4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
25 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
26 * Use is subject to license terms.
29 #pragma ident "%Z%%M% %I% %E% SMI"
32 * DTrace - Dynamic Tracing for Solaris
34 * This is the implementation of the Solaris Dynamic Tracing framework
35 * (DTrace). The user-visible interface to DTrace is described at length in
36 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
37 * library, the in-kernel DTrace framework, and the DTrace providers are
38 * described in the block comments in the <sys/dtrace.h> header file. The
39 * internal architecture of DTrace is described in the block comments in the
40 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
41 * implementation very much assume mastery of all of these sources; if one has
42 * an unanswered question about the implementation, one should consult them
45 * The functions here are ordered roughly as follows:
47 * - Probe context functions
48 * - Probe hashing functions
49 * - Non-probe context utility functions
50 * - Matching functions
51 * - Provider-to-Framework API functions
52 * - Probe management functions
53 * - DIF object functions
55 * - Predicate functions
58 * - Enabling functions
60 * - Anonymous enabling functions
61 * - Consumer state functions
64 * - Driver cookbook functions
66 * Each group of functions begins with a block comment labelled the "DTrace
67 * [Group] Functions", allowing one to find each block by searching forward
68 * on capital-f functions.
70 #include <sys/errno.h>
75 #include <sys/modctl.h>
77 #include <sys/systm.h>
80 #include <sys/sunddi.h>
82 #include <sys/cpuvar.h>
85 #include <sys/strsubr.h>
87 #include <sys/sysmacros.h>
88 #include <sys/dtrace_impl.h>
89 #include <sys/atomic.h>
90 #include <sys/cmn_err.h>
92 #include <sys/mutex_impl.h>
93 #include <sys/rwlock_impl.h>
95 #include <sys/ctf_api.h>
97 #include <sys/panic.h>
98 #include <sys/priv_impl.h>
100 #include <sys/policy.h>
102 #include <sys/cred_impl.h>
103 #include <sys/procfs_isa.h>
105 #include <sys/taskq.h>
107 #include <sys/mkdev.h>
110 #include <sys/zone.h>
111 #include <sys/socket.h>
112 #include <netinet/in.h>
114 /* FreeBSD includes: */
116 #include <sys/callout.h>
117 #include <sys/ctype.h>
118 #include <sys/limits.h>
120 #include <sys/kernel.h>
121 #include <sys/malloc.h>
122 #include <sys/sysctl.h>
123 #include <sys/lock.h>
124 #include <sys/mutex.h>
125 #include <sys/rwlock.h>
127 #include <sys/dtrace_bsd.h>
128 #include <netinet/in.h>
129 #include "dtrace_cddl.h"
130 #include "dtrace_debug.c"
134 * DTrace Tunable Variables
136 * The following variables may be tuned by adding a line to /etc/system that
137 * includes both the name of the DTrace module ("dtrace") and the name of the
138 * variable. For example:
140 * set dtrace:dtrace_destructive_disallow = 1
142 * In general, the only variables that one should be tuning this way are those
143 * that affect system-wide DTrace behavior, and for which the default behavior
144 * is undesirable. Most of these variables are tunable on a per-consumer
145 * basis using DTrace options, and need not be tuned on a system-wide basis.
146 * When tuning these variables, avoid pathological values; while some attempt
147 * is made to verify the integrity of these variables, they are not considered
148 * part of the supported interface to DTrace, and they are therefore not
149 * checked comprehensively. Further, these variables should not be tuned
150 * dynamically via "mdb -kw" or other means; they should only be tuned via
153 int dtrace_destructive_disallow = 0;
154 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
155 size_t dtrace_difo_maxsize = (256 * 1024);
156 dtrace_optval_t dtrace_dof_maxsize = (256 * 1024);
157 size_t dtrace_global_maxsize = (16 * 1024);
158 size_t dtrace_actions_max = (16 * 1024);
159 size_t dtrace_retain_max = 1024;
160 dtrace_optval_t dtrace_helper_actions_max = 32;
161 dtrace_optval_t dtrace_helper_providers_max = 32;
162 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
163 size_t dtrace_strsize_default = 256;
164 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
165 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
166 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
167 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
168 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
169 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
170 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
171 dtrace_optval_t dtrace_nspec_default = 1;
172 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
173 dtrace_optval_t dtrace_stackframes_default = 20;
174 dtrace_optval_t dtrace_ustackframes_default = 20;
175 dtrace_optval_t dtrace_jstackframes_default = 50;
176 dtrace_optval_t dtrace_jstackstrsize_default = 512;
177 int dtrace_msgdsize_max = 128;
178 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
179 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
180 int dtrace_devdepth_max = 32;
181 int dtrace_err_verbose;
182 hrtime_t dtrace_deadman_interval = NANOSEC;
183 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
184 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
187 * DTrace External Variables
189 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
190 * available to DTrace consumers via the backtick (`) syntax. One of these,
191 * dtrace_zero, is made deliberately so: it is provided as a source of
192 * well-known, zero-filled memory. While this variable is not documented,
193 * it is used by some translators as an implementation detail.
