 * Return a string. In the event that the user lacks the privilege to access
2738 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2739 * don't fail access checking.
2741 * dtrace_dif_variable() uses this routine as a helper for various
2742 * builtin values such as 'execname' and 'probefunc.'
2745 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2746 dtrace_mstate_t *mstate)
2748 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2753 * The easy case: this probe is allowed to read all of memory, so
2754 * we can just return this as a vanilla pointer.
2756 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2760 * This is the tougher case: we copy the string in question from
2761 * kernel memory into scratch memory and return it that way: this
2762 * ensures that we won't trip up when access checking tests the
2763 * BYREF return value.
2765 strsz = dtrace_strlen((char *)addr, size) + 1;
2767 if (mstate->dtms_scratch_ptr + strsz >
2768 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2769 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2773 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2775 ret = mstate->dtms_scratch_ptr;
2776 mstate->dtms_scratch_ptr += strsz;
2781 * Return a string from a memoy address which is known to have one or
2782 * more concatenated, individually zero terminated, sub-strings.
2783 * In the event that the user lacks the privilege to access
2784 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2785 * don't fail access checking.
2787 * dtrace_dif_variable() uses this routine as a helper for various
2788 * builtin values such as 'execargs'.
2791 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2792 dtrace_mstate_t *mstate)
2798 if (mstate->dtms_scratch_ptr + strsz >
2799 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2800 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2804 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2807 /* Replace sub-string termination characters with a space. */
2808 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2813 ret = mstate->dtms_scratch_ptr;
2814 mstate->dtms_scratch_ptr += strsz;
2819 * This function implements the DIF emulator's variable lookups. The emulator
2820 * passes a reserved variable identifier and optional built-in array index.
2823 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2827 * If we're accessing one of the uncached arguments, we'll turn this
2828 * into a reference in the args array.
2830 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2831 ndx = v - DIF_VAR_ARG0;
2837 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2838 if (ndx >= sizeof (mstate->dtms_arg) /
2839 sizeof (mstate->dtms_arg[0])) {
2840 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2841 dtrace_provider_t *pv;
2844 pv = mstate->dtms_probe->dtpr_provider;
2845 if (pv->dtpv_pops.dtps_getargval != NULL)
2846 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2847 mstate->dtms_probe->dtpr_id,
2848 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2850 val = dtrace_getarg(ndx, aframes);
2853 * This is regrettably required to keep the compiler
2854 * from tail-optimizing the call to dtrace_getarg().
2855 * The condition always evaluates to true, but the
2856 * compiler has no way of figuring that out a priori.
2857 * (None of this would be necessary if the compiler
2858 * could be relied upon to _always_ tail-optimize
2859 * the call to dtrace_getarg() -- but it can't.)
2861 if (mstate->dtms_probe != NULL)
2867 return (mstate->dtms_arg[ndx]);
2870 case DIF_VAR_UREGS: {
2873 if (!dtrace_priv_proc(state))
2876 if ((lwp = curthread->t_lwp) == NULL) {
2877 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2878 cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2882 return (dtrace_getreg(lwp->lwp_regs, ndx));
2886 case DIF_VAR_UREGS: {
2887 struct trapframe *tframe;
2889 if (!dtrace_priv_proc(state))
2892 if ((tframe = curthread->td_frame) == NULL) {
2893 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2894 cpu_core[curcpu].cpuc_dtrace_illval = 0;
2898 return (dtrace_getreg(tframe, ndx));
2902 case DIF_VAR_CURTHREAD:
2903 if (!dtrace_priv_kernel(state))
2905 return ((uint64_t)(uintptr_t)curthread);
2907 case DIF_VAR_TIMESTAMP:
2908 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2909 mstate->dtms_timestamp = dtrace_gethrtime();
2910 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2912 return (mstate->dtms_timestamp);
2914 case DIF_VAR_VTIMESTAMP:
2915 ASSERT(dtrace_vtime_references != 0);
2916 return (curthread->t_dtrace_vtime);
2918 case DIF_VAR_WALLTIMESTAMP:
2919 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2920 mstate->dtms_walltimestamp = dtrace_gethrestime();
2921 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2923 return (mstate->dtms_walltimestamp);
2927 if (!dtrace_priv_kernel(state))
2929 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2930 mstate->dtms_ipl = dtrace_getipl();
2931 mstate->dtms_present |= DTRACE_MSTATE_IPL;
2933 return (mstate->dtms_ipl);
2937 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2938 return (mstate->dtms_epid);
2941 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2942 return (mstate->dtms_probe->dtpr_id);
2944 case DIF_VAR_STACKDEPTH:
2945 if (!dtrace_priv_kernel(state))
2947 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2948 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2950 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2951 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2953 return (mstate->dtms_stackdepth);
2955 case DIF_VAR_USTACKDEPTH:
2956 if (!dtrace_priv_proc(state))
2958 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2960 * See comment in DIF_VAR_PID.
2962 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2964 mstate->dtms_ustackdepth = 0;
2966 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2967 mstate->dtms_ustackdepth =
2968 dtrace_getustackdepth();
2969 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2971 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2973 return (mstate->dtms_ustackdepth);
2975 case DIF_VAR_CALLER:
2976 if (!dtrace_priv_kernel(state))
2978 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2979 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2981 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2983 * If this is an unanchored probe, we are
2984 * required to go through the slow path:
2985 * dtrace_caller() only guarantees correct
2986 * results for anchored probes.
2988 pc_t caller[2] = {0, 0};
2990 dtrace_getpcstack(caller, 2, aframes,
2991 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2992 mstate->dtms_caller = caller[1];
2993 } else if ((mstate->dtms_caller =
2994 dtrace_caller(aframes)) == -1) {
2996 * We have failed to do this the quick way;
2997 * we must resort to the slower approach of
2998 * calling dtrace_getpcstack().
3002 dtrace_getpcstack(&caller, 1, aframes, NULL);
3003 mstate->dtms_caller = caller;
3006 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3008 return (mstate->dtms_caller);
3010 case DIF_VAR_UCALLER:
3011 if (!dtrace_priv_proc(state))
3014 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3018 * dtrace_getupcstack() fills in the first uint64_t
3019 * with the current PID. The second uint64_t will
3020 * be the program counter at user-level. The third
3021 * uint64_t will contain the caller, which is what
3025 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3026 dtrace_getupcstack(ustack, 3);
3027 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3028 mstate->dtms_ucaller = ustack[2];
3029 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3032 return (mstate->dtms_ucaller);
3034 case DIF_VAR_PROBEPROV:
3035 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3036 return (dtrace_dif_varstr(
3037 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3040 case DIF_VAR_PROBEMOD:
3041 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3042 return (dtrace_dif_varstr(
3043 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3046 case DIF_VAR_PROBEFUNC:
3047 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3048 return (dtrace_dif_varstr(
3049 (uintptr_t)mstate->dtms_probe->dtpr_func,
3052 case DIF_VAR_PROBENAME:
3053 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3054 return (dtrace_dif_varstr(
3055 (uintptr_t)mstate->dtms_probe->dtpr_name,
3059 if (!dtrace_priv_proc(state))
3064 * Note that we are assuming that an unanchored probe is
3065 * always due to a high-level interrupt. (And we're assuming
3066 * that there is only a single high level interrupt.)
3068 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3069 return (pid0.pid_id);
3072 * It is always safe to dereference one's own t_procp pointer:
3073 * it always points to a valid, allocated proc structure.
3074 * Further, it is always safe to dereference the p_pidp member
3075 * of one's own proc structure. (These are truisms becuase
3076 * threads and processes don't clean up their own state --
3077 * they leave that task to whomever reaps them.)
3079 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3081 return ((uint64_t)curproc->p_pid);
3085 if (!dtrace_priv_proc(state))
3090 * See comment in DIF_VAR_PID.
3092 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3093 return (pid0.pid_id);
3096 * It is always safe to dereference one's own t_procp pointer:
3097 * it always points to a valid, allocated proc structure.
3098 * (This is true because threads don't clean up their own
3099 * state -- they leave that task to whomever reaps them.)
3101 return ((uint64_t)curthread->t_procp->p_ppid);
3103 return ((uint64_t)curproc->p_pptr->p_pid);
3109 * See comment in DIF_VAR_PID.
3111 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3115 return ((uint64_t)curthread->t_tid);
3117 case DIF_VAR_EXECARGS: {
3118 struct pargs *p_args = curthread->td_proc->p_args;
3123 return (dtrace_dif_varstrz(
3124 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3127 case DIF_VAR_EXECNAME:
3129 if (!dtrace_priv_proc(state))
3133 * See comment in DIF_VAR_PID.
3135 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3136 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3139 * It is always safe to dereference one's own t_procp pointer:
3140 * it always points to a valid, allocated proc structure.
3141 * (This is true because threads don't clean up their own
3142 * state -- they leave that task to whomever reaps them.)
3144 return (dtrace_dif_varstr(
3145 (uintptr_t)curthread->t_procp->p_user.u_comm,
3148 return (dtrace_dif_varstr(
3149 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3152 case DIF_VAR_ZONENAME:
3154 if (!dtrace_priv_proc(state))
3158 * See comment in DIF_VAR_PID.
3160 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3161 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3164 * It is always safe to dereference one's own t_procp pointer:
3165 * it always points to a valid, allocated proc structure.
3166 * (This is true because threads don't clean up their own
3167 * state -- they leave that task to whomever reaps them.)
3169 return (dtrace_dif_varstr(
3170 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3177 if (!dtrace_priv_proc(state))
3182 * See comment in DIF_VAR_PID.
3184 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3185 return ((uint64_t)p0.p_cred->cr_uid);
3189 * It is always safe to dereference one's own t_procp pointer:
3190 * it always points to a valid, allocated proc structure.
3191 * (This is true because threads don't clean up their own
3192 * state -- they leave that task to whomever reaps them.)
3194 * Additionally, it is safe to dereference one's own process
3195 * credential, since this is never NULL after process birth.
3197 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3200 if (!dtrace_priv_proc(state))
3205 * See comment in DIF_VAR_PID.
3207 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3208 return ((uint64_t)p0.p_cred->cr_gid);
3212 * It is always safe to dereference one's own t_procp pointer:
3213 * it always points to a valid, allocated proc structure.
3214 * (This is true because threads don't clean up their own
3215 * state -- they leave that task to whomever reaps them.)
3217 * Additionally, it is safe to dereference one's own process
3218 * credential, since this is never NULL after process birth.
3220 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3222 case DIF_VAR_ERRNO: {
3225 if (!dtrace_priv_proc(state))
3229 * See comment in DIF_VAR_PID.
3231 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3235 * It is always safe to dereference one's own t_lwp pointer in
3236 * the event that this pointer is non-NULL. (This is true
3237 * because threads and lwps don't clean up their own state --
3238 * they leave that task to whomever reaps them.)
3240 if ((lwp = curthread->t_lwp) == NULL)
3243 return ((uint64_t)lwp->lwp_errno);
3245 return (curthread->td_errno);
3254 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3260 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3261 * Notice that we don't bother validating the proper number of arguments or
3262 * their types in the tuple stack. This isn't needed because all argument
3263 * interpretation is safe because of our load safety -- the worst that can
3264 * happen is that a bogus program can obtain bogus results.
3267 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3268 dtrace_key_t *tupregs, int nargs,
3269 dtrace_mstate_t *mstate, dtrace_state_t *state)
3271 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3272 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3273 dtrace_vstate_t *vstate = &state->dts_vstate;
3286 struct thread *lowner;
3288 struct lock_object *li;
3295 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3299 case DIF_SUBR_MUTEX_OWNED:
3300 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3306 m.mx = dtrace_load64(tupregs[0].dttk_value);
3307 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3308 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3310 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3313 case DIF_SUBR_MUTEX_OWNER:
3314 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3320 m.mx = dtrace_load64(tupregs[0].dttk_value);
3321 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3322 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3323 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3328 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3329 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3335 m.mx = dtrace_load64(tupregs[0].dttk_value);
3336 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3339 case DIF_SUBR_MUTEX_TYPE_SPIN:
3340 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3346 m.mx = dtrace_load64(tupregs[0].dttk_value);
3347 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3350 case DIF_SUBR_RW_READ_HELD: {
3353 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3359 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3360 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3364 case DIF_SUBR_RW_WRITE_HELD:
3365 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3371 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3372 regs[rd] = _RW_WRITE_HELD(&r.ri);
3375 case DIF_SUBR_RW_ISWRITER:
3376 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3382 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3383 regs[rd] = _RW_ISWRITER(&r.ri);
3387 case DIF_SUBR_MUTEX_OWNED:
3388 if (!dtrace_canload(tupregs[0].dttk_value,
3389 sizeof (struct lock_object), mstate, vstate)) {
3393 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3394 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3397 case DIF_SUBR_MUTEX_OWNER:
3398 if (!dtrace_canload(tupregs[0].dttk_value,
3399 sizeof (struct lock_object), mstate, vstate)) {
3403 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3404 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3405 regs[rd] = (uintptr_t)lowner;
3408 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3409 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3414 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3415 /* XXX - should be only LC_SLEEPABLE? */
3416 regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3417 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3420 case DIF_SUBR_MUTEX_TYPE_SPIN:
3421 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3426 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3427 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3430 case DIF_SUBR_RW_READ_HELD:
3431 case DIF_SUBR_SX_SHARED_HELD:
3432 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3437 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3438 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3442 case DIF_SUBR_RW_WRITE_HELD:
3443 case DIF_SUBR_SX_EXCLUSIVE_HELD:
3444 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3449 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3450 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3451 regs[rd] = (lowner == curthread);
3454 case DIF_SUBR_RW_ISWRITER:
3455 case DIF_SUBR_SX_ISEXCLUSIVE:
3456 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3461 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3462 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3465 #endif /* ! defined(sun) */
3467 case DIF_SUBR_BCOPY: {
3469 * We need to be sure that the destination is in the scratch
3470 * region -- no other region is allowed.
3472 uintptr_t src = tupregs[0].dttk_value;
3473 uintptr_t dest = tupregs[1].dttk_value;
3474 size_t size = tupregs[2].dttk_value;
3476 if (!dtrace_inscratch(dest, size, mstate)) {
3477 *flags |= CPU_DTRACE_BADADDR;
3482 if (!dtrace_canload(src, size, mstate, vstate)) {
3487 dtrace_bcopy((void *)src, (void *)dest, size);
3491 case DIF_SUBR_ALLOCA:
3492 case DIF_SUBR_COPYIN: {
3493 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3495 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3496 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3499 * This action doesn't require any credential checks since
3500 * probes will not activate in user contexts to which the
3501 * enabling user does not have permissions.
3505 * Rounding up the user allocation size could have overflowed
3506 * a large, bogus allocation (like -1ULL) to 0.
3508 if (scratch_size < size ||
3509 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3510 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3515 if (subr == DIF_SUBR_COPYIN) {
3516 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3517 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3518 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3521 mstate->dtms_scratch_ptr += scratch_size;
3526 case DIF_SUBR_COPYINTO: {
3527 uint64_t size = tupregs[1].dttk_value;
3528 uintptr_t dest = tupregs[2].dttk_value;
3531 * This action doesn't require any credential checks since
3532 * probes will not activate in user contexts to which the
3533 * enabling user does not have permissions.
3535 if (!dtrace_inscratch(dest, size, mstate)) {
3536 *flags |= CPU_DTRACE_BADADDR;
3541 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3542 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3543 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3547 case DIF_SUBR_COPYINSTR: {
3548 uintptr_t dest = mstate->dtms_scratch_ptr;
3549 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3551 if (nargs > 1 && tupregs[1].dttk_value < size)
3552 size = tupregs[1].dttk_value + 1;
3555 * This action doesn't require any credential checks since
3556 * probes will not activate in user contexts to which the
3557 * enabling user does not have permissions.
3559 if (!DTRACE_INSCRATCH(mstate, size)) {
3560 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3565 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3566 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3567 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3569 ((char *)dest)[size - 1] = '\0';
3570 mstate->dtms_scratch_ptr += size;
3576 case DIF_SUBR_MSGSIZE:
3577 case DIF_SUBR_MSGDSIZE: {
3578 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3579 uintptr_t wptr, rptr;
3583 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3585 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3591 wptr = dtrace_loadptr(baddr +
3592 offsetof(mblk_t, b_wptr));
3594 rptr = dtrace_loadptr(baddr +
3595 offsetof(mblk_t, b_rptr));
3598 *flags |= CPU_DTRACE_BADADDR;
3599 *illval = tupregs[0].dttk_value;
3603 daddr = dtrace_loadptr(baddr +
3604 offsetof(mblk_t, b_datap));
3606 baddr = dtrace_loadptr(baddr +
3607 offsetof(mblk_t, b_cont));
3610 * We want to prevent against denial-of-service here,
3611 * so we're only going to search the list for
3612 * dtrace_msgdsize_max mblks.
3614 if (cont++ > dtrace_msgdsize_max) {
3615 *flags |= CPU_DTRACE_ILLOP;
3619 if (subr == DIF_SUBR_MSGDSIZE) {
3620 if (dtrace_load8(daddr +
3621 offsetof(dblk_t, db_type)) != M_DATA)
3625 count += wptr - rptr;
3628 if (!(*flags & CPU_DTRACE_FAULT))
3635 case DIF_SUBR_PROGENYOF: {
3636 pid_t pid = tupregs[0].dttk_value;
3640 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3642 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3644 if (p->p_pidp->pid_id == pid) {
3646 if (p->p_pid == pid) {
3653 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3659 case DIF_SUBR_SPECULATION:
3660 regs[rd] = dtrace_speculation(state);
3663 case DIF_SUBR_COPYOUT: {
3664 uintptr_t kaddr = tupregs[0].dttk_value;
3665 uintptr_t uaddr = tupregs[1].dttk_value;
3666 uint64_t size = tupregs[2].dttk_value;
3668 if (!dtrace_destructive_disallow &&
3669 dtrace_priv_proc_control(state) &&
3670 !dtrace_istoxic(kaddr, size)) {
3671 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3672 dtrace_copyout(kaddr, uaddr, size, flags);
3673 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3678 case DIF_SUBR_COPYOUTSTR: {
3679 uintptr_t kaddr = tupregs[0].dttk_value;
3680 uintptr_t uaddr = tupregs[1].dttk_value;
3681 uint64_t size = tupregs[2].dttk_value;
3683 if (!dtrace_destructive_disallow &&
3684 dtrace_priv_proc_control(state) &&
3685 !dtrace_istoxic(kaddr, size)) {
3686 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3687 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3688 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3693 case DIF_SUBR_STRLEN: {
3695 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3696 sz = dtrace_strlen((char *)addr,
3697 state->dts_options[DTRACEOPT_STRSIZE]);
3699 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3709 case DIF_SUBR_STRCHR:
3710 case DIF_SUBR_STRRCHR: {
3712 * We're going to iterate over the string looking for the
3713 * specified character. We will iterate until we have reached
3714 * the string length or we have found the character. If this
3715 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3716 * of the specified character instead of the first.
3718 uintptr_t saddr = tupregs[0].dttk_value;
3719 uintptr_t addr = tupregs[0].dttk_value;
3720 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3721 char c, target = (char)tupregs[1].dttk_value;
3723 for (regs[rd] = 0; addr < limit; addr++) {
3724 if ((c = dtrace_load8(addr)) == target) {
3727 if (subr == DIF_SUBR_STRCHR)
3735 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3743 case DIF_SUBR_STRSTR:
3744 case DIF_SUBR_INDEX:
3745 case DIF_SUBR_RINDEX: {
3747 * We're going to iterate over the string looking for the
3748 * specified string. We will iterate until we have reached
3749 * the string length or we have found the string. (Yes, this
3750 * is done in the most naive way possible -- but considering
3751 * that the string we're searching for is likely to be
3752 * relatively short, the complexity of Rabin-Karp or similar
3753 * hardly seems merited.)
3755 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3756 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3757 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3758 size_t len = dtrace_strlen(addr, size);
3759 size_t sublen = dtrace_strlen(substr, size);
3760 char *limit = addr + len, *orig = addr;
3761 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3764 regs[rd] = notfound;
3766 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3771 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3778 * strstr() and index()/rindex() have similar semantics if
3779 * both strings are the empty string: strstr() returns a
3780 * pointer to the (empty) string, and index() and rindex()
3781 * both return index 0 (regardless of any position argument).
3783 if (sublen == 0 && len == 0) {
3784 if (subr == DIF_SUBR_STRSTR)
3785 regs[rd] = (uintptr_t)addr;
3791 if (subr != DIF_SUBR_STRSTR) {
3792 if (subr == DIF_SUBR_RINDEX) {
3799 * Both index() and rindex() take an optional position
3800 * argument that denotes the starting position.
3803 int64_t pos = (int64_t)tupregs[2].dttk_value;
3806 * If the position argument to index() is
3807 * negative, Perl implicitly clamps it at
3808 * zero. This semantic is a little surprising
3809 * given the special meaning of negative
3810 * positions to similar Perl functions like
3811 * substr(), but it appears to reflect a
3812 * notion that index() can start from a
3813 * negative index and increment its way up to
3814 * the string. Given this notion, Perl's
3815 * rindex() is at least self-consistent in
3816 * that it implicitly clamps positions greater
3817 * than the string length to be the string
3818 * length. Where Perl completely loses
3819 * coherence, however, is when the specified
3820 * substring is the empty string (""). In
3821 * this case, even if the position is
3822 * negative, rindex() returns 0 -- and even if
3823 * the position is greater than the length,
3824 * index() returns the string length. These
3825 * semantics violate the notion that index()
3826 * should never return a value less than the
3827 * specified position and that rindex() should
3828 * never return a value greater than the
3829 * specified position. (One assumes that
3830 * these semantics are artifacts of Perl's
3831 * implementation and not the results of
3832 * deliberate design -- it beggars belief that
3833 * even Larry Wall could desire such oddness.)
3834 * While in the abstract one would wish for
3835 * consistent position semantics across
3836 * substr(), index() and rindex() -- or at the
3837 * very least self-consistent position
3838 * semantics for index() and rindex() -- we
3839 * instead opt to keep with the extant Perl
3840 * semantics, in all their broken glory. (Do
3841 * we have more desire to maintain Perl's
3842 * semantics than Perl does? Probably.)
3844 if (subr == DIF_SUBR_RINDEX) {
3868 for (regs[rd] = notfound; addr != limit; addr += inc) {
3869 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3870 if (subr != DIF_SUBR_STRSTR) {
3872 * As D index() and rindex() are
3873 * modeled on Perl (and not on awk),
3874 * we return a zero-based (and not a
3875 * one-based) index. (For you Perl
3876 * weenies: no, we're not going to add
3877 * $[ -- and shouldn't you be at a con
3880 regs[rd] = (uintptr_t)(addr - orig);
3884 ASSERT(subr == DIF_SUBR_STRSTR);
3885 regs[rd] = (uintptr_t)addr;
3893 case DIF_SUBR_STRTOK: {
3894 uintptr_t addr = tupregs[0].dttk_value;
3895 uintptr_t tokaddr = tupregs[1].dttk_value;
3896 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3897 uintptr_t limit, toklimit = tokaddr + size;
3898 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
3899 char *dest = (char *)mstate->dtms_scratch_ptr;
3903 * Check both the token buffer and (later) the input buffer,
3904 * since both could be non-scratch addresses.
3906 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3911 if (!DTRACE_INSCRATCH(mstate, size)) {
3912 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3919 * If the address specified is NULL, we use our saved
3920 * strtok pointer from the mstate. Note that this
3921 * means that the saved strtok pointer is _only_
3922 * valid within multiple enablings of the same probe --
3923 * it behaves like an implicit clause-local variable.
3925 addr = mstate->dtms_strtok;
3928 * If the user-specified address is non-NULL we must
3929 * access check it. This is the only time we have
3930 * a chance to do so, since this address may reside
3931 * in the string table of this clause-- future calls
3932 * (when we fetch addr from mstate->dtms_strtok)
3933 * would fail this access check.
3935 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3942 * First, zero the token map, and then process the token
3943 * string -- setting a bit in the map for every character
3944 * found in the token string.
3946 for (i = 0; i < sizeof (tokmap); i++)
3949 for (; tokaddr < toklimit; tokaddr++) {
3950 if ((c = dtrace_load8(tokaddr)) == '\0')
3953 ASSERT((c >> 3) < sizeof (tokmap));
3954 tokmap[c >> 3] |= (1 << (c & 0x7));
3957 for (limit = addr + size; addr < limit; addr++) {
3959 * We're looking for a character that is _not_ contained
3960 * in the token string.
3962 if ((c = dtrace_load8(addr)) == '\0')
3965 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3971 * We reached the end of the string without finding
3972 * any character that was not in the token string.
3973 * We return NULL in this case, and we set the saved
3974 * address to NULL as well.
3977 mstate->dtms_strtok = 0;
3982 * From here on, we're copying into the destination string.
3984 for (i = 0; addr < limit && i < size - 1; addr++) {
3985 if ((c = dtrace_load8(addr)) == '\0')
3988 if (tokmap[c >> 3] & (1 << (c & 0x7)))
3997 regs[rd] = (uintptr_t)dest;
3998 mstate->dtms_scratch_ptr += size;
3999 mstate->dtms_strtok = addr;
4003 case DIF_SUBR_SUBSTR: {
4004 uintptr_t s = tupregs[0].dttk_value;
4005 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4006 char *d = (char *)mstate->dtms_scratch_ptr;
4007 int64_t index = (int64_t)tupregs[1].dttk_value;
4008 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4009 size_t len = dtrace_strlen((char *)s, size);
4012 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4017 if (!DTRACE_INSCRATCH(mstate, size)) {
4018 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4024 remaining = (int64_t)size;
4029 if (index < 0 && index + remaining > 0) {
4035 if (index >= len || index < 0) {
4037 } else if (remaining < 0) {
4038 remaining += len - index;
4039 } else if (index + remaining > size) {
4040 remaining = size - index;
4043 for (i = 0; i < remaining; i++) {
4044 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4050 mstate->dtms_scratch_ptr += size;
4051 regs[rd] = (uintptr_t)d;
4055 case DIF_SUBR_TOUPPER:
4056 case DIF_SUBR_TOLOWER: {
4057 uintptr_t s = tupregs[0].dttk_value;
4058 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4059 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4060 size_t len = dtrace_strlen((char *)s, size);
4061 char lower, upper, convert;
4064 if (subr == DIF_SUBR_TOUPPER) {
4074 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4079 if (!DTRACE_INSCRATCH(mstate, size)) {
4080 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4085 for (i = 0; i < size - 1; i++) {
4086 if ((c = dtrace_load8(s + i)) == '\0')
4089 if (c >= lower && c <= upper)
4090 c = convert + (c - lower);
4097 regs[rd] = (uintptr_t)dest;
4098 mstate->dtms_scratch_ptr += size;
4103 case DIF_SUBR_GETMAJOR:
4105 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4107 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4111 case DIF_SUBR_GETMINOR:
4113 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4115 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4119 case DIF_SUBR_DDI_PATHNAME: {
4121 * This one is a galactic mess. We are going to roughly
4122 * emulate ddi_pathname(), but it's made more complicated
4123 * by the fact that we (a) want to include the minor name and
4124 * (b) must proceed iteratively instead of recursively.
4126 uintptr_t dest = mstate->dtms_scratch_ptr;
4127 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4128 char *start = (char *)dest, *end = start + size - 1;
4129 uintptr_t daddr = tupregs[0].dttk_value;
4130 int64_t minor = (int64_t)tupregs[1].dttk_value;
4132 int i, len, depth = 0;
4135 * Due to all the pointer jumping we do and context we must
4136 * rely upon, we just mandate that the user must have kernel
4137 * read privileges to use this routine.
4139 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4140 *flags |= CPU_DTRACE_KPRIV;
4145 if (!DTRACE_INSCRATCH(mstate, size)) {
4146 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4154 * We want to have a name for the minor. In order to do this,
4155 * we need to walk the minor list from the devinfo. We want
4156 * to be sure that we don't infinitely walk a circular list,
4157 * so we check for circularity by sending a scout pointer
4158 * ahead two elements for every element that we iterate over;
4159 * if the list is circular, these will ultimately point to the
4160 * same element. You may recognize this little trick as the
4161 * answer to a stupid interview question -- one that always
4162 * seems to be asked by those who had to have it laboriously
4163 * explained to them, and who can't even concisely describe
4164 * the conditions under which one would be forced to resort to
4165 * this technique. Needless to say, those conditions are
4166 * found here -- and probably only here. Is this the only use
4167 * of this infamous trick in shipping, production code? If it
4168 * isn't, it probably should be...
4171 uintptr_t maddr = dtrace_loadptr(daddr +
4172 offsetof(struct dev_info, devi_minor));
4174 uintptr_t next = offsetof(struct ddi_minor_data, next);
4175 uintptr_t name = offsetof(struct ddi_minor_data,
4176 d_minor) + offsetof(struct ddi_minor, name);
4177 uintptr_t dev = offsetof(struct ddi_minor_data,
4178 d_minor) + offsetof(struct ddi_minor, dev);
4182 scout = dtrace_loadptr(maddr + next);
4184 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4187 m = dtrace_load64(maddr + dev) & MAXMIN64;
4189 m = dtrace_load32(maddr + dev) & MAXMIN;
4192 maddr = dtrace_loadptr(maddr + next);
4197 scout = dtrace_loadptr(scout + next);
4202 scout = dtrace_loadptr(scout + next);
4207 if (scout == maddr) {
4208 *flags |= CPU_DTRACE_ILLOP;
4216 * We have the minor data. Now we need to
4217 * copy the minor's name into the end of the
4220 s = (char *)dtrace_loadptr(maddr + name);
4221 len = dtrace_strlen(s, size);
4223 if (*flags & CPU_DTRACE_FAULT)
4227 if ((end -= (len + 1)) < start)
4233 for (i = 1; i <= len; i++)
4234 end[i] = dtrace_load8((uintptr_t)s++);
4239 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4240 ddi_node_state_t devi_state;
4242 devi_state = dtrace_load32(daddr +
4243 offsetof(struct dev_info, devi_node_state));
4245 if (*flags & CPU_DTRACE_FAULT)
4248 if (devi_state >= DS_INITIALIZED) {
4249 s = (char *)dtrace_loadptr(daddr +
4250 offsetof(struct dev_info, devi_addr));
4251 len = dtrace_strlen(s, size);
4253 if (*flags & CPU_DTRACE_FAULT)
4257 if ((end -= (len + 1)) < start)
4263 for (i = 1; i <= len; i++)
4264 end[i] = dtrace_load8((uintptr_t)s++);
4268 * Now for the node name...
4270 s = (char *)dtrace_loadptr(daddr +
4271 offsetof(struct dev_info, devi_node_name));
4273 daddr = dtrace_loadptr(daddr +
4274 offsetof(struct dev_info, devi_parent));
4277 * If our parent is NULL (that is, if we're the root
4278 * node), we're going to use the special path
4284 len = dtrace_strlen(s, size);
4285 if (*flags & CPU_DTRACE_FAULT)
4288 if ((end -= (len + 1)) < start)
4291 for (i = 1; i <= len; i++)
4292 end[i] = dtrace_load8((uintptr_t)s++);
4295 if (depth++ > dtrace_devdepth_max) {
4296 *flags |= CPU_DTRACE_ILLOP;
4302 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4305 regs[rd] = (uintptr_t)end;
4306 mstate->dtms_scratch_ptr += size;
4313 case DIF_SUBR_STRJOIN: {
4314 char *d = (char *)mstate->dtms_scratch_ptr;
4315 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4316 uintptr_t s1 = tupregs[0].dttk_value;
4317 uintptr_t s2 = tupregs[1].dttk_value;
4320 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4321 !dtrace_strcanload(s2, size, mstate, vstate)) {
4326 if (!DTRACE_INSCRATCH(mstate, size)) {
4327 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4334 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4339 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4347 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4352 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4357 mstate->dtms_scratch_ptr += i;
4358 regs[rd] = (uintptr_t)d;
4364 case DIF_SUBR_LLTOSTR: {
4365 int64_t i = (int64_t)tupregs[0].dttk_value;
4366 uint64_t val, digit;
4367 uint64_t size = 65; /* enough room for 2^64 in binary */
4368 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4372 if ((base = tupregs[1].dttk_value) <= 1 ||
4373 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4374 *flags |= CPU_DTRACE_ILLOP;
4379 val = (base == 10 && i < 0) ? i * -1 : i;
4381 if (!DTRACE_INSCRATCH(mstate, size)) {
4382 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4387 for (*end-- = '\0'; val; val /= base) {
4388 if ((digit = val % base) <= '9' - '0') {
4389 *end-- = '0' + digit;
4391 *end-- = 'a' + (digit - ('9' - '0') - 1);
4395 if (i == 0 && base == 16)
4401 if (i == 0 || base == 8 || base == 16)
4404 if (i < 0 && base == 10)
4407 regs[rd] = (uintptr_t)end + 1;
4408 mstate->dtms_scratch_ptr += size;
4412 case DIF_SUBR_HTONS:
4413 case DIF_SUBR_NTOHS:
4414 #if BYTE_ORDER == BIG_ENDIAN
4415 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4417 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4422 case DIF_SUBR_HTONL:
4423 case DIF_SUBR_NTOHL:
4424 #if BYTE_ORDER == BIG_ENDIAN
4425 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4427 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4432 case DIF_SUBR_HTONLL:
4433 case DIF_SUBR_NTOHLL:
4434 #if BYTE_ORDER == BIG_ENDIAN
4435 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4437 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4442 case DIF_SUBR_DIRNAME:
4443 case DIF_SUBR_BASENAME: {
4444 char *dest = (char *)mstate->dtms_scratch_ptr;
4445 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4446 uintptr_t src = tupregs[0].dttk_value;
4447 int i, j, len = dtrace_strlen((char *)src, size);
4448 int lastbase = -1, firstbase = -1, lastdir = -1;
4451 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4456 if (!DTRACE_INSCRATCH(mstate, size)) {
4457 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4463 * The basename and dirname for a zero-length string is
4468 src = (uintptr_t)".";
4472 * Start from the back of the string, moving back toward the
4473 * front until we see a character that isn't a slash. That
4474 * character is the last character in the basename.
4476 for (i = len - 1; i >= 0; i--) {
4477 if (dtrace_load8(src + i) != '/')
4485 * Starting from the last character in the basename, move
4486 * towards the front until we find a slash. The character
4487 * that we processed immediately before that is the first
4488 * character in the basename.
4490 for (; i >= 0; i--) {
4491 if (dtrace_load8(src + i) == '/')
4499 * Now keep going until we find a non-slash character. That
4500 * character is the last character in the dirname.
4502 for (; i >= 0; i--) {
4503 if (dtrace_load8(src + i) != '/')
4510 ASSERT(!(lastbase == -1 && firstbase != -1));
4511 ASSERT(!(firstbase == -1 && lastdir != -1));
4513 if (lastbase == -1) {
4515 * We didn't find a non-slash character. We know that
4516 * the length is non-zero, so the whole string must be
4517 * slashes. In either the dirname or the basename
4518 * case, we return '/'.
4520 ASSERT(firstbase == -1);
4521 firstbase = lastbase = lastdir = 0;
4524 if (firstbase == -1) {
4526 * The entire string consists only of a basename
4527 * component. If we're looking for dirname, we need
4528 * to change our string to be just "."; if we're
4529 * looking for a basename, we'll just set the first
4530 * character of the basename to be 0.
4532 if (subr == DIF_SUBR_DIRNAME) {
4533 ASSERT(lastdir == -1);
4534 src = (uintptr_t)".";
4541 if (subr == DIF_SUBR_DIRNAME) {
4542 if (lastdir == -1) {
4544 * We know that we have a slash in the name --
4545 * or lastdir would be set to 0, above. And
4546 * because lastdir is -1, we know that this
4547 * slash must be the first character. (That
4548 * is, the full string must be of the form
4549 * "/basename".) In this case, the last
4550 * character of the directory name is 0.
4558 ASSERT(subr == DIF_SUBR_BASENAME);
4559 ASSERT(firstbase != -1 && lastbase != -1);
4564 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4565 dest[j] = dtrace_load8(src + i);
4568 regs[rd] = (uintptr_t)dest;
4569 mstate->dtms_scratch_ptr += size;
4573 case DIF_SUBR_CLEANPATH: {
4574 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4575 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4576 uintptr_t src = tupregs[0].dttk_value;
4579 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4584 if (!DTRACE_INSCRATCH(mstate, size)) {
4585 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4591 * Move forward, loading each character.
4594 c = dtrace_load8(src + i++);
4596 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4604 c = dtrace_load8(src + i++);
4608 * We have two slashes -- we can just advance
4609 * to the next character.