195 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
198 * DTrace Internal Variables
201 static dev_info_t *dtrace_devi; /* device info */
204 static vmem_t *dtrace_arena; /* probe ID arena */
205 static vmem_t *dtrace_minor; /* minor number arena */
206 static taskq_t *dtrace_taskq; /* task queue */
208 static struct unrhdr *dtrace_arena; /* Probe ID number. */
210 static dtrace_probe_t **dtrace_probes; /* array of all probes */
211 static int dtrace_nprobes; /* number of probes */
212 static dtrace_provider_t *dtrace_provider; /* provider list */
213 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
214 static int dtrace_opens; /* number of opens */
215 static int dtrace_helpers; /* number of helpers */
217 static void *dtrace_softstate; /* softstate pointer */
219 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
220 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
221 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
222 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
223 static int dtrace_toxranges; /* number of toxic ranges */
224 static int dtrace_toxranges_max; /* size of toxic range array */
225 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
226 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
227 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
228 static kthread_t *dtrace_panicked; /* panicking thread */
229 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
230 static dtrace_genid_t dtrace_probegen; /* current probe generation */
231 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
232 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
233 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
235 static struct mtx dtrace_unr_mtx;
236 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
237 int dtrace_in_probe; /* non-zero if executing a probe */
238 #if defined(__i386__) || defined(__amd64__)
239 uintptr_t dtrace_in_probe_addr; /* Address of invop when already in probe */
245 * DTrace is protected by three (relatively coarse-grained) locks:
247 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
248 * including enabling state, probes, ECBs, consumer state, helper state,
249 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
250 * probe context is lock-free -- synchronization is handled via the
251 * dtrace_sync() cross call mechanism.
253 * (2) dtrace_provider_lock is required when manipulating provider state, or
254 * when provider state must be held constant.
256 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
257 * when meta provider state must be held constant.
259 * The lock ordering between these three locks is dtrace_meta_lock before
260 * dtrace_provider_lock before dtrace_lock. (In particular, there are
261 * several places where dtrace_provider_lock is held by the framework as it
262 * calls into the providers -- which then call back into the framework,
263 * grabbing dtrace_lock.)
265 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
266 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
267 * role as a coarse-grained lock; it is acquired before both of these locks.
268 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
269 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
270 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
271 * acquired _between_ dtrace_provider_lock and dtrace_lock.
273 static kmutex_t dtrace_lock; /* probe state lock */
274 static kmutex_t dtrace_provider_lock; /* provider state lock */
275 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
278 /* XXX FreeBSD hacks. */
279 static kmutex_t mod_lock;
281 #define cr_suid cr_svuid
282 #define cr_sgid cr_svgid
283 #define ipaddr_t in_addr_t
284 #define mod_modname pathname
285 #define vuprintf vprintf
286 #define ttoproc(_a) ((_a)->td_proc)
287 #define crgetzoneid(_a) 0
290 #define CPU_ON_INTR(_a) 0
292 #define PRIV_EFFECTIVE (1 << 0)
293 #define PRIV_DTRACE_KERNEL (1 << 1)
294 #define PRIV_DTRACE_PROC (1 << 2)
295 #define PRIV_DTRACE_USER (1 << 3)
296 #define PRIV_PROC_OWNER (1 << 4)
297 #define PRIV_PROC_ZONE (1 << 5)
300 SYSCTL_NODE(_debug, OID_AUTO, dtrace, CTLFLAG_RD, 0, "DTrace Information");
304 #define curcpu CPU->cpu_id
309 * DTrace Provider Variables
311 * These are the variables relating to DTrace as a provider (that is, the
312 * provider of the BEGIN, END, and ERROR probes).
314 static dtrace_pattr_t dtrace_provider_attr = {
315 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
316 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
317 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
318 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
319 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
326 static dtrace_pops_t dtrace_provider_ops = {
327 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
328 (void (*)(void *, modctl_t *))dtrace_nullop,
329 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
330 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
331 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
332 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
336 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
339 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
340 static dtrace_id_t dtrace_probeid_end; /* special END probe */
341 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
344 * DTrace Helper Tracing Variables
346 uint32_t dtrace_helptrace_next = 0;
347 uint32_t dtrace_helptrace_nlocals;
348 char *dtrace_helptrace_buffer;
349 int dtrace_helptrace_bufsize = 512 * 1024;
352 int dtrace_helptrace_enabled = 1;
354 int dtrace_helptrace_enabled = 0;
358 * DTrace Error Hashing
360 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
361 * table. This is very useful for checking coverage of tests that are
362 * expected to induce DIF or DOF processing errors, and may be useful for
363 * debugging problems in the DIF code generator or in DOF generation . The
364 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
367 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
368 static const char *dtrace_errlast;
369 static kthread_t *dtrace_errthread;
370 static kmutex_t dtrace_errlock;
374 * DTrace Macros and Constants
376 * These are various macros that are useful in various spots in the
377 * implementation, along with a few random constants that have no meaning
378 * outside of the implementation. There is no real structure to this cpp
379 * mishmash -- but is there ever?
381 #define DTRACE_HASHSTR(hash, probe) \
382 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
384 #define DTRACE_HASHNEXT(hash, probe) \
385 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
387 #define DTRACE_HASHPREV(hash, probe) \
388 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
390 #define DTRACE_HASHEQ(hash, lhs, rhs) \
391 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
392 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
394 #define DTRACE_AGGHASHSIZE_SLEW 17
396 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
399 * The key for a thread-local variable consists of the lower 61 bits of the
400 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
401 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
402 * equal to a variable identifier. This is necessary (but not sufficient) to
403 * assure that global associative arrays never collide with thread-local
404 * variables. To guarantee that they cannot collide, we must also define the
405 * order for keying dynamic variables. That order is:
407 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
409 * Because the variable-key and the tls-key are in orthogonal spaces, there is
410 * no way for a global variable key signature to match a thread-local key
414 #define DTRACE_TLS_THRKEY(where) { \
416 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
417 for (; actv; actv >>= 1) \
419 ASSERT(intr < (1 << 3)); \
420 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
421 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
424 #define DTRACE_TLS_THRKEY(where) { \
425 solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
427 uint_t actv = _c->cpu_intr_actv; \
428 for (; actv; actv >>= 1) \
430 ASSERT(intr < (1 << 3)); \
431 (where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
432 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
436 #define DT_BSWAP_8(x) ((x) & 0xff)
437 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
438 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
439 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
441 #define DT_MASK_LO 0x00000000FFFFFFFFULL
443 #define DTRACE_STORE(type, tomax, offset, what) \
444 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
447 #define DTRACE_ALIGNCHECK(addr, size, flags) \
448 if (addr & (size - 1)) { \
449 *flags |= CPU_DTRACE_BADALIGN; \
450 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
454 #define DTRACE_ALIGNCHECK(addr, size, flags)
458 * Test whether a range of memory starting at testaddr of size testsz falls
459 * within the range of memory described by addr, sz. We take care to avoid
460 * problems with overflow and underflow of the unsigned quantities, and
461 * disallow all negative sizes. Ranges of size 0 are allowed.
463 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
464 ((testaddr) - (baseaddr) < (basesz) && \
465 (testaddr) + (testsz) - (baseaddr) <= (basesz) && \
466 (testaddr) + (testsz) >= (testaddr))
469 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
470 * alloc_sz on the righthand side of the comparison in order to avoid overflow
471 * or underflow in the comparison with it. This is simpler than the INRANGE
472 * check above, because we know that the dtms_scratch_ptr is valid in the
473 * range. Allocations of size zero are allowed.