4616 * This is not "." and it's not ".." -- we can
4617 * just store the "/" and this character and
4625 c = dtrace_load8(src + i++);
4629 * This is a "/./" component. We're not going
4630 * to store anything in the destination buffer;
4631 * we're just going to go to the next component.
4638 * This is not ".." -- we can just store the
4639 * "/." and this character and continue
4648 c = dtrace_load8(src + i++);
4650 if (c != '/' && c != '\0') {
4652 * This is not ".." -- it's "..[mumble]".
4653 * We'll store the "/.." and this character
4654 * and continue processing.
4664 * This is "/../" or "/..\0". We need to back up
4665 * our destination pointer until we find a "/".
4668 while (j != 0 && dest[--j] != '/')
4673 } while (c != '\0');
4676 regs[rd] = (uintptr_t)dest;
4677 mstate->dtms_scratch_ptr += size;
4681 case DIF_SUBR_INET_NTOA:
4682 case DIF_SUBR_INET_NTOA6:
4683 case DIF_SUBR_INET_NTOP: {
4688 if (subr == DIF_SUBR_INET_NTOP) {
4689 af = (int)tupregs[0].dttk_value;
4692 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4696 if (af == AF_INET) {
4701 * Safely load the IPv4 address.
4703 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4706 * Check an IPv4 string will fit in scratch.
4708 size = INET_ADDRSTRLEN;
4709 if (!DTRACE_INSCRATCH(mstate, size)) {
4710 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4714 base = (char *)mstate->dtms_scratch_ptr;
4715 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4718 * Stringify as a dotted decimal quad.
4721 ptr8 = (uint8_t *)&ip4;
4722 for (i = 3; i >= 0; i--) {
4728 for (; val; val /= 10) {
4729 *end-- = '0' + (val % 10);
4736 ASSERT(end + 1 >= base);
4738 } else if (af == AF_INET6) {
4739 struct in6_addr ip6;
4740 int firstzero, tryzero, numzero, v6end;
4742 const char digits[] = "0123456789abcdef";
4745 * Stringify using RFC 1884 convention 2 - 16 bit
4746 * hexadecimal values with a zero-run compression.
4747 * Lower case hexadecimal digits are used.
4748 * eg, fe80::214:4fff:fe0b:76c8.
4749 * The IPv4 embedded form is returned for inet_ntop,
4750 * just the IPv4 string is returned for inet_ntoa6.
4754 * Safely load the IPv6 address.
4757 (void *)(uintptr_t)tupregs[argi].dttk_value,
4758 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4761 * Check an IPv6 string will fit in scratch.
4763 size = INET6_ADDRSTRLEN;
4764 if (!DTRACE_INSCRATCH(mstate, size)) {
4765 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4769 base = (char *)mstate->dtms_scratch_ptr;
4770 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4774 * Find the longest run of 16 bit zero values
4775 * for the single allowed zero compression - "::".
4780 for (i = 0; i < sizeof (struct in6_addr); i++) {
4782 if (ip6._S6_un._S6_u8[i] == 0 &&
4784 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4786 tryzero == -1 && i % 2 == 0) {
4791 if (tryzero != -1 &&
4793 (ip6._S6_un._S6_u8[i] != 0 ||
4795 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4797 i == sizeof (struct in6_addr) - 1)) {
4799 if (i - tryzero <= numzero) {
4804 firstzero = tryzero;
4805 numzero = i - i % 2 - tryzero;
4809 if (ip6._S6_un._S6_u8[i] == 0 &&
4811 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4813 i == sizeof (struct in6_addr) - 1)
4817 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4820 * Check for an IPv4 embedded address.
4822 v6end = sizeof (struct in6_addr) - 2;
4823 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4824 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4825 for (i = sizeof (struct in6_addr) - 1;
4826 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4827 ASSERT(end >= base);
4830 val = ip6._S6_un._S6_u8[i];
4832 val = ip6.__u6_addr.__u6_addr8[i];
4838 for (; val; val /= 10) {
4839 *end-- = '0' + val % 10;
4843 if (i > DTRACE_V4MAPPED_OFFSET)
4847 if (subr == DIF_SUBR_INET_NTOA6)
4851 * Set v6end to skip the IPv4 address that
4852 * we have already stringified.
4858 * Build the IPv6 string by working through the
4859 * address in reverse.
4861 for (i = v6end; i >= 0; i -= 2) {
4862 ASSERT(end >= base);
4864 if (i == firstzero + numzero - 2) {
4871 if (i < 14 && i != firstzero - 2)
4875 val = (ip6._S6_un._S6_u8[i] << 8) +
4876 ip6._S6_un._S6_u8[i + 1];
4878 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4879 ip6.__u6_addr.__u6_addr8[i + 1];
4885 for (; val; val /= 16) {
4886 *end-- = digits[val % 16];
4890 ASSERT(end + 1 >= base);
4894 * The user didn't use AH_INET or AH_INET6.
4896 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4901 inetout: regs[rd] = (uintptr_t)end + 1;
4902 mstate->dtms_scratch_ptr += size;
4906 case DIF_SUBR_MEMREF: {
4907 uintptr_t size = 2 * sizeof(uintptr_t);
4908 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4909 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4911 /* address and length */
4912 memref[0] = tupregs[0].dttk_value;
4913 memref[1] = tupregs[1].dttk_value;
4915 regs[rd] = (uintptr_t) memref;
4916 mstate->dtms_scratch_ptr += scratch_size;
4920 case DIF_SUBR_TYPEREF: {
4921 uintptr_t size = 4 * sizeof(uintptr_t);
4922 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4923 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4925 /* address, num_elements, type_str, type_len */
4926 typeref[0] = tupregs[0].dttk_value;
4927 typeref[1] = tupregs[1].dttk_value;
4928 typeref[2] = tupregs[2].dttk_value;
4929 typeref[3] = tupregs[3].dttk_value;
4931 regs[rd] = (uintptr_t) typeref;
4932 mstate->dtms_scratch_ptr += scratch_size;
4939 * Emulate the execution of DTrace IR instructions specified by the given
4940 * DIF object. This function is deliberately void of assertions as all of
4941 * the necessary checks are handled by a call to dtrace_difo_validate().
4944 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4945 dtrace_vstate_t *vstate, dtrace_state_t *state)
4947 const dif_instr_t *text = difo->dtdo_buf;
4948 const uint_t textlen = difo->dtdo_len;
4949 const char *strtab = difo->dtdo_strtab;
4950 const uint64_t *inttab = difo->dtdo_inttab;
4953 dtrace_statvar_t *svar;
4954 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4956 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4957 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4959 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4960 uint64_t regs[DIF_DIR_NREGS];
4963 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4965 uint_t pc = 0, id, opc = 0;
4971 * We stash the current DIF object into the machine state: we need it
4972 * for subsequent access checking.
4974 mstate->dtms_difo = difo;
4976 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
4978 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4982 r1 = DIF_INSTR_R1(instr);
4983 r2 = DIF_INSTR_R2(instr);
4984 rd = DIF_INSTR_RD(instr);
4986 switch (DIF_INSTR_OP(instr)) {
4988 regs[rd] = regs[r1] | regs[r2];
4991 regs[rd] = regs[r1] ^ regs[r2];
4994 regs[rd] = regs[r1] & regs[r2];
4997 regs[rd] = regs[r1] << regs[r2];
5000 regs[rd] = regs[r1] >> regs[r2];
5003 regs[rd] = regs[r1] - regs[r2];
5006 regs[rd] = regs[r1] + regs[r2];
5009 regs[rd] = regs[r1] * regs[r2];
5012 if (regs[r2] == 0) {
5014 *flags |= CPU_DTRACE_DIVZERO;
5016 regs[rd] = (int64_t)regs[r1] /
5022 if (regs[r2] == 0) {
5024 *flags |= CPU_DTRACE_DIVZERO;
5026 regs[rd] = regs[r1] / regs[r2];
5031 if (regs[r2] == 0) {
5033 *flags |= CPU_DTRACE_DIVZERO;
5035 regs[rd] = (int64_t)regs[r1] %
5041 if (regs[r2] == 0) {
5043 *flags |= CPU_DTRACE_DIVZERO;
5045 regs[rd] = regs[r1] % regs[r2];
5050 regs[rd] = ~regs[r1];
5053 regs[rd] = regs[r1];
5056 cc_r = regs[r1] - regs[r2];
5060 cc_c = regs[r1] < regs[r2];
5063 cc_n = cc_v = cc_c = 0;
5064 cc_z = regs[r1] == 0;
5067 pc = DIF_INSTR_LABEL(instr);
5071 pc = DIF_INSTR_LABEL(instr);
5075 pc = DIF_INSTR_LABEL(instr);
5078 if ((cc_z | (cc_n ^ cc_v)) == 0)
5079 pc = DIF_INSTR_LABEL(instr);
5082 if ((cc_c | cc_z) == 0)
5083 pc = DIF_INSTR_LABEL(instr);
5086 if ((cc_n ^ cc_v) == 0)
5087 pc = DIF_INSTR_LABEL(instr);
5091 pc = DIF_INSTR_LABEL(instr);
5095 pc = DIF_INSTR_LABEL(instr);
5099 pc = DIF_INSTR_LABEL(instr);
5102 if (cc_z | (cc_n ^ cc_v))
5103 pc = DIF_INSTR_LABEL(instr);
5107 pc = DIF_INSTR_LABEL(instr);
5110 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5111 *flags |= CPU_DTRACE_KPRIV;
5117 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5120 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5121 *flags |= CPU_DTRACE_KPRIV;
5127 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5130 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5131 *flags |= CPU_DTRACE_KPRIV;
5137 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5140 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5141 *flags |= CPU_DTRACE_KPRIV;
5147 regs[rd] = dtrace_load8(regs[r1]);
5150 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5151 *flags |= CPU_DTRACE_KPRIV;
5157 regs[rd] = dtrace_load16(regs[r1]);
5160 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5161 *flags |= CPU_DTRACE_KPRIV;
5167 regs[rd] = dtrace_load32(regs[r1]);
5170 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5171 *flags |= CPU_DTRACE_KPRIV;
5177 regs[rd] = dtrace_load64(regs[r1]);
5181 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5184 regs[rd] = (int16_t)
5185 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5188 regs[rd] = (int32_t)
5189 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5193 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5197 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5201 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5205 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5214 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5217 regs[rd] = (uint64_t)(uintptr_t)
5218 (strtab + DIF_INSTR_STRING(instr));
5221 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5222 uintptr_t s1 = regs[r1];
5223 uintptr_t s2 = regs[r2];
5226 !dtrace_strcanload(s1, sz, mstate, vstate))
5229 !dtrace_strcanload(s2, sz, mstate, vstate))
5232 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5240 regs[rd] = dtrace_dif_variable(mstate, state,
5244 id = DIF_INSTR_VAR(instr);
5246 if (id >= DIF_VAR_OTHER_UBASE) {
5249 id -= DIF_VAR_OTHER_UBASE;
5250 svar = vstate->dtvs_globals[id];
5251 ASSERT(svar != NULL);
5252 v = &svar->dtsv_var;
5254 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5255 regs[rd] = svar->dtsv_data;
5259 a = (uintptr_t)svar->dtsv_data;
5261 if (*(uint8_t *)a == UINT8_MAX) {
5263 * If the 0th byte is set to UINT8_MAX
5264 * then this is to be treated as a
5265 * reference to a NULL variable.
5269 regs[rd] = a + sizeof (uint64_t);
5275 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5279 id = DIF_INSTR_VAR(instr);
5281 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5282 id -= DIF_VAR_OTHER_UBASE;
5284 svar = vstate->dtvs_globals[id];
5285 ASSERT(svar != NULL);
5286 v = &svar->dtsv_var;
5288 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5289 uintptr_t a = (uintptr_t)svar->dtsv_data;
5292 ASSERT(svar->dtsv_size != 0);
5294 if (regs[rd] == 0) {
5295 *(uint8_t *)a = UINT8_MAX;
5299 a += sizeof (uint64_t);
5301 if (!dtrace_vcanload(
5302 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5306 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5307 (void *)a, &v->dtdv_type);
5311 svar->dtsv_data = regs[rd];
5316 * There are no DTrace built-in thread-local arrays at
5317 * present. This opcode is saved for future work.
5319 *flags |= CPU_DTRACE_ILLOP;
5324 id = DIF_INSTR_VAR(instr);
5326 if (id < DIF_VAR_OTHER_UBASE) {
5328 * For now, this has no meaning.
5334 id -= DIF_VAR_OTHER_UBASE;
5336 ASSERT(id < vstate->dtvs_nlocals);
5337 ASSERT(vstate->dtvs_locals != NULL);
5339 svar = vstate->dtvs_locals[id];
5340 ASSERT(svar != NULL);
5341 v = &svar->dtsv_var;
5343 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5344 uintptr_t a = (uintptr_t)svar->dtsv_data;
5345 size_t sz = v->dtdv_type.dtdt_size;
5347 sz += sizeof (uint64_t);
5348 ASSERT(svar->dtsv_size == NCPU * sz);
5351 if (*(uint8_t *)a == UINT8_MAX) {
5353 * If the 0th byte is set to UINT8_MAX
5354 * then this is to be treated as a
5355 * reference to a NULL variable.
5359 regs[rd] = a + sizeof (uint64_t);
5365 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5366 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5367 regs[rd] = tmp[curcpu];
5371 id = DIF_INSTR_VAR(instr);
5373 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5374 id -= DIF_VAR_OTHER_UBASE;
5375 ASSERT(id < vstate->dtvs_nlocals);
5377 ASSERT(vstate->dtvs_locals != NULL);
5378 svar = vstate->dtvs_locals[id];
5379 ASSERT(svar != NULL);
5380 v = &svar->dtsv_var;
5382 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5383 uintptr_t a = (uintptr_t)svar->dtsv_data;
5384 size_t sz = v->dtdv_type.dtdt_size;
5386 sz += sizeof (uint64_t);
5387 ASSERT(svar->dtsv_size == NCPU * sz);
5390 if (regs[rd] == 0) {
5391 *(uint8_t *)a = UINT8_MAX;
5395 a += sizeof (uint64_t);
5398 if (!dtrace_vcanload(
5399 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5403 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5404 (void *)a, &v->dtdv_type);
5408 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5409 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5410 tmp[curcpu] = regs[rd];
5414 dtrace_dynvar_t *dvar;
5417 id = DIF_INSTR_VAR(instr);
5418 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5419 id -= DIF_VAR_OTHER_UBASE;
5420 v = &vstate->dtvs_tlocals[id];
5422 key = &tupregs[DIF_DTR_NREGS];
5423 key[0].dttk_value = (uint64_t)id;
5424 key[0].dttk_size = 0;
5425 DTRACE_TLS_THRKEY(key[1].dttk_value);
5426 key[1].dttk_size = 0;
5428 dvar = dtrace_dynvar(dstate, 2, key,
5429 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5437 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5438 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5440 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5447 dtrace_dynvar_t *dvar;
5450 id = DIF_INSTR_VAR(instr);
5451 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5452 id -= DIF_VAR_OTHER_UBASE;
5454 key = &tupregs[DIF_DTR_NREGS];
5455 key[0].dttk_value = (uint64_t)id;
5456 key[0].dttk_size = 0;
5457 DTRACE_TLS_THRKEY(key[1].dttk_value);
5458 key[1].dttk_size = 0;
5459 v = &vstate->dtvs_tlocals[id];
5461 dvar = dtrace_dynvar(dstate, 2, key,
5462 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5463 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5464 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5465 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5468 * Given that we're storing to thread-local data,
5469 * we need to flush our predicate cache.
5471 curthread->t_predcache = 0;
5476 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5477 if (!dtrace_vcanload(
5478 (void *)(uintptr_t)regs[rd],
5479 &v->dtdv_type, mstate, vstate))
5482 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5483 dvar->dtdv_data, &v->dtdv_type);
5485 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5492 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5496 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5497 regs, tupregs, ttop, mstate, state);
5501 if (ttop == DIF_DTR_NREGS) {
5502 *flags |= CPU_DTRACE_TUPOFLOW;
5506 if (r1 == DIF_TYPE_STRING) {
5508 * If this is a string type and the size is 0,
5509 * we'll use the system-wide default string
5510 * size. Note that we are _not_ looking at
5511 * the value of the DTRACEOPT_STRSIZE option;
5512 * had this been set, we would expect to have
5513 * a non-zero size value in the "pushtr".
5515 tupregs[ttop].dttk_size =
5516 dtrace_strlen((char *)(uintptr_t)regs[rd],
5517 regs[r2] ? regs[r2] :
5518 dtrace_strsize_default) + 1;
5520 tupregs[ttop].dttk_size = regs[r2];
5523 tupregs[ttop++].dttk_value = regs[rd];
5527 if (ttop == DIF_DTR_NREGS) {
5528 *flags |= CPU_DTRACE_TUPOFLOW;
5532 tupregs[ttop].dttk_value = regs[rd];
5533 tupregs[ttop++].dttk_size = 0;
5541 case DIF_OP_FLUSHTS:
5546 case DIF_OP_LDTAA: {
5547 dtrace_dynvar_t *dvar;
5548 dtrace_key_t *key = tupregs;
5549 uint_t nkeys = ttop;
5551 id = DIF_INSTR_VAR(instr);
5552 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5553 id -= DIF_VAR_OTHER_UBASE;
5555 key[nkeys].dttk_value = (uint64_t)id;
5556 key[nkeys++].dttk_size = 0;
5558 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5559 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5560 key[nkeys++].dttk_size = 0;
5561 v = &vstate->dtvs_tlocals[id];
5563 v = &vstate->dtvs_globals[id]->dtsv_var;
5566 dvar = dtrace_dynvar(dstate, nkeys, key,
5567 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5568 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5569 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5576 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5577 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5579 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5586 case DIF_OP_STTAA: {
5587 dtrace_dynvar_t *dvar;
5588 dtrace_key_t *key = tupregs;
5589 uint_t nkeys = ttop;
5591 id = DIF_INSTR_VAR(instr);
5592 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5593 id -= DIF_VAR_OTHER_UBASE;
5595 key[nkeys].dttk_value = (uint64_t)id;
5596 key[nkeys++].dttk_size = 0;
5598 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5599 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5600 key[nkeys++].dttk_size = 0;
5601 v = &vstate->dtvs_tlocals[id];
5603 v = &vstate->dtvs_globals[id]->dtsv_var;
5606 dvar = dtrace_dynvar(dstate, nkeys, key,
5607 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5608 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5609 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5610 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5615 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5616 if (!dtrace_vcanload(
5617 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5621 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5622 dvar->dtdv_data, &v->dtdv_type);
5624 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5630 case DIF_OP_ALLOCS: {
5631 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5632 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5635 * Rounding up the user allocation size could have
5636 * overflowed large, bogus allocations (like -1ULL) to
5639 if (size < regs[r1] ||
5640 !DTRACE_INSCRATCH(mstate, size)) {
5641 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5646 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5647 mstate->dtms_scratch_ptr += size;
5653 if (!dtrace_canstore(regs[rd], regs[r2],
5655 *flags |= CPU_DTRACE_BADADDR;
5660 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5663 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5664 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5668 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5669 *flags |= CPU_DTRACE_BADADDR;
5673 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5677 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5678 *flags |= CPU_DTRACE_BADADDR;
5683 *flags |= CPU_DTRACE_BADALIGN;
5687 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5691 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5692 *flags |= CPU_DTRACE_BADADDR;
5697 *flags |= CPU_DTRACE_BADALIGN;
5701 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5705 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5706 *flags |= CPU_DTRACE_BADADDR;
5711 *flags |= CPU_DTRACE_BADALIGN;
5715 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5720 if (!(*flags & CPU_DTRACE_FAULT))
5723 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5724 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5730 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5732 dtrace_probe_t *probe = ecb->dte_probe;
5733 dtrace_provider_t *prov = probe->dtpr_provider;
5734 char c[DTRACE_FULLNAMELEN + 80], *str;
5735 char *msg = "dtrace: breakpoint action at probe ";
5736 char *ecbmsg = " (ecb ";
5737 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5738 uintptr_t val = (uintptr_t)ecb;
5739 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5741 if (dtrace_destructive_disallow)
5745 * It's impossible to be taking action on the NULL probe.
5747 ASSERT(probe != NULL);
5750 * This is a poor man's (destitute man's?) sprintf(): we want to
5751 * print the provider name, module name, function name and name of
5752 * the probe, along with the hex address of the ECB with the breakpoint
5753 * action -- all of which we must place in the character buffer by
5756 while (*msg != '\0')
5759 for (str = prov->dtpv_name; *str != '\0'; str++)
5763 for (str = probe->dtpr_mod; *str != '\0'; str++)
5767 for (str = probe->dtpr_func; *str != '\0'; str++)
5771 for (str = probe->dtpr_name; *str != '\0'; str++)
5774 while (*ecbmsg != '\0')
5777 while (shift >= 0) {
5778 mask = (uintptr_t)0xf << shift;
5780 if (val >= ((uintptr_t)1 << shift))
5781 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5791 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5796 dtrace_action_panic(dtrace_ecb_t *ecb)
5798 dtrace_probe_t *probe = ecb->dte_probe;
5801 * It's impossible to be taking action on the NULL probe.
5803 ASSERT(probe != NULL);
5805 if (dtrace_destructive_disallow)
5808 if (dtrace_panicked != NULL)
5811 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5815 * We won the right to panic. (We want to be sure that only one
5816 * thread calls panic() from dtrace_probe(), and that panic() is
5817 * called exactly once.)
5819 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5820 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5821 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5825 dtrace_action_raise(uint64_t sig)
5827 if (dtrace_destructive_disallow)
5831 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5837 * raise() has a queue depth of 1 -- we ignore all subsequent
5838 * invocations of the raise() action.
5840 if (curthread->t_dtrace_sig == 0)
5841 curthread->t_dtrace_sig = (uint8_t)sig;
5843 curthread->t_sig_check = 1;
5846 struct proc *p = curproc;
5848 kern_psignal(p, sig);
5854 dtrace_action_stop(void)
5856 if (dtrace_destructive_disallow)
5860 if (!curthread->t_dtrace_stop) {
5861 curthread->t_dtrace_stop = 1;
5862 curthread->t_sig_check = 1;
5866 struct proc *p = curproc;
5868 kern_psignal(p, SIGSTOP);
5874 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5877 volatile uint16_t *flags;
5881 cpu_t *cpu = &solaris_cpu[curcpu];
5884 if (dtrace_destructive_disallow)
5887 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5889 now = dtrace_gethrtime();
5891 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5893 * We need to advance the mark to the current time.
5895 cpu->cpu_dtrace_chillmark = now;
5896 cpu->cpu_dtrace_chilled = 0;
5900 * Now check to see if the requested chill time would take us over
5901 * the maximum amount of time allowed in the chill interval. (Or
5902 * worse, if the calculation itself induces overflow.)
5904 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5905 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5906 *flags |= CPU_DTRACE_ILLOP;
5910 while (dtrace_gethrtime() - now < val)
5914 * Normally, we assure that the value of the variable "timestamp" does
5915 * not change within an ECB. The presence of chill() represents an
5916 * exception to this rule, however.
5918 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5919 cpu->cpu_dtrace_chilled += val;
5923 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5924 uint64_t *buf, uint64_t arg)
5926 int nframes = DTRACE_USTACK_NFRAMES(arg);
5927 int strsize = DTRACE_USTACK_STRSIZE(arg);
5928 uint64_t *pcs = &buf[1], *fps;
5929 char *str = (char *)&pcs[nframes];
5930 int size, offs = 0, i, j;
5931 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5932 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5936 * Should be taking a faster path if string space has not been
5939 ASSERT(strsize != 0);
5942 * We will first allocate some temporary space for the frame pointers.
5944 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5945 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5946 (nframes * sizeof (uint64_t));
5948 if (!DTRACE_INSCRATCH(mstate, size)) {
5950 * Not enough room for our frame pointers -- need to indicate
5951 * that we ran out of scratch space.
5953 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5957 mstate->dtms_scratch_ptr += size;
5958 saved = mstate->dtms_scratch_ptr;
5961 * Now get a stack with both program counters and frame pointers.
5963 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5964 dtrace_getufpstack(buf, fps, nframes + 1);
5965 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5968 * If that faulted, we're cooked.
5970 if (*flags & CPU_DTRACE_FAULT)
5974 * Now we want to walk up the stack, calling the USTACK helper. For
5975 * each iteration, we restore the scratch pointer.
5977 for (i = 0; i < nframes; i++) {
5978 mstate->dtms_scratch_ptr = saved;
5980 if (offs >= strsize)
5983 sym = (char *)(uintptr_t)dtrace_helper(
5984 DTRACE_HELPER_ACTION_USTACK,
5985 mstate, state, pcs[i], fps[i]);
5988 * If we faulted while running the helper, we're going to
5989 * clear the fault and null out the corresponding string.
5991 if (*flags & CPU_DTRACE_FAULT) {
5992 *flags &= ~CPU_DTRACE_FAULT;
6002 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6005 * Now copy in the string that the helper returned to us.
6007 for (j = 0; offs + j < strsize; j++) {
6008 if ((str[offs + j] = sym[j]) == '\0')
6012 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6017 if (offs >= strsize) {
6019 * If we didn't have room for all of the strings, we don't
6020 * abort processing -- this needn't be a fatal error -- but we
6021 * still want to increment a counter (dts_stkstroverflows) to
6022 * allow this condition to be warned about. (If this is from
6023 * a jstack() action, it is easily tuned via jstackstrsize.)
6025 dtrace_error(&state->dts_stkstroverflows);
6028 while (offs < strsize)
6032 mstate->dtms_scratch_ptr = old;
6036 * If you're looking for the epicenter of DTrace, you just found it. This
6037 * is the function called by the provider to fire a probe -- from which all
6038 * subsequent probe-context DTrace activity emanates.
6041 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6042 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6044 processorid_t cpuid;
6045 dtrace_icookie_t cookie;
6046 dtrace_probe_t *probe;
6047 dtrace_mstate_t mstate;
6049 dtrace_action_t *act;
6053 volatile uint16_t *flags;
6056 if (panicstr != NULL)
6061 * Kick out immediately if this CPU is still being born (in which case
6062 * curthread will be set to -1) or the current thread can't allow
6063 * probes in its current context.
6065 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6069 cookie = dtrace_interrupt_disable();
6070 probe = dtrace_probes[id - 1];
6072 onintr = CPU_ON_INTR(CPU);
6074 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6075 probe->dtpr_predcache == curthread->t_predcache) {
6077 * We have hit in the predicate cache; we know that
6078 * this predicate would evaluate to be false.
6080 dtrace_interrupt_enable(cookie);
6085 if (panic_quiesce) {
6087 if (panicstr != NULL) {
6090 * We don't trace anything if we're panicking.
6092 dtrace_interrupt_enable(cookie);
6096 now = dtrace_gethrtime();
6097 vtime = dtrace_vtime_references != 0;
6099 if (vtime && curthread->t_dtrace_start)
6100 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6102 mstate.dtms_difo = NULL;
6103 mstate.dtms_probe = probe;
6104 mstate.dtms_strtok = 0;
6105 mstate.dtms_arg[0] = arg0;
6106 mstate.dtms_arg[1] = arg1;
6107 mstate.dtms_arg[2] = arg2;
6108 mstate.dtms_arg[3] = arg3;
6109 mstate.dtms_arg[4] = arg4;
6111 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6113 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6114 dtrace_predicate_t *pred = ecb->dte_predicate;
6115 dtrace_state_t *state = ecb->dte_state;
6116 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6117 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6118 dtrace_vstate_t *vstate = &state->dts_vstate;
6119 dtrace_provider_t *prov = probe->dtpr_provider;
6120 uint64_t tracememsize = 0;
6125 * A little subtlety with the following (seemingly innocuous)
6126 * declaration of the automatic 'val': by looking at the
6127 * code, you might think that it could be declared in the
6128 * action processing loop, below. (That is, it's only used in
6129 * the action processing loop.) However, it must be declared
6130 * out of that scope because in the case of DIF expression
6131 * arguments to aggregating actions, one iteration of the
6132 * action loop will use the last iteration's value.
6136 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6137 *flags &= ~CPU_DTRACE_ERROR;
6139 if (prov == dtrace_provider) {
6141 * If dtrace itself is the provider of this probe,
6142 * we're only going to continue processing the ECB if
6143 * arg0 (the dtrace_state_t) is equal to the ECB's
6144 * creating state. (This prevents disjoint consumers
6145 * from seeing one another's metaprobes.)
6147 if (arg0 != (uint64_t)(uintptr_t)state)
6151 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6153 * We're not currently active. If our provider isn't
6154 * the dtrace pseudo provider, we're not interested.
6156 if (prov != dtrace_provider)
6160 * Now we must further check if we are in the BEGIN
6161 * probe. If we are, we will only continue processing
6162 * if we're still in WARMUP -- if one BEGIN enabling
6163 * has invoked the exit() action, we don't want to
6164 * evaluate subsequent BEGIN enablings.
6166 if (probe->dtpr_id == dtrace_probeid_begin &&
6167 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6168 ASSERT(state->dts_activity ==
6169 DTRACE_ACTIVITY_DRAINING);
6174 if (ecb->dte_cond) {
6176 * If the dte_cond bits indicate that this
6177 * consumer is only allowed to see user-mode firings
6178 * of this probe, call the provider's dtps_usermode()
6179 * entry point to check that the probe was fired
6180 * while in a user context. Skip this ECB if that's
6183 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6184 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6185 probe->dtpr_id, probe->dtpr_arg) == 0)
6190 * This is more subtle than it looks. We have to be
6191 * absolutely certain that CRED() isn't going to
6192 * change out from under us so it's only legit to
6193 * examine that structure if we're in constrained
6194 * situations. Currently, the only times we'll this
6195 * check is if a non-super-user has enabled the
6196 * profile or syscall providers -- providers that
6197 * allow visibility of all processes. For the
6198 * profile case, the check above will ensure that
6199 * we're examining a user context.
6201 if (ecb->dte_cond & DTRACE_COND_OWNER) {
6204 ecb->dte_state->dts_cred.dcr_cred;
6207 ASSERT(s_cr != NULL);
6209 if ((cr = CRED()) == NULL ||
6210 s_cr->cr_uid != cr->cr_uid ||
6211 s_cr->cr_uid != cr->cr_ruid ||
6212 s_cr->cr_uid != cr->cr_suid ||
6213 s_cr->cr_gid != cr->cr_gid ||
6214 s_cr->cr_gid != cr->cr_rgid ||
6215 s_cr->cr_gid != cr->cr_sgid ||
6216 (proc = ttoproc(curthread)) == NULL ||
6217 (proc->p_flag & SNOCD))
6221 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6224 ecb->dte_state->dts_cred.dcr_cred;
6226 ASSERT(s_cr != NULL);
6228 if ((cr = CRED()) == NULL ||
6229 s_cr->cr_zone->zone_id !=
6230 cr->cr_zone->zone_id)
6236 if (now - state->dts_alive > dtrace_deadman_timeout) {
6238 * We seem to be dead. Unless we (a) have kernel
6239 * destructive permissions (b) have explicitly enabled
6240 * destructive actions and (c) destructive actions have
6241 * not been disabled, we're going to transition into
6242 * the KILLED state, from which no further processing
6243 * on this state will be performed.
6245 if (!dtrace_priv_kernel_destructive(state) ||
6246 !state->dts_cred.dcr_destructive ||
6247 dtrace_destructive_disallow) {
6248 void *activity = &state->dts_activity;
6249 dtrace_activity_t current;
6252 current = state->dts_activity;
6253 } while (dtrace_cas32(activity, current,
6254 DTRACE_ACTIVITY_KILLED) != current);
6260 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6261 ecb->dte_alignment, state, &mstate)) < 0)
6264 tomax = buf->dtb_tomax;
6265 ASSERT(tomax != NULL);
6267 if (ecb->dte_size != 0) {
6268 dtrace_rechdr_t dtrh;
6269 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6270 mstate.dtms_timestamp = dtrace_gethrtime();
6271 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6273 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6274 dtrh.dtrh_epid = ecb->dte_epid;
6275 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6276 mstate.dtms_timestamp);
6277 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6280 mstate.dtms_epid = ecb->dte_epid;
6281 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6283 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6284 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6286 mstate.dtms_access = 0;
6289 dtrace_difo_t *dp = pred->dtp_difo;
6292 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6294 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6295 dtrace_cacheid_t cid = probe->dtpr_predcache;
6297 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6299 * Update the predicate cache...
6301 ASSERT(cid == pred->dtp_cacheid);
6302 curthread->t_predcache = cid;
6309 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6310 act != NULL; act = act->dta_next) {
6313 dtrace_recdesc_t *rec = &act->dta_rec;
6315 size = rec->dtrd_size;
6316 valoffs = offs + rec->dtrd_offset;
6318 if (DTRACEACT_ISAGG(act->dta_kind)) {
6320 dtrace_aggregation_t *agg;
6322 agg = (dtrace_aggregation_t *)act;
6324 if ((dp = act->dta_difo) != NULL)
6325 v = dtrace_dif_emulate(dp,
6326 &mstate, vstate, state);
6328 if (*flags & CPU_DTRACE_ERROR)
6332 * Note that we always pass the expression
6333 * value from the previous iteration of the
6334 * action loop. This value will only be used
6335 * if there is an expression argument to the
6336 * aggregating action, denoted by the
6337 * dtag_hasarg field.
6339 dtrace_aggregate(agg, buf,
6340 offs, aggbuf, v, val);
6344 switch (act->dta_kind) {
6345 case DTRACEACT_STOP:
6346 if (dtrace_priv_proc_destructive(state))
6347 dtrace_action_stop();
6350 case DTRACEACT_BREAKPOINT:
6351 if (dtrace_priv_kernel_destructive(state))
6352 dtrace_action_breakpoint(ecb);
6355 case DTRACEACT_PANIC:
6356 if (dtrace_priv_kernel_destructive(state))
6357 dtrace_action_panic(ecb);
6360 case DTRACEACT_STACK:
6361 if (!dtrace_priv_kernel(state))
6364 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6365 size / sizeof (pc_t), probe->dtpr_aframes,
6366 DTRACE_ANCHORED(probe) ? NULL :
6370 case DTRACEACT_JSTACK:
6371 case DTRACEACT_USTACK:
6372 if (!dtrace_priv_proc(state))
6376 * See comment in DIF_VAR_PID.
6378 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6380 int depth = DTRACE_USTACK_NFRAMES(
6383 dtrace_bzero((void *)(tomax + valoffs),
6384 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6385 + depth * sizeof (uint64_t));
6390 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6391 curproc->p_dtrace_helpers != NULL) {
6393 * This is the slow path -- we have
6394 * allocated string space, and we're
6395 * getting the stack of a process that
6396 * has helpers. Call into a separate
6397 * routine to perform this processing.
6399 dtrace_action_ustack(&mstate, state,
6400 (uint64_t *)(tomax + valoffs),
6405 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6406 dtrace_getupcstack((uint64_t *)
6408 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6409 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6419 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6421 if (*flags & CPU_DTRACE_ERROR)
6424 switch (act->dta_kind) {
6425 case DTRACEACT_SPECULATE: {
6426 dtrace_rechdr_t *dtrh;
6428 ASSERT(buf == &state->dts_buffer[cpuid]);
6429 buf = dtrace_speculation_buffer(state,
6433 *flags |= CPU_DTRACE_DROP;
6437 offs = dtrace_buffer_reserve(buf,
6438 ecb->dte_needed, ecb->dte_alignment,
6442 *flags |= CPU_DTRACE_DROP;
6446 tomax = buf->dtb_tomax;
6447 ASSERT(tomax != NULL);
6449 if (ecb->dte_size == 0)
6452 ASSERT3U(ecb->dte_size, >=,
6453 sizeof (dtrace_rechdr_t));
6454 dtrh = ((void *)(tomax + offs));
6455 dtrh->dtrh_epid = ecb->dte_epid;
6457 * When the speculation is committed, all of
6458 * the records in the speculative buffer will
6459 * have their timestamps set to the commit
6460 * time. Until then, it is set to a sentinel
6461 * value, for debugability.
6463 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6467 case DTRACEACT_PRINTM: {
6468 /* The DIF returns a 'memref'. */
6469 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6471 /* Get the size from the memref. */
6475 * Check if the size exceeds the allocated
6478 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6480 *flags |= CPU_DTRACE_DROP;
6484 /* Store the size in the buffer first. */
6485 DTRACE_STORE(uintptr_t, tomax,
6489 * Offset the buffer address to the start
6492 valoffs += sizeof(uintptr_t);
6495 * Reset to the memory address rather than
6496 * the memref array, then let the BYREF
6497 * code below do the work to store the
6498 * memory data in the buffer.
6504 case DTRACEACT_PRINTT: {
6505 /* The DIF returns a 'typeref'. */
6506 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6511 * Get the type string length and round it
6512 * up so that the data that follows is
6513 * aligned for easy access.
6515 size_t typs = strlen((char *) typeref[2]) + 1;
6516 typs = roundup(typs, sizeof(uintptr_t));
6519 *Get the size from the typeref using the
6520 * number of elements and the type size.
6522 size = typeref[1] * typeref[3];
6525 * Check if the size exceeds the allocated
6528 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6530 *flags |= CPU_DTRACE_DROP;
6534 /* Store the size in the buffer first. */
6535 DTRACE_STORE(uintptr_t, tomax,
6537 valoffs += sizeof(uintptr_t);
6539 /* Store the type size in the buffer. */
6540 DTRACE_STORE(uintptr_t, tomax,
6541 valoffs, typeref[3]);
6542 valoffs += sizeof(uintptr_t);
6546 for (s = 0; s < typs; s++) {
6548 c = dtrace_load8(val++);
6550 DTRACE_STORE(uint8_t, tomax,
6555 * Reset to the memory address rather than
6556 * the typeref array, then let the BYREF
6557 * code below do the work to store the
6558 * memory data in the buffer.
6564 case DTRACEACT_CHILL:
6565 if (dtrace_priv_kernel_destructive(state))
6566 dtrace_action_chill(&mstate, val);
6569 case DTRACEACT_RAISE:
6570 if (dtrace_priv_proc_destructive(state))
6571 dtrace_action_raise(val);
6574 case DTRACEACT_COMMIT:
6578 * We need to commit our buffer state.
6581 buf->dtb_offset = offs + ecb->dte_size;
6582 buf = &state->dts_buffer[cpuid];
6583 dtrace_speculation_commit(state, cpuid, val);
6587 case DTRACEACT_DISCARD:
6588 dtrace_speculation_discard(state, cpuid, val);
6591 case DTRACEACT_DIFEXPR:
6592 case DTRACEACT_LIBACT:
6593 case DTRACEACT_PRINTF:
6594 case DTRACEACT_PRINTA:
6595 case DTRACEACT_SYSTEM:
6596 case DTRACEACT_FREOPEN:
6597 case DTRACEACT_TRACEMEM:
6600 case DTRACEACT_TRACEMEM_DYNSIZE:
6606 if (!dtrace_priv_kernel(state))
6610 case DTRACEACT_USYM:
6611 case DTRACEACT_UMOD:
6612 case DTRACEACT_UADDR: {
6614 struct pid *pid = curthread->t_procp->p_pidp;
6617 if (!dtrace_priv_proc(state))
6620 DTRACE_STORE(uint64_t, tomax,
6622 valoffs, (uint64_t)pid->pid_id);
6624 valoffs, (uint64_t) curproc->p_pid);
6626 DTRACE_STORE(uint64_t, tomax,
6627 valoffs + sizeof (uint64_t), val);
6632 case DTRACEACT_EXIT: {
6634 * For the exit action, we are going to attempt
6635 * to atomically set our activity to be
6636 * draining. If this fails (either because
6637 * another CPU has beat us to the exit action,
6638 * or because our current activity is something
6639 * other than ACTIVE or WARMUP), we will
6640 * continue. This assures that the exit action
6641 * can be successfully recorded at most once
6642 * when we're in the ACTIVE state. If we're
6643 * encountering the exit() action while in
6644 * COOLDOWN, however, we want to honor the new
6645 * status code. (We know that we're the only
6646 * thread in COOLDOWN, so there is no race.)
6648 void *activity = &state->dts_activity;
6649 dtrace_activity_t current = state->dts_activity;
6651 if (current == DTRACE_ACTIVITY_COOLDOWN)
6654 if (current != DTRACE_ACTIVITY_WARMUP)
6655 current = DTRACE_ACTIVITY_ACTIVE;
6657 if (dtrace_cas32(activity, current,
6658 DTRACE_ACTIVITY_DRAINING) != current) {
6659 *flags |= CPU_DTRACE_DROP;
6670 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6671 uintptr_t end = valoffs + size;
6673 if (tracememsize != 0 &&
6674 valoffs + tracememsize < end) {
6675 end = valoffs + tracememsize;
6679 if (!dtrace_vcanload((void *)(uintptr_t)val,
6680 &dp->dtdo_rtype, &mstate, vstate))
6684 * If this is a string, we're going to only
6685 * load until we find the zero byte -- after
6686 * which we'll store zero bytes.
6688 if (dp->dtdo_rtype.dtdt_kind ==
6691 int intuple = act->dta_intuple;
6694 for (s = 0; s < size; s++) {
6696 c = dtrace_load8(val++);
6698 DTRACE_STORE(uint8_t, tomax,
6701 if (c == '\0' && intuple)
6708 while (valoffs < end) {
6709 DTRACE_STORE(uint8_t, tomax, valoffs++,
6710 dtrace_load8(val++));
6720 case sizeof (uint8_t):
6721 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6723 case sizeof (uint16_t):
6724 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6726 case sizeof (uint32_t):
6727 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6729 case sizeof (uint64_t):
6730 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6734 * Any other size should have been returned by
6735 * reference, not by value.
6742 if (*flags & CPU_DTRACE_DROP)
6745 if (*flags & CPU_DTRACE_FAULT) {
6747 dtrace_action_t *err;
6751 if (probe->dtpr_id == dtrace_probeid_error) {
6753 * There's nothing we can do -- we had an
6754 * error on the error probe. We bump an
6755 * error counter to at least indicate that
6756 * this condition happened.
6758 dtrace_error(&state->dts_dblerrors);
6764 * Before recursing on dtrace_probe(), we
6765 * need to explicitly clear out our start
6766 * time to prevent it from being accumulated
6767 * into t_dtrace_vtime.
6769 curthread->t_dtrace_start = 0;
6773 * Iterate over the actions to figure out which action
6774 * we were processing when we experienced the error.
6775 * Note that act points _past_ the faulting action; if
6776 * act is ecb->dte_action, the fault was in the
6777 * predicate, if it's ecb->dte_action->dta_next it's
6778 * in action #1, and so on.
6780 for (err = ecb->dte_action, ndx = 0;
6781 err != act; err = err->dta_next, ndx++)
6784 dtrace_probe_error(state, ecb->dte_epid, ndx,
6785 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6786 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6787 cpu_core[cpuid].cpuc_dtrace_illval);
6793 buf->dtb_offset = offs + ecb->dte_size;
6797 curthread->t_dtrace_start = dtrace_gethrtime();
6799 dtrace_interrupt_enable(cookie);
6803 * DTrace Probe Hashing Functions
6805 * The functions in this section (and indeed, the functions in remaining
6806 * sections) are not _called_ from probe context. (Any exceptions to this are
6807 * marked with a "Note:".) Rather, they are called from elsewhere in the
6808 * DTrace framework to look-up probes in, add probes to and remove probes from
6809 * the DTrace probe hashes. (Each probe is hashed by each element of the
6810 * probe tuple -- allowing for fast lookups, regardless of what was
6814 dtrace_hash_str(const char *p)
6820 hval = (hval << 4) + *p++;
6821 if ((g = (hval & 0xf0000000)) != 0)
6828 static dtrace_hash_t *
6829 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6831 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6833 hash->dth_stroffs = stroffs;
6834 hash->dth_nextoffs = nextoffs;
6835 hash->dth_prevoffs = prevoffs;
6838 hash->dth_mask = hash->dth_size - 1;
6840 hash->dth_tab = kmem_zalloc(hash->dth_size *
6841 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6847 dtrace_hash_destroy(dtrace_hash_t *hash)
6852 for (i = 0; i < hash->dth_size; i++)
6853 ASSERT(hash->dth_tab[i] == NULL);
6856 kmem_free(hash->dth_tab,
6857 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6858 kmem_free(hash, sizeof (dtrace_hash_t));
6862 dtrace_hash_resize(dtrace_hash_t *hash)
6864 int size = hash->dth_size, i, ndx;
6865 int new_size = hash->dth_size << 1;
6866 int new_mask = new_size - 1;
6867 dtrace_hashbucket_t **new_tab, *bucket, *next;
6869 ASSERT((new_size & new_mask) == 0);
6871 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6873 for (i = 0; i < size; i++) {
6874 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6875 dtrace_probe_t *probe = bucket->dthb_chain;
6877 ASSERT(probe != NULL);
6878 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6880 next = bucket->dthb_next;
6881 bucket->dthb_next = new_tab[ndx];
6882 new_tab[ndx] = bucket;
6886 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6887 hash->dth_tab = new_tab;
6888 hash->dth_size = new_size;
6889 hash->dth_mask = new_mask;
6893 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6895 int hashval = DTRACE_HASHSTR(hash, new);
6896 int ndx = hashval & hash->dth_mask;
6897 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6898 dtrace_probe_t **nextp, **prevp;
6900 for (; bucket != NULL; bucket = bucket->dthb_next) {
6901 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6905 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6906 dtrace_hash_resize(hash);
6907 dtrace_hash_add(hash, new);
6911 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6912 bucket->dthb_next = hash->dth_tab[ndx];
6913 hash->dth_tab[ndx] = bucket;
6914 hash->dth_nbuckets++;
6917 nextp = DTRACE_HASHNEXT(hash, new);
6918 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6919 *nextp = bucket->dthb_chain;
6921 if (bucket->dthb_chain != NULL) {
6922 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6923 ASSERT(*prevp == NULL);
6927 bucket->dthb_chain = new;
6931 static dtrace_probe_t *
6932 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6934 int hashval = DTRACE_HASHSTR(hash, template);
6935 int ndx = hashval & hash->dth_mask;
6936 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6938 for (; bucket != NULL; bucket = bucket->dthb_next) {
6939 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6940 return (bucket->dthb_chain);
6947 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6949 int hashval = DTRACE_HASHSTR(hash, template);
6950 int ndx = hashval & hash->dth_mask;
6951 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6953 for (; bucket != NULL; bucket = bucket->dthb_next) {
6954 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6955 return (bucket->dthb_len);
6962 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6964 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6965 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6967 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6968 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6971 * Find the bucket that we're removing this probe from.
6973 for (; bucket != NULL; bucket = bucket->dthb_next) {
6974 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6978 ASSERT(bucket != NULL);
6980 if (*prevp == NULL) {
6981 if (*nextp == NULL) {
6983 * The removed probe was the only probe on this
6984 * bucket; we need to remove the bucket.
6986 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6988 ASSERT(bucket->dthb_chain == probe);
6992 hash->dth_tab[ndx] = bucket->dthb_next;
6994 while (b->dthb_next != bucket)
6996 b->dthb_next = bucket->dthb_next;
6999 ASSERT(hash->dth_nbuckets > 0);
7000 hash->dth_nbuckets--;
7001 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7005 bucket->dthb_chain = *nextp;
7007 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7011 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7015 * DTrace Utility Functions
7017 * These are random utility functions that are _not_ called from probe context.
7020 dtrace_badattr(const dtrace_attribute_t *a)
7022 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7023 a->dtat_data > DTRACE_STABILITY_MAX ||
7024 a->dtat_class > DTRACE_CLASS_MAX);
7028 * Return a duplicate copy of a string. If the specified string is NULL,
7029 * this function returns a zero-length string.
7032 dtrace_strdup(const char *str)
7034 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7037 (void) strcpy(new, str);
7042 #define DTRACE_ISALPHA(c) \
7043 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7046 dtrace_badname(const char *s)
7050 if (s == NULL || (c = *s++) == '\0')
7053 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7056 while ((c = *s++) != '\0') {
7057 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7058 c != '-' && c != '_' && c != '.' && c != '`')
7066 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7071 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7073 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7075 priv = DTRACE_PRIV_ALL;
7077 *uidp = crgetuid(cr);
7078 *zoneidp = crgetzoneid(cr);
7081 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7082 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7083 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7084 priv |= DTRACE_PRIV_USER;
7085 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7086 priv |= DTRACE_PRIV_PROC;
7087 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7088 priv |= DTRACE_PRIV_OWNER;
7089 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7090 priv |= DTRACE_PRIV_ZONEOWNER;
7093 priv = DTRACE_PRIV_ALL;
7099 #ifdef DTRACE_ERRDEBUG
7101 dtrace_errdebug(const char *str)
7103 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
7106 mutex_enter(&dtrace_errlock);
7107 dtrace_errlast = str;
7108 dtrace_errthread = curthread;
7110 while (occupied++ < DTRACE_ERRHASHSZ) {
7111 if (dtrace_errhash[hval].dter_msg == str) {
7112 dtrace_errhash[hval].dter_count++;
7116 if (dtrace_errhash[hval].dter_msg != NULL) {
7117 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7121 dtrace_errhash[hval].dter_msg = str;
7122 dtrace_errhash[hval].dter_count = 1;
7126 panic("dtrace: undersized error hash");
7128 mutex_exit(&dtrace_errlock);
7133 * DTrace Matching Functions
7135 * These functions are used to match groups of probes, given some elements of
7136 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7139 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7142 if (priv != DTRACE_PRIV_ALL) {
7143 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7144 uint32_t match = priv & ppriv;
7147 * No PRIV_DTRACE_* privileges...
7149 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7150 DTRACE_PRIV_KERNEL)) == 0)
7154 * No matching bits, but there were bits to match...
7156 if (match == 0 && ppriv != 0)
7160 * Need to have permissions to the process, but don't...
7162 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7163 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7168 * Need to be in the same zone unless we possess the
7169 * privilege to examine all zones.
7171 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7172 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7181 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7182 * consists of input pattern strings and an ops-vector to evaluate them.
7183 * This function returns >0 for match, 0 for no match, and <0 for error.
7186 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7187 uint32_t priv, uid_t uid, zoneid_t zoneid)
7189 dtrace_provider_t *pvp = prp->dtpr_provider;
7192 if (pvp->dtpv_defunct)
7195 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7198 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7201 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7204 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7207 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7214 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7215 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7216 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7217 * In addition, all of the recursion cases except for '*' matching have been
7218 * unwound. For '*', we still implement recursive evaluation, but a depth
7219 * counter is maintained and matching is aborted if we recurse too deep.
7220 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7223 dtrace_match_glob(const char *s, const char *p, int depth)
7229 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7233 s = ""; /* treat NULL as empty string */
7242 if ((c = *p++) == '\0')
7243 return (s1 == '\0');
7247 int ok = 0, notflag = 0;
7258 if ((c = *p++) == '\0')
7262 if (c == '-' && lc != '\0' && *p != ']') {
7263 if ((c = *p++) == '\0')
7265 if (c == '\\' && (c = *p++) == '\0')
7269 if (s1 < lc || s1 > c)
7273 } else if (lc <= s1 && s1 <= c)
7276 } else if (c == '\\' && (c = *p++) == '\0')
7279 lc = c; /* save left-hand 'c' for next iteration */
7289 if ((c = *p++) == '\0')
7301 if ((c = *p++) == '\0')
7317 p++; /* consecutive *'s are identical to a single one */
7322 for (s = olds; *s != '\0'; s++) {
7323 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7333 dtrace_match_string(const char *s, const char *p, int depth)
7335 return (s != NULL && strcmp(s, p) == 0);
7340 dtrace_match_nul(const char *s, const char *p, int depth)
7342 return (1); /* always match the empty pattern */
7347 dtrace_match_nonzero(const char *s, const char *p, int depth)
7349 return (s != NULL && s[0] != '\0');
7353 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7354 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7356 dtrace_probe_t template, *probe;
7357 dtrace_hash_t *hash = NULL;
7358 int len, best = INT_MAX, nmatched = 0;
7361 ASSERT(MUTEX_HELD(&dtrace_lock));
7364 * If the probe ID is specified in the key, just lookup by ID and
7365 * invoke the match callback once if a matching probe is found.
7367 if (pkp->dtpk_id != DTRACE_IDNONE) {
7368 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7369 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7370 (void) (*matched)(probe, arg);
7376 template.dtpr_mod = (char *)pkp->dtpk_mod;
7377 template.dtpr_func = (char *)pkp->dtpk_func;
7378 template.dtpr_name = (char *)pkp->dtpk_name;
7381 * We want to find the most distinct of the module name, function
7382 * name, and name. So for each one that is not a glob pattern or
7383 * empty string, we perform a lookup in the corresponding hash and
7384 * use the hash table with the fewest collisions to do our search.
7386 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7387 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7389 hash = dtrace_bymod;
7392 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7393 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7395 hash = dtrace_byfunc;
7398 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7399 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7401 hash = dtrace_byname;
7405 * If we did not select a hash table, iterate over every probe and
7406 * invoke our callback for each one that matches our input probe key.
7409 for (i = 0; i < dtrace_nprobes; i++) {
7410 if ((probe = dtrace_probes[i]) == NULL ||
7411 dtrace_match_probe(probe, pkp, priv, uid,
7417 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7425 * If we selected a hash table, iterate over each probe of the same key
7426 * name and invoke the callback for every probe that matches the other
7427 * attributes of our input probe key.
7429 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7430 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7432 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7437 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7445 * Return the function pointer dtrace_probecmp() should use to compare the
7446 * specified pattern with a string. For NULL or empty patterns, we select
7447 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7448 * For non-empty non-glob strings, we use dtrace_match_string().
7450 static dtrace_probekey_f *
7451 dtrace_probekey_func(const char *p)
7455 if (p == NULL || *p == '\0')
7456 return (&dtrace_match_nul);
7458 while ((c = *p++) != '\0') {
7459 if (c == '[' || c == '?' || c == '*' || c == '\\')
7460 return (&dtrace_match_glob);
7463 return (&dtrace_match_string);
7467 * Build a probe comparison key for use with dtrace_match_probe() from the
7468 * given probe description. By convention, a null key only matches anchored
7469 * probes: if each field is the empty string, reset dtpk_fmatch to
7470 * dtrace_match_nonzero().
7473 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7475 pkp->dtpk_prov = pdp->dtpd_provider;
7476 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7478 pkp->dtpk_mod = pdp->dtpd_mod;
7479 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7481 pkp->dtpk_func = pdp->dtpd_func;
7482 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7484 pkp->dtpk_name = pdp->dtpd_name;
7485 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7487 pkp->dtpk_id = pdp->dtpd_id;
7489 if (pkp->dtpk_id == DTRACE_IDNONE &&
7490 pkp->dtpk_pmatch == &dtrace_match_nul &&
7491 pkp->dtpk_mmatch == &dtrace_match_nul &&
7492 pkp->dtpk_fmatch == &dtrace_match_nul &&
7493 pkp->dtpk_nmatch == &dtrace_match_nul)
7494 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7498 * DTrace Provider-to-Framework API Functions
7500 * These functions implement much of the Provider-to-Framework API, as
7501 * described in <sys/dtrace.h>. The parts of the API not in this section are
7502 * the functions in the API for probe management (found below), and
7503 * dtrace_probe() itself (found above).
7507 * Register the calling provider with the DTrace framework. This should
7508 * generally be called by DTrace providers in their attach(9E) entry point.
7511 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7512 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7514 dtrace_provider_t *provider;
7516 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7517 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7518 "arguments", name ? name : "<NULL>");
7522 if (name[0] == '\0' || dtrace_badname(name)) {
7523 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7524 "provider name", name);
7528 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7529 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7530 pops->dtps_destroy == NULL ||
7531 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7532 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7533 "provider ops", name);
7537 if (dtrace_badattr(&pap->dtpa_provider) ||
7538 dtrace_badattr(&pap->dtpa_mod) ||
7539 dtrace_badattr(&pap->dtpa_func) ||
7540 dtrace_badattr(&pap->dtpa_name) ||
7541 dtrace_badattr(&pap->dtpa_args)) {
7542 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7543 "provider attributes", name);
7547 if (priv & ~DTRACE_PRIV_ALL) {
7548 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7549 "privilege attributes", name);
7553 if ((priv & DTRACE_PRIV_KERNEL) &&
7554 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7555 pops->dtps_usermode == NULL) {
7556 cmn_err(CE_WARN, "failed to register provider '%s': need "
7557 "dtps_usermode() op for given privilege attributes", name);
7561 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7562 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7563 (void) strcpy(provider->dtpv_name, name);
7565 provider->dtpv_attr = *pap;
7566 provider->dtpv_priv.dtpp_flags = priv;
7568 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7569 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7571 provider->dtpv_pops = *pops;
7573 if (pops->dtps_provide == NULL) {
7574 ASSERT(pops->dtps_provide_module != NULL);
7575 provider->dtpv_pops.dtps_provide =
7576 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7579 if (pops->dtps_provide_module == NULL) {
7580 ASSERT(pops->dtps_provide != NULL);
7581 provider->dtpv_pops.dtps_provide_module =
7582 (void (*)(void *, modctl_t *))dtrace_nullop;
7585 if (pops->dtps_suspend == NULL) {
7586 ASSERT(pops->dtps_resume == NULL);
7587 provider->dtpv_pops.dtps_suspend =
7588 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7589 provider->dtpv_pops.dtps_resume =
7590 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7593 provider->dtpv_arg = arg;
7594 *idp = (dtrace_provider_id_t)provider;
7596 if (pops == &dtrace_provider_ops) {
7597 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7598 ASSERT(MUTEX_HELD(&dtrace_lock));
7599 ASSERT(dtrace_anon.dta_enabling == NULL);
7602 * We make sure that the DTrace provider is at the head of
7603 * the provider chain.
7605 provider->dtpv_next = dtrace_provider;
7606 dtrace_provider = provider;
7610 mutex_enter(&dtrace_provider_lock);
7611 mutex_enter(&dtrace_lock);
7614 * If there is at least one provider registered, we'll add this
7615 * provider after the first provider.
7617 if (dtrace_provider != NULL) {
7618 provider->dtpv_next = dtrace_provider->dtpv_next;
7619 dtrace_provider->dtpv_next = provider;
7621 dtrace_provider = provider;
7624 if (dtrace_retained != NULL) {
7625 dtrace_enabling_provide(provider);
7628 * Now we need to call dtrace_enabling_matchall() -- which
7629 * will acquire cpu_lock and dtrace_lock. We therefore need
7630 * to drop all of our locks before calling into it...
7632 mutex_exit(&dtrace_lock);
7633 mutex_exit(&dtrace_provider_lock);
7634 dtrace_enabling_matchall();
7639 mutex_exit(&dtrace_lock);
7640 mutex_exit(&dtrace_provider_lock);
7646 * Unregister the specified provider from the DTrace framework. This should
7647 * generally be called by DTrace providers in their detach(9E) entry point.
7650 dtrace_unregister(dtrace_provider_id_t id)
7652 dtrace_provider_t *old = (dtrace_provider_t *)id;
7653 dtrace_provider_t *prev = NULL;
7654 int i, self = 0, noreap = 0;
7655 dtrace_probe_t *probe, *first = NULL;
7657 if (old->dtpv_pops.dtps_enable ==
7658 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7660 * If DTrace itself is the provider, we're called with locks
7663 ASSERT(old == dtrace_provider);
7665 ASSERT(dtrace_devi != NULL);
7667 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7668 ASSERT(MUTEX_HELD(&dtrace_lock));
7671 if (dtrace_provider->dtpv_next != NULL) {
7673 * There's another provider here; return failure.
7678 mutex_enter(&dtrace_provider_lock);
7680 mutex_enter(&mod_lock);
7682 mutex_enter(&dtrace_lock);
7686 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7687 * probes, we refuse to let providers slither away, unless this
7688 * provider has already been explicitly invalidated.
7690 if (!old->dtpv_defunct &&
7691 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7692 dtrace_anon.dta_state->dts_necbs > 0))) {
7694 mutex_exit(&dtrace_lock);
7696 mutex_exit(&mod_lock);
7698 mutex_exit(&dtrace_provider_lock);
7704 * Attempt to destroy the probes associated with this provider.
7706 for (i = 0; i < dtrace_nprobes; i++) {
7707 if ((probe = dtrace_probes[i]) == NULL)
7710 if (probe->dtpr_provider != old)
7713 if (probe->dtpr_ecb == NULL)
7717 * If we are trying to unregister a defunct provider, and the
7718 * provider was made defunct within the interval dictated by
7719 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7720 * attempt to reap our enablings. To denote that the provider
7721 * should reattempt to unregister itself at some point in the
7722 * future, we will return a differentiable error code (EAGAIN
7723 * instead of EBUSY) in this case.
7725 if (dtrace_gethrtime() - old->dtpv_defunct >
7726 dtrace_unregister_defunct_reap)
7730 mutex_exit(&dtrace_lock);
7732 mutex_exit(&mod_lock);
7734 mutex_exit(&dtrace_provider_lock);
7740 (void) taskq_dispatch(dtrace_taskq,
7741 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7747 * All of the probes for this provider are disabled; we can safely
7748 * remove all of them from their hash chains and from the probe array.
7750 for (i = 0; i < dtrace_nprobes; i++) {
7751 if ((probe = dtrace_probes[i]) == NULL)
7754 if (probe->dtpr_provider != old)
7757 dtrace_probes[i] = NULL;
7759 dtrace_hash_remove(dtrace_bymod, probe);
7760 dtrace_hash_remove(dtrace_byfunc, probe);
7761 dtrace_hash_remove(dtrace_byname, probe);
7763 if (first == NULL) {
7765 probe->dtpr_nextmod = NULL;
7767 probe->dtpr_nextmod = first;
7773 * The provider's probes have been removed from the hash chains and
7774 * from the probe array. Now issue a dtrace_sync() to be sure that
7775 * everyone has cleared out from any probe array processing.
7779 for (probe = first; probe != NULL; probe = first) {
7780 first = probe->dtpr_nextmod;
7782 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7784 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7785 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7786 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7788 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7790 free_unr(dtrace_arena, probe->dtpr_id);
7792 kmem_free(probe, sizeof (dtrace_probe_t));
7795 if ((prev = dtrace_provider) == old) {
7797 ASSERT(self || dtrace_devi == NULL);
7798 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7800 dtrace_provider = old->dtpv_next;
7802 while (prev != NULL && prev->dtpv_next != old)
7803 prev = prev->dtpv_next;
7806 panic("attempt to unregister non-existent "
7807 "dtrace provider %p\n", (void *)id);
7810 prev->dtpv_next = old->dtpv_next;
7814 mutex_exit(&dtrace_lock);
7816 mutex_exit(&mod_lock);
7818 mutex_exit(&dtrace_provider_lock);
7821 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7822 kmem_free(old, sizeof (dtrace_provider_t));
7828 * Invalidate the specified provider. All subsequent probe lookups for the
7829 * specified provider will fail, but its probes will not be removed.
7832 dtrace_invalidate(dtrace_provider_id_t id)
7834 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7836 ASSERT(pvp->dtpv_pops.dtps_enable !=
7837 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7839 mutex_enter(&dtrace_provider_lock);
7840 mutex_enter(&dtrace_lock);
7842 pvp->dtpv_defunct = dtrace_gethrtime();
7844 mutex_exit(&dtrace_lock);
7845 mutex_exit(&dtrace_provider_lock);
7849 * Indicate whether or not DTrace has attached.
7852 dtrace_attached(void)
7855 * dtrace_provider will be non-NULL iff the DTrace driver has
7856 * attached. (It's non-NULL because DTrace is always itself a
7859 return (dtrace_provider != NULL);
7863 * Remove all the unenabled probes for the given provider. This function is
7864 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7865 * -- just as many of its associated probes as it can.
7868 dtrace_condense(dtrace_provider_id_t id)
7870 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7872 dtrace_probe_t *probe;
7875 * Make sure this isn't the dtrace provider itself.
7877 ASSERT(prov->dtpv_pops.dtps_enable !=
7878 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7880 mutex_enter(&dtrace_provider_lock);
7881 mutex_enter(&dtrace_lock);
7884 * Attempt to destroy the probes associated with this provider.
7886 for (i = 0; i < dtrace_nprobes; i++) {
7887 if ((probe = dtrace_probes[i]) == NULL)
7890 if (probe->dtpr_provider != prov)
7893 if (probe->dtpr_ecb != NULL)
7896 dtrace_probes[i] = NULL;
7898 dtrace_hash_remove(dtrace_bymod, probe);
7899 dtrace_hash_remove(dtrace_byfunc, probe);
7900 dtrace_hash_remove(dtrace_byname, probe);
7902 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7904 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7905 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7906 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7907 kmem_free(probe, sizeof (dtrace_probe_t));
7909 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7911 free_unr(dtrace_arena, i + 1);
7915 mutex_exit(&dtrace_lock);
7916 mutex_exit(&dtrace_provider_lock);
7922 * DTrace Probe Management Functions
7924 * The functions in this section perform the DTrace probe management,
7925 * including functions to create probes, look-up probes, and call into the
7926 * providers to request that probes be provided. Some of these functions are
7927 * in the Provider-to-Framework API; these functions can be identified by the
7928 * fact that they are not declared "static".
7932 * Create a probe with the specified module name, function name, and name.
7935 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7936 const char *func, const char *name, int aframes, void *arg)
7938 dtrace_probe_t *probe, **probes;
7939 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7942 if (provider == dtrace_provider) {
7943 ASSERT(MUTEX_HELD(&dtrace_lock));
7945 mutex_enter(&dtrace_lock);
7949 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7950 VM_BESTFIT | VM_SLEEP);
7952 id = alloc_unr(dtrace_arena);
7954 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7956 probe->dtpr_id = id;
7957 probe->dtpr_gen = dtrace_probegen++;
7958 probe->dtpr_mod = dtrace_strdup(mod);
7959 probe->dtpr_func = dtrace_strdup(func);
7960 probe->dtpr_name = dtrace_strdup(name);
7961 probe->dtpr_arg = arg;
7962 probe->dtpr_aframes = aframes;
7963 probe->dtpr_provider = provider;
7965 dtrace_hash_add(dtrace_bymod, probe);
7966 dtrace_hash_add(dtrace_byfunc, probe);
7967 dtrace_hash_add(dtrace_byname, probe);
7969 if (id - 1 >= dtrace_nprobes) {
7970 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7971 size_t nsize = osize << 1;
7975 ASSERT(dtrace_probes == NULL);
7976 nsize = sizeof (dtrace_probe_t *);
7979 probes = kmem_zalloc(nsize, KM_SLEEP);
7981 if (dtrace_probes == NULL) {
7983 dtrace_probes = probes;
7986 dtrace_probe_t **oprobes = dtrace_probes;
7988 bcopy(oprobes, probes, osize);
7989 dtrace_membar_producer();
7990 dtrace_probes = probes;
7995 * All CPUs are now seeing the new probes array; we can
7996 * safely free the old array.
7998 kmem_free(oprobes, osize);
7999 dtrace_nprobes <<= 1;
8002 ASSERT(id - 1 < dtrace_nprobes);
8005 ASSERT(dtrace_probes[id - 1] == NULL);
8006 dtrace_probes[id - 1] = probe;
8008 if (provider != dtrace_provider)
8009 mutex_exit(&dtrace_lock);
8014 static dtrace_probe_t *
8015 dtrace_probe_lookup_id(dtrace_id_t id)
8017 ASSERT(MUTEX_HELD(&dtrace_lock));
8019 if (id == 0 || id > dtrace_nprobes)
8022 return (dtrace_probes[id - 1]);
8026 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8028 *((dtrace_id_t *)arg) = probe->dtpr_id;
8030 return (DTRACE_MATCH_DONE);
8034 * Look up a probe based on provider and one or more of module name, function
8035 * name and probe name.
8038 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
8039 char *func, char *name)
8041 dtrace_probekey_t pkey;
8045 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8046 pkey.dtpk_pmatch = &dtrace_match_string;
8047 pkey.dtpk_mod = mod;
8048 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8049 pkey.dtpk_func = func;
8050 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8051 pkey.dtpk_name = name;
8052 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8053 pkey.dtpk_id = DTRACE_IDNONE;
8055 mutex_enter(&dtrace_lock);
8056 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8057 dtrace_probe_lookup_match, &id);
8058 mutex_exit(&dtrace_lock);
8060 ASSERT(match == 1 || match == 0);
8061 return (match ? id : 0);
8065 * Returns the probe argument associated with the specified probe.
8068 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8070 dtrace_probe_t *probe;
8073 mutex_enter(&dtrace_lock);
8075 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8076 probe->dtpr_provider == (dtrace_provider_t *)id)
8077 rval = probe->dtpr_arg;
8079 mutex_exit(&dtrace_lock);
8085 * Copy a probe into a probe description.
8088 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8090 bzero(pdp, sizeof (dtrace_probedesc_t));
8091 pdp->dtpd_id = prp->dtpr_id;
8093 (void) strncpy(pdp->dtpd_provider,
8094 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8096 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8097 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8098 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8103 dtrace_probe_provide_cb(linker_file_t lf, void *arg)
8105 dtrace_provider_t *prv = (dtrace_provider_t *) arg;
8107 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, lf);
8115 * Called to indicate that a probe -- or probes -- should be provided by a
8116 * specfied provider. If the specified description is NULL, the provider will
8117 * be told to provide all of its probes. (This is done whenever a new
8118 * consumer comes along, or whenever a retained enabling is to be matched.) If
8119 * the specified description is non-NULL, the provider is given the
8120 * opportunity to dynamically provide the specified probe, allowing providers
8121 * to support the creation of probes on-the-fly. (So-called _autocreated_
8122 * probes.) If the provider is NULL, the operations will be applied to all
8123 * providers; if the provider is non-NULL the operations will only be applied
8124 * to the specified provider. The dtrace_provider_lock must be held, and the
8125 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8126 * will need to grab the dtrace_lock when it reenters the framework through
8127 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8130 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8137 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8141 prv = dtrace_provider;
8146 * First, call the blanket provide operation.
8148 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8152 * Now call the per-module provide operation. We will grab
8153 * mod_lock to prevent the list from being modified. Note
8154 * that this also prevents the mod_busy bits from changing.
8155 * (mod_busy can only be changed with mod_lock held.)
8157 mutex_enter(&mod_lock);
8161 if (ctl->mod_busy || ctl->mod_mp == NULL)
8164 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8166 } while ((ctl = ctl->mod_next) != &modules);
8168 mutex_exit(&mod_lock);
8170 (void) linker_file_foreach(dtrace_probe_provide_cb, prv);
8172 } while (all && (prv = prv->dtpv_next) != NULL);
8177 * Iterate over each probe, and call the Framework-to-Provider API function
8181 dtrace_probe_foreach(uintptr_t offs)
8183 dtrace_provider_t *prov;
8184 void (*func)(void *, dtrace_id_t, void *);
8185 dtrace_probe_t *probe;
8186 dtrace_icookie_t cookie;
8190 * We disable interrupts to walk through the probe array. This is
8191 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8192 * won't see stale data.
8194 cookie = dtrace_interrupt_disable();
8196 for (i = 0; i < dtrace_nprobes; i++) {
8197 if ((probe = dtrace_probes[i]) == NULL)
8200 if (probe->dtpr_ecb == NULL) {
8202 * This probe isn't enabled -- don't call the function.
8207 prov = probe->dtpr_provider;
8208 func = *((void(**)(void *, dtrace_id_t, void *))
8209 ((uintptr_t)&prov->dtpv_pops + offs));
8211 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8214 dtrace_interrupt_enable(cookie);
8219 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8221 dtrace_probekey_t pkey;
8226 ASSERT(MUTEX_HELD(&dtrace_lock));
8227 dtrace_ecb_create_cache = NULL;
8231 * If we're passed a NULL description, we're being asked to
8232 * create an ECB with a NULL probe.
8234 (void) dtrace_ecb_create_enable(NULL, enab);
8238 dtrace_probekey(desc, &pkey);
8239 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8240 &priv, &uid, &zoneid);
8242 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8247 * DTrace Helper Provider Functions
8250 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8252 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8253 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8254 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8258 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8259 const dof_provider_t *dofprov, char *strtab)
8261 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8262 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8263 dofprov->dofpv_provattr);
8264 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8265 dofprov->dofpv_modattr);
8266 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8267 dofprov->dofpv_funcattr);
8268 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8269 dofprov->dofpv_nameattr);
8270 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8271 dofprov->dofpv_argsattr);
8275 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8277 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8278 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8279 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8280 dof_provider_t *provider;
8282 uint32_t *off, *enoff;
8286 dtrace_helper_provdesc_t dhpv;
8287 dtrace_helper_probedesc_t dhpb;
8288 dtrace_meta_t *meta = dtrace_meta_pid;
8289 dtrace_mops_t *mops = &meta->dtm_mops;
8292 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8293 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8294 provider->dofpv_strtab * dof->dofh_secsize);
8295 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8296 provider->dofpv_probes * dof->dofh_secsize);
8297 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8298 provider->dofpv_prargs * dof->dofh_secsize);
8299 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8300 provider->dofpv_proffs * dof->dofh_secsize);
8302 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8303 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8304 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8308 * See dtrace_helper_provider_validate().