475 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
476 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
477 (mstate)->dtms_scratch_ptr >= (alloc_sz))
479 #define DTRACE_LOADFUNC(bits) \
482 dtrace_load##bits(uintptr_t addr) \
484 size_t size = bits / NBBY; \
486 uint##bits##_t rval; \
488 volatile uint16_t *flags = (volatile uint16_t *) \
489 &cpu_core[curcpu].cpuc_dtrace_flags; \
491 DTRACE_ALIGNCHECK(addr, size, flags); \
493 for (i = 0; i < dtrace_toxranges; i++) { \
494 if (addr >= dtrace_toxrange[i].dtt_limit) \
497 if (addr + size <= dtrace_toxrange[i].dtt_base) \
501 * This address falls within a toxic region; return 0. \
503 *flags |= CPU_DTRACE_BADADDR; \
504 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
508 *flags |= CPU_DTRACE_NOFAULT; \
510 rval = *((volatile uint##bits##_t *)addr); \
511 *flags &= ~CPU_DTRACE_NOFAULT; \
513 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
517 #define dtrace_loadptr dtrace_load64
519 #define dtrace_loadptr dtrace_load32
522 #define DTRACE_DYNHASH_FREE 0
523 #define DTRACE_DYNHASH_SINK 1
524 #define DTRACE_DYNHASH_VALID 2
526 #define DTRACE_MATCH_NEXT 0
527 #define DTRACE_MATCH_DONE 1
528 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
529 #define DTRACE_STATE_ALIGN 64
531 #define DTRACE_FLAGS2FLT(flags) \
532 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
533 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
534 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
535 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
536 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
537 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
538 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
539 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
540 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
543 #define DTRACEACT_ISSTRING(act) \
544 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
545 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
547 /* Function prototype definitions: */
548 static size_t dtrace_strlen(const char *, size_t);
549 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
550 static void dtrace_enabling_provide(dtrace_provider_t *);
551 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
552 static void dtrace_enabling_matchall(void);
553 static dtrace_state_t *dtrace_anon_grab(void);
554 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
555 dtrace_state_t *, uint64_t, uint64_t);
556 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
557 static void dtrace_buffer_drop(dtrace_buffer_t *);
558 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
559 dtrace_state_t *, dtrace_mstate_t *);
560 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
562 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
563 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
564 uint16_t dtrace_load16(uintptr_t);
565 uint32_t dtrace_load32(uintptr_t);
566 uint64_t dtrace_load64(uintptr_t);
567 uint8_t dtrace_load8(uintptr_t);
568 void dtrace_dynvar_clean(dtrace_dstate_t *);
569 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
570 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
571 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
574 * DTrace Probe Context Functions
576 * These functions are called from probe context. Because probe context is
577 * any context in which C may be called, arbitrarily locks may be held,
578 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
579 * As a result, functions called from probe context may only call other DTrace
580 * support functions -- they may not interact at all with the system at large.
581 * (Note that the ASSERT macro is made probe-context safe by redefining it in
582 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
583 * loads are to be performed from probe context, they _must_ be in terms of
584 * the safe dtrace_load*() variants.
586 * Some functions in this block are not actually called from probe context;
587 * for these functions, there will be a comment above the function reading
588 * "Note: not called from probe context."
591 dtrace_panic(const char *format, ...)
595 va_start(alist, format);
596 dtrace_vpanic(format, alist);
601 dtrace_assfail(const char *a, const char *f, int l)
603 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
606 * We just need something here that even the most clever compiler
607 * cannot optimize away.
609 return (a[(uintptr_t)f]);
613 * Atomically increment a specified error counter from probe context.
616 dtrace_error(uint32_t *counter)
619 * Most counters stored to in probe context are per-CPU counters.
620 * However, there are some error conditions that are sufficiently
621 * arcane that they don't merit per-CPU storage. If these counters
622 * are incremented concurrently on different CPUs, scalability will be
623 * adversely affected -- but we don't expect them to be white-hot in a
624 * correctly constructed enabling...