8310 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8311 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8312 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8313 provider->dofpv_prenoffs * dof->dofh_secsize);
8314 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8317 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8320 * Create the provider.
8322 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8324 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8330 * Create the probes.
8332 for (i = 0; i < nprobes; i++) {
8333 probe = (dof_probe_t *)(uintptr_t)(daddr +
8334 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8336 dhpb.dthpb_mod = dhp->dofhp_mod;
8337 dhpb.dthpb_func = strtab + probe->dofpr_func;
8338 dhpb.dthpb_name = strtab + probe->dofpr_name;
8339 dhpb.dthpb_base = probe->dofpr_addr;
8340 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8341 dhpb.dthpb_noffs = probe->dofpr_noffs;
8342 if (enoff != NULL) {
8343 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8344 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8346 dhpb.dthpb_enoffs = NULL;
8347 dhpb.dthpb_nenoffs = 0;
8349 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8350 dhpb.dthpb_nargc = probe->dofpr_nargc;
8351 dhpb.dthpb_xargc = probe->dofpr_xargc;
8352 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8353 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8355 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8360 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8362 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8363 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8366 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8368 for (i = 0; i < dof->dofh_secnum; i++) {
8369 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8370 dof->dofh_secoff + i * dof->dofh_secsize);
8372 if (sec->dofs_type != DOF_SECT_PROVIDER)
8375 dtrace_helper_provide_one(dhp, sec, pid);
8379 * We may have just created probes, so we must now rematch against
8380 * any retained enablings. Note that this call will acquire both
8381 * cpu_lock and dtrace_lock; the fact that we are holding
8382 * dtrace_meta_lock now is what defines the ordering with respect to
8383 * these three locks.
8385 dtrace_enabling_matchall();
8389 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8391 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8392 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8394 dof_provider_t *provider;
8396 dtrace_helper_provdesc_t dhpv;
8397 dtrace_meta_t *meta = dtrace_meta_pid;
8398 dtrace_mops_t *mops = &meta->dtm_mops;
8400 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8401 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8402 provider->dofpv_strtab * dof->dofh_secsize);
8404 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8407 * Create the provider.
8409 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8411 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8417 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8419 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8420 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8423 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8425 for (i = 0; i < dof->dofh_secnum; i++) {
8426 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8427 dof->dofh_secoff + i * dof->dofh_secsize);
8429 if (sec->dofs_type != DOF_SECT_PROVIDER)
8432 dtrace_helper_provider_remove_one(dhp, sec, pid);
8437 * DTrace Meta Provider-to-Framework API Functions
8439 * These functions implement the Meta Provider-to-Framework API, as described
8440 * in <sys/dtrace.h>.
8443 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8444 dtrace_meta_provider_id_t *idp)
8446 dtrace_meta_t *meta;
8447 dtrace_helpers_t *help, *next;
8450 *idp = DTRACE_METAPROVNONE;
8453 * We strictly don't need the name, but we hold onto it for
8454 * debuggability. All hail error queues!
8457 cmn_err(CE_WARN, "failed to register meta-provider: "
8463 mops->dtms_create_probe == NULL ||
8464 mops->dtms_provide_pid == NULL ||
8465 mops->dtms_remove_pid == NULL) {
8466 cmn_err(CE_WARN, "failed to register meta-register %s: "
8467 "invalid ops", name);
8471 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8472 meta->dtm_mops = *mops;
8473 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8474 (void) strcpy(meta->dtm_name, name);
8475 meta->dtm_arg = arg;
8477 mutex_enter(&dtrace_meta_lock);
8478 mutex_enter(&dtrace_lock);
8480 if (dtrace_meta_pid != NULL) {
8481 mutex_exit(&dtrace_lock);
8482 mutex_exit(&dtrace_meta_lock);
8483 cmn_err(CE_WARN, "failed to register meta-register %s: "
8484 "user-land meta-provider exists", name);
8485 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8486 kmem_free(meta, sizeof (dtrace_meta_t));
8490 dtrace_meta_pid = meta;
8491 *idp = (dtrace_meta_provider_id_t)meta;
8494 * If there are providers and probes ready to go, pass them
8495 * off to the new meta provider now.
8498 help = dtrace_deferred_pid;
8499 dtrace_deferred_pid = NULL;
8501 mutex_exit(&dtrace_lock);
8503 while (help != NULL) {
8504 for (i = 0; i < help->dthps_nprovs; i++) {
8505 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8509 next = help->dthps_next;
8510 help->dthps_next = NULL;
8511 help->dthps_prev = NULL;
8512 help->dthps_deferred = 0;
8516 mutex_exit(&dtrace_meta_lock);
8522 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8524 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8526 mutex_enter(&dtrace_meta_lock);
8527 mutex_enter(&dtrace_lock);
8529 if (old == dtrace_meta_pid) {
8530 pp = &dtrace_meta_pid;
8532 panic("attempt to unregister non-existent "
8533 "dtrace meta-provider %p\n", (void *)old);
8536 if (old->dtm_count != 0) {
8537 mutex_exit(&dtrace_lock);
8538 mutex_exit(&dtrace_meta_lock);
8544 mutex_exit(&dtrace_lock);
8545 mutex_exit(&dtrace_meta_lock);
8547 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8548 kmem_free(old, sizeof (dtrace_meta_t));
8555 * DTrace DIF Object Functions
8558 dtrace_difo_err(uint_t pc, const char *format, ...)
8560 if (dtrace_err_verbose) {
8563 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8564 va_start(alist, format);
8565 (void) vuprintf(format, alist);
8569 #ifdef DTRACE_ERRDEBUG
8570 dtrace_errdebug(format);
8576 * Validate a DTrace DIF object by checking the IR instructions. The following
8577 * rules are currently enforced by dtrace_difo_validate():
8579 * 1. Each instruction must have a valid opcode
8580 * 2. Each register, string, variable, or subroutine reference must be valid
8581 * 3. No instruction can modify register %r0 (must be zero)
8582 * 4. All instruction reserved bits must be set to zero
8583 * 5. The last instruction must be a "ret" instruction
8584 * 6. All branch targets must reference a valid instruction _after_ the branch
8587 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8591 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8595 kcheckload = cr == NULL ||
8596 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8598 dp->dtdo_destructive = 0;
8600 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8601 dif_instr_t instr = dp->dtdo_buf[pc];
8603 uint_t r1 = DIF_INSTR_R1(instr);
8604 uint_t r2 = DIF_INSTR_R2(instr);
8605 uint_t rd = DIF_INSTR_RD(instr);
8606 uint_t rs = DIF_INSTR_RS(instr);
8607 uint_t label = DIF_INSTR_LABEL(instr);
8608 uint_t v = DIF_INSTR_VAR(instr);
8609 uint_t subr = DIF_INSTR_SUBR(instr);
8610 uint_t type = DIF_INSTR_TYPE(instr);
8611 uint_t op = DIF_INSTR_OP(instr);
8629 err += efunc(pc, "invalid register %u\n", r1);
8631 err += efunc(pc, "invalid register %u\n", r2);
8633 err += efunc(pc, "invalid register %u\n", rd);
8635 err += efunc(pc, "cannot write to %r0\n");
8641 err += efunc(pc, "invalid register %u\n", r1);
8643 err += efunc(pc, "non-zero reserved bits\n");
8645 err += efunc(pc, "invalid register %u\n", rd);
8647 err += efunc(pc, "cannot write to %r0\n");
8657 err += efunc(pc, "invalid register %u\n", r1);
8659 err += efunc(pc, "non-zero reserved bits\n");
8661 err += efunc(pc, "invalid register %u\n", rd);
8663 err += efunc(pc, "cannot write to %r0\n");
8665 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8666 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8676 err += efunc(pc, "invalid register %u\n", r1);
8678 err += efunc(pc, "non-zero reserved bits\n");
8680 err += efunc(pc, "invalid register %u\n", rd);
8682 err += efunc(pc, "cannot write to %r0\n");
8692 err += efunc(pc, "invalid register %u\n", r1);
8694 err += efunc(pc, "non-zero reserved bits\n");
8696 err += efunc(pc, "invalid register %u\n", rd);
8698 err += efunc(pc, "cannot write to %r0\n");
8705 err += efunc(pc, "invalid register %u\n", r1);
8707 err += efunc(pc, "non-zero reserved bits\n");
8709 err += efunc(pc, "invalid register %u\n", rd);
8711 err += efunc(pc, "cannot write to 0 address\n");
8716 err += efunc(pc, "invalid register %u\n", r1);
8718 err += efunc(pc, "invalid register %u\n", r2);
8720 err += efunc(pc, "non-zero reserved bits\n");
8724 err += efunc(pc, "invalid register %u\n", r1);
8725 if (r2 != 0 || rd != 0)
8726 err += efunc(pc, "non-zero reserved bits\n");
8739 if (label >= dp->dtdo_len) {
8740 err += efunc(pc, "invalid branch target %u\n",
8744 err += efunc(pc, "backward branch to %u\n",
8749 if (r1 != 0 || r2 != 0)
8750 err += efunc(pc, "non-zero reserved bits\n");
8752 err += efunc(pc, "invalid register %u\n", rd);
8756 case DIF_OP_FLUSHTS:
8757 if (r1 != 0 || r2 != 0 || rd != 0)
8758 err += efunc(pc, "non-zero reserved bits\n");
8761 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8762 err += efunc(pc, "invalid integer ref %u\n",
8763 DIF_INSTR_INTEGER(instr));
8766 err += efunc(pc, "invalid register %u\n", rd);
8768 err += efunc(pc, "cannot write to %r0\n");
8771 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8772 err += efunc(pc, "invalid string ref %u\n",
8773 DIF_INSTR_STRING(instr));
8776 err += efunc(pc, "invalid register %u\n", rd);
8778 err += efunc(pc, "cannot write to %r0\n");
8782 if (r1 > DIF_VAR_ARRAY_MAX)
8783 err += efunc(pc, "invalid array %u\n", r1);
8785 err += efunc(pc, "invalid register %u\n", r2);
8787 err += efunc(pc, "invalid register %u\n", rd);
8789 err += efunc(pc, "cannot write to %r0\n");
8796 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8797 err += efunc(pc, "invalid variable %u\n", v);
8799 err += efunc(pc, "invalid register %u\n", rd);
8801 err += efunc(pc, "cannot write to %r0\n");
8808 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8809 err += efunc(pc, "invalid variable %u\n", v);
8811 err += efunc(pc, "invalid register %u\n", rd);
8814 if (subr > DIF_SUBR_MAX)
8815 err += efunc(pc, "invalid subr %u\n", subr);
8817 err += efunc(pc, "invalid register %u\n", rd);
8819 err += efunc(pc, "cannot write to %r0\n");
8821 if (subr == DIF_SUBR_COPYOUT ||
8822 subr == DIF_SUBR_COPYOUTSTR) {
8823 dp->dtdo_destructive = 1;
8827 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8828 err += efunc(pc, "invalid ref type %u\n", type);
8830 err += efunc(pc, "invalid register %u\n", r2);
8832 err += efunc(pc, "invalid register %u\n", rs);
8835 if (type != DIF_TYPE_CTF)
8836 err += efunc(pc, "invalid val type %u\n", type);
8838 err += efunc(pc, "invalid register %u\n", r2);
8840 err += efunc(pc, "invalid register %u\n", rs);
8843 err += efunc(pc, "invalid opcode %u\n",
8844 DIF_INSTR_OP(instr));
8848 if (dp->dtdo_len != 0 &&
8849 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8850 err += efunc(dp->dtdo_len - 1,
8851 "expected 'ret' as last DIF instruction\n");
8854 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8856 * If we're not returning by reference, the size must be either
8857 * 0 or the size of one of the base types.
8859 switch (dp->dtdo_rtype.dtdt_size) {
8861 case sizeof (uint8_t):
8862 case sizeof (uint16_t):
8863 case sizeof (uint32_t):
8864 case sizeof (uint64_t):
8868 err += efunc(dp->dtdo_len - 1, "bad return size");
8872 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8873 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8874 dtrace_diftype_t *vt, *et;
8877 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8878 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8879 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8880 err += efunc(i, "unrecognized variable scope %d\n",
8885 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8886 v->dtdv_kind != DIFV_KIND_SCALAR) {
8887 err += efunc(i, "unrecognized variable type %d\n",
8892 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8893 err += efunc(i, "%d exceeds variable id limit\n", id);
8897 if (id < DIF_VAR_OTHER_UBASE)
8901 * For user-defined variables, we need to check that this
8902 * definition is identical to any previous definition that we
8905 ndx = id - DIF_VAR_OTHER_UBASE;
8907 switch (v->dtdv_scope) {
8908 case DIFV_SCOPE_GLOBAL:
8909 if (ndx < vstate->dtvs_nglobals) {
8910 dtrace_statvar_t *svar;
8912 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8913 existing = &svar->dtsv_var;
8918 case DIFV_SCOPE_THREAD:
8919 if (ndx < vstate->dtvs_ntlocals)
8920 existing = &vstate->dtvs_tlocals[ndx];
8923 case DIFV_SCOPE_LOCAL:
8924 if (ndx < vstate->dtvs_nlocals) {
8925 dtrace_statvar_t *svar;
8927 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8928 existing = &svar->dtsv_var;
8936 if (vt->dtdt_flags & DIF_TF_BYREF) {
8937 if (vt->dtdt_size == 0) {
8938 err += efunc(i, "zero-sized variable\n");
8942 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8943 vt->dtdt_size > dtrace_global_maxsize) {
8944 err += efunc(i, "oversized by-ref global\n");
8949 if (existing == NULL || existing->dtdv_id == 0)
8952 ASSERT(existing->dtdv_id == v->dtdv_id);
8953 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8955 if (existing->dtdv_kind != v->dtdv_kind)
8956 err += efunc(i, "%d changed variable kind\n", id);
8958 et = &existing->dtdv_type;
8960 if (vt->dtdt_flags != et->dtdt_flags) {
8961 err += efunc(i, "%d changed variable type flags\n", id);
8965 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8966 err += efunc(i, "%d changed variable type size\n", id);
8975 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
8976 * are much more constrained than normal DIFOs. Specifically, they may
8979 * 1. Make calls to subroutines other than copyin(), copyinstr() or
8980 * miscellaneous string routines
8981 * 2. Access DTrace variables other than the args[] array, and the
8982 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8983 * 3. Have thread-local variables.
8984 * 4. Have dynamic variables.
8987 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8989 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8993 for (pc = 0; pc < dp->dtdo_len; pc++) {
8994 dif_instr_t instr = dp->dtdo_buf[pc];
8996 uint_t v = DIF_INSTR_VAR(instr);
8997 uint_t subr = DIF_INSTR_SUBR(instr);
8998 uint_t op = DIF_INSTR_OP(instr);
9053 case DIF_OP_FLUSHTS:
9065 if (v >= DIF_VAR_OTHER_UBASE)
9068 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9071 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9072 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9073 v == DIF_VAR_EXECARGS ||
9074 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9075 v == DIF_VAR_UID || v == DIF_VAR_GID)
9078 err += efunc(pc, "illegal variable %u\n", v);
9085 err += efunc(pc, "illegal dynamic variable load\n");
9091 err += efunc(pc, "illegal dynamic variable store\n");
9095 if (subr == DIF_SUBR_ALLOCA ||
9096 subr == DIF_SUBR_BCOPY ||
9097 subr == DIF_SUBR_COPYIN ||
9098 subr == DIF_SUBR_COPYINTO ||
9099 subr == DIF_SUBR_COPYINSTR ||
9100 subr == DIF_SUBR_INDEX ||
9101 subr == DIF_SUBR_INET_NTOA ||
9102 subr == DIF_SUBR_INET_NTOA6 ||
9103 subr == DIF_SUBR_INET_NTOP ||
9104 subr == DIF_SUBR_LLTOSTR ||
9105 subr == DIF_SUBR_RINDEX ||
9106 subr == DIF_SUBR_STRCHR ||
9107 subr == DIF_SUBR_STRJOIN ||
9108 subr == DIF_SUBR_STRRCHR ||
9109 subr == DIF_SUBR_STRSTR ||
9110 subr == DIF_SUBR_HTONS ||
9111 subr == DIF_SUBR_HTONL ||
9112 subr == DIF_SUBR_HTONLL ||
9113 subr == DIF_SUBR_NTOHS ||
9114 subr == DIF_SUBR_NTOHL ||
9115 subr == DIF_SUBR_NTOHLL ||
9116 subr == DIF_SUBR_MEMREF ||
9117 subr == DIF_SUBR_TYPEREF)
9120 err += efunc(pc, "invalid subr %u\n", subr);
9124 err += efunc(pc, "invalid opcode %u\n",
9125 DIF_INSTR_OP(instr));
9133 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9137 dtrace_difo_cacheable(dtrace_difo_t *dp)
9144 for (i = 0; i < dp->dtdo_varlen; i++) {
9145 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9147 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9150 switch (v->dtdv_id) {
9151 case DIF_VAR_CURTHREAD:
9154 case DIF_VAR_EXECARGS:
9155 case DIF_VAR_EXECNAME:
9156 case DIF_VAR_ZONENAME:
9165 * This DIF object may be cacheable. Now we need to look for any
9166 * array loading instructions, any memory loading instructions, or
9167 * any stores to thread-local variables.
9169 for (i = 0; i < dp->dtdo_len; i++) {
9170 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9172 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9173 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9174 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9175 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9183 dtrace_difo_hold(dtrace_difo_t *dp)
9187 ASSERT(MUTEX_HELD(&dtrace_lock));
9190 ASSERT(dp->dtdo_refcnt != 0);
9193 * We need to check this DIF object for references to the variable
9194 * DIF_VAR_VTIMESTAMP.
9196 for (i = 0; i < dp->dtdo_varlen; i++) {
9197 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9199 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9202 if (dtrace_vtime_references++ == 0)
9203 dtrace_vtime_enable();
9208 * This routine calculates the dynamic variable chunksize for a given DIF
9209 * object. The calculation is not fool-proof, and can probably be tricked by
9210 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9211 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9212 * if a dynamic variable size exceeds the chunksize.
9215 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9218 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9219 const dif_instr_t *text = dp->dtdo_buf;
9225 for (pc = 0; pc < dp->dtdo_len; pc++) {
9226 dif_instr_t instr = text[pc];
9227 uint_t op = DIF_INSTR_OP(instr);
9228 uint_t rd = DIF_INSTR_RD(instr);
9229 uint_t r1 = DIF_INSTR_R1(instr);
9233 dtrace_key_t *key = tupregs;
9237 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9242 key = &tupregs[DIF_DTR_NREGS];
9243 key[0].dttk_size = 0;
9244 key[1].dttk_size = 0;
9246 scope = DIFV_SCOPE_THREAD;
9253 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9254 key[nkeys++].dttk_size = 0;
9256 key[nkeys++].dttk_size = 0;
9258 if (op == DIF_OP_STTAA) {
9259 scope = DIFV_SCOPE_THREAD;
9261 scope = DIFV_SCOPE_GLOBAL;
9267 if (ttop == DIF_DTR_NREGS)
9270 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9272 * If the register for the size of the "pushtr"
9273 * is %r0 (or the value is 0) and the type is
9274 * a string, we'll use the system-wide default
9277 tupregs[ttop++].dttk_size =
9278 dtrace_strsize_default;
9283 tupregs[ttop++].dttk_size = sval;
9289 if (ttop == DIF_DTR_NREGS)
9292 tupregs[ttop++].dttk_size = 0;
9295 case DIF_OP_FLUSHTS:
9312 * We have a dynamic variable allocation; calculate its size.
9314 for (ksize = 0, i = 0; i < nkeys; i++)
9315 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9317 size = sizeof (dtrace_dynvar_t);
9318 size += sizeof (dtrace_key_t) * (nkeys - 1);
9322 * Now we need to determine the size of the stored data.
9324 id = DIF_INSTR_VAR(instr);
9326 for (i = 0; i < dp->dtdo_varlen; i++) {
9327 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9329 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9330 size += v->dtdv_type.dtdt_size;
9335 if (i == dp->dtdo_varlen)
9339 * We have the size. If this is larger than the chunk size
9340 * for our dynamic variable state, reset the chunk size.
9342 size = P2ROUNDUP(size, sizeof (uint64_t));
9344 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9345 vstate->dtvs_dynvars.dtds_chunksize = size;
9350 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9352 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9355 ASSERT(MUTEX_HELD(&dtrace_lock));
9356 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9358 for (i = 0; i < dp->dtdo_varlen; i++) {
9359 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9360 dtrace_statvar_t *svar, ***svarp = NULL;
9362 uint8_t scope = v->dtdv_scope;
9365 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9368 id -= DIF_VAR_OTHER_UBASE;
9371 case DIFV_SCOPE_THREAD:
9372 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9373 dtrace_difv_t *tlocals;
9375 if ((ntlocals = (otlocals << 1)) == 0)
9378 osz = otlocals * sizeof (dtrace_difv_t);
9379 nsz = ntlocals * sizeof (dtrace_difv_t);
9381 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9384 bcopy(vstate->dtvs_tlocals,
9386 kmem_free(vstate->dtvs_tlocals, osz);
9389 vstate->dtvs_tlocals = tlocals;
9390 vstate->dtvs_ntlocals = ntlocals;
9393 vstate->dtvs_tlocals[id] = *v;
9396 case DIFV_SCOPE_LOCAL:
9397 np = &vstate->dtvs_nlocals;
9398 svarp = &vstate->dtvs_locals;
9400 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9401 dsize = NCPU * (v->dtdv_type.dtdt_size +
9404 dsize = NCPU * sizeof (uint64_t);
9408 case DIFV_SCOPE_GLOBAL:
9409 np = &vstate->dtvs_nglobals;
9410 svarp = &vstate->dtvs_globals;
9412 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9413 dsize = v->dtdv_type.dtdt_size +
9422 while (id >= (oldsvars = *np)) {
9423 dtrace_statvar_t **statics;
9424 int newsvars, oldsize, newsize;
9426 if ((newsvars = (oldsvars << 1)) == 0)
9429 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9430 newsize = newsvars * sizeof (dtrace_statvar_t *);
9432 statics = kmem_zalloc(newsize, KM_SLEEP);
9435 bcopy(*svarp, statics, oldsize);
9436 kmem_free(*svarp, oldsize);
9443 if ((svar = (*svarp)[id]) == NULL) {
9444 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9445 svar->dtsv_var = *v;
9447 if ((svar->dtsv_size = dsize) != 0) {
9448 svar->dtsv_data = (uint64_t)(uintptr_t)
9449 kmem_zalloc(dsize, KM_SLEEP);
9452 (*svarp)[id] = svar;
9455 svar->dtsv_refcnt++;
9458 dtrace_difo_chunksize(dp, vstate);
9459 dtrace_difo_hold(dp);
9462 static dtrace_difo_t *
9463 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9468 ASSERT(dp->dtdo_buf != NULL);
9469 ASSERT(dp->dtdo_refcnt != 0);
9471 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9473 ASSERT(dp->dtdo_buf != NULL);
9474 sz = dp->dtdo_len * sizeof (dif_instr_t);
9475 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9476 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9477 new->dtdo_len = dp->dtdo_len;
9479 if (dp->dtdo_strtab != NULL) {
9480 ASSERT(dp->dtdo_strlen != 0);
9481 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9482 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9483 new->dtdo_strlen = dp->dtdo_strlen;
9486 if (dp->dtdo_inttab != NULL) {
9487 ASSERT(dp->dtdo_intlen != 0);
9488 sz = dp->dtdo_intlen * sizeof (uint64_t);
9489 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9490 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9491 new->dtdo_intlen = dp->dtdo_intlen;
9494 if (dp->dtdo_vartab != NULL) {
9495 ASSERT(dp->dtdo_varlen != 0);
9496 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9497 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9498 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9499 new->dtdo_varlen = dp->dtdo_varlen;
9502 dtrace_difo_init(new, vstate);
9507 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9511 ASSERT(dp->dtdo_refcnt == 0);
9513 for (i = 0; i < dp->dtdo_varlen; i++) {
9514 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9515 dtrace_statvar_t *svar, **svarp = NULL;
9517 uint8_t scope = v->dtdv_scope;
9521 case DIFV_SCOPE_THREAD:
9524 case DIFV_SCOPE_LOCAL:
9525 np = &vstate->dtvs_nlocals;
9526 svarp = vstate->dtvs_locals;
9529 case DIFV_SCOPE_GLOBAL:
9530 np = &vstate->dtvs_nglobals;
9531 svarp = vstate->dtvs_globals;
9538 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9541 id -= DIF_VAR_OTHER_UBASE;
9545 ASSERT(svar != NULL);
9546 ASSERT(svar->dtsv_refcnt > 0);
9548 if (--svar->dtsv_refcnt > 0)
9551 if (svar->dtsv_size != 0) {
9552 ASSERT(svar->dtsv_data != 0);
9553 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9557 kmem_free(svar, sizeof (dtrace_statvar_t));
9561 if (dp->dtdo_buf != NULL)
9562 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9563 if (dp->dtdo_inttab != NULL)
9564 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9565 if (dp->dtdo_strtab != NULL)
9566 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9567 if (dp->dtdo_vartab != NULL)
9568 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9570 kmem_free(dp, sizeof (dtrace_difo_t));
9574 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9578 ASSERT(MUTEX_HELD(&dtrace_lock));
9579 ASSERT(dp->dtdo_refcnt != 0);
9581 for (i = 0; i < dp->dtdo_varlen; i++) {
9582 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9584 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9587 ASSERT(dtrace_vtime_references > 0);
9588 if (--dtrace_vtime_references == 0)
9589 dtrace_vtime_disable();
9592 if (--dp->dtdo_refcnt == 0)
9593 dtrace_difo_destroy(dp, vstate);
9597 * DTrace Format Functions
9600 dtrace_format_add(dtrace_state_t *state, char *str)
9603 uint16_t ndx, len = strlen(str) + 1;
9605 fmt = kmem_zalloc(len, KM_SLEEP);
9606 bcopy(str, fmt, len);
9608 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9609 if (state->dts_formats[ndx] == NULL) {
9610 state->dts_formats[ndx] = fmt;
9615 if (state->dts_nformats == USHRT_MAX) {
9617 * This is only likely if a denial-of-service attack is being
9618 * attempted. As such, it's okay to fail silently here.
9620 kmem_free(fmt, len);
9625 * For simplicity, we always resize the formats array to be exactly the
9626 * number of formats.
9628 ndx = state->dts_nformats++;
9629 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9631 if (state->dts_formats != NULL) {
9633 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9634 kmem_free(state->dts_formats, ndx * sizeof (char *));
9637 state->dts_formats = new;
9638 state->dts_formats[ndx] = fmt;
9644 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9648 ASSERT(state->dts_formats != NULL);
9649 ASSERT(format <= state->dts_nformats);
9650 ASSERT(state->dts_formats[format - 1] != NULL);
9652 fmt = state->dts_formats[format - 1];
9653 kmem_free(fmt, strlen(fmt) + 1);
9654 state->dts_formats[format - 1] = NULL;
9658 dtrace_format_destroy(dtrace_state_t *state)
9662 if (state->dts_nformats == 0) {
9663 ASSERT(state->dts_formats == NULL);
9667 ASSERT(state->dts_formats != NULL);
9669 for (i = 0; i < state->dts_nformats; i++) {
9670 char *fmt = state->dts_formats[i];
9675 kmem_free(fmt, strlen(fmt) + 1);
9678 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9679 state->dts_nformats = 0;
9680 state->dts_formats = NULL;
9684 * DTrace Predicate Functions
9686 static dtrace_predicate_t *
9687 dtrace_predicate_create(dtrace_difo_t *dp)
9689 dtrace_predicate_t *pred;
9691 ASSERT(MUTEX_HELD(&dtrace_lock));
9692 ASSERT(dp->dtdo_refcnt != 0);
9694 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9695 pred->dtp_difo = dp;
9696 pred->dtp_refcnt = 1;
9698 if (!dtrace_difo_cacheable(dp))
9701 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9703 * This is only theoretically possible -- we have had 2^32
9704 * cacheable predicates on this machine. We cannot allow any
9705 * more predicates to become cacheable: as unlikely as it is,
9706 * there may be a thread caching a (now stale) predicate cache
9707 * ID. (N.B.: the temptation is being successfully resisted to
9708 * have this cmn_err() "Holy shit -- we executed this code!")
9713 pred->dtp_cacheid = dtrace_predcache_id++;
9719 dtrace_predicate_hold(dtrace_predicate_t *pred)
9721 ASSERT(MUTEX_HELD(&dtrace_lock));
9722 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9723 ASSERT(pred->dtp_refcnt > 0);
9729 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9731 dtrace_difo_t *dp = pred->dtp_difo;
9733 ASSERT(MUTEX_HELD(&dtrace_lock));
9734 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9735 ASSERT(pred->dtp_refcnt > 0);
9737 if (--pred->dtp_refcnt == 0) {
9738 dtrace_difo_release(pred->dtp_difo, vstate);
9739 kmem_free(pred, sizeof (dtrace_predicate_t));
9744 * DTrace Action Description Functions
9746 static dtrace_actdesc_t *
9747 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9748 uint64_t uarg, uint64_t arg)
9750 dtrace_actdesc_t *act;
9753 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9754 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9757 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9758 act->dtad_kind = kind;
9759 act->dtad_ntuple = ntuple;
9760 act->dtad_uarg = uarg;
9761 act->dtad_arg = arg;
9762 act->dtad_refcnt = 1;
9768 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9770 ASSERT(act->dtad_refcnt >= 1);
9775 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9777 dtrace_actkind_t kind = act->dtad_kind;
9780 ASSERT(act->dtad_refcnt >= 1);
9782 if (--act->dtad_refcnt != 0)
9785 if ((dp = act->dtad_difo) != NULL)
9786 dtrace_difo_release(dp, vstate);
9788 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9789 char *str = (char *)(uintptr_t)act->dtad_arg;
9792 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9793 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9797 kmem_free(str, strlen(str) + 1);
9800 kmem_free(act, sizeof (dtrace_actdesc_t));
9804 * DTrace ECB Functions
9806 static dtrace_ecb_t *
9807 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9812 ASSERT(MUTEX_HELD(&dtrace_lock));
9814 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9815 ecb->dte_predicate = NULL;
9816 ecb->dte_probe = probe;
9819 * The default size is the size of the default action: recording
9822 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
9823 ecb->dte_alignment = sizeof (dtrace_epid_t);
9825 epid = state->dts_epid++;
9827 if (epid - 1 >= state->dts_necbs) {
9828 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9829 int necbs = state->dts_necbs << 1;
9831 ASSERT(epid == state->dts_necbs + 1);
9834 ASSERT(oecbs == NULL);
9838 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9841 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9843 dtrace_membar_producer();
9844 state->dts_ecbs = ecbs;
9846 if (oecbs != NULL) {
9848 * If this state is active, we must dtrace_sync()
9849 * before we can free the old dts_ecbs array: we're
9850 * coming in hot, and there may be active ring
9851 * buffer processing (which indexes into the dts_ecbs
9852 * array) on another CPU.
9854 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9857 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9860 dtrace_membar_producer();
9861 state->dts_necbs = necbs;
9864 ecb->dte_state = state;
9866 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9867 dtrace_membar_producer();
9868 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9874 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9876 dtrace_probe_t *probe = ecb->dte_probe;
9878 ASSERT(MUTEX_HELD(&cpu_lock));
9879 ASSERT(MUTEX_HELD(&dtrace_lock));
9880 ASSERT(ecb->dte_next == NULL);
9882 if (probe == NULL) {
9884 * This is the NULL probe -- there's nothing to do.
9889 if (probe->dtpr_ecb == NULL) {
9890 dtrace_provider_t *prov = probe->dtpr_provider;
9893 * We're the first ECB on this probe.
9895 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9897 if (ecb->dte_predicate != NULL)
9898 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9900 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9901 probe->dtpr_id, probe->dtpr_arg);
9904 * This probe is already active. Swing the last pointer to
9905 * point to the new ECB, and issue a dtrace_sync() to assure
9906 * that all CPUs have seen the change.
9908 ASSERT(probe->dtpr_ecb_last != NULL);
9909 probe->dtpr_ecb_last->dte_next = ecb;
9910 probe->dtpr_ecb_last = ecb;
9911 probe->dtpr_predcache = 0;
9918 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9920 dtrace_action_t *act;
9921 uint32_t curneeded = UINT32_MAX;
9922 uint32_t aggbase = UINT32_MAX;
9925 * If we record anything, we always record the dtrace_rechdr_t. (And
9926 * we always record it first.)
9928 ecb->dte_size = sizeof (dtrace_rechdr_t);
9929 ecb->dte_alignment = sizeof (dtrace_epid_t);
9931 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9932 dtrace_recdesc_t *rec = &act->dta_rec;
9933 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
9935 ecb->dte_alignment = MAX(ecb->dte_alignment,
9936 rec->dtrd_alignment);
9938 if (DTRACEACT_ISAGG(act->dta_kind)) {
9939 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9941 ASSERT(rec->dtrd_size != 0);
9942 ASSERT(agg->dtag_first != NULL);
9943 ASSERT(act->dta_prev->dta_intuple);
9944 ASSERT(aggbase != UINT32_MAX);
9945 ASSERT(curneeded != UINT32_MAX);
9947 agg->dtag_base = aggbase;
9949 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9950 rec->dtrd_offset = curneeded;
9951 curneeded += rec->dtrd_size;
9952 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
9954 aggbase = UINT32_MAX;
9955 curneeded = UINT32_MAX;
9956 } else if (act->dta_intuple) {
9957 if (curneeded == UINT32_MAX) {
9959 * This is the first record in a tuple. Align
9960 * curneeded to be at offset 4 in an 8-byte
9963 ASSERT(act->dta_prev == NULL ||
9964 !act->dta_prev->dta_intuple);
9965 ASSERT3U(aggbase, ==, UINT32_MAX);
9966 curneeded = P2PHASEUP(ecb->dte_size,
9967 sizeof (uint64_t), sizeof (dtrace_aggid_t));
9969 aggbase = curneeded - sizeof (dtrace_aggid_t);
9970 ASSERT(IS_P2ALIGNED(aggbase,
9971 sizeof (uint64_t)));
9973 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9974 rec->dtrd_offset = curneeded;
9975 curneeded += rec->dtrd_size;
9977 /* tuples must be followed by an aggregation */
9978 ASSERT(act->dta_prev == NULL ||
9979 !act->dta_prev->dta_intuple);
9981 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
9982 rec->dtrd_alignment);
9983 rec->dtrd_offset = ecb->dte_size;
9984 ecb->dte_size += rec->dtrd_size;
9985 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
9989 if ((act = ecb->dte_action) != NULL &&
9990 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9991 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
9993 * If the size is still sizeof (dtrace_rechdr_t), then all
9994 * actions store no data; set the size to 0.
9999 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10000 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10001 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10005 static dtrace_action_t *
10006 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10008 dtrace_aggregation_t *agg;
10009 size_t size = sizeof (uint64_t);
10010 int ntuple = desc->dtad_ntuple;
10011 dtrace_action_t *act;
10012 dtrace_recdesc_t *frec;
10013 dtrace_aggid_t aggid;
10014 dtrace_state_t *state = ecb->dte_state;
10016 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10017 agg->dtag_ecb = ecb;
10019 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10021 switch (desc->dtad_kind) {
10022 case DTRACEAGG_MIN:
10023 agg->dtag_initial = INT64_MAX;
10024 agg->dtag_aggregate = dtrace_aggregate_min;
10027 case DTRACEAGG_MAX:
10028 agg->dtag_initial = INT64_MIN;
10029 agg->dtag_aggregate = dtrace_aggregate_max;
10032 case DTRACEAGG_COUNT:
10033 agg->dtag_aggregate = dtrace_aggregate_count;
10036 case DTRACEAGG_QUANTIZE:
10037 agg->dtag_aggregate = dtrace_aggregate_quantize;
10038 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10042 case DTRACEAGG_LQUANTIZE: {
10043 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10044 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10046 agg->dtag_initial = desc->dtad_arg;
10047 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10049 if (step == 0 || levels == 0)
10052 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10056 case DTRACEAGG_LLQUANTIZE: {
10057 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10058 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10059 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10060 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10063 agg->dtag_initial = desc->dtad_arg;
10064 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10066 if (factor < 2 || low >= high || nsteps < factor)
10070 * Now check that the number of steps evenly divides a power
10071 * of the factor. (This assures both integer bucket size and
10072 * linearity within each magnitude.)
10074 for (v = factor; v < nsteps; v *= factor)
10077 if ((v % nsteps) || (nsteps % factor))
10080 size = (dtrace_aggregate_llquantize_bucket(factor,
10081 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10085 case DTRACEAGG_AVG:
10086 agg->dtag_aggregate = dtrace_aggregate_avg;
10087 size = sizeof (uint64_t) * 2;
10090 case DTRACEAGG_STDDEV:
10091 agg->dtag_aggregate = dtrace_aggregate_stddev;
10092 size = sizeof (uint64_t) * 4;
10095 case DTRACEAGG_SUM:
10096 agg->dtag_aggregate = dtrace_aggregate_sum;
10103 agg->dtag_action.dta_rec.dtrd_size = size;
10109 * We must make sure that we have enough actions for the n-tuple.
10111 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10112 if (DTRACEACT_ISAGG(act->dta_kind))
10115 if (--ntuple == 0) {
10117 * This is the action with which our n-tuple begins.
10119 agg->dtag_first = act;
10125 * This n-tuple is short by ntuple elements. Return failure.
10127 ASSERT(ntuple != 0);
10129 kmem_free(agg, sizeof (dtrace_aggregation_t));
10134 * If the last action in the tuple has a size of zero, it's actually
10135 * an expression argument for the aggregating action.
10137 ASSERT(ecb->dte_action_last != NULL);
10138 act = ecb->dte_action_last;
10140 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10141 ASSERT(act->dta_difo != NULL);
10143 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10144 agg->dtag_hasarg = 1;
10148 * We need to allocate an id for this aggregation.
10151 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10152 VM_BESTFIT | VM_SLEEP);
10154 aggid = alloc_unr(state->dts_aggid_arena);
10157 if (aggid - 1 >= state->dts_naggregations) {
10158 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10159 dtrace_aggregation_t **aggs;
10160 int naggs = state->dts_naggregations << 1;
10161 int onaggs = state->dts_naggregations;
10163 ASSERT(aggid == state->dts_naggregations + 1);
10166 ASSERT(oaggs == NULL);
10170 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10172 if (oaggs != NULL) {
10173 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10174 kmem_free(oaggs, onaggs * sizeof (*aggs));
10177 state->dts_aggregations = aggs;
10178 state->dts_naggregations = naggs;
10181 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10182 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10184 frec = &agg->dtag_first->dta_rec;
10185 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10186 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10188 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10189 ASSERT(!act->dta_intuple);
10190 act->dta_intuple = 1;
10193 return (&agg->dtag_action);
10197 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10199 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10200 dtrace_state_t *state = ecb->dte_state;
10201 dtrace_aggid_t aggid = agg->dtag_id;
10203 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10205 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10207 free_unr(state->dts_aggid_arena, aggid);
10210 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10211 state->dts_aggregations[aggid - 1] = NULL;
10213 kmem_free(agg, sizeof (dtrace_aggregation_t));
10217 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10219 dtrace_action_t *action, *last;
10220 dtrace_difo_t *dp = desc->dtad_difo;
10221 uint32_t size = 0, align = sizeof (uint8_t), mask;
10222 uint16_t format = 0;
10223 dtrace_recdesc_t *rec;
10224 dtrace_state_t *state = ecb->dte_state;
10225 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
10226 uint64_t arg = desc->dtad_arg;
10228 ASSERT(MUTEX_HELD(&dtrace_lock));
10229 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10231 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10233 * If this is an aggregating action, there must be neither
10234 * a speculate nor a commit on the action chain.
10236 dtrace_action_t *act;
10238 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10239 if (act->dta_kind == DTRACEACT_COMMIT)
10242 if (act->dta_kind == DTRACEACT_SPECULATE)
10246 action = dtrace_ecb_aggregation_create(ecb, desc);
10248 if (action == NULL)
10251 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10252 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10253 dp != NULL && dp->dtdo_destructive)) {
10254 state->dts_destructive = 1;
10257 switch (desc->dtad_kind) {
10258 case DTRACEACT_PRINTF:
10259 case DTRACEACT_PRINTA:
10260 case DTRACEACT_SYSTEM:
10261 case DTRACEACT_FREOPEN:
10262 case DTRACEACT_DIFEXPR:
10264 * We know that our arg is a string -- turn it into a
10268 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10269 desc->dtad_kind == DTRACEACT_DIFEXPR);
10274 ASSERT(arg > KERNELBASE);
10276 format = dtrace_format_add(state,
10277 (char *)(uintptr_t)arg);
10281 case DTRACEACT_LIBACT:
10282 case DTRACEACT_TRACEMEM:
10283 case DTRACEACT_TRACEMEM_DYNSIZE:
10287 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10290 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10291 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10294 size = opt[DTRACEOPT_STRSIZE];
10299 case DTRACEACT_STACK:
10300 if ((nframes = arg) == 0) {
10301 nframes = opt[DTRACEOPT_STACKFRAMES];
10302 ASSERT(nframes > 0);
10306 size = nframes * sizeof (pc_t);
10309 case DTRACEACT_JSTACK:
10310 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10311 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10313 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10314 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10316 arg = DTRACE_USTACK_ARG(nframes, strsize);
10319 case DTRACEACT_USTACK:
10320 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10321 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10322 strsize = DTRACE_USTACK_STRSIZE(arg);
10323 nframes = opt[DTRACEOPT_USTACKFRAMES];
10324 ASSERT(nframes > 0);
10325 arg = DTRACE_USTACK_ARG(nframes, strsize);
10329 * Save a slot for the pid.
10331 size = (nframes + 1) * sizeof (uint64_t);
10332 size += DTRACE_USTACK_STRSIZE(arg);
10333 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10337 case DTRACEACT_SYM:
10338 case DTRACEACT_MOD:
10339 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10340 sizeof (uint64_t)) ||
10341 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10345 case DTRACEACT_USYM:
10346 case DTRACEACT_UMOD:
10347 case DTRACEACT_UADDR:
10349 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10350 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10354 * We have a slot for the pid, plus a slot for the
10355 * argument. To keep things simple (aligned with
10356 * bitness-neutral sizing), we store each as a 64-bit
10359 size = 2 * sizeof (uint64_t);
10362 case DTRACEACT_STOP:
10363 case DTRACEACT_BREAKPOINT:
10364 case DTRACEACT_PANIC:
10367 case DTRACEACT_CHILL:
10368 case DTRACEACT_DISCARD:
10369 case DTRACEACT_RAISE:
10374 case DTRACEACT_EXIT:
10376 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10377 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10381 case DTRACEACT_SPECULATE:
10382 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10388 state->dts_speculates = 1;
10391 case DTRACEACT_PRINTM:
10392 size = dp->dtdo_rtype.dtdt_size;
10395 case DTRACEACT_PRINTT:
10396 size = dp->dtdo_rtype.dtdt_size;
10399 case DTRACEACT_COMMIT: {
10400 dtrace_action_t *act = ecb->dte_action;
10402 for (; act != NULL; act = act->dta_next) {
10403 if (act->dta_kind == DTRACEACT_COMMIT)
10416 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10418 * If this is a data-storing action or a speculate,
10419 * we must be sure that there isn't a commit on the
10422 dtrace_action_t *act = ecb->dte_action;
10424 for (; act != NULL; act = act->dta_next) {
10425 if (act->dta_kind == DTRACEACT_COMMIT)
10430 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10431 action->dta_rec.dtrd_size = size;
10434 action->dta_refcnt = 1;
10435 rec = &action->dta_rec;
10436 size = rec->dtrd_size;
10438 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10439 if (!(size & mask)) {
10445 action->dta_kind = desc->dtad_kind;
10447 if ((action->dta_difo = dp) != NULL)
10448 dtrace_difo_hold(dp);
10450 rec->dtrd_action = action->dta_kind;
10451 rec->dtrd_arg = arg;
10452 rec->dtrd_uarg = desc->dtad_uarg;
10453 rec->dtrd_alignment = (uint16_t)align;
10454 rec->dtrd_format = format;
10456 if ((last = ecb->dte_action_last) != NULL) {
10457 ASSERT(ecb->dte_action != NULL);
10458 action->dta_prev = last;
10459 last->dta_next = action;
10461 ASSERT(ecb->dte_action == NULL);
10462 ecb->dte_action = action;
10465 ecb->dte_action_last = action;
10471 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10473 dtrace_action_t *act = ecb->dte_action, *next;
10474 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10478 if (act != NULL && act->dta_refcnt > 1) {
10479 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10482 for (; act != NULL; act = next) {
10483 next = act->dta_next;
10484 ASSERT(next != NULL || act == ecb->dte_action_last);
10485 ASSERT(act->dta_refcnt == 1);
10487 if ((format = act->dta_rec.dtrd_format) != 0)
10488 dtrace_format_remove(ecb->dte_state, format);
10490 if ((dp = act->dta_difo) != NULL)
10491 dtrace_difo_release(dp, vstate);
10493 if (DTRACEACT_ISAGG(act->dta_kind)) {
10494 dtrace_ecb_aggregation_destroy(ecb, act);
10496 kmem_free(act, sizeof (dtrace_action_t));
10501 ecb->dte_action = NULL;
10502 ecb->dte_action_last = NULL;
10507 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10510 * We disable the ECB by removing it from its probe.
10512 dtrace_ecb_t *pecb, *prev = NULL;
10513 dtrace_probe_t *probe = ecb->dte_probe;
10515 ASSERT(MUTEX_HELD(&dtrace_lock));
10517 if (probe == NULL) {
10519 * This is the NULL probe; there is nothing to disable.
10524 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10530 ASSERT(pecb != NULL);
10532 if (prev == NULL) {
10533 probe->dtpr_ecb = ecb->dte_next;
10535 prev->dte_next = ecb->dte_next;
10538 if (ecb == probe->dtpr_ecb_last) {
10539 ASSERT(ecb->dte_next == NULL);
10540 probe->dtpr_ecb_last = prev;
10544 * The ECB has been disconnected from the probe; now sync to assure
10545 * that all CPUs have seen the change before returning.
10549 if (probe->dtpr_ecb == NULL) {
10551 * That was the last ECB on the probe; clear the predicate
10552 * cache ID for the probe, disable it and sync one more time
10553 * to assure that we'll never hit it again.
10555 dtrace_provider_t *prov = probe->dtpr_provider;
10557 ASSERT(ecb->dte_next == NULL);
10558 ASSERT(probe->dtpr_ecb_last == NULL);
10559 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10560 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10561 probe->dtpr_id, probe->dtpr_arg);
10565 * There is at least one ECB remaining on the probe. If there
10566 * is _exactly_ one, set the probe's predicate cache ID to be
10567 * the predicate cache ID of the remaining ECB.
10569 ASSERT(probe->dtpr_ecb_last != NULL);
10570 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10572 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10573 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10575 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10578 probe->dtpr_predcache = p->dtp_cacheid;
10581 ecb->dte_next = NULL;
10586 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10588 dtrace_state_t *state = ecb->dte_state;
10589 dtrace_vstate_t *vstate = &state->dts_vstate;
10590 dtrace_predicate_t *pred;
10591 dtrace_epid_t epid = ecb->dte_epid;
10593 ASSERT(MUTEX_HELD(&dtrace_lock));
10594 ASSERT(ecb->dte_next == NULL);
10595 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10597 if ((pred = ecb->dte_predicate) != NULL)
10598 dtrace_predicate_release(pred, vstate);
10600 dtrace_ecb_action_remove(ecb);
10602 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10603 state->dts_ecbs[epid - 1] = NULL;
10605 kmem_free(ecb, sizeof (dtrace_ecb_t));
10608 static dtrace_ecb_t *
10609 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10610 dtrace_enabling_t *enab)
10613 dtrace_predicate_t *pred;
10614 dtrace_actdesc_t *act;
10615 dtrace_provider_t *prov;
10616 dtrace_ecbdesc_t *desc = enab->dten_current;
10618 ASSERT(MUTEX_HELD(&dtrace_lock));
10619 ASSERT(state != NULL);
10621 ecb = dtrace_ecb_add(state, probe);
10622 ecb->dte_uarg = desc->dted_uarg;
10624 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10625 dtrace_predicate_hold(pred);
10626 ecb->dte_predicate = pred;
10629 if (probe != NULL) {
10631 * If the provider shows more leg than the consumer is old
10632 * enough to see, we need to enable the appropriate implicit
10633 * predicate bits to prevent the ecb from activating at
10636 * Providers specifying DTRACE_PRIV_USER at register time
10637 * are stating that they need the /proc-style privilege
10638 * model to be enforced, and this is what DTRACE_COND_OWNER
10639 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10641 prov = probe->dtpr_provider;
10642 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10643 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10644 ecb->dte_cond |= DTRACE_COND_OWNER;
10646 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10647 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10648 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10651 * If the provider shows us kernel innards and the user
10652 * is lacking sufficient privilege, enable the
10653 * DTRACE_COND_USERMODE implicit predicate.
10655 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10656 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10657 ecb->dte_cond |= DTRACE_COND_USERMODE;
10660 if (dtrace_ecb_create_cache != NULL) {
10662 * If we have a cached ecb, we'll use its action list instead
10663 * of creating our own (saving both time and space).
10665 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10666 dtrace_action_t *act = cached->dte_action;
10669 ASSERT(act->dta_refcnt > 0);
10671 ecb->dte_action = act;
10672 ecb->dte_action_last = cached->dte_action_last;
10673 ecb->dte_needed = cached->dte_needed;
10674 ecb->dte_size = cached->dte_size;
10675 ecb->dte_alignment = cached->dte_alignment;
10681 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10682 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10683 dtrace_ecb_destroy(ecb);
10688 dtrace_ecb_resize(ecb);
10690 return (dtrace_ecb_create_cache = ecb);
10694 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10697 dtrace_enabling_t *enab = arg;
10698 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10700 ASSERT(state != NULL);
10702 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10704 * This probe was created in a generation for which this
10705 * enabling has previously created ECBs; we don't want to
10706 * enable it again, so just kick out.
10708 return (DTRACE_MATCH_NEXT);
10711 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10712 return (DTRACE_MATCH_DONE);
10714 dtrace_ecb_enable(ecb);
10715 return (DTRACE_MATCH_NEXT);
10718 static dtrace_ecb_t *
10719 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10723 ASSERT(MUTEX_HELD(&dtrace_lock));
10725 if (id == 0 || id > state->dts_necbs)
10728 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10729 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10731 return (state->dts_ecbs[id - 1]);
10734 static dtrace_aggregation_t *
10735 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10737 dtrace_aggregation_t *agg;
10739 ASSERT(MUTEX_HELD(&dtrace_lock));
10741 if (id == 0 || id > state->dts_naggregations)
10744 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10745 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10746 agg->dtag_id == id);
10748 return (state->dts_aggregations[id - 1]);
10752 * DTrace Buffer Functions
10754 * The following functions manipulate DTrace buffers. Most of these functions
10755 * are called in the context of establishing or processing consumer state;
10756 * exceptions are explicitly noted.
10760 * Note: called from cross call context. This function switches the two
10761 * buffers on a given CPU. The atomicity of this operation is assured by
10762 * disabling interrupts while the actual switch takes place; the disabling of
10763 * interrupts serializes the execution with any execution of dtrace_probe() on
10767 dtrace_buffer_switch(dtrace_buffer_t *buf)
10769 caddr_t tomax = buf->dtb_tomax;
10770 caddr_t xamot = buf->dtb_xamot;
10771 dtrace_icookie_t cookie;
10774 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10775 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10777 cookie = dtrace_interrupt_disable();
10778 now = dtrace_gethrtime();
10779 buf->dtb_tomax = xamot;
10780 buf->dtb_xamot = tomax;
10781 buf->dtb_xamot_drops = buf->dtb_drops;
10782 buf->dtb_xamot_offset = buf->dtb_offset;
10783 buf->dtb_xamot_errors = buf->dtb_errors;
10784 buf->dtb_xamot_flags = buf->dtb_flags;
10785 buf->dtb_offset = 0;
10786 buf->dtb_drops = 0;
10787 buf->dtb_errors = 0;
10788 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10789 buf->dtb_interval = now - buf->dtb_switched;
10790 buf->dtb_switched = now;
10791 dtrace_interrupt_enable(cookie);
10795 * Note: called from cross call context. This function activates a buffer
10796 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10797 * is guaranteed by the disabling of interrupts.
10800 dtrace_buffer_activate(dtrace_state_t *state)
10802 dtrace_buffer_t *buf;
10803 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10805 buf = &state->dts_buffer[curcpu];
10807 if (buf->dtb_tomax != NULL) {
10809 * We might like to assert that the buffer is marked inactive,
10810 * but this isn't necessarily true: the buffer for the CPU
10811 * that processes the BEGIN probe has its buffer activated
10812 * manually. In this case, we take the (harmless) action
10813 * re-clearing the bit INACTIVE bit.
10815 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10818 dtrace_interrupt_enable(cookie);
10822 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10828 dtrace_buffer_t *buf;
10831 ASSERT(MUTEX_HELD(&cpu_lock));
10832 ASSERT(MUTEX_HELD(&dtrace_lock));
10834 if (size > dtrace_nonroot_maxsize &&
10835 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10841 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10844 buf = &bufs[cp->cpu_id];
10847 * If there is already a buffer allocated for this CPU, it
10848 * is only possible that this is a DR event. In this case,
10850 if (buf->dtb_tomax != NULL) {
10851 ASSERT(buf->dtb_size == size);
10855 ASSERT(buf->dtb_xamot == NULL);
10857 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10860 buf->dtb_size = size;
10861 buf->dtb_flags = flags;
10862 buf->dtb_offset = 0;
10863 buf->dtb_drops = 0;
10865 if (flags & DTRACEBUF_NOSWITCH)
10868 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10870 } while ((cp = cp->cpu_next) != cpu_list);
10878 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10881 buf = &bufs[cp->cpu_id];
10883 if (buf->dtb_xamot != NULL) {
10884 ASSERT(buf->dtb_tomax != NULL);
10885 ASSERT(buf->dtb_size == size);
10886 kmem_free(buf->dtb_xamot, size);
10889 if (buf->dtb_tomax != NULL) {
10890 ASSERT(buf->dtb_size == size);
10891 kmem_free(buf->dtb_tomax, size);
10894 buf->dtb_tomax = NULL;
10895 buf->dtb_xamot = NULL;
10897 } while ((cp = cp->cpu_next) != cpu_list);
10903 #if defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
10905 * FreeBSD isn't good at limiting the amount of memory we
10906 * ask to malloc, so let's place a limit here before trying
10907 * to do something that might well end in tears at bedtime.
10909 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10913 ASSERT(MUTEX_HELD(&dtrace_lock));
10915 if (cpu != DTRACE_CPUALL && cpu != i)
10921 * If there is already a buffer allocated for this CPU, it
10922 * is only possible that this is a DR event. In this case,
10923 * the buffer size must match our specified size.
10925 if (buf->dtb_tomax != NULL) {
10926 ASSERT(buf->dtb_size == size);
10930 ASSERT(buf->dtb_xamot == NULL);
10932 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10935 buf->dtb_size = size;
10936 buf->dtb_flags = flags;
10937 buf->dtb_offset = 0;
10938 buf->dtb_drops = 0;
10940 if (flags & DTRACEBUF_NOSWITCH)
10943 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10951 * Error allocating memory, so free the buffers that were
10952 * allocated before the failed allocation.
10955 if (cpu != DTRACE_CPUALL && cpu != i)
10960 if (buf->dtb_xamot != NULL) {
10961 ASSERT(buf->dtb_tomax != NULL);
10962 ASSERT(buf->dtb_size == size);
10963 kmem_free(buf->dtb_xamot, size);
10966 if (buf->dtb_tomax != NULL) {
10967 ASSERT(buf->dtb_size == size);
10968 kmem_free(buf->dtb_tomax, size);
10971 buf->dtb_tomax = NULL;
10972 buf->dtb_xamot = NULL;
10982 * Note: called from probe context. This function just increments the drop
10983 * count on a buffer. It has been made a function to allow for the
10984 * possibility of understanding the source of mysterious drop counts. (A
10985 * problem for which one may be particularly disappointed that DTrace cannot
10986 * be used to understand DTrace.)
10989 dtrace_buffer_drop(dtrace_buffer_t *buf)
10995 * Note: called from probe context. This function is called to reserve space
10996 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
10997 * mstate. Returns the new offset in the buffer, or a negative value if an
10998 * error has occurred.
11001 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11002 dtrace_state_t *state, dtrace_mstate_t *mstate)
11004 intptr_t offs = buf->dtb_offset, soffs;
11009 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11012 if ((tomax = buf->dtb_tomax) == NULL) {
11013 dtrace_buffer_drop(buf);
11017 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11018 while (offs & (align - 1)) {
11020 * Assert that our alignment is off by a number which
11021 * is itself sizeof (uint32_t) aligned.
11023 ASSERT(!((align - (offs & (align - 1))) &
11024 (sizeof (uint32_t) - 1)));
11025 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11026 offs += sizeof (uint32_t);
11029 if ((soffs = offs + needed) > buf->dtb_size) {
11030 dtrace_buffer_drop(buf);
11034 if (mstate == NULL)
11037 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11038 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11039 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11044 if (buf->dtb_flags & DTRACEBUF_FILL) {
11045 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11046 (buf->dtb_flags & DTRACEBUF_FULL))
11051 total = needed + (offs & (align - 1));
11054 * For a ring buffer, life is quite a bit more complicated. Before
11055 * we can store any padding, we need to adjust our wrapping offset.
11056 * (If we've never before wrapped or we're not about to, no adjustment
11059 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11060 offs + total > buf->dtb_size) {
11061 woffs = buf->dtb_xamot_offset;
11063 if (offs + total > buf->dtb_size) {
11065 * We can't fit in the end of the buffer. First, a
11066 * sanity check that we can fit in the buffer at all.
11068 if (total > buf->dtb_size) {
11069 dtrace_buffer_drop(buf);
11074 * We're going to be storing at the top of the buffer,
11075 * so now we need to deal with the wrapped offset. We
11076 * only reset our wrapped offset to 0 if it is
11077 * currently greater than the current offset. If it
11078 * is less than the current offset, it is because a
11079 * previous allocation induced a wrap -- but the
11080 * allocation didn't subsequently take the space due
11081 * to an error or false predicate evaluation. In this
11082 * case, we'll just leave the wrapped offset alone: if
11083 * the wrapped offset hasn't been advanced far enough
11084 * for this allocation, it will be adjusted in the
11087 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11095 * Now we know that we're going to be storing to the
11096 * top of the buffer and that there is room for us
11097 * there. We need to clear the buffer from the current
11098 * offset to the end (there may be old gunk there).
11100 while (offs < buf->dtb_size)
11104 * We need to set our offset to zero. And because we
11105 * are wrapping, we need to set the bit indicating as
11106 * much. We can also adjust our needed space back
11107 * down to the space required by the ECB -- we know
11108 * that the top of the buffer is aligned.
11112 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11115 * There is room for us in the buffer, so we simply
11116 * need to check the wrapped offset.
11118 if (woffs < offs) {
11120 * The wrapped offset is less than the offset.
11121 * This can happen if we allocated buffer space
11122 * that induced a wrap, but then we didn't
11123 * subsequently take the space due to an error
11124 * or false predicate evaluation. This is
11125 * okay; we know that _this_ allocation isn't
11126 * going to induce a wrap. We still can't
11127 * reset the wrapped offset to be zero,
11128 * however: the space may have been trashed in
11129 * the previous failed probe attempt. But at
11130 * least the wrapped offset doesn't need to
11131 * be adjusted at all...
11137 while (offs + total > woffs) {
11138 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11141 if (epid == DTRACE_EPIDNONE) {
11142 size = sizeof (uint32_t);
11144 ASSERT3U(epid, <=, state->dts_necbs);
11145 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11147 size = state->dts_ecbs[epid - 1]->dte_size;
11150 ASSERT(woffs + size <= buf->dtb_size);
11153 if (woffs + size == buf->dtb_size) {
11155 * We've reached the end of the buffer; we want
11156 * to set the wrapped offset to 0 and break
11157 * out. However, if the offs is 0, then we're
11158 * in a strange edge-condition: the amount of
11159 * space that we want to reserve plus the size
11160 * of the record that we're overwriting is
11161 * greater than the size of the buffer. This
11162 * is problematic because if we reserve the
11163 * space but subsequently don't consume it (due
11164 * to a failed predicate or error) the wrapped
11165 * offset will be 0 -- yet the EPID at offset 0
11166 * will not be committed. This situation is
11167 * relatively easy to deal with: if we're in
11168 * this case, the buffer is indistinguishable
11169 * from one that hasn't wrapped; we need only
11170 * finish the job by clearing the wrapped bit,
11171 * explicitly setting the offset to be 0, and
11172 * zero'ing out the old data in the buffer.
11175 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11176 buf->dtb_offset = 0;
11179 while (woffs < buf->dtb_size)
11180 tomax[woffs++] = 0;
11191 * We have a wrapped offset. It may be that the wrapped offset
11192 * has become zero -- that's okay.
11194 buf->dtb_xamot_offset = woffs;
11199 * Now we can plow the buffer with any necessary padding.
11201 while (offs & (align - 1)) {
11203 * Assert that our alignment is off by a number which
11204 * is itself sizeof (uint32_t) aligned.
11206 ASSERT(!((align - (offs & (align - 1))) &
11207 (sizeof (uint32_t) - 1)));
11208 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11209 offs += sizeof (uint32_t);
11212 if (buf->dtb_flags & DTRACEBUF_FILL) {
11213 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11214 buf->dtb_flags |= DTRACEBUF_FULL;
11219 if (mstate == NULL)
11223 * For ring buffers and fill buffers, the scratch space is always
11224 * the inactive buffer.
11226 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11227 mstate->dtms_scratch_size = buf->dtb_size;
11228 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11234 dtrace_buffer_polish(dtrace_buffer_t *buf)
11236 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11237 ASSERT(MUTEX_HELD(&dtrace_lock));
11239 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11243 * We need to polish the ring buffer. There are three cases:
11245 * - The first (and presumably most common) is that there is no gap
11246 * between the buffer offset and the wrapped offset. In this case,
11247 * there is nothing in the buffer that isn't valid data; we can
11248 * mark the buffer as polished and return.
11250 * - The second (less common than the first but still more common
11251 * than the third) is that there is a gap between the buffer offset
11252 * and the wrapped offset, and the wrapped offset is larger than the
11253 * buffer offset. This can happen because of an alignment issue, or
11254 * can happen because of a call to dtrace_buffer_reserve() that
11255 * didn't subsequently consume the buffer space. In this case,
11256 * we need to zero the data from the buffer offset to the wrapped
11259 * - The third (and least common) is that there is a gap between the
11260 * buffer offset and the wrapped offset, but the wrapped offset is
11261 * _less_ than the buffer offset. This can only happen because a
11262 * call to dtrace_buffer_reserve() induced a wrap, but the space
11263 * was not subsequently consumed. In this case, we need to zero the
11264 * space from the offset to the end of the buffer _and_ from the
11265 * top of the buffer to the wrapped offset.
11267 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11268 bzero(buf->dtb_tomax + buf->dtb_offset,
11269 buf->dtb_xamot_offset - buf->dtb_offset);
11272 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11273 bzero(buf->dtb_tomax + buf->dtb_offset,
11274 buf->dtb_size - buf->dtb_offset);
11275 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11280 * This routine determines if data generated at the specified time has likely
11281 * been entirely consumed at user-level. This routine is called to determine
11282 * if an ECB on a defunct probe (but for an active enabling) can be safely
11283 * disabled and destroyed.
11286 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11290 for (i = 0; i < NCPU; i++) {
11291 dtrace_buffer_t *buf = &bufs[i];
11293 if (buf->dtb_size == 0)
11296 if (buf->dtb_flags & DTRACEBUF_RING)
11299 if (!buf->dtb_switched && buf->dtb_offset != 0)
11302 if (buf->dtb_switched - buf->dtb_interval < when)
11310 dtrace_buffer_free(dtrace_buffer_t *bufs)
11314 for (i = 0; i < NCPU; i++) {
11315 dtrace_buffer_t *buf = &bufs[i];
11317 if (buf->dtb_tomax == NULL) {
11318 ASSERT(buf->dtb_xamot == NULL);
11319 ASSERT(buf->dtb_size == 0);
11323 if (buf->dtb_xamot != NULL) {
11324 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11325 kmem_free(buf->dtb_xamot, buf->dtb_size);
11328 kmem_free(buf->dtb_tomax, buf->dtb_size);
11330 buf->dtb_tomax = NULL;
11331 buf->dtb_xamot = NULL;
11336 * DTrace Enabling Functions
11338 static dtrace_enabling_t *
11339 dtrace_enabling_create(dtrace_vstate_t *vstate)
11341 dtrace_enabling_t *enab;
11343 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11344 enab->dten_vstate = vstate;
11350 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11352 dtrace_ecbdesc_t **ndesc;
11353 size_t osize, nsize;
11356 * We can't add to enablings after we've enabled them, or after we've
11359 ASSERT(enab->dten_probegen == 0);
11360 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11362 if (enab->dten_ndesc < enab->dten_maxdesc) {
11363 enab->dten_desc[enab->dten_ndesc++] = ecb;
11367 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11369 if (enab->dten_maxdesc == 0) {
11370 enab->dten_maxdesc = 1;
11372 enab->dten_maxdesc <<= 1;
11375 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11377 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11378 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11379 bcopy(enab->dten_desc, ndesc, osize);
11380 if (enab->dten_desc != NULL)
11381 kmem_free(enab->dten_desc, osize);
11383 enab->dten_desc = ndesc;
11384 enab->dten_desc[enab->dten_ndesc++] = ecb;
11388 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11389 dtrace_probedesc_t *pd)
11391 dtrace_ecbdesc_t *new;
11392 dtrace_predicate_t *pred;
11393 dtrace_actdesc_t *act;
11396 * We're going to create a new ECB description that matches the
11397 * specified ECB in every way, but has the specified probe description.
11399 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11401 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11402 dtrace_predicate_hold(pred);
11404 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11405 dtrace_actdesc_hold(act);
11407 new->dted_action = ecb->dted_action;
11408 new->dted_pred = ecb->dted_pred;
11409 new->dted_probe = *pd;
11410 new->dted_uarg = ecb->dted_uarg;
11412 dtrace_enabling_add(enab, new);
11416 dtrace_enabling_dump(dtrace_enabling_t *enab)
11420 for (i = 0; i < enab->dten_ndesc; i++) {
11421 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11423 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11424 desc->dtpd_provider, desc->dtpd_mod,
11425 desc->dtpd_func, desc->dtpd_name);
11430 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11433 dtrace_ecbdesc_t *ep;
11434 dtrace_vstate_t *vstate = enab->dten_vstate;
11436 ASSERT(MUTEX_HELD(&dtrace_lock));
11438 for (i = 0; i < enab->dten_ndesc; i++) {
11439 dtrace_actdesc_t *act, *next;
11440 dtrace_predicate_t *pred;
11442 ep = enab->dten_desc[i];
11444 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11445 dtrace_predicate_release(pred, vstate);
11447 for (act = ep->dted_action; act != NULL; act = next) {
11448 next = act->dtad_next;
11449 dtrace_actdesc_release(act, vstate);
11452 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11455 if (enab->dten_desc != NULL)
11456 kmem_free(enab->dten_desc,
11457 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11460 * If this was a retained enabling, decrement the dts_nretained count
11461 * and take it off of the dtrace_retained list.
11463 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11464 dtrace_retained == enab) {
11465 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11466 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11467 enab->dten_vstate->dtvs_state->dts_nretained--;
11470 if (enab->dten_prev == NULL) {
11471 if (dtrace_retained == enab) {
11472 dtrace_retained = enab->dten_next;
11474 if (dtrace_retained != NULL)
11475 dtrace_retained->dten_prev = NULL;
11478 ASSERT(enab != dtrace_retained);
11479 ASSERT(dtrace_retained != NULL);
11480 enab->dten_prev->dten_next = enab->dten_next;
11483 if (enab->dten_next != NULL) {
11484 ASSERT(dtrace_retained != NULL);
11485 enab->dten_next->dten_prev = enab->dten_prev;
11488 kmem_free(enab, sizeof (dtrace_enabling_t));
11492 dtrace_enabling_retain(dtrace_enabling_t *enab)
11494 dtrace_state_t *state;
11496 ASSERT(MUTEX_HELD(&dtrace_lock));
11497 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11498 ASSERT(enab->dten_vstate != NULL);
11500 state = enab->dten_vstate->dtvs_state;
11501 ASSERT(state != NULL);
11504 * We only allow each state to retain dtrace_retain_max enablings.
11506 if (state->dts_nretained >= dtrace_retain_max)
11509 state->dts_nretained++;
11511 if (dtrace_retained == NULL) {
11512 dtrace_retained = enab;
11516 enab->dten_next = dtrace_retained;
11517 dtrace_retained->dten_prev = enab;
11518 dtrace_retained = enab;
11524 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11525 dtrace_probedesc_t *create)
11527 dtrace_enabling_t *new, *enab;
11528 int found = 0, err = ENOENT;
11530 ASSERT(MUTEX_HELD(&dtrace_lock));
11531 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11532 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11533 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11534 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11536 new = dtrace_enabling_create(&state->dts_vstate);
11539 * Iterate over all retained enablings, looking for enablings that
11540 * match the specified state.
11542 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11546 * dtvs_state can only be NULL for helper enablings -- and
11547 * helper enablings can't be retained.
11549 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11551 if (enab->dten_vstate->dtvs_state != state)
11555 * Now iterate over each probe description; we're looking for
11556 * an exact match to the specified probe description.
11558 for (i = 0; i < enab->dten_ndesc; i++) {
11559 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11560 dtrace_probedesc_t *pd = &ep->dted_probe;
11562 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11565 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11568 if (strcmp(pd->dtpd_func, match->dtpd_func))
11571 if (strcmp(pd->dtpd_name, match->dtpd_name))
11575 * We have a winning probe! Add it to our growing
11579 dtrace_enabling_addlike(new, ep, create);
11583 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11584 dtrace_enabling_destroy(new);
11592 dtrace_enabling_retract(dtrace_state_t *state)
11594 dtrace_enabling_t *enab, *next;
11596 ASSERT(MUTEX_HELD(&dtrace_lock));
11599 * Iterate over all retained enablings, destroy the enablings retained
11600 * for the specified state.
11602 for (enab = dtrace_retained; enab != NULL; enab = next) {
11603 next = enab->dten_next;
11606 * dtvs_state can only be NULL for helper enablings -- and
11607 * helper enablings can't be retained.
11609 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11611 if (enab->dten_vstate->dtvs_state == state) {
11612 ASSERT(state->dts_nretained > 0);
11613 dtrace_enabling_destroy(enab);
11617 ASSERT(state->dts_nretained == 0);
11621 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11626 ASSERT(MUTEX_HELD(&cpu_lock));
11627 ASSERT(MUTEX_HELD(&dtrace_lock));
11629 for (i = 0; i < enab->dten_ndesc; i++) {
11630 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11632 enab->dten_current = ep;
11633 enab->dten_error = 0;
11635 matched += dtrace_probe_enable(&ep->dted_probe, enab);
11637 if (enab->dten_error != 0) {
11639 * If we get an error half-way through enabling the
11640 * probes, we kick out -- perhaps with some number of
11641 * them enabled. Leaving enabled probes enabled may
11642 * be slightly confusing for user-level, but we expect
11643 * that no one will attempt to actually drive on in
11644 * the face of such errors. If this is an anonymous
11645 * enabling (indicated with a NULL nmatched pointer),
11646 * we cmn_err() a message. We aren't expecting to
11647 * get such an error -- such as it can exist at all,
11648 * it would be a result of corrupted DOF in the driver
11651 if (nmatched == NULL) {
11652 cmn_err(CE_WARN, "dtrace_enabling_match() "
11653 "error on %p: %d", (void *)ep,
11657 return (enab->dten_error);
11661 enab->dten_probegen = dtrace_probegen;
11662 if (nmatched != NULL)
11663 *nmatched = matched;
11669 dtrace_enabling_matchall(void)
11671 dtrace_enabling_t *enab;
11673 mutex_enter(&cpu_lock);
11674 mutex_enter(&dtrace_lock);
11677 * Iterate over all retained enablings to see if any probes match
11678 * against them. We only perform this operation on enablings for which
11679 * we have sufficient permissions by virtue of being in the global zone
11680 * or in the same zone as the DTrace client. Because we can be called
11681 * after dtrace_detach() has been called, we cannot assert that there
11682 * are retained enablings. We can safely load from dtrace_retained,
11683 * however: the taskq_destroy() at the end of dtrace_detach() will
11684 * block pending our completion.
11686 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11688 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11690 if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11692 (void) dtrace_enabling_match(enab, NULL);
11695 mutex_exit(&dtrace_lock);
11696 mutex_exit(&cpu_lock);
11700 * If an enabling is to be enabled without having matched probes (that is, if
11701 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11702 * enabling must be _primed_ by creating an ECB for every ECB description.