631 if ((nval = oval + 1) == 0) {
633 * If the counter would wrap, set it to 1 -- assuring
634 * that the counter is never zero when we have seen
635 * errors. (The counter must be 32-bits because we
636 * aren't guaranteed a 64-bit compare&swap operation.)
637 * To save this code both the infamy of being fingered
638 * by a priggish news story and the indignity of being
639 * the target of a neo-puritan witch trial, we're
640 * carefully avoiding any colorful description of the
641 * likelihood of this condition -- but suffice it to
642 * say that it is only slightly more likely than the
643 * overflow of predicate cache IDs, as discussed in
644 * dtrace_predicate_create().
648 } while (dtrace_cas32(counter, oval, nval) != oval);
652 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
653 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
661 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
663 if (dest < mstate->dtms_scratch_base)
666 if (dest + size < dest)
669 if (dest + size > mstate->dtms_scratch_ptr)
676 dtrace_canstore_statvar(uint64_t addr, size_t sz,
677 dtrace_statvar_t **svars, int nsvars)
681 for (i = 0; i < nsvars; i++) {
682 dtrace_statvar_t *svar = svars[i];
684 if (svar == NULL || svar->dtsv_size == 0)
687 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
695 * Check to see if the address is within a memory region to which a store may
696 * be issued. This includes the DTrace scratch areas, and any DTrace variable
697 * region. The caller of dtrace_canstore() is responsible for performing any
698 * alignment checks that are needed before stores are actually executed.
701 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
702 dtrace_vstate_t *vstate)
705 * First, check to see if the address is in scratch space...
707 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
708 mstate->dtms_scratch_size))
712 * Now check to see if it's a dynamic variable. This check will pick
713 * up both thread-local variables and any global dynamically-allocated
716 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
717 vstate->dtvs_dynvars.dtds_size)) {
718 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
719 uintptr_t base = (uintptr_t)dstate->dtds_base +
720 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
724 * Before we assume that we can store here, we need to make
725 * sure that it isn't in our metadata -- storing to our
726 * dynamic variable metadata would corrupt our state. For
727 * the range to not include any dynamic variable metadata,
730 * (1) Start above the hash table that is at the base of
731 * the dynamic variable space
733 * (2) Have a starting chunk offset that is beyond the
734 * dtrace_dynvar_t that is at the base of every chunk
736 * (3) Not span a chunk boundary
742 chunkoffs = (addr - base) % dstate->dtds_chunksize;
744 if (chunkoffs < sizeof (dtrace_dynvar_t))
747 if (chunkoffs + sz > dstate->dtds_chunksize)
754 * Finally, check the static local and global variables. These checks
755 * take the longest, so we perform them last.
757 if (dtrace_canstore_statvar(addr, sz,
758 vstate->dtvs_locals, vstate->dtvs_nlocals))
761 if (dtrace_canstore_statvar(addr, sz,
762 vstate->dtvs_globals, vstate->dtvs_nglobals))
770 * Convenience routine to check to see if the address is within a memory
771 * region in which a load may be issued given the user's privilege level;
772 * if not, it sets the appropriate error flags and loads 'addr' into the
773 * illegal value slot.
775 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
776 * appropriate memory access protection.
779 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
780 dtrace_vstate_t *vstate)
782 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
785 * If we hold the privilege to read from kernel memory, then
786 * everything is readable.
788 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
792 * You can obviously read that which you can store.
794 if (dtrace_canstore(addr, sz, mstate, vstate))
798 * We're allowed to read from our own string table.
800 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
801 mstate->dtms_difo->dtdo_strlen))
804 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
810 * Convenience routine to check to see if a given string is within a memory
811 * region in which a load may be issued given the user's privilege level;
812 * this exists so that we don't need to issue unnecessary dtrace_strlen()
813 * calls in the event that the user has all privileges.
816 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
817 dtrace_vstate_t *vstate)
822 * If we hold the privilege to read from kernel memory, then
823 * everything is readable.
825 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
828 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
829 if (dtrace_canload(addr, strsz, mstate, vstate))
836 * Convenience routine to check to see if a given variable is within a memory
837 * region in which a load may be issued given the user's privilege level.