11703 * This must be done to assure that we know the number of speculations, the
11704 * number of aggregations, the minimum buffer size needed, etc. before we
11705 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11706 * enabling any probes, we create ECBs for every ECB decription, but with a
11707 * NULL probe -- which is exactly what this function does.
11710 dtrace_enabling_prime(dtrace_state_t *state)
11712 dtrace_enabling_t *enab;
11715 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11716 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11718 if (enab->dten_vstate->dtvs_state != state)
11722 * We don't want to prime an enabling more than once, lest
11723 * we allow a malicious user to induce resource exhaustion.
11724 * (The ECBs that result from priming an enabling aren't
11725 * leaked -- but they also aren't deallocated until the
11726 * consumer state is destroyed.)
11728 if (enab->dten_primed)
11731 for (i = 0; i < enab->dten_ndesc; i++) {
11732 enab->dten_current = enab->dten_desc[i];
11733 (void) dtrace_probe_enable(NULL, enab);
11736 enab->dten_primed = 1;
11741 * Called to indicate that probes should be provided due to retained
11742 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11743 * must take an initial lap through the enabling calling the dtps_provide()
11744 * entry point explicitly to allow for autocreated probes.
11747 dtrace_enabling_provide(dtrace_provider_t *prv)
11750 dtrace_probedesc_t desc;
11752 ASSERT(MUTEX_HELD(&dtrace_lock));
11753 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11757 prv = dtrace_provider;
11761 dtrace_enabling_t *enab = dtrace_retained;
11762 void *parg = prv->dtpv_arg;
11764 for (; enab != NULL; enab = enab->dten_next) {
11765 for (i = 0; i < enab->dten_ndesc; i++) {
11766 desc = enab->dten_desc[i]->dted_probe;
11767 mutex_exit(&dtrace_lock);
11768 prv->dtpv_pops.dtps_provide(parg, &desc);
11769 mutex_enter(&dtrace_lock);
11772 } while (all && (prv = prv->dtpv_next) != NULL);
11774 mutex_exit(&dtrace_lock);
11775 dtrace_probe_provide(NULL, all ? NULL : prv);
11776 mutex_enter(&dtrace_lock);
11780 * Called to reap ECBs that are attached to probes from defunct providers.
11783 dtrace_enabling_reap(void)
11785 dtrace_provider_t *prov;
11786 dtrace_probe_t *probe;
11791 mutex_enter(&cpu_lock);
11792 mutex_enter(&dtrace_lock);
11794 for (i = 0; i < dtrace_nprobes; i++) {
11795 if ((probe = dtrace_probes[i]) == NULL)
11798 if (probe->dtpr_ecb == NULL)
11801 prov = probe->dtpr_provider;
11803 if ((when = prov->dtpv_defunct) == 0)
11807 * We have ECBs on a defunct provider: we want to reap these
11808 * ECBs to allow the provider to unregister. The destruction
11809 * of these ECBs must be done carefully: if we destroy the ECB
11810 * and the consumer later wishes to consume an EPID that
11811 * corresponds to the destroyed ECB (and if the EPID metadata
11812 * has not been previously consumed), the consumer will abort
11813 * processing on the unknown EPID. To reduce (but not, sadly,
11814 * eliminate) the possibility of this, we will only destroy an
11815 * ECB for a defunct provider if, for the state that
11816 * corresponds to the ECB:
11818 * (a) There is no speculative tracing (which can effectively
11819 * cache an EPID for an arbitrary amount of time).
11821 * (b) The principal buffers have been switched twice since the
11822 * provider became defunct.
11824 * (c) The aggregation buffers are of zero size or have been
11825 * switched twice since the provider became defunct.
11827 * We use dts_speculates to determine (a) and call a function
11828 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
11829 * that as soon as we've been unable to destroy one of the ECBs
11830 * associated with the probe, we quit trying -- reaping is only
11831 * fruitful in as much as we can destroy all ECBs associated
11832 * with the defunct provider's probes.
11834 while ((ecb = probe->dtpr_ecb) != NULL) {
11835 dtrace_state_t *state = ecb->dte_state;
11836 dtrace_buffer_t *buf = state->dts_buffer;
11837 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11839 if (state->dts_speculates)
11842 if (!dtrace_buffer_consumed(buf, when))
11845 if (!dtrace_buffer_consumed(aggbuf, when))
11848 dtrace_ecb_disable(ecb);
11849 ASSERT(probe->dtpr_ecb != ecb);
11850 dtrace_ecb_destroy(ecb);
11854 mutex_exit(&dtrace_lock);
11855 mutex_exit(&cpu_lock);
11859 * DTrace DOF Functions
11863 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11865 if (dtrace_err_verbose)
11866 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11868 #ifdef DTRACE_ERRDEBUG
11869 dtrace_errdebug(str);
11874 * Create DOF out of a currently enabled state. Right now, we only create
11875 * DOF containing the run-time options -- but this could be expanded to create
11876 * complete DOF representing the enabled state.
11879 dtrace_dof_create(dtrace_state_t *state)
11883 dof_optdesc_t *opt;
11884 int i, len = sizeof (dof_hdr_t) +
11885 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11886 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11888 ASSERT(MUTEX_HELD(&dtrace_lock));
11890 dof = kmem_zalloc(len, KM_SLEEP);
11891 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11892 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11893 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11894 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11896 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11897 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11898 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11899 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11900 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11901 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11903 dof->dofh_flags = 0;
11904 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11905 dof->dofh_secsize = sizeof (dof_sec_t);
11906 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11907 dof->dofh_secoff = sizeof (dof_hdr_t);
11908 dof->dofh_loadsz = len;
11909 dof->dofh_filesz = len;
11913 * Fill in the option section header...
11915 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11916 sec->dofs_type = DOF_SECT_OPTDESC;
11917 sec->dofs_align = sizeof (uint64_t);
11918 sec->dofs_flags = DOF_SECF_LOAD;
11919 sec->dofs_entsize = sizeof (dof_optdesc_t);
11921 opt = (dof_optdesc_t *)((uintptr_t)sec +
11922 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11924 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11925 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11927 for (i = 0; i < DTRACEOPT_MAX; i++) {
11928 opt[i].dofo_option = i;
11929 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11930 opt[i].dofo_value = state->dts_options[i];
11937 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11939 dof_hdr_t hdr, *dof;
11941 ASSERT(!MUTEX_HELD(&dtrace_lock));
11944 * First, we're going to copyin() the sizeof (dof_hdr_t).
11946 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11947 dtrace_dof_error(NULL, "failed to copyin DOF header");
11953 * Now we'll allocate the entire DOF and copy it in -- provided
11954 * that the length isn't outrageous.
11956 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11957 dtrace_dof_error(&hdr, "load size exceeds maximum");
11962 if (hdr.dofh_loadsz < sizeof (hdr)) {
11963 dtrace_dof_error(&hdr, "invalid load size");
11968 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11970 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
11971 kmem_free(dof, hdr.dofh_loadsz);
11980 static __inline uchar_t
11981 dtrace_dof_char(char c) {
12000 return (c - 'A' + 10);
12007 return (c - 'a' + 10);
12009 /* Should not reach here. */
12015 dtrace_dof_property(const char *name)
12019 unsigned int len, i;
12024 * Unfortunately, array of values in .conf files are always (and
12025 * only) interpreted to be integer arrays. We must read our DOF
12026 * as an integer array, and then squeeze it into a byte array.
12028 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12029 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12032 for (i = 0; i < len; i++)
12033 buf[i] = (uchar_t)(((int *)buf)[i]);
12035 if (len < sizeof (dof_hdr_t)) {
12036 ddi_prop_free(buf);
12037 dtrace_dof_error(NULL, "truncated header");
12041 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12042 ddi_prop_free(buf);
12043 dtrace_dof_error(NULL, "truncated DOF");
12047 if (loadsz >= dtrace_dof_maxsize) {
12048 ddi_prop_free(buf);
12049 dtrace_dof_error(NULL, "oversized DOF");
12053 dof = kmem_alloc(loadsz, KM_SLEEP);
12054 bcopy(buf, dof, loadsz);
12055 ddi_prop_free(buf);
12060 if ((p_env = getenv(name)) == NULL)
12063 len = strlen(p_env) / 2;
12065 buf = kmem_alloc(len, KM_SLEEP);
12067 dof = (dof_hdr_t *) buf;
12071 for (i = 0; i < len; i++) {
12072 buf[i] = (dtrace_dof_char(p[0]) << 4) |
12073 dtrace_dof_char(p[1]);
12079 if (len < sizeof (dof_hdr_t)) {
12081 dtrace_dof_error(NULL, "truncated header");
12085 if (len < (loadsz = dof->dofh_loadsz)) {
12087 dtrace_dof_error(NULL, "truncated DOF");
12091 if (loadsz >= dtrace_dof_maxsize) {
12093 dtrace_dof_error(NULL, "oversized DOF");
12102 dtrace_dof_destroy(dof_hdr_t *dof)
12104 kmem_free(dof, dof->dofh_loadsz);
12108 * Return the dof_sec_t pointer corresponding to a given section index. If the
12109 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12110 * a type other than DOF_SECT_NONE is specified, the header is checked against
12111 * this type and NULL is returned if the types do not match.
12114 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12116 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12117 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12119 if (i >= dof->dofh_secnum) {
12120 dtrace_dof_error(dof, "referenced section index is invalid");
12124 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12125 dtrace_dof_error(dof, "referenced section is not loadable");
12129 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12130 dtrace_dof_error(dof, "referenced section is the wrong type");
12137 static dtrace_probedesc_t *
12138 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12140 dof_probedesc_t *probe;
12142 uintptr_t daddr = (uintptr_t)dof;
12146 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12147 dtrace_dof_error(dof, "invalid probe section");
12151 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12152 dtrace_dof_error(dof, "bad alignment in probe description");
12156 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12157 dtrace_dof_error(dof, "truncated probe description");
12161 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12162 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12164 if (strtab == NULL)
12167 str = daddr + strtab->dofs_offset;
12168 size = strtab->dofs_size;
12170 if (probe->dofp_provider >= strtab->dofs_size) {
12171 dtrace_dof_error(dof, "corrupt probe provider");
12175 (void) strncpy(desc->dtpd_provider,
12176 (char *)(str + probe->dofp_provider),
12177 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12179 if (probe->dofp_mod >= strtab->dofs_size) {
12180 dtrace_dof_error(dof, "corrupt probe module");
12184 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12185 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12187 if (probe->dofp_func >= strtab->dofs_size) {
12188 dtrace_dof_error(dof, "corrupt probe function");
12192 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12193 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12195 if (probe->dofp_name >= strtab->dofs_size) {
12196 dtrace_dof_error(dof, "corrupt probe name");
12200 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12201 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12206 static dtrace_difo_t *
12207 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12212 dof_difohdr_t *dofd;
12213 uintptr_t daddr = (uintptr_t)dof;
12214 size_t max = dtrace_difo_maxsize;
12217 static const struct {
12225 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12226 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12227 sizeof (dif_instr_t), "multiple DIF sections" },
12229 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12230 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12231 sizeof (uint64_t), "multiple integer tables" },
12233 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12234 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12235 sizeof (char), "multiple string tables" },
12237 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12238 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12239 sizeof (uint_t), "multiple variable tables" },
12241 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
12244 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12245 dtrace_dof_error(dof, "invalid DIFO header section");
12249 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12250 dtrace_dof_error(dof, "bad alignment in DIFO header");
12254 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12255 sec->dofs_size % sizeof (dof_secidx_t)) {
12256 dtrace_dof_error(dof, "bad size in DIFO header");
12260 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12261 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12263 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12264 dp->dtdo_rtype = dofd->dofd_rtype;
12266 for (l = 0; l < n; l++) {
12271 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12272 dofd->dofd_links[l])) == NULL)
12273 goto err; /* invalid section link */
12275 if (ttl + subsec->dofs_size > max) {
12276 dtrace_dof_error(dof, "exceeds maximum size");
12280 ttl += subsec->dofs_size;
12282 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12283 if (subsec->dofs_type != difo[i].section)
12286 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12287 dtrace_dof_error(dof, "section not loaded");
12291 if (subsec->dofs_align != difo[i].align) {
12292 dtrace_dof_error(dof, "bad alignment");
12296 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12297 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12299 if (*bufp != NULL) {
12300 dtrace_dof_error(dof, difo[i].msg);
12304 if (difo[i].entsize != subsec->dofs_entsize) {
12305 dtrace_dof_error(dof, "entry size mismatch");
12309 if (subsec->dofs_entsize != 0 &&
12310 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12311 dtrace_dof_error(dof, "corrupt entry size");
12315 *lenp = subsec->dofs_size;
12316 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12317 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12318 *bufp, subsec->dofs_size);
12320 if (subsec->dofs_entsize != 0)
12321 *lenp /= subsec->dofs_entsize;
12327 * If we encounter a loadable DIFO sub-section that is not
12328 * known to us, assume this is a broken program and fail.
12330 if (difo[i].section == DOF_SECT_NONE &&
12331 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12332 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12337 if (dp->dtdo_buf == NULL) {
12339 * We can't have a DIF object without DIF text.
12341 dtrace_dof_error(dof, "missing DIF text");
12346 * Before we validate the DIF object, run through the variable table
12347 * looking for the strings -- if any of their size are under, we'll set
12348 * their size to be the system-wide default string size. Note that
12349 * this should _not_ happen if the "strsize" option has been set --
12350 * in this case, the compiler should have set the size to reflect the
12351 * setting of the option.
12353 for (i = 0; i < dp->dtdo_varlen; i++) {
12354 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12355 dtrace_diftype_t *t = &v->dtdv_type;
12357 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12360 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12361 t->dtdt_size = dtrace_strsize_default;
12364 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12367 dtrace_difo_init(dp, vstate);
12371 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12372 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12373 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12374 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12376 kmem_free(dp, sizeof (dtrace_difo_t));
12380 static dtrace_predicate_t *
12381 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12386 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12389 return (dtrace_predicate_create(dp));
12392 static dtrace_actdesc_t *
12393 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12396 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12397 dof_actdesc_t *desc;
12398 dof_sec_t *difosec;
12400 uintptr_t daddr = (uintptr_t)dof;
12402 dtrace_actkind_t kind;
12404 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12405 dtrace_dof_error(dof, "invalid action section");
12409 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12410 dtrace_dof_error(dof, "truncated action description");
12414 if (sec->dofs_align != sizeof (uint64_t)) {
12415 dtrace_dof_error(dof, "bad alignment in action description");
12419 if (sec->dofs_size < sec->dofs_entsize) {
12420 dtrace_dof_error(dof, "section entry size exceeds total size");
12424 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12425 dtrace_dof_error(dof, "bad entry size in action description");
12429 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12430 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12434 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12435 desc = (dof_actdesc_t *)(daddr +
12436 (uintptr_t)sec->dofs_offset + offs);
12437 kind = (dtrace_actkind_t)desc->dofa_kind;
12439 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12440 (kind != DTRACEACT_PRINTA ||
12441 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12442 (kind == DTRACEACT_DIFEXPR &&
12443 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12449 * The argument to these actions is an index into the
12450 * DOF string table. For printf()-like actions, this
12451 * is the format string. For print(), this is the
12452 * CTF type of the expression result.
12454 if ((strtab = dtrace_dof_sect(dof,
12455 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12458 str = (char *)((uintptr_t)dof +
12459 (uintptr_t)strtab->dofs_offset);
12461 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12462 if (str[i] == '\0')
12466 if (i >= strtab->dofs_size) {
12467 dtrace_dof_error(dof, "bogus format string");
12471 if (i == desc->dofa_arg) {
12472 dtrace_dof_error(dof, "empty format string");
12476 i -= desc->dofa_arg;
12477 fmt = kmem_alloc(i + 1, KM_SLEEP);
12478 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12479 arg = (uint64_t)(uintptr_t)fmt;
12481 if (kind == DTRACEACT_PRINTA) {
12482 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12485 arg = desc->dofa_arg;
12489 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12490 desc->dofa_uarg, arg);
12492 if (last != NULL) {
12493 last->dtad_next = act;
12500 if (desc->dofa_difo == DOF_SECIDX_NONE)
12503 if ((difosec = dtrace_dof_sect(dof,
12504 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12507 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12509 if (act->dtad_difo == NULL)
12513 ASSERT(first != NULL);
12517 for (act = first; act != NULL; act = next) {
12518 next = act->dtad_next;
12519 dtrace_actdesc_release(act, vstate);
12525 static dtrace_ecbdesc_t *
12526 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12529 dtrace_ecbdesc_t *ep;
12530 dof_ecbdesc_t *ecb;
12531 dtrace_probedesc_t *desc;
12532 dtrace_predicate_t *pred = NULL;
12534 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12535 dtrace_dof_error(dof, "truncated ECB description");
12539 if (sec->dofs_align != sizeof (uint64_t)) {
12540 dtrace_dof_error(dof, "bad alignment in ECB description");
12544 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12545 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12550 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12551 ep->dted_uarg = ecb->dofe_uarg;
12552 desc = &ep->dted_probe;
12554 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12557 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12558 if ((sec = dtrace_dof_sect(dof,
12559 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12562 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12565 ep->dted_pred.dtpdd_predicate = pred;
12568 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12569 if ((sec = dtrace_dof_sect(dof,
12570 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12573 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12575 if (ep->dted_action == NULL)
12583 dtrace_predicate_release(pred, vstate);
12584 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12589 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12590 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12591 * site of any user SETX relocations to account for load object base address.
12592 * In the future, if we need other relocations, this function can be extended.
12595 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12597 uintptr_t daddr = (uintptr_t)dof;
12598 dof_relohdr_t *dofr =
12599 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12600 dof_sec_t *ss, *rs, *ts;
12604 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12605 sec->dofs_align != sizeof (dof_secidx_t)) {
12606 dtrace_dof_error(dof, "invalid relocation header");
12610 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12611 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12612 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12614 if (ss == NULL || rs == NULL || ts == NULL)
12615 return (-1); /* dtrace_dof_error() has been called already */
12617 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12618 rs->dofs_align != sizeof (uint64_t)) {
12619 dtrace_dof_error(dof, "invalid relocation section");
12623 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12624 n = rs->dofs_size / rs->dofs_entsize;
12626 for (i = 0; i < n; i++) {
12627 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12629 switch (r->dofr_type) {
12630 case DOF_RELO_NONE:
12632 case DOF_RELO_SETX:
12633 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12634 sizeof (uint64_t) > ts->dofs_size) {
12635 dtrace_dof_error(dof, "bad relocation offset");
12639 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12640 dtrace_dof_error(dof, "misaligned setx relo");
12644 *(uint64_t *)taddr += ubase;
12647 dtrace_dof_error(dof, "invalid relocation type");
12651 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12658 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12659 * header: it should be at the front of a memory region that is at least
12660 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12661 * size. It need not be validated in any other way.
12664 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12665 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12667 uint64_t len = dof->dofh_loadsz, seclen;
12668 uintptr_t daddr = (uintptr_t)dof;
12669 dtrace_ecbdesc_t *ep;
12670 dtrace_enabling_t *enab;
12673 ASSERT(MUTEX_HELD(&dtrace_lock));
12674 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12677 * Check the DOF header identification bytes. In addition to checking
12678 * valid settings, we also verify that unused bits/bytes are zeroed so
12679 * we can use them later without fear of regressing existing binaries.
12681 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12682 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12683 dtrace_dof_error(dof, "DOF magic string mismatch");
12687 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12688 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12689 dtrace_dof_error(dof, "DOF has invalid data model");
12693 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12694 dtrace_dof_error(dof, "DOF encoding mismatch");
12698 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12699 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12700 dtrace_dof_error(dof, "DOF version mismatch");
12704 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12705 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12709 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12710 dtrace_dof_error(dof, "DOF uses too many integer registers");
12714 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12715 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12719 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12720 if (dof->dofh_ident[i] != 0) {
12721 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12726 if (dof->dofh_flags & ~DOF_FL_VALID) {
12727 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12731 if (dof->dofh_secsize == 0) {
12732 dtrace_dof_error(dof, "zero section header size");
12737 * Check that the section headers don't exceed the amount of DOF
12738 * data. Note that we cast the section size and number of sections
12739 * to uint64_t's to prevent possible overflow in the multiplication.
12741 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12743 if (dof->dofh_secoff > len || seclen > len ||
12744 dof->dofh_secoff + seclen > len) {
12745 dtrace_dof_error(dof, "truncated section headers");
12749 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12750 dtrace_dof_error(dof, "misaligned section headers");
12754 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12755 dtrace_dof_error(dof, "misaligned section size");
12760 * Take an initial pass through the section headers to be sure that
12761 * the headers don't have stray offsets. If the 'noprobes' flag is
12762 * set, do not permit sections relating to providers, probes, or args.
12764 for (i = 0; i < dof->dofh_secnum; i++) {
12765 dof_sec_t *sec = (dof_sec_t *)(daddr +
12766 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12769 switch (sec->dofs_type) {
12770 case DOF_SECT_PROVIDER:
12771 case DOF_SECT_PROBES:
12772 case DOF_SECT_PRARGS:
12773 case DOF_SECT_PROFFS:
12774 dtrace_dof_error(dof, "illegal sections "
12780 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12781 continue; /* just ignore non-loadable sections */
12783 if (sec->dofs_align & (sec->dofs_align - 1)) {
12784 dtrace_dof_error(dof, "bad section alignment");
12788 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12789 dtrace_dof_error(dof, "misaligned section");
12793 if (sec->dofs_offset > len || sec->dofs_size > len ||
12794 sec->dofs_offset + sec->dofs_size > len) {
12795 dtrace_dof_error(dof, "corrupt section header");
12799 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12800 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12801 dtrace_dof_error(dof, "non-terminating string table");
12807 * Take a second pass through the sections and locate and perform any
12808 * relocations that are present. We do this after the first pass to
12809 * be sure that all sections have had their headers validated.
12811 for (i = 0; i < dof->dofh_secnum; i++) {
12812 dof_sec_t *sec = (dof_sec_t *)(daddr +
12813 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12815 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12816 continue; /* skip sections that are not loadable */
12818 switch (sec->dofs_type) {
12819 case DOF_SECT_URELHDR:
12820 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12826 if ((enab = *enabp) == NULL)
12827 enab = *enabp = dtrace_enabling_create(vstate);
12829 for (i = 0; i < dof->dofh_secnum; i++) {
12830 dof_sec_t *sec = (dof_sec_t *)(daddr +
12831 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12833 if (sec->dofs_type != DOF_SECT_ECBDESC)
12836 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12837 dtrace_enabling_destroy(enab);
12842 dtrace_enabling_add(enab, ep);
12849 * Process DOF for any options. This routine assumes that the DOF has been
12850 * at least processed by dtrace_dof_slurp().
12853 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12858 dof_optdesc_t *desc;
12860 for (i = 0; i < dof->dofh_secnum; i++) {
12861 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12862 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12864 if (sec->dofs_type != DOF_SECT_OPTDESC)
12867 if (sec->dofs_align != sizeof (uint64_t)) {
12868 dtrace_dof_error(dof, "bad alignment in "
12869 "option description");
12873 if ((entsize = sec->dofs_entsize) == 0) {
12874 dtrace_dof_error(dof, "zeroed option entry size");
12878 if (entsize < sizeof (dof_optdesc_t)) {
12879 dtrace_dof_error(dof, "bad option entry size");
12883 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12884 desc = (dof_optdesc_t *)((uintptr_t)dof +
12885 (uintptr_t)sec->dofs_offset + offs);
12887 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12888 dtrace_dof_error(dof, "non-zero option string");
12892 if (desc->dofo_value == DTRACEOPT_UNSET) {
12893 dtrace_dof_error(dof, "unset option");
12897 if ((rval = dtrace_state_option(state,
12898 desc->dofo_option, desc->dofo_value)) != 0) {
12899 dtrace_dof_error(dof, "rejected option");
12909 * DTrace Consumer State Functions
12912 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12914 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12917 dtrace_dynvar_t *dvar, *next, *start;
12920 ASSERT(MUTEX_HELD(&dtrace_lock));
12921 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12923 bzero(dstate, sizeof (dtrace_dstate_t));
12925 if ((dstate->dtds_chunksize = chunksize) == 0)
12926 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12928 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12931 if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
12934 dstate->dtds_size = size;
12935 dstate->dtds_base = base;
12936 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12937 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12939 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12941 if (hashsize != 1 && (hashsize & 1))
12944 dstate->dtds_hashsize = hashsize;
12945 dstate->dtds_hash = dstate->dtds_base;
12948 * Set all of our hash buckets to point to the single sink, and (if
12949 * it hasn't already been set), set the sink's hash value to be the
12950 * sink sentinel value. The sink is needed for dynamic variable
12951 * lookups to know that they have iterated over an entire, valid hash
12954 for (i = 0; i < hashsize; i++)
12955 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12957 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12958 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12961 * Determine number of active CPUs. Divide free list evenly among
12964 start = (dtrace_dynvar_t *)
12965 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12966 limit = (uintptr_t)base + size;
12968 maxper = (limit - (uintptr_t)start) / NCPU;
12969 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12974 for (i = 0; i < NCPU; i++) {
12976 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12979 * If we don't even have enough chunks to make it once through
12980 * NCPUs, we're just going to allocate everything to the first
12981 * CPU. And if we're on the last CPU, we're going to allocate
12982 * whatever is left over. In either case, we set the limit to
12983 * be the limit of the dynamic variable space.
12985 if (maxper == 0 || i == NCPU - 1) {
12986 limit = (uintptr_t)base + size;
12989 limit = (uintptr_t)start + maxper;
12990 start = (dtrace_dynvar_t *)limit;
12993 ASSERT(limit <= (uintptr_t)base + size);
12996 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12997 dstate->dtds_chunksize);
12999 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13002 dvar->dtdv_next = next;
13014 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13016 ASSERT(MUTEX_HELD(&cpu_lock));
13018 if (dstate->dtds_base == NULL)
13021 kmem_free(dstate->dtds_base, dstate->dtds_size);
13022 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13026 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13029 * Logical XOR, where are you?
13031 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13033 if (vstate->dtvs_nglobals > 0) {
13034 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13035 sizeof (dtrace_statvar_t *));
13038 if (vstate->dtvs_ntlocals > 0) {
13039 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13040 sizeof (dtrace_difv_t));
13043 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13045 if (vstate->dtvs_nlocals > 0) {
13046 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13047 sizeof (dtrace_statvar_t *));
13053 dtrace_state_clean(dtrace_state_t *state)
13055 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13058 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13059 dtrace_speculation_clean(state);
13063 dtrace_state_deadman(dtrace_state_t *state)
13069 now = dtrace_gethrtime();
13071 if (state != dtrace_anon.dta_state &&
13072 now - state->dts_laststatus >= dtrace_deadman_user)
13076 * We must be sure that dts_alive never appears to be less than the
13077 * value upon entry to dtrace_state_deadman(), and because we lack a
13078 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13079 * store INT64_MAX to it, followed by a memory barrier, followed by
13080 * the new value. This assures that dts_alive never appears to be
13081 * less than its true value, regardless of the order in which the
13082 * stores to the underlying storage are issued.
13084 state->dts_alive = INT64_MAX;
13085 dtrace_membar_producer();
13086 state->dts_alive = now;
13090 dtrace_state_clean(void *arg)
13092 dtrace_state_t *state = arg;
13093 dtrace_optval_t *opt = state->dts_options;
13095 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13098 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13099 dtrace_speculation_clean(state);
13101 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13102 dtrace_state_clean, state);
13106 dtrace_state_deadman(void *arg)
13108 dtrace_state_t *state = arg;
13113 dtrace_debug_output();
13115 now = dtrace_gethrtime();
13117 if (state != dtrace_anon.dta_state &&
13118 now - state->dts_laststatus >= dtrace_deadman_user)
13122 * We must be sure that dts_alive never appears to be less than the
13123 * value upon entry to dtrace_state_deadman(), and because we lack a
13124 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13125 * store INT64_MAX to it, followed by a memory barrier, followed by
13126 * the new value. This assures that dts_alive never appears to be
13127 * less than its true value, regardless of the order in which the
13128 * stores to the underlying storage are issued.
13130 state->dts_alive = INT64_MAX;
13131 dtrace_membar_producer();
13132 state->dts_alive = now;
13134 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13135 dtrace_state_deadman, state);
13139 static dtrace_state_t *
13141 dtrace_state_create(dev_t *devp, cred_t *cr)
13143 dtrace_state_create(struct cdev *dev)
13154 dtrace_state_t *state;
13155 dtrace_optval_t *opt;
13156 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13158 ASSERT(MUTEX_HELD(&dtrace_lock));
13159 ASSERT(MUTEX_HELD(&cpu_lock));
13162 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13163 VM_BESTFIT | VM_SLEEP);
13165 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13166 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13170 state = ddi_get_soft_state(dtrace_softstate, minor);
13177 /* Allocate memory for the state. */
13178 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
13181 state->dts_epid = DTRACE_EPIDNONE + 1;
13183 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
13185 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13186 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13188 if (devp != NULL) {
13189 major = getemajor(*devp);
13191 major = ddi_driver_major(dtrace_devi);
13194 state->dts_dev = makedevice(major, minor);
13197 *devp = state->dts_dev;
13199 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
13200 state->dts_dev = dev;
13204 * We allocate NCPU buffers. On the one hand, this can be quite
13205 * a bit of memory per instance (nearly 36K on a Starcat). On the
13206 * other hand, it saves an additional memory reference in the probe
13209 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13210 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13213 state->dts_cleaner = CYCLIC_NONE;
13214 state->dts_deadman = CYCLIC_NONE;
13216 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
13217 callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
13219 state->dts_vstate.dtvs_state = state;
13221 for (i = 0; i < DTRACEOPT_MAX; i++)
13222 state->dts_options[i] = DTRACEOPT_UNSET;
13225 * Set the default options.
13227 opt = state->dts_options;
13228 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13229 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13230 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13231 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13232 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13233 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13234 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13235 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13236 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13237 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13238 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13239 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13240 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13241 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13243 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13246 * Depending on the user credentials, we set flag bits which alter probe
13247 * visibility or the amount of destructiveness allowed. In the case of
13248 * actual anonymous tracing, or the possession of all privileges, all of
13249 * the normal checks are bypassed.
13251 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13252 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13253 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13256 * Set up the credentials for this instantiation. We take a
13257 * hold on the credential to prevent it from disappearing on
13258 * us; this in turn prevents the zone_t referenced by this
13259 * credential from disappearing. This means that we can
13260 * examine the credential and the zone from probe context.
13263 state->dts_cred.dcr_cred = cr;
13266 * CRA_PROC means "we have *some* privilege for dtrace" and
13267 * unlocks the use of variables like pid, zonename, etc.
13269 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13270 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13271 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13275 * dtrace_user allows use of syscall and profile providers.
13276 * If the user also has proc_owner and/or proc_zone, we
13277 * extend the scope to include additional visibility and
13278 * destructive power.
13280 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13281 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13282 state->dts_cred.dcr_visible |=
13283 DTRACE_CRV_ALLPROC;
13285 state->dts_cred.dcr_action |=
13286 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13289 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13290 state->dts_cred.dcr_visible |=
13291 DTRACE_CRV_ALLZONE;
13293 state->dts_cred.dcr_action |=
13294 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13298 * If we have all privs in whatever zone this is,
13299 * we can do destructive things to processes which
13300 * have altered credentials.
13303 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13304 cr->cr_zone->zone_privset)) {
13305 state->dts_cred.dcr_action |=
13306 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13312 * Holding the dtrace_kernel privilege also implies that
13313 * the user has the dtrace_user privilege from a visibility
13314 * perspective. But without further privileges, some
13315 * destructive actions are not available.
13317 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13319 * Make all probes in all zones visible. However,
13320 * this doesn't mean that all actions become available
13323 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13324 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13326 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13329 * Holding proc_owner means that destructive actions
13330 * for *this* zone are allowed.
13332 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13333 state->dts_cred.dcr_action |=
13334 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13337 * Holding proc_zone means that destructive actions
13338 * for this user/group ID in all zones is allowed.
13340 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13341 state->dts_cred.dcr_action |=
13342 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13346 * If we have all privs in whatever zone this is,
13347 * we can do destructive things to processes which
13348 * have altered credentials.
13350 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13351 cr->cr_zone->zone_privset)) {
13352 state->dts_cred.dcr_action |=
13353 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13359 * Holding the dtrace_proc privilege gives control over fasttrap
13360 * and pid providers. We need to grant wider destructive
13361 * privileges in the event that the user has proc_owner and/or
13364 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13365 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13366 state->dts_cred.dcr_action |=
13367 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13369 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13370 state->dts_cred.dcr_action |=
13371 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13379 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13381 dtrace_optval_t *opt = state->dts_options, size;
13382 processorid_t cpu = 0;;
13383 int flags = 0, rval;
13385 ASSERT(MUTEX_HELD(&dtrace_lock));
13386 ASSERT(MUTEX_HELD(&cpu_lock));
13387 ASSERT(which < DTRACEOPT_MAX);
13388 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13389 (state == dtrace_anon.dta_state &&
13390 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13392 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13395 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13396 cpu = opt[DTRACEOPT_CPU];
13398 if (which == DTRACEOPT_SPECSIZE)
13399 flags |= DTRACEBUF_NOSWITCH;
13401 if (which == DTRACEOPT_BUFSIZE) {
13402 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13403 flags |= DTRACEBUF_RING;
13405 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13406 flags |= DTRACEBUF_FILL;
13408 if (state != dtrace_anon.dta_state ||
13409 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13410 flags |= DTRACEBUF_INACTIVE;
13413 for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
13415 * The size must be 8-byte aligned. If the size is not 8-byte
13416 * aligned, drop it down by the difference.
13418 if (size & (sizeof (uint64_t) - 1))
13419 size -= size & (sizeof (uint64_t) - 1);
13421 if (size < state->dts_reserve) {
13423 * Buffers always must be large enough to accommodate
13424 * their prereserved space. We return E2BIG instead
13425 * of ENOMEM in this case to allow for user-level
13426 * software to differentiate the cases.
13431 rval = dtrace_buffer_alloc(buf, size, flags, cpu);
13433 if (rval != ENOMEM) {
13438 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13446 dtrace_state_buffers(dtrace_state_t *state)
13448 dtrace_speculation_t *spec = state->dts_speculations;
13451 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13452 DTRACEOPT_BUFSIZE)) != 0)
13455 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13456 DTRACEOPT_AGGSIZE)) != 0)
13459 for (i = 0; i < state->dts_nspeculations; i++) {
13460 if ((rval = dtrace_state_buffer(state,
13461 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13469 dtrace_state_prereserve(dtrace_state_t *state)
13472 dtrace_probe_t *probe;
13474 state->dts_reserve = 0;
13476 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13480 * If our buffer policy is a "fill" buffer policy, we need to set the
13481 * prereserved space to be the space required by the END probes.
13483 probe = dtrace_probes[dtrace_probeid_end - 1];
13484 ASSERT(probe != NULL);
13486 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13487 if (ecb->dte_state != state)
13490 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13495 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13497 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13498 dtrace_speculation_t *spec;
13499 dtrace_buffer_t *buf;
13501 cyc_handler_t hdlr;
13504 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13505 dtrace_icookie_t cookie;
13507 mutex_enter(&cpu_lock);
13508 mutex_enter(&dtrace_lock);
13510 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13516 * Before we can perform any checks, we must prime all of the
13517 * retained enablings that correspond to this state.
13519 dtrace_enabling_prime(state);
13521 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13526 dtrace_state_prereserve(state);
13529 * Now we want to do is try to allocate our speculations.
13530 * We do not automatically resize the number of speculations; if
13531 * this fails, we will fail the operation.
13533 nspec = opt[DTRACEOPT_NSPEC];
13534 ASSERT(nspec != DTRACEOPT_UNSET);
13536 if (nspec > INT_MAX) {
13541 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
13543 if (spec == NULL) {
13548 state->dts_speculations = spec;
13549 state->dts_nspeculations = (int)nspec;
13551 for (i = 0; i < nspec; i++) {
13552 if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
13557 spec[i].dtsp_buffer = buf;
13560 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13561 if (dtrace_anon.dta_state == NULL) {
13566 if (state->dts_necbs != 0) {
13571 state->dts_anon = dtrace_anon_grab();
13572 ASSERT(state->dts_anon != NULL);
13573 state = state->dts_anon;
13576 * We want "grabanon" to be set in the grabbed state, so we'll
13577 * copy that option value from the grabbing state into the
13580 state->dts_options[DTRACEOPT_GRABANON] =
13581 opt[DTRACEOPT_GRABANON];
13583 *cpu = dtrace_anon.dta_beganon;
13586 * If the anonymous state is active (as it almost certainly
13587 * is if the anonymous enabling ultimately matched anything),
13588 * we don't allow any further option processing -- but we
13589 * don't return failure.
13591 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13595 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13596 opt[DTRACEOPT_AGGSIZE] != 0) {
13597 if (state->dts_aggregations == NULL) {
13599 * We're not going to create an aggregation buffer
13600 * because we don't have any ECBs that contain
13601 * aggregations -- set this option to 0.
13603 opt[DTRACEOPT_AGGSIZE] = 0;
13606 * If we have an aggregation buffer, we must also have
13607 * a buffer to use as scratch.
13609 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13610 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13611 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13616 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13617 opt[DTRACEOPT_SPECSIZE] != 0) {
13618 if (!state->dts_speculates) {
13620 * We're not going to create speculation buffers
13621 * because we don't have any ECBs that actually
13622 * speculate -- set the speculation size to 0.
13624 opt[DTRACEOPT_SPECSIZE] = 0;
13629 * The bare minimum size for any buffer that we're actually going to
13630 * do anything to is sizeof (uint64_t).
13632 sz = sizeof (uint64_t);
13634 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13635 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13636 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13638 * A buffer size has been explicitly set to 0 (or to a size
13639 * that will be adjusted to 0) and we need the space -- we
13640 * need to return failure. We return ENOSPC to differentiate
13641 * it from failing to allocate a buffer due to failure to meet
13642 * the reserve (for which we return E2BIG).
13648 if ((rval = dtrace_state_buffers(state)) != 0)
13651 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13652 sz = dtrace_dstate_defsize;
13655 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13660 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13662 } while (sz >>= 1);
13664 opt[DTRACEOPT_DYNVARSIZE] = sz;
13669 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13670 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13672 if (opt[DTRACEOPT_CLEANRATE] == 0)
13673 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13675 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13676 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13678 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13679 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13681 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13683 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13684 hdlr.cyh_arg = state;
13685 hdlr.cyh_level = CY_LOW_LEVEL;
13688 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13690 state->dts_cleaner = cyclic_add(&hdlr, &when);
13692 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13693 hdlr.cyh_arg = state;
13694 hdlr.cyh_level = CY_LOW_LEVEL;
13697 when.cyt_interval = dtrace_deadman_interval;
13699 state->dts_deadman = cyclic_add(&hdlr, &when);
13701 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13702 dtrace_state_clean, state);
13703 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13704 dtrace_state_deadman, state);
13707 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13710 * Now it's time to actually fire the BEGIN probe. We need to disable
13711 * interrupts here both to record the CPU on which we fired the BEGIN
13712 * probe (the data from this CPU will be processed first at user
13713 * level) and to manually activate the buffer for this CPU.
13715 cookie = dtrace_interrupt_disable();
13717 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13718 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13720 dtrace_probe(dtrace_probeid_begin,
13721 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13722 dtrace_interrupt_enable(cookie);
13724 * We may have had an exit action from a BEGIN probe; only change our
13725 * state to ACTIVE if we're still in WARMUP.
13727 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13728 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13730 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13731 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13734 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13735 * want each CPU to transition its principal buffer out of the
13736 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13737 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13738 * atomically transition from processing none of a state's ECBs to
13739 * processing all of them.
13741 dtrace_xcall(DTRACE_CPUALL,
13742 (dtrace_xcall_t)dtrace_buffer_activate, state);
13746 dtrace_buffer_free(state->dts_buffer);
13747 dtrace_buffer_free(state->dts_aggbuffer);
13749 if ((nspec = state->dts_nspeculations) == 0) {
13750 ASSERT(state->dts_speculations == NULL);
13754 spec = state->dts_speculations;
13755 ASSERT(spec != NULL);
13757 for (i = 0; i < state->dts_nspeculations; i++) {
13758 if ((buf = spec[i].dtsp_buffer) == NULL)
13761 dtrace_buffer_free(buf);
13762 kmem_free(buf, bufsize);
13765 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13766 state->dts_nspeculations = 0;
13767 state->dts_speculations = NULL;
13770 mutex_exit(&dtrace_lock);
13771 mutex_exit(&cpu_lock);
13777 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13779 dtrace_icookie_t cookie;
13781 ASSERT(MUTEX_HELD(&dtrace_lock));
13783 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13784 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13788 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13789 * to be sure that every CPU has seen it. See below for the details
13790 * on why this is done.
13792 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13796 * By this point, it is impossible for any CPU to be still processing
13797 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13798 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13799 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13800 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13801 * iff we're in the END probe.
13803 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13805 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13808 * Finally, we can release the reserve and call the END probe. We
13809 * disable interrupts across calling the END probe to allow us to
13810 * return the CPU on which we actually called the END probe. This
13811 * allows user-land to be sure that this CPU's principal buffer is
13814 state->dts_reserve = 0;
13816 cookie = dtrace_interrupt_disable();
13818 dtrace_probe(dtrace_probeid_end,
13819 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13820 dtrace_interrupt_enable(cookie);
13822 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13829 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13830 dtrace_optval_t val)
13832 ASSERT(MUTEX_HELD(&dtrace_lock));
13834 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13837 if (option >= DTRACEOPT_MAX)
13840 if (option != DTRACEOPT_CPU && val < 0)
13844 case DTRACEOPT_DESTRUCTIVE:
13845 if (dtrace_destructive_disallow)
13848 state->dts_cred.dcr_destructive = 1;
13851 case DTRACEOPT_BUFSIZE:
13852 case DTRACEOPT_DYNVARSIZE:
13853 case DTRACEOPT_AGGSIZE:
13854 case DTRACEOPT_SPECSIZE:
13855 case DTRACEOPT_STRSIZE:
13859 if (val >= LONG_MAX) {
13861 * If this is an otherwise negative value, set it to
13862 * the highest multiple of 128m less than LONG_MAX.
13863 * Technically, we're adjusting the size without
13864 * regard to the buffer resizing policy, but in fact,
13865 * this has no effect -- if we set the buffer size to
13866 * ~LONG_MAX and the buffer policy is ultimately set to
13867 * be "manual", the buffer allocation is guaranteed to
13868 * fail, if only because the allocation requires two
13869 * buffers. (We set the the size to the highest
13870 * multiple of 128m because it ensures that the size
13871 * will remain a multiple of a megabyte when
13872 * repeatedly halved -- all the way down to 15m.)
13874 val = LONG_MAX - (1 << 27) + 1;
13878 state->dts_options[option] = val;
13884 dtrace_state_destroy(dtrace_state_t *state)
13887 dtrace_vstate_t *vstate = &state->dts_vstate;
13889 minor_t minor = getminor(state->dts_dev);
13891 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13892 dtrace_speculation_t *spec = state->dts_speculations;
13893 int nspec = state->dts_nspeculations;
13896 ASSERT(MUTEX_HELD(&dtrace_lock));
13897 ASSERT(MUTEX_HELD(&cpu_lock));
13900 * First, retract any retained enablings for this state.
13902 dtrace_enabling_retract(state);
13903 ASSERT(state->dts_nretained == 0);
13905 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13906 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13908 * We have managed to come into dtrace_state_destroy() on a
13909 * hot enabling -- almost certainly because of a disorderly
13910 * shutdown of a consumer. (That is, a consumer that is
13911 * exiting without having called dtrace_stop().) In this case,
13912 * we're going to set our activity to be KILLED, and then
13913 * issue a sync to be sure that everyone is out of probe
13914 * context before we start blowing away ECBs.
13916 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13921 * Release the credential hold we took in dtrace_state_create().
13923 if (state->dts_cred.dcr_cred != NULL)
13924 crfree(state->dts_cred.dcr_cred);
13927 * Now we can safely disable and destroy any enabled probes. Because
13928 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13929 * (especially if they're all enabled), we take two passes through the
13930 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13931 * in the second we disable whatever is left over.
13933 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13934 for (i = 0; i < state->dts_necbs; i++) {
13935 if ((ecb = state->dts_ecbs[i]) == NULL)
13938 if (match && ecb->dte_probe != NULL) {
13939 dtrace_probe_t *probe = ecb->dte_probe;
13940 dtrace_provider_t *prov = probe->dtpr_provider;
13942 if (!(prov->dtpv_priv.dtpp_flags & match))
13946 dtrace_ecb_disable(ecb);
13947 dtrace_ecb_destroy(ecb);
13955 * Before we free the buffers, perform one more sync to assure that
13956 * every CPU is out of probe context.
13960 dtrace_buffer_free(state->dts_buffer);
13961 dtrace_buffer_free(state->dts_aggbuffer);
13963 for (i = 0; i < nspec; i++)
13964 dtrace_buffer_free(spec[i].dtsp_buffer);
13967 if (state->dts_cleaner != CYCLIC_NONE)
13968 cyclic_remove(state->dts_cleaner);
13970 if (state->dts_deadman != CYCLIC_NONE)
13971 cyclic_remove(state->dts_deadman);
13973 callout_stop(&state->dts_cleaner);
13974 callout_drain(&state->dts_cleaner);
13975 callout_stop(&state->dts_deadman);
13976 callout_drain(&state->dts_deadman);
13979 dtrace_dstate_fini(&vstate->dtvs_dynvars);
13980 dtrace_vstate_fini(vstate);
13981 if (state->dts_ecbs != NULL)
13982 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13984 if (state->dts_aggregations != NULL) {
13986 for (i = 0; i < state->dts_naggregations; i++)
13987 ASSERT(state->dts_aggregations[i] == NULL);
13989 ASSERT(state->dts_naggregations > 0);
13990 kmem_free(state->dts_aggregations,
13991 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13994 kmem_free(state->dts_buffer, bufsize);
13995 kmem_free(state->dts_aggbuffer, bufsize);
13997 for (i = 0; i < nspec; i++)
13998 kmem_free(spec[i].dtsp_buffer, bufsize);
14001 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14003 dtrace_format_destroy(state);
14005 if (state->dts_aggid_arena != NULL) {
14007 vmem_destroy(state->dts_aggid_arena);
14009 delete_unrhdr(state->dts_aggid_arena);
14011 state->dts_aggid_arena = NULL;
14014 ddi_soft_state_free(dtrace_softstate, minor);
14015 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14020 * DTrace Anonymous Enabling Functions
14022 static dtrace_state_t *
14023 dtrace_anon_grab(void)
14025 dtrace_state_t *state;
14027 ASSERT(MUTEX_HELD(&dtrace_lock));
14029 if ((state = dtrace_anon.dta_state) == NULL) {
14030 ASSERT(dtrace_anon.dta_enabling == NULL);
14034 ASSERT(dtrace_anon.dta_enabling != NULL);
14035 ASSERT(dtrace_retained != NULL);
14037 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14038 dtrace_anon.dta_enabling = NULL;
14039 dtrace_anon.dta_state = NULL;
14045 dtrace_anon_property(void)
14048 dtrace_state_t *state;
14050 char c[32]; /* enough for "dof-data-" + digits */
14052 ASSERT(MUTEX_HELD(&dtrace_lock));
14053 ASSERT(MUTEX_HELD(&cpu_lock));
14055 for (i = 0; ; i++) {
14056 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14058 dtrace_err_verbose = 1;
14060 if ((dof = dtrace_dof_property(c)) == NULL) {
14061 dtrace_err_verbose = 0;
14067 * We want to create anonymous state, so we need to transition
14068 * the kernel debugger to indicate that DTrace is active. If
14069 * this fails (e.g. because the debugger has modified text in
14070 * some way), we won't continue with the processing.
14072 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14073 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14074 "enabling ignored.");
14075 dtrace_dof_destroy(dof);
14081 * If we haven't allocated an anonymous state, we'll do so now.
14083 if ((state = dtrace_anon.dta_state) == NULL) {
14085 state = dtrace_state_create(NULL, NULL);
14087 state = dtrace_state_create(NULL);
14089 dtrace_anon.dta_state = state;
14091 if (state == NULL) {
14093 * This basically shouldn't happen: the only
14094 * failure mode from dtrace_state_create() is a
14095 * failure of ddi_soft_state_zalloc() that
14096 * itself should never happen. Still, the
14097 * interface allows for a failure mode, and
14098 * we want to fail as gracefully as possible:
14099 * we'll emit an error message and cease
14100 * processing anonymous state in this case.
14102 cmn_err(CE_WARN, "failed to create "
14103 "anonymous state");
14104 dtrace_dof_destroy(dof);
14109 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14110 &dtrace_anon.dta_enabling, 0, B_TRUE);
14113 rv = dtrace_dof_options(dof, state);
14115 dtrace_err_verbose = 0;
14116 dtrace_dof_destroy(dof);
14120 * This is malformed DOF; chuck any anonymous state
14123 ASSERT(dtrace_anon.dta_enabling == NULL);
14124 dtrace_state_destroy(state);
14125 dtrace_anon.dta_state = NULL;
14129 ASSERT(dtrace_anon.dta_enabling != NULL);
14132 if (dtrace_anon.dta_enabling != NULL) {
14136 * dtrace_enabling_retain() can only fail because we are
14137 * trying to retain more enablings than are allowed -- but
14138 * we only have one anonymous enabling, and we are guaranteed
14139 * to be allowed at least one retained enabling; we assert
14140 * that dtrace_enabling_retain() returns success.
14142 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14145 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14150 * DTrace Helper Functions
14153 dtrace_helper_trace(dtrace_helper_action_t *helper,
14154 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14156 uint32_t size, next, nnext, i;
14157 dtrace_helptrace_t *ent;
14158 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
14160 if (!dtrace_helptrace_enabled)
14163 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14166 * What would a tracing framework be without its own tracing
14167 * framework? (Well, a hell of a lot simpler, for starters...)
14169 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14170 sizeof (uint64_t) - sizeof (uint64_t);
14173 * Iterate until we can allocate a slot in the trace buffer.
14176 next = dtrace_helptrace_next;
14178 if (next + size < dtrace_helptrace_bufsize) {
14179 nnext = next + size;
14183 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14186 * We have our slot; fill it in.
14191 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
14192 ent->dtht_helper = helper;
14193 ent->dtht_where = where;
14194 ent->dtht_nlocals = vstate->dtvs_nlocals;
14196 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14197 mstate->dtms_fltoffs : -1;
14198 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14199 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
14201 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14202 dtrace_statvar_t *svar;
14204 if ((svar = vstate->dtvs_locals[i]) == NULL)
14207 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14208 ent->dtht_locals[i] =
14209 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
14214 dtrace_helper(int which, dtrace_mstate_t *mstate,
14215 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14217 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
14218 uint64_t sarg0 = mstate->dtms_arg[0];
14219 uint64_t sarg1 = mstate->dtms_arg[1];
14221 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14222 dtrace_helper_action_t *helper;
14223 dtrace_vstate_t *vstate;
14224 dtrace_difo_t *pred;
14225 int i, trace = dtrace_helptrace_enabled;
14227 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14229 if (helpers == NULL)
14232 if ((helper = helpers->dthps_actions[which]) == NULL)
14235 vstate = &helpers->dthps_vstate;
14236 mstate->dtms_arg[0] = arg0;
14237 mstate->dtms_arg[1] = arg1;
14240 * Now iterate over each helper. If its predicate evaluates to 'true',
14241 * we'll call the corresponding actions. Note that the below calls
14242 * to dtrace_dif_emulate() may set faults in machine state. This is
14243 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14244 * the stored DIF offset with its own (which is the desired behavior).
14245 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14246 * from machine state; this is okay, too.
14248 for (; helper != NULL; helper = helper->dtha_next) {
14249 if ((pred = helper->dtha_predicate) != NULL) {
14251 dtrace_helper_trace(helper, mstate, vstate, 0);
14253 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14256 if (*flags & CPU_DTRACE_FAULT)
14260 for (i = 0; i < helper->dtha_nactions; i++) {
14262 dtrace_helper_trace(helper,
14263 mstate, vstate, i + 1);
14265 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14266 mstate, vstate, state);
14268 if (*flags & CPU_DTRACE_FAULT)
14274 dtrace_helper_trace(helper, mstate, vstate,
14275 DTRACE_HELPTRACE_NEXT);
14279 dtrace_helper_trace(helper, mstate, vstate,
14280 DTRACE_HELPTRACE_DONE);
14283 * Restore the arg0 that we saved upon entry.
14285 mstate->dtms_arg[0] = sarg0;
14286 mstate->dtms_arg[1] = sarg1;
14292 dtrace_helper_trace(helper, mstate, vstate,
14293 DTRACE_HELPTRACE_ERR);
14296 * Restore the arg0 that we saved upon entry.
14298 mstate->dtms_arg[0] = sarg0;
14299 mstate->dtms_arg[1] = sarg1;
14305 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14306 dtrace_vstate_t *vstate)
14310 if (helper->dtha_predicate != NULL)
14311 dtrace_difo_release(helper->dtha_predicate, vstate);
14313 for (i = 0; i < helper->dtha_nactions; i++) {
14314 ASSERT(helper->dtha_actions[i] != NULL);
14315 dtrace_difo_release(helper->dtha_actions[i], vstate);
14318 kmem_free(helper->dtha_actions,
14319 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14320 kmem_free(helper, sizeof (dtrace_helper_action_t));
14324 dtrace_helper_destroygen(int gen)
14326 proc_t *p = curproc;
14327 dtrace_helpers_t *help = p->p_dtrace_helpers;
14328 dtrace_vstate_t *vstate;
14331 ASSERT(MUTEX_HELD(&dtrace_lock));
14333 if (help == NULL || gen > help->dthps_generation)
14336 vstate = &help->dthps_vstate;
14338 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14339 dtrace_helper_action_t *last = NULL, *h, *next;
14341 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14342 next = h->dtha_next;
14344 if (h->dtha_generation == gen) {
14345 if (last != NULL) {
14346 last->dtha_next = next;
14348 help->dthps_actions[i] = next;
14351 dtrace_helper_action_destroy(h, vstate);
14359 * Interate until we've cleared out all helper providers with the
14360 * given generation number.
14363 dtrace_helper_provider_t *prov;
14366 * Look for a helper provider with the right generation. We
14367 * have to start back at the beginning of the list each time
14368 * because we drop dtrace_lock. It's unlikely that we'll make
14369 * more than two passes.
14371 for (i = 0; i < help->dthps_nprovs; i++) {
14372 prov = help->dthps_provs[i];
14374 if (prov->dthp_generation == gen)
14379 * If there were no matches, we're done.
14381 if (i == help->dthps_nprovs)
14385 * Move the last helper provider into this slot.
14387 help->dthps_nprovs--;
14388 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14389 help->dthps_provs[help->dthps_nprovs] = NULL;
14391 mutex_exit(&dtrace_lock);
14394 * If we have a meta provider, remove this helper provider.
14396 mutex_enter(&dtrace_meta_lock);
14397 if (dtrace_meta_pid != NULL) {
14398 ASSERT(dtrace_deferred_pid == NULL);
14399 dtrace_helper_provider_remove(&prov->dthp_prov,
14402 mutex_exit(&dtrace_meta_lock);
14404 dtrace_helper_provider_destroy(prov);
14406 mutex_enter(&dtrace_lock);
14413 dtrace_helper_validate(dtrace_helper_action_t *helper)
14418 if ((dp = helper->dtha_predicate) != NULL)
14419 err += dtrace_difo_validate_helper(dp);
14421 for (i = 0; i < helper->dtha_nactions; i++)
14422 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14428 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14430 dtrace_helpers_t *help;
14431 dtrace_helper_action_t *helper, *last;
14432 dtrace_actdesc_t *act;
14433 dtrace_vstate_t *vstate;
14434 dtrace_predicate_t *pred;
14435 int count = 0, nactions = 0, i;
14437 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14440 help = curproc->p_dtrace_helpers;
14441 last = help->dthps_actions[which];
14442 vstate = &help->dthps_vstate;
14444 for (count = 0; last != NULL; last = last->dtha_next) {
14446 if (last->dtha_next == NULL)
14451 * If we already have dtrace_helper_actions_max helper actions for this
14452 * helper action type, we'll refuse to add a new one.
14454 if (count >= dtrace_helper_actions_max)
14457 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14458 helper->dtha_generation = help->dthps_generation;
14460 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14461 ASSERT(pred->dtp_difo != NULL);
14462 dtrace_difo_hold(pred->dtp_difo);
14463 helper->dtha_predicate = pred->dtp_difo;
14466 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14467 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14470 if (act->dtad_difo == NULL)
14476 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14477 (helper->dtha_nactions = nactions), KM_SLEEP);
14479 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14480 dtrace_difo_hold(act->dtad_difo);
14481 helper->dtha_actions[i++] = act->dtad_difo;
14484 if (!dtrace_helper_validate(helper))
14487 if (last == NULL) {
14488 help->dthps_actions[which] = helper;
14490 last->dtha_next = helper;
14493 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14494 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14495 dtrace_helptrace_next = 0;
14500 dtrace_helper_action_destroy(helper, vstate);
14505 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14506 dof_helper_t *dofhp)
14508 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14510 mutex_enter(&dtrace_meta_lock);
14511 mutex_enter(&dtrace_lock);
14513 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14515 * If the dtrace module is loaded but not attached, or if
14516 * there aren't isn't a meta provider registered to deal with
14517 * these provider descriptions, we need to postpone creating
14518 * the actual providers until later.
14521 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14522 dtrace_deferred_pid != help) {
14523 help->dthps_deferred = 1;
14524 help->dthps_pid = p->p_pid;
14525 help->dthps_next = dtrace_deferred_pid;
14526 help->dthps_prev = NULL;
14527 if (dtrace_deferred_pid != NULL)
14528 dtrace_deferred_pid->dthps_prev = help;
14529 dtrace_deferred_pid = help;
14532 mutex_exit(&dtrace_lock);
14534 } else if (dofhp != NULL) {
14536 * If the dtrace module is loaded and we have a particular
14537 * helper provider description, pass that off to the
14541 mutex_exit(&dtrace_lock);
14543 dtrace_helper_provide(dofhp, p->p_pid);
14547 * Otherwise, just pass all the helper provider descriptions
14548 * off to the meta provider.
14552 mutex_exit(&dtrace_lock);
14554 for (i = 0; i < help->dthps_nprovs; i++) {
14555 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14560 mutex_exit(&dtrace_meta_lock);
14564 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14566 dtrace_helpers_t *help;
14567 dtrace_helper_provider_t *hprov, **tmp_provs;
14568 uint_t tmp_maxprovs, i;
14570 ASSERT(MUTEX_HELD(&dtrace_lock));
14572 help = curproc->p_dtrace_helpers;
14573 ASSERT(help != NULL);
14576 * If we already have dtrace_helper_providers_max helper providers,
14577 * we're refuse to add a new one.
14579 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14583 * Check to make sure this isn't a duplicate.
14585 for (i = 0; i < help->dthps_nprovs; i++) {
14586 if (dofhp->dofhp_addr ==
14587 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14591 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14592 hprov->dthp_prov = *dofhp;
14593 hprov->dthp_ref = 1;
14594 hprov->dthp_generation = gen;
14597 * Allocate a bigger table for helper providers if it's already full.
14599 if (help->dthps_maxprovs == help->dthps_nprovs) {
14600 tmp_maxprovs = help->dthps_maxprovs;
14601 tmp_provs = help->dthps_provs;
14603 if (help->dthps_maxprovs == 0)
14604 help->dthps_maxprovs = 2;
14606 help->dthps_maxprovs *= 2;
14607 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14608 help->dthps_maxprovs = dtrace_helper_providers_max;
14610 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14612 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14613 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14615 if (tmp_provs != NULL) {
14616 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14617 sizeof (dtrace_helper_provider_t *));
14618 kmem_free(tmp_provs, tmp_maxprovs *
14619 sizeof (dtrace_helper_provider_t *));
14623 help->dthps_provs[help->dthps_nprovs] = hprov;
14624 help->dthps_nprovs++;
14630 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14632 mutex_enter(&dtrace_lock);
14634 if (--hprov->dthp_ref == 0) {
14636 mutex_exit(&dtrace_lock);
14637 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14638 dtrace_dof_destroy(dof);
14639 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14641 mutex_exit(&dtrace_lock);
14646 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14648 uintptr_t daddr = (uintptr_t)dof;
14649 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14650 dof_provider_t *provider;
14651 dof_probe_t *probe;
14653 char *strtab, *typestr;
14654 dof_stridx_t typeidx;
14656 uint_t nprobes, j, k;
14658 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14660 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14661 dtrace_dof_error(dof, "misaligned section offset");
14666 * The section needs to be large enough to contain the DOF provider
14667 * structure appropriate for the given version.
14669 if (sec->dofs_size <
14670 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14671 offsetof(dof_provider_t, dofpv_prenoffs) :
14672 sizeof (dof_provider_t))) {
14673 dtrace_dof_error(dof, "provider section too small");
14677 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14678 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14679 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14680 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14681 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14683 if (str_sec == NULL || prb_sec == NULL ||
14684 arg_sec == NULL || off_sec == NULL)
14689 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14690 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14691 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14692 provider->dofpv_prenoffs)) == NULL)
14695 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14697 if (provider->dofpv_name >= str_sec->dofs_size ||
14698 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14699 dtrace_dof_error(dof, "invalid provider name");
14703 if (prb_sec->dofs_entsize == 0 ||
14704 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14705 dtrace_dof_error(dof, "invalid entry size");
14709 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14710 dtrace_dof_error(dof, "misaligned entry size");
14714 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14715 dtrace_dof_error(dof, "invalid entry size");
14719 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14720 dtrace_dof_error(dof, "misaligned section offset");
14724 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14725 dtrace_dof_error(dof, "invalid entry size");
14729 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14731 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14734 * Take a pass through the probes to check for errors.
14736 for (j = 0; j < nprobes; j++) {
14737 probe = (dof_probe_t *)(uintptr_t)(daddr +
14738 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14740 if (probe->dofpr_func >= str_sec->dofs_size) {
14741 dtrace_dof_error(dof, "invalid function name");
14745 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14746 dtrace_dof_error(dof, "function name too long");
14750 if (probe->dofpr_name >= str_sec->dofs_size ||
14751 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14752 dtrace_dof_error(dof, "invalid probe name");
14757 * The offset count must not wrap the index, and the offsets
14758 * must also not overflow the section's data.
14760 if (probe->dofpr_offidx + probe->dofpr_noffs <
14761 probe->dofpr_offidx ||
14762 (probe->dofpr_offidx + probe->dofpr_noffs) *
14763 off_sec->dofs_entsize > off_sec->dofs_size) {
14764 dtrace_dof_error(dof, "invalid probe offset");
14768 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14770 * If there's no is-enabled offset section, make sure
14771 * there aren't any is-enabled offsets. Otherwise
14772 * perform the same checks as for probe offsets
14773 * (immediately above).
14775 if (enoff_sec == NULL) {
14776 if (probe->dofpr_enoffidx != 0 ||
14777 probe->dofpr_nenoffs != 0) {
14778 dtrace_dof_error(dof, "is-enabled "
14779 "offsets with null section");
14782 } else if (probe->dofpr_enoffidx +
14783 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14784 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14785 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14786 dtrace_dof_error(dof, "invalid is-enabled "
14791 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14792 dtrace_dof_error(dof, "zero probe and "
14793 "is-enabled offsets");
14796 } else if (probe->dofpr_noffs == 0) {
14797 dtrace_dof_error(dof, "zero probe offsets");
14801 if (probe->dofpr_argidx + probe->dofpr_xargc <
14802 probe->dofpr_argidx ||
14803 (probe->dofpr_argidx + probe->dofpr_xargc) *
14804 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14805 dtrace_dof_error(dof, "invalid args");
14809 typeidx = probe->dofpr_nargv;
14810 typestr = strtab + probe->dofpr_nargv;
14811 for (k = 0; k < probe->dofpr_nargc; k++) {
14812 if (typeidx >= str_sec->dofs_size) {
14813 dtrace_dof_error(dof, "bad "
14814 "native argument type");
14818 typesz = strlen(typestr) + 1;
14819 if (typesz > DTRACE_ARGTYPELEN) {
14820 dtrace_dof_error(dof, "native "
14821 "argument type too long");
14828 typeidx = probe->dofpr_xargv;
14829 typestr = strtab + probe->dofpr_xargv;
14830 for (k = 0; k < probe->dofpr_xargc; k++) {
14831 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14832 dtrace_dof_error(dof, "bad "
14833 "native argument index");
14837 if (typeidx >= str_sec->dofs_size) {
14838 dtrace_dof_error(dof, "bad "
14839 "translated argument type");
14843 typesz = strlen(typestr) + 1;
14844 if (typesz > DTRACE_ARGTYPELEN) {
14845 dtrace_dof_error(dof, "translated argument "
14859 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14861 dtrace_helpers_t *help;
14862 dtrace_vstate_t *vstate;
14863 dtrace_enabling_t *enab = NULL;
14864 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14865 uintptr_t daddr = (uintptr_t)dof;
14867 ASSERT(MUTEX_HELD(&dtrace_lock));
14869 if ((help = curproc->p_dtrace_helpers) == NULL)
14870 help = dtrace_helpers_create(curproc);
14872 vstate = &help->dthps_vstate;
14874 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14875 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14876 dtrace_dof_destroy(dof);
14881 * Look for helper providers and validate their descriptions.
14884 for (i = 0; i < dof->dofh_secnum; i++) {
14885 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14886 dof->dofh_secoff + i * dof->dofh_secsize);
14888 if (sec->dofs_type != DOF_SECT_PROVIDER)
14891 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14892 dtrace_enabling_destroy(enab);
14893 dtrace_dof_destroy(dof);
14902 * Now we need to walk through the ECB descriptions in the enabling.
14904 for (i = 0; i < enab->dten_ndesc; i++) {
14905 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14906 dtrace_probedesc_t *desc = &ep->dted_probe;
14908 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14911 if (strcmp(desc->dtpd_mod, "helper") != 0)
14914 if (strcmp(desc->dtpd_func, "ustack") != 0)
14917 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14920 * Adding this helper action failed -- we are now going
14921 * to rip out the entire generation and return failure.
14923 (void) dtrace_helper_destroygen(help->dthps_generation);
14924 dtrace_enabling_destroy(enab);
14925 dtrace_dof_destroy(dof);
14932 if (nhelpers < enab->dten_ndesc)
14933 dtrace_dof_error(dof, "unmatched helpers");
14935 gen = help->dthps_generation++;
14936 dtrace_enabling_destroy(enab);
14938 if (dhp != NULL && nprovs > 0) {
14939 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14940 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14941 mutex_exit(&dtrace_lock);
14942 dtrace_helper_provider_register(curproc, help, dhp);
14943 mutex_enter(&dtrace_lock);
14950 dtrace_dof_destroy(dof);
14955 static dtrace_helpers_t *
14956 dtrace_helpers_create(proc_t *p)
14958 dtrace_helpers_t *help;
14960 ASSERT(MUTEX_HELD(&dtrace_lock));
14961 ASSERT(p->p_dtrace_helpers == NULL);
14963 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14964 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14965 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14967 p->p_dtrace_helpers = help;
14977 dtrace_helpers_destroy(proc_t *p)
14979 dtrace_helpers_t *help;
14980 dtrace_vstate_t *vstate;
14982 proc_t *p = curproc;
14986 mutex_enter(&dtrace_lock);
14988 ASSERT(p->p_dtrace_helpers != NULL);
14989 ASSERT(dtrace_helpers > 0);
14991 help = p->p_dtrace_helpers;
14992 vstate = &help->dthps_vstate;
14995 * We're now going to lose the help from this process.
14997 p->p_dtrace_helpers = NULL;
15001 * Destory the helper actions.
15003 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15004 dtrace_helper_action_t *h, *next;
15006 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15007 next = h->dtha_next;
15008 dtrace_helper_action_destroy(h, vstate);
15013 mutex_exit(&dtrace_lock);
15016 * Destroy the helper providers.
15018 if (help->dthps_maxprovs > 0) {
15019 mutex_enter(&dtrace_meta_lock);
15020 if (dtrace_meta_pid != NULL) {
15021 ASSERT(dtrace_deferred_pid == NULL);
15023 for (i = 0; i < help->dthps_nprovs; i++) {
15024 dtrace_helper_provider_remove(
15025 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15028 mutex_enter(&dtrace_lock);
15029 ASSERT(help->dthps_deferred == 0 ||
15030 help->dthps_next != NULL ||
15031 help->dthps_prev != NULL ||
15032 help == dtrace_deferred_pid);
15035 * Remove the helper from the deferred list.
15037 if (help->dthps_next != NULL)
15038 help->dthps_next->dthps_prev = help->dthps_prev;
15039 if (help->dthps_prev != NULL)
15040 help->dthps_prev->dthps_next = help->dthps_next;
15041 if (dtrace_deferred_pid == help) {
15042 dtrace_deferred_pid = help->dthps_next;
15043 ASSERT(help->dthps_prev == NULL);
15046 mutex_exit(&dtrace_lock);
15049 mutex_exit(&dtrace_meta_lock);
15051 for (i = 0; i < help->dthps_nprovs; i++) {
15052 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15055 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15056 sizeof (dtrace_helper_provider_t *));
15059 mutex_enter(&dtrace_lock);
15061 dtrace_vstate_fini(&help->dthps_vstate);
15062 kmem_free(help->dthps_actions,
15063 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15064 kmem_free(help, sizeof (dtrace_helpers_t));
15067 mutex_exit(&dtrace_lock);
15074 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15076 dtrace_helpers_t *help, *newhelp;
15077 dtrace_helper_action_t *helper, *new, *last;
15079 dtrace_vstate_t *vstate;
15080 int i, j, sz, hasprovs = 0;
15082 mutex_enter(&dtrace_lock);
15083 ASSERT(from->p_dtrace_helpers != NULL);
15084 ASSERT(dtrace_helpers > 0);
15086 help = from->p_dtrace_helpers;
15087 newhelp = dtrace_helpers_create(to);
15088 ASSERT(to->p_dtrace_helpers != NULL);
15090 newhelp->dthps_generation = help->dthps_generation;
15091 vstate = &newhelp->dthps_vstate;
15094 * Duplicate the helper actions.
15096 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15097 if ((helper = help->dthps_actions[i]) == NULL)
15100 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15101 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15103 new->dtha_generation = helper->dtha_generation;
15105 if ((dp = helper->dtha_predicate) != NULL) {
15106 dp = dtrace_difo_duplicate(dp, vstate);
15107 new->dtha_predicate = dp;
15110 new->dtha_nactions = helper->dtha_nactions;
15111 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15112 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15114 for (j = 0; j < new->dtha_nactions; j++) {
15115 dtrace_difo_t *dp = helper->dtha_actions[j];
15117 ASSERT(dp != NULL);
15118 dp = dtrace_difo_duplicate(dp, vstate);
15119 new->dtha_actions[j] = dp;
15122 if (last != NULL) {
15123 last->dtha_next = new;
15125 newhelp->dthps_actions[i] = new;
15133 * Duplicate the helper providers and register them with the
15134 * DTrace framework.
15136 if (help->dthps_nprovs > 0) {
15137 newhelp->dthps_nprovs = help->dthps_nprovs;
15138 newhelp->dthps_maxprovs = help->dthps_nprovs;
15139 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15140 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15141 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15142 newhelp->dthps_provs[i] = help->dthps_provs[i];
15143 newhelp->dthps_provs[i]->dthp_ref++;
15149 mutex_exit(&dtrace_lock);
15152 dtrace_helper_provider_register(to, newhelp, NULL);
15157 * DTrace Hook Functions
15160 dtrace_module_loaded(modctl_t *ctl)
15162 dtrace_provider_t *prv;
15164 mutex_enter(&dtrace_provider_lock);
15166 mutex_enter(&mod_lock);
15169 ASSERT(ctl->mod_busy);
15172 * We're going to call each providers per-module provide operation
15173 * specifying only this module.
15175 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15176 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15179 mutex_exit(&mod_lock);
15181 mutex_exit(&dtrace_provider_lock);
15184 * If we have any retained enablings, we need to match against them.
15185 * Enabling probes requires that cpu_lock be held, and we cannot hold
15186 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15187 * module. (In particular, this happens when loading scheduling
15188 * classes.) So if we have any retained enablings, we need to dispatch
15189 * our task queue to do the match for us.
15191 mutex_enter(&dtrace_lock);
15193 if (dtrace_retained == NULL) {
15194 mutex_exit(&dtrace_lock);
15198 (void) taskq_dispatch(dtrace_taskq,
15199 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15201 mutex_exit(&dtrace_lock);
15204 * And now, for a little heuristic sleaze: in general, we want to
15205 * match modules as soon as they load. However, we cannot guarantee
15206 * this, because it would lead us to the lock ordering violation
15207 * outlined above. The common case, of course, is that cpu_lock is
15208 * _not_ held -- so we delay here for a clock tick, hoping that that's
15209 * long enough for the task queue to do its work. If it's not, it's
15210 * not a serious problem -- it just means that the module that we
15211 * just loaded may not be immediately instrumentable.
15217 dtrace_module_unloaded(modctl_t *ctl)
15219 dtrace_probe_t template, *probe, *first, *next;
15220 dtrace_provider_t *prov;
15222 template.dtpr_mod = ctl->mod_modname;
15224 mutex_enter(&dtrace_provider_lock);
15226 mutex_enter(&mod_lock);
15228 mutex_enter(&dtrace_lock);
15230 if (dtrace_bymod == NULL) {
15232 * The DTrace module is loaded (obviously) but not attached;
15233 * we don't have any work to do.
15235 mutex_exit(&dtrace_provider_lock);
15237 mutex_exit(&mod_lock);
15239 mutex_exit(&dtrace_lock);
15243 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15244 probe != NULL; probe = probe->dtpr_nextmod) {
15245 if (probe->dtpr_ecb != NULL) {
15246 mutex_exit(&dtrace_provider_lock);
15248 mutex_exit(&mod_lock);
15250 mutex_exit(&dtrace_lock);
15253 * This shouldn't _actually_ be possible -- we're
15254 * unloading a module that has an enabled probe in it.
15255 * (It's normally up to the provider to make sure that
15256 * this can't happen.) However, because dtps_enable()
15257 * doesn't have a failure mode, there can be an
15258 * enable/unload race. Upshot: we don't want to
15259 * assert, but we're not going to disable the
15262 if (dtrace_err_verbose) {
15263 cmn_err(CE_WARN, "unloaded module '%s' had "
15264 "enabled probes", ctl->mod_modname);
15273 for (first = NULL; probe != NULL; probe = next) {
15274 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15276 dtrace_probes[probe->dtpr_id - 1] = NULL;
15278 next = probe->dtpr_nextmod;
15279 dtrace_hash_remove(dtrace_bymod, probe);
15280 dtrace_hash_remove(dtrace_byfunc, probe);
15281 dtrace_hash_remove(dtrace_byname, probe);
15283 if (first == NULL) {
15285 probe->dtpr_nextmod = NULL;
15287 probe->dtpr_nextmod = first;
15293 * We've removed all of the module's probes from the hash chains and
15294 * from the probe array. Now issue a dtrace_sync() to be sure that
15295 * everyone has cleared out from any probe array processing.
15299 for (probe = first; probe != NULL; probe = first) {
15300 first = probe->dtpr_nextmod;
15301 prov = probe->dtpr_provider;
15302 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15304 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15305 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15306 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15307 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15308 kmem_free(probe, sizeof (dtrace_probe_t));
15311 mutex_exit(&dtrace_lock);
15313 mutex_exit(&mod_lock);
15315 mutex_exit(&dtrace_provider_lock);
15319 dtrace_suspend(void)
15321 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15325 dtrace_resume(void)
15327 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15332 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15334 ASSERT(MUTEX_HELD(&cpu_lock));
15335 mutex_enter(&dtrace_lock);
15339 dtrace_state_t *state;
15340 dtrace_optval_t *opt, rs, c;
15343 * For now, we only allocate a new buffer for anonymous state.
15345 if ((state = dtrace_anon.dta_state) == NULL)
15348 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15351 opt = state->dts_options;
15352 c = opt[DTRACEOPT_CPU];
15354 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15358 * Regardless of what the actual policy is, we're going to
15359 * temporarily set our resize policy to be manual. We're
15360 * also going to temporarily set our CPU option to denote
15361 * the newly configured CPU.
15363 rs = opt[DTRACEOPT_BUFRESIZE];
15364 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15365 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15367 (void) dtrace_state_buffers(state);
15369 opt[DTRACEOPT_BUFRESIZE] = rs;
15370 opt[DTRACEOPT_CPU] = c;
15377 * We don't free the buffer in the CPU_UNCONFIG case. (The
15378 * buffer will be freed when the consumer exits.)
15386 mutex_exit(&dtrace_lock);
15392 dtrace_cpu_setup_initial(processorid_t cpu)
15394 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15399 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15401 if (dtrace_toxranges >= dtrace_toxranges_max) {
15403 dtrace_toxrange_t *range;
15405 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15408 ASSERT(dtrace_toxrange == NULL);
15409 ASSERT(dtrace_toxranges_max == 0);
15410 dtrace_toxranges_max = 1;
15412 dtrace_toxranges_max <<= 1;
15415 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15416 range = kmem_zalloc(nsize, KM_SLEEP);
15418 if (dtrace_toxrange != NULL) {
15419 ASSERT(osize != 0);
15420 bcopy(dtrace_toxrange, range, osize);
15421 kmem_free(dtrace_toxrange, osize);
15424 dtrace_toxrange = range;
15427 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15428 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15430 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15431 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15432 dtrace_toxranges++;
15436 * DTrace Driver Cookbook Functions
15441 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15443 dtrace_provider_id_t id;
15444 dtrace_state_t *state = NULL;
15445 dtrace_enabling_t *enab;
15447 mutex_enter(&cpu_lock);
15448 mutex_enter(&dtrace_provider_lock);
15449 mutex_enter(&dtrace_lock);
15451 if (ddi_soft_state_init(&dtrace_softstate,
15452 sizeof (dtrace_state_t), 0) != 0) {
15453 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15454 mutex_exit(&cpu_lock);
15455 mutex_exit(&dtrace_provider_lock);
15456 mutex_exit(&dtrace_lock);
15457 return (DDI_FAILURE);
15460 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15461 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15462 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15463 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15464 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15465 ddi_remove_minor_node(devi, NULL);
15466 ddi_soft_state_fini(&dtrace_softstate);
15467 mutex_exit(&cpu_lock);
15468 mutex_exit(&dtrace_provider_lock);
15469 mutex_exit(&dtrace_lock);
15470 return (DDI_FAILURE);
15473 ddi_report_dev(devi);
15474 dtrace_devi = devi;
15476 dtrace_modload = dtrace_module_loaded;
15477 dtrace_modunload = dtrace_module_unloaded;
15478 dtrace_cpu_init = dtrace_cpu_setup_initial;
15479 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15480 dtrace_helpers_fork = dtrace_helpers_duplicate;
15481 dtrace_cpustart_init = dtrace_suspend;
15482 dtrace_cpustart_fini = dtrace_resume;
15483 dtrace_debugger_init = dtrace_suspend;
15484 dtrace_debugger_fini = dtrace_resume;
15486 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15488 ASSERT(MUTEX_HELD(&cpu_lock));
15490 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15491 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15492 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15493 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15494 VM_SLEEP | VMC_IDENTIFIER);
15495 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15498 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15499 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15500 NULL, NULL, NULL, NULL, NULL, 0);
15502 ASSERT(MUTEX_HELD(&cpu_lock));
15503 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15504 offsetof(dtrace_probe_t, dtpr_nextmod),
15505 offsetof(dtrace_probe_t, dtpr_prevmod));
15507 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15508 offsetof(dtrace_probe_t, dtpr_nextfunc),
15509 offsetof(dtrace_probe_t, dtpr_prevfunc));
15511 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15512 offsetof(dtrace_probe_t, dtpr_nextname),
15513 offsetof(dtrace_probe_t, dtpr_prevname));
15515 if (dtrace_retain_max < 1) {
15516 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15517 "setting to 1", dtrace_retain_max);
15518 dtrace_retain_max = 1;
15522 * Now discover our toxic ranges.
15524 dtrace_toxic_ranges(dtrace_toxrange_add);
15527 * Before we register ourselves as a provider to our own framework,
15528 * we would like to assert that dtrace_provider is NULL -- but that's
15529 * not true if we were loaded as a dependency of a DTrace provider.
15530 * Once we've registered, we can assert that dtrace_provider is our
15533 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15534 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15536 ASSERT(dtrace_provider != NULL);
15537 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15539 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15540 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15541 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15542 dtrace_provider, NULL, NULL, "END", 0, NULL);
15543 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15544 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15546 dtrace_anon_property();
15547 mutex_exit(&cpu_lock);
15550 * If DTrace helper tracing is enabled, we need to allocate the
15551 * trace buffer and initialize the values.
15553 if (dtrace_helptrace_enabled) {
15554 ASSERT(dtrace_helptrace_buffer == NULL);
15555 dtrace_helptrace_buffer =
15556 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15557 dtrace_helptrace_next = 0;
15561 * If there are already providers, we must ask them to provide their
15562 * probes, and then match any anonymous enabling against them. Note
15563 * that there should be no other retained enablings at this time:
15564 * the only retained enablings at this time should be the anonymous
15567 if (dtrace_anon.dta_enabling != NULL) {
15568 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15570 dtrace_enabling_provide(NULL);
15571 state = dtrace_anon.dta_state;
15574 * We couldn't hold cpu_lock across the above call to
15575 * dtrace_enabling_provide(), but we must hold it to actually
15576 * enable the probes. We have to drop all of our locks, pick
15577 * up cpu_lock, and regain our locks before matching the
15578 * retained anonymous enabling.
15580 mutex_exit(&dtrace_lock);
15581 mutex_exit(&dtrace_provider_lock);
15583 mutex_enter(&cpu_lock);
15584 mutex_enter(&dtrace_provider_lock);
15585 mutex_enter(&dtrace_lock);
15587 if ((enab = dtrace_anon.dta_enabling) != NULL)
15588 (void) dtrace_enabling_match(enab, NULL);
15590 mutex_exit(&cpu_lock);
15593 mutex_exit(&dtrace_lock);
15594 mutex_exit(&dtrace_provider_lock);
15596 if (state != NULL) {
15598 * If we created any anonymous state, set it going now.
15600 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15603 return (DDI_SUCCESS);
15608 #if __FreeBSD_version >= 800039
15609 static void dtrace_dtr(void *);
15616 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15618 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15621 dtrace_state_t *state;
15627 if (getminor(*devp) == DTRACEMNRN_HELPER)
15631 * If this wasn't an open with the "helper" minor, then it must be
15632 * the "dtrace" minor.
15634 ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15636 cred_t *cred_p = NULL;
15638 #if __FreeBSD_version < 800039
15640 * The first minor device is the one that is cloned so there is
15641 * nothing more to do here.
15643 if (dev2unit(dev) == 0)
15647 * Devices are cloned, so if the DTrace state has already
15648 * been allocated, that means this device belongs to a
15649 * different client. Each client should open '/dev/dtrace'
15650 * to get a cloned device.
15652 if (dev->si_drv1 != NULL)
15656 cred_p = dev->si_cred;
15660 * If no DTRACE_PRIV_* bits are set in the credential, then the
15661 * caller lacks sufficient permission to do anything with DTrace.
15663 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15664 if (priv == DTRACE_PRIV_NONE) {
15666 #if __FreeBSD_version < 800039
15667 /* Destroy the cloned device. */
15676 * Ask all providers to provide all their probes.
15678 mutex_enter(&dtrace_provider_lock);
15679 dtrace_probe_provide(NULL, NULL);
15680 mutex_exit(&dtrace_provider_lock);
15682 mutex_enter(&cpu_lock);
15683 mutex_enter(&dtrace_lock);
15685 dtrace_membar_producer();
15689 * If the kernel debugger is active (that is, if the kernel debugger
15690 * modified text in some way), we won't allow the open.
15692 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15694 mutex_exit(&cpu_lock);
15695 mutex_exit(&dtrace_lock);
15699 state = dtrace_state_create(devp, cred_p);
15701 state = dtrace_state_create(dev);
15702 #if __FreeBSD_version < 800039
15703 dev->si_drv1 = state;
15705 devfs_set_cdevpriv(state, dtrace_dtr);
15707 /* This code actually belongs in dtrace_attach() */
15708 if (dtrace_opens == 1)
15709 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15713 mutex_exit(&cpu_lock);
15715 if (state == NULL) {
15717 if (--dtrace_opens == 0)
15718 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15722 mutex_exit(&dtrace_lock);
15724 #if __FreeBSD_version < 800039
15725 /* Destroy the cloned device. */
15732 mutex_exit(&dtrace_lock);
15740 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15741 #elif __FreeBSD_version < 800039
15743 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15746 dtrace_dtr(void *data)
15750 minor_t minor = getminor(dev);
15751 dtrace_state_t *state;
15753 if (minor == DTRACEMNRN_HELPER)
15756 state = ddi_get_soft_state(dtrace_softstate, minor);
15758 #if __FreeBSD_version < 800039
15759 dtrace_state_t *state = dev->si_drv1;
15761 /* Check if this is not a cloned device. */
15762 if (dev2unit(dev) == 0)
15765 dtrace_state_t *state = data;
15770 mutex_enter(&cpu_lock);
15771 mutex_enter(&dtrace_lock);
15773 if (state != NULL) {
15774 if (state->dts_anon) {
15776 * There is anonymous state. Destroy that first.
15778 ASSERT(dtrace_anon.dta_state == NULL);
15779 dtrace_state_destroy(state->dts_anon);
15782 dtrace_state_destroy(state);
15785 kmem_free(state, 0);
15786 #if __FreeBSD_version < 800039
15787 dev->si_drv1 = NULL;
15792 ASSERT(dtrace_opens > 0);
15794 if (--dtrace_opens == 0)
15795 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15798 /* This code actually belongs in dtrace_detach() */
15799 if ((dtrace_opens == 0) && (dtrace_taskq != NULL)) {
15800 taskq_destroy(dtrace_taskq);
15801 dtrace_taskq = NULL;
15805 mutex_exit(&dtrace_lock);
15806 mutex_exit(&cpu_lock);
15808 #if __FreeBSD_version < 800039
15809 /* Schedule this cloned device to be destroyed. */
15810 destroy_dev_sched(dev);
15813 #if defined(sun) || __FreeBSD_version < 800039
15821 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15824 dof_helper_t help, *dhp = NULL;
15827 case DTRACEHIOC_ADDDOF:
15828 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15829 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15834 arg = (intptr_t)help.dofhp_dof;
15837 case DTRACEHIOC_ADD: {
15838 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15843 mutex_enter(&dtrace_lock);
15846 * dtrace_helper_slurp() takes responsibility for the dof --
15847 * it may free it now or it may save it and free it later.
15849 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15856 mutex_exit(&dtrace_lock);
15860 case DTRACEHIOC_REMOVE: {
15861 mutex_enter(&dtrace_lock);
15862 rval = dtrace_helper_destroygen(arg);
15863 mutex_exit(&dtrace_lock);
15877 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15879 minor_t minor = getminor(dev);
15880 dtrace_state_t *state;
15883 if (minor == DTRACEMNRN_HELPER)
15884 return (dtrace_ioctl_helper(cmd, arg, rv));
15886 state = ddi_get_soft_state(dtrace_softstate, minor);
15888 if (state->dts_anon) {
15889 ASSERT(dtrace_anon.dta_state == NULL);
15890 state = state->dts_anon;
15894 case DTRACEIOC_PROVIDER: {
15895 dtrace_providerdesc_t pvd;
15896 dtrace_provider_t *pvp;
15898 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15901 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15902 mutex_enter(&dtrace_provider_lock);
15904 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15905 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15909 mutex_exit(&dtrace_provider_lock);
15914 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15915 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15917 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15923 case DTRACEIOC_EPROBE: {
15924 dtrace_eprobedesc_t epdesc;
15926 dtrace_action_t *act;
15932 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15935 mutex_enter(&dtrace_lock);
15937 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15938 mutex_exit(&dtrace_lock);
15942 if (ecb->dte_probe == NULL) {
15943 mutex_exit(&dtrace_lock);
15947 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15948 epdesc.dtepd_uarg = ecb->dte_uarg;
15949 epdesc.dtepd_size = ecb->dte_size;
15951 nrecs = epdesc.dtepd_nrecs;
15952 epdesc.dtepd_nrecs = 0;
15953 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15954 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15957 epdesc.dtepd_nrecs++;
15961 * Now that we have the size, we need to allocate a temporary
15962 * buffer in which to store the complete description. We need
15963 * the temporary buffer to be able to drop dtrace_lock()
15964 * across the copyout(), below.
15966 size = sizeof (dtrace_eprobedesc_t) +
15967 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15969 buf = kmem_alloc(size, KM_SLEEP);
15970 dest = (uintptr_t)buf;
15972 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15973 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15975 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15976 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15982 bcopy(&act->dta_rec, (void *)dest,
15983 sizeof (dtrace_recdesc_t));
15984 dest += sizeof (dtrace_recdesc_t);
15987 mutex_exit(&dtrace_lock);
15989 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15990 kmem_free(buf, size);
15994 kmem_free(buf, size);
15998 case DTRACEIOC_AGGDESC: {
15999 dtrace_aggdesc_t aggdesc;
16000 dtrace_action_t *act;
16001 dtrace_aggregation_t *agg;
16004 dtrace_recdesc_t *lrec;
16009 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16012 mutex_enter(&dtrace_lock);
16014 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16015 mutex_exit(&dtrace_lock);
16019 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16021 nrecs = aggdesc.dtagd_nrecs;
16022 aggdesc.dtagd_nrecs = 0;
16024 offs = agg->dtag_base;
16025 lrec = &agg->dtag_action.dta_rec;
16026 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16028 for (act = agg->dtag_first; ; act = act->dta_next) {
16029 ASSERT(act->dta_intuple ||
16030 DTRACEACT_ISAGG(act->dta_kind));
16033 * If this action has a record size of zero, it
16034 * denotes an argument to the aggregating action.
16035 * Because the presence of this record doesn't (or
16036 * shouldn't) affect the way the data is interpreted,
16037 * we don't copy it out to save user-level the
16038 * confusion of dealing with a zero-length record.
16040 if (act->dta_rec.dtrd_size == 0) {
16041 ASSERT(agg->dtag_hasarg);
16045 aggdesc.dtagd_nrecs++;
16047 if (act == &agg->dtag_action)
16052 * Now that we have the size, we need to allocate a temporary
16053 * buffer in which to store the complete description. We need
16054 * the temporary buffer to be able to drop dtrace_lock()
16055 * across the copyout(), below.
16057 size = sizeof (dtrace_aggdesc_t) +
16058 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16060 buf = kmem_alloc(size, KM_SLEEP);
16061 dest = (uintptr_t)buf;
16063 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16064 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16066 for (act = agg->dtag_first; ; act = act->dta_next) {
16067 dtrace_recdesc_t rec = act->dta_rec;
16070 * See the comment in the above loop for why we pass
16071 * over zero-length records.
16073 if (rec.dtrd_size == 0) {
16074 ASSERT(agg->dtag_hasarg);
16081 rec.dtrd_offset -= offs;
16082 bcopy(&rec, (void *)dest, sizeof (rec));
16083 dest += sizeof (dtrace_recdesc_t);
16085 if (act == &agg->dtag_action)
16089 mutex_exit(&dtrace_lock);
16091 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16092 kmem_free(buf, size);
16096 kmem_free(buf, size);
16100 case DTRACEIOC_ENABLE: {
16102 dtrace_enabling_t *enab = NULL;
16103 dtrace_vstate_t *vstate;
16109 * If a NULL argument has been passed, we take this as our
16110 * cue to reevaluate our enablings.
16113 dtrace_enabling_matchall();
16118 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16121 mutex_enter(&cpu_lock);
16122 mutex_enter(&dtrace_lock);
16123 vstate = &state->dts_vstate;
16125 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16126 mutex_exit(&dtrace_lock);
16127 mutex_exit(&cpu_lock);
16128 dtrace_dof_destroy(dof);
16132 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16133 mutex_exit(&dtrace_lock);
16134 mutex_exit(&cpu_lock);
16135 dtrace_dof_destroy(dof);
16139 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16140 dtrace_enabling_destroy(enab);
16141 mutex_exit(&dtrace_lock);
16142 mutex_exit(&cpu_lock);
16143 dtrace_dof_destroy(dof);
16147 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16148 err = dtrace_enabling_retain(enab);
16150 dtrace_enabling_destroy(enab);
16153 mutex_exit(&cpu_lock);
16154 mutex_exit(&dtrace_lock);
16155 dtrace_dof_destroy(dof);
16160 case DTRACEIOC_REPLICATE: {
16161 dtrace_repldesc_t desc;
16162 dtrace_probedesc_t *match = &desc.dtrpd_match;
16163 dtrace_probedesc_t *create = &desc.dtrpd_create;
16166 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16169 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16170 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16171 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16172 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16174 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16175 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16176 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16177 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16179 mutex_enter(&dtrace_lock);
16180 err = dtrace_enabling_replicate(state, match, create);
16181 mutex_exit(&dtrace_lock);
16186 case DTRACEIOC_PROBEMATCH:
16187 case DTRACEIOC_PROBES: {
16188 dtrace_probe_t *probe = NULL;
16189 dtrace_probedesc_t desc;
16190 dtrace_probekey_t pkey;
16197 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16200 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16201 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16202 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16203 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16206 * Before we attempt to match this probe, we want to give
16207 * all providers the opportunity to provide it.
16209 if (desc.dtpd_id == DTRACE_IDNONE) {
16210 mutex_enter(&dtrace_provider_lock);
16211 dtrace_probe_provide(&desc, NULL);
16212 mutex_exit(&dtrace_provider_lock);
16216 if (cmd == DTRACEIOC_PROBEMATCH) {
16217 dtrace_probekey(&desc, &pkey);
16218 pkey.dtpk_id = DTRACE_IDNONE;
16221 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16223 mutex_enter(&dtrace_lock);
16225 if (cmd == DTRACEIOC_PROBEMATCH) {
16226 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16227 if ((probe = dtrace_probes[i - 1]) != NULL &&
16228 (m = dtrace_match_probe(probe, &pkey,
16229 priv, uid, zoneid)) != 0)
16234 mutex_exit(&dtrace_lock);
16239 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16240 if ((probe = dtrace_probes[i - 1]) != NULL &&
16241 dtrace_match_priv(probe, priv, uid, zoneid))
16246 if (probe == NULL) {
16247 mutex_exit(&dtrace_lock);
16251 dtrace_probe_description(probe, &desc);
16252 mutex_exit(&dtrace_lock);
16254 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16260 case DTRACEIOC_PROBEARG: {
16261 dtrace_argdesc_t desc;
16262 dtrace_probe_t *probe;
16263 dtrace_provider_t *prov;
16265 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16268 if (desc.dtargd_id == DTRACE_IDNONE)
16271 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16274 mutex_enter(&dtrace_provider_lock);
16275 mutex_enter(&mod_lock);
16276 mutex_enter(&dtrace_lock);
16278 if (desc.dtargd_id > dtrace_nprobes) {
16279 mutex_exit(&dtrace_lock);
16280 mutex_exit(&mod_lock);
16281 mutex_exit(&dtrace_provider_lock);
16285 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16286 mutex_exit(&dtrace_lock);
16287 mutex_exit(&mod_lock);
16288 mutex_exit(&dtrace_provider_lock);
16292 mutex_exit(&dtrace_lock);
16294 prov = probe->dtpr_provider;
16296 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16298 * There isn't any typed information for this probe.
16299 * Set the argument number to DTRACE_ARGNONE.
16301 desc.dtargd_ndx = DTRACE_ARGNONE;
16303 desc.dtargd_native[0] = '\0';
16304 desc.dtargd_xlate[0] = '\0';
16305 desc.dtargd_mapping = desc.dtargd_ndx;
16307 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16308 probe->dtpr_id, probe->dtpr_arg, &desc);
16311 mutex_exit(&mod_lock);
16312 mutex_exit(&dtrace_provider_lock);
16314 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16320 case DTRACEIOC_GO: {
16321 processorid_t cpuid;
16322 rval = dtrace_state_go(state, &cpuid);
16327 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16333 case DTRACEIOC_STOP: {
16334 processorid_t cpuid;
16336 mutex_enter(&dtrace_lock);
16337 rval = dtrace_state_stop(state, &cpuid);
16338 mutex_exit(&dtrace_lock);
16343 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16349 case DTRACEIOC_DOFGET: {
16350 dof_hdr_t hdr, *dof;
16353 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16356 mutex_enter(&dtrace_lock);
16357 dof = dtrace_dof_create(state);
16358 mutex_exit(&dtrace_lock);
16360 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16361 rval = copyout(dof, (void *)arg, len);
16362 dtrace_dof_destroy(dof);
16364 return (rval == 0 ? 0 : EFAULT);
16367 case DTRACEIOC_AGGSNAP:
16368 case DTRACEIOC_BUFSNAP: {
16369 dtrace_bufdesc_t desc;
16371 dtrace_buffer_t *buf;
16373 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16376 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16379 mutex_enter(&dtrace_lock);
16381 if (cmd == DTRACEIOC_BUFSNAP) {
16382 buf = &state->dts_buffer[desc.dtbd_cpu];
16384 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16387 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16388 size_t sz = buf->dtb_offset;
16390 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16391 mutex_exit(&dtrace_lock);
16396 * If this buffer has already been consumed, we're
16397 * going to indicate that there's nothing left here
16400 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16401 mutex_exit(&dtrace_lock);
16403 desc.dtbd_size = 0;
16404 desc.dtbd_drops = 0;
16405 desc.dtbd_errors = 0;
16406 desc.dtbd_oldest = 0;
16407 sz = sizeof (desc);
16409 if (copyout(&desc, (void *)arg, sz) != 0)
16416 * If this is a ring buffer that has wrapped, we want
16417 * to copy the whole thing out.
16419 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16420 dtrace_buffer_polish(buf);
16421 sz = buf->dtb_size;
16424 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16425 mutex_exit(&dtrace_lock);
16429 desc.dtbd_size = sz;
16430 desc.dtbd_drops = buf->dtb_drops;
16431 desc.dtbd_errors = buf->dtb_errors;
16432 desc.dtbd_oldest = buf->dtb_xamot_offset;
16433 desc.dtbd_timestamp = dtrace_gethrtime();
16435 mutex_exit(&dtrace_lock);
16437 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16440 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16445 if (buf->dtb_tomax == NULL) {
16446 ASSERT(buf->dtb_xamot == NULL);
16447 mutex_exit(&dtrace_lock);
16451 cached = buf->dtb_tomax;
16452 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16454 dtrace_xcall(desc.dtbd_cpu,
16455 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16457 state->dts_errors += buf->dtb_xamot_errors;
16460 * If the buffers did not actually switch, then the cross call
16461 * did not take place -- presumably because the given CPU is
16462 * not in the ready set. If this is the case, we'll return
16465 if (buf->dtb_tomax == cached) {
16466 ASSERT(buf->dtb_xamot != cached);
16467 mutex_exit(&dtrace_lock);
16471 ASSERT(cached == buf->dtb_xamot);
16474 * We have our snapshot; now copy it out.
16476 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16477 buf->dtb_xamot_offset) != 0) {
16478 mutex_exit(&dtrace_lock);
16482 desc.dtbd_size = buf->dtb_xamot_offset;
16483 desc.dtbd_drops = buf->dtb_xamot_drops;
16484 desc.dtbd_errors = buf->dtb_xamot_errors;
16485 desc.dtbd_oldest = 0;
16486 desc.dtbd_timestamp = buf->dtb_switched;
16488 mutex_exit(&dtrace_lock);
16491 * Finally, copy out the buffer description.
16493 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16499 case DTRACEIOC_CONF: {
16500 dtrace_conf_t conf;
16502 bzero(&conf, sizeof (conf));
16503 conf.dtc_difversion = DIF_VERSION;
16504 conf.dtc_difintregs = DIF_DIR_NREGS;
16505 conf.dtc_diftupregs = DIF_DTR_NREGS;
16506 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16508 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16514 case DTRACEIOC_STATUS: {
16515 dtrace_status_t stat;
16516 dtrace_dstate_t *dstate;
16521 * See the comment in dtrace_state_deadman() for the reason
16522 * for setting dts_laststatus to INT64_MAX before setting
16523 * it to the correct value.
16525 state->dts_laststatus = INT64_MAX;
16526 dtrace_membar_producer();
16527 state->dts_laststatus = dtrace_gethrtime();
16529 bzero(&stat, sizeof (stat));
16531 mutex_enter(&dtrace_lock);
16533 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16534 mutex_exit(&dtrace_lock);
16538 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16539 stat.dtst_exiting = 1;
16541 nerrs = state->dts_errors;
16542 dstate = &state->dts_vstate.dtvs_dynvars;
16544 for (i = 0; i < NCPU; i++) {
16545 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16547 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16548 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16549 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16551 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16552 stat.dtst_filled++;
16554 nerrs += state->dts_buffer[i].dtb_errors;
16556 for (j = 0; j < state->dts_nspeculations; j++) {
16557 dtrace_speculation_t *spec;
16558 dtrace_buffer_t *buf;
16560 spec = &state->dts_speculations[j];
16561 buf = &spec->dtsp_buffer[i];
16562 stat.dtst_specdrops += buf->dtb_xamot_drops;
16566 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16567 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16568 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16569 stat.dtst_dblerrors = state->dts_dblerrors;
16571 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16572 stat.dtst_errors = nerrs;
16574 mutex_exit(&dtrace_lock);
16576 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16582 case DTRACEIOC_FORMAT: {
16583 dtrace_fmtdesc_t fmt;
16587 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16590 mutex_enter(&dtrace_lock);
16592 if (fmt.dtfd_format == 0 ||
16593 fmt.dtfd_format > state->dts_nformats) {
16594 mutex_exit(&dtrace_lock);
16599 * Format strings are allocated contiguously and they are
16600 * never freed; if a format index is less than the number
16601 * of formats, we can assert that the format map is non-NULL
16602 * and that the format for the specified index is non-NULL.
16604 ASSERT(state->dts_formats != NULL);
16605 str = state->dts_formats[fmt.dtfd_format - 1];
16606 ASSERT(str != NULL);
16608 len = strlen(str) + 1;
16610 if (len > fmt.dtfd_length) {
16611 fmt.dtfd_length = len;
16613 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16614 mutex_exit(&dtrace_lock);
16618 if (copyout(str, fmt.dtfd_string, len) != 0) {
16619 mutex_exit(&dtrace_lock);
16624 mutex_exit(&dtrace_lock);
16637 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16639 dtrace_state_t *state;
16646 return (DDI_SUCCESS);
16649 return (DDI_FAILURE);
16652 mutex_enter(&cpu_lock);
16653 mutex_enter(&dtrace_provider_lock);
16654 mutex_enter(&dtrace_lock);
16656 ASSERT(dtrace_opens == 0);
16658 if (dtrace_helpers > 0) {
16659 mutex_exit(&dtrace_provider_lock);
16660 mutex_exit(&dtrace_lock);
16661 mutex_exit(&cpu_lock);
16662 return (DDI_FAILURE);
16665 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16666 mutex_exit(&dtrace_provider_lock);
16667 mutex_exit(&dtrace_lock);
16668 mutex_exit(&cpu_lock);
16669 return (DDI_FAILURE);
16672 dtrace_provider = NULL;
16674 if ((state = dtrace_anon_grab()) != NULL) {
16676 * If there were ECBs on this state, the provider should
16677 * have not been allowed to detach; assert that there is
16680 ASSERT(state->dts_necbs == 0);
16681 dtrace_state_destroy(state);
16684 * If we're being detached with anonymous state, we need to
16685 * indicate to the kernel debugger that DTrace is now inactive.
16687 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16690 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16691 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16692 dtrace_cpu_init = NULL;
16693 dtrace_helpers_cleanup = NULL;
16694 dtrace_helpers_fork = NULL;
16695 dtrace_cpustart_init = NULL;
16696 dtrace_cpustart_fini = NULL;
16697 dtrace_debugger_init = NULL;
16698 dtrace_debugger_fini = NULL;
16699 dtrace_modload = NULL;
16700 dtrace_modunload = NULL;
16702 mutex_exit(&cpu_lock);
16704 if (dtrace_helptrace_enabled) {
16705 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16706 dtrace_helptrace_buffer = NULL;
16709 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16710 dtrace_probes = NULL;
16711 dtrace_nprobes = 0;
16713 dtrace_hash_destroy(dtrace_bymod);
16714 dtrace_hash_destroy(dtrace_byfunc);
16715 dtrace_hash_destroy(dtrace_byname);
16716 dtrace_bymod = NULL;
16717 dtrace_byfunc = NULL;
16718 dtrace_byname = NULL;
16720 kmem_cache_destroy(dtrace_state_cache);
16721 vmem_destroy(dtrace_minor);
16722 vmem_destroy(dtrace_arena);
16724 if (dtrace_toxrange != NULL) {
16725 kmem_free(dtrace_toxrange,
16726 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16727 dtrace_toxrange = NULL;
16728 dtrace_toxranges = 0;
16729 dtrace_toxranges_max = 0;
16732 ddi_remove_minor_node(dtrace_devi, NULL);
16733 dtrace_devi = NULL;
16735 ddi_soft_state_fini(&dtrace_softstate);
16737 ASSERT(dtrace_vtime_references == 0);
16738 ASSERT(dtrace_opens == 0);
16739 ASSERT(dtrace_retained == NULL);
16741 mutex_exit(&dtrace_lock);
16742 mutex_exit(&dtrace_provider_lock);
16745 * We don't destroy the task queue until after we have dropped our
16746 * locks (taskq_destroy() may block on running tasks). To prevent
16747 * attempting to do work after we have effectively detached but before
16748 * the task queue has been destroyed, all tasks dispatched via the
16749 * task queue must check that DTrace is still attached before
16750 * performing any operation.
16752 taskq_destroy(dtrace_taskq);
16753 dtrace_taskq = NULL;
16755 return (DDI_SUCCESS);
16762 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16767 case DDI_INFO_DEVT2DEVINFO:
16768 *result = (void *)dtrace_devi;
16769 error = DDI_SUCCESS;
16771 case DDI_INFO_DEVT2INSTANCE:
16772 *result = (void *)0;
16773 error = DDI_SUCCESS;
16776 error = DDI_FAILURE;
16783 static struct cb_ops dtrace_cb_ops = {
16784 dtrace_open, /* open */
16785 dtrace_close, /* close */
16786 nulldev, /* strategy */
16787 nulldev, /* print */
16791 dtrace_ioctl, /* ioctl */
16792 nodev, /* devmap */
16794 nodev, /* segmap */
16795 nochpoll, /* poll */
16796 ddi_prop_op, /* cb_prop_op */
16798 D_NEW | D_MP /* Driver compatibility flag */
16801 static struct dev_ops dtrace_ops = {
16802 DEVO_REV, /* devo_rev */
16804 dtrace_info, /* get_dev_info */
16805 nulldev, /* identify */
16806 nulldev, /* probe */
16807 dtrace_attach, /* attach */
16808 dtrace_detach, /* detach */
16810 &dtrace_cb_ops, /* driver operations */
16811 NULL, /* bus operations */
16812 nodev /* dev power */
16815 static struct modldrv modldrv = {
16816 &mod_driverops, /* module type (this is a pseudo driver) */
16817 "Dynamic Tracing", /* name of module */
16818 &dtrace_ops, /* driver ops */
16821 static struct modlinkage modlinkage = {
16830 return (mod_install(&modlinkage));
16834 _info(struct modinfo *modinfop)
16836 return (mod_info(&modlinkage, modinfop));
16842 return (mod_remove(&modlinkage));
16846 static d_ioctl_t dtrace_ioctl;
16847 static d_ioctl_t dtrace_ioctl_helper;
16848 static void dtrace_load(void *);
16849 static int dtrace_unload(void);
16850 #if __FreeBSD_version < 800039
16851 static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16852 static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */
16853 static eventhandler_tag eh_tag; /* Event handler tag. */
16855 static struct cdev *dtrace_dev;
16856 static struct cdev *helper_dev;
16859 void dtrace_invop_init(void);
16860 void dtrace_invop_uninit(void);
16862 static struct cdevsw dtrace_cdevsw = {
16863 .d_version = D_VERSION,
16864 #if __FreeBSD_version < 800039
16865 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16866 .d_close = dtrace_close,
16868 .d_ioctl = dtrace_ioctl,
16869 .d_open = dtrace_open,
16870 .d_name = "dtrace",
16873 static struct cdevsw helper_cdevsw = {
16874 .d_version = D_VERSION,
16875 .d_ioctl = dtrace_ioctl_helper,
16876 .d_name = "helper",
16879 #include <dtrace_anon.c>
16880 #if __FreeBSD_version < 800039
16881 #include <dtrace_clone.c>
16883 #include <dtrace_ioctl.c>
16884 #include <dtrace_load.c>
16885 #include <dtrace_modevent.c>
16886 #include <dtrace_sysctl.c>
16887 #include <dtrace_unload.c>
16888 #include <dtrace_vtime.c>
16889 #include <dtrace_hacks.c>
16890 #include <dtrace_isa.c>
16892 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16893 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16894 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16896 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16897 MODULE_VERSION(dtrace, 1);
16898 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
16899 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);