840 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
841 dtrace_vstate_t *vstate)
844 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
847 * If we hold the privilege to read from kernel memory, then
848 * everything is readable.
850 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
853 if (type->dtdt_kind == DIF_TYPE_STRING)
854 sz = dtrace_strlen(src,
855 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
857 sz = type->dtdt_size;
859 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
863 * Compare two strings using safe loads.
866 dtrace_strncmp(char *s1, char *s2, size_t limit)
869 volatile uint16_t *flags;
871 if (s1 == s2 || limit == 0)
874 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
880 c1 = dtrace_load8((uintptr_t)s1++);
886 c2 = dtrace_load8((uintptr_t)s2++);
891 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
897 * Compute strlen(s) for a string using safe memory accesses. The additional
898 * len parameter is used to specify a maximum length to ensure completion.
901 dtrace_strlen(const char *s, size_t lim)
905 for (len = 0; len != lim; len++) {
906 if (dtrace_load8((uintptr_t)s++) == '\0')
914 * Check if an address falls within a toxic region.
917 dtrace_istoxic(uintptr_t kaddr, size_t size)
919 uintptr_t taddr, tsize;
922 for (i = 0; i < dtrace_toxranges; i++) {
923 taddr = dtrace_toxrange[i].dtt_base;
924 tsize = dtrace_toxrange[i].dtt_limit - taddr;
926 if (kaddr - taddr < tsize) {
927 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
928 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
932 if (taddr - kaddr < size) {
933 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
934 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
943 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
944 * memory specified by the DIF program. The dst is assumed to be safe memory
945 * that we can store to directly because it is managed by DTrace. As with
946 * standard bcopy, overlapping copies are handled properly.
949 dtrace_bcopy(const void *src, void *dst, size_t len)
953 const uint8_t *s2 = src;
957 *s1++ = dtrace_load8((uintptr_t)s2++);
958 } while (--len != 0);
964 *--s1 = dtrace_load8((uintptr_t)--s2);
965 } while (--len != 0);
971 * Copy src to dst using safe memory accesses, up to either the specified
972 * length, or the point that a nul byte is encountered. The src is assumed to
973 * be unsafe memory specified by the DIF program. The dst is assumed to be
974 * safe memory that we can store to directly because it is managed by DTrace.
975 * Unlike dtrace_bcopy(), overlapping regions are not handled.
978 dtrace_strcpy(const void *src, void *dst, size_t len)
981 uint8_t *s1 = dst, c;
982 const uint8_t *s2 = src;
985 *s1++ = c = dtrace_load8((uintptr_t)s2++);
986 } while (--len != 0 && c != '\0');
991 * Copy src to dst, deriving the size and type from the specified (BYREF)
992 * variable type. The src is assumed to be unsafe memory specified by the DIF
993 * program. The dst is assumed to be DTrace variable memory that is of the
994 * specified type; we assume that we can store to directly.
997 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
999 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1001 if (type->dtdt_kind == DIF_TYPE_STRING) {
1002 dtrace_strcpy(src, dst, type->dtdt_size);
1004 dtrace_bcopy(src, dst, type->dtdt_size);
1009 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1010 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1011 * safe memory that we can access directly because it is managed by DTrace.
1014 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1016 volatile uint16_t *flags;
1018 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1023 if (s1 == NULL || s2 == NULL)
1026 if (s1 != s2 && len != 0) {
1027 const uint8_t *ps1 = s1;
1028 const uint8_t *ps2 = s2;
1031 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1033 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1039 * Zero the specified region using a simple byte-by-byte loop. Note that this
1040 * is for safe DTrace-managed memory only.
1043 dtrace_bzero(void *dst, size_t len)
1047 for (cp = dst; len != 0; len--)
1052 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1056 result[0] = addend1[0] + addend2[0];
1057 result[1] = addend1[1] + addend2[1] +
1058 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1065 * Shift the 128-bit value in a by b. If b is positive, shift left.
1066 * If b is negative, shift right.
1069 dtrace_shift_128(uint64_t *a, int b)
1079 a[0] = a[1] >> (b - 64);
1083 mask = 1LL << (64 - b);
1085 a[0] |= ((a[1] & mask) << (64 - b));
1090 a[1] = a[0] << (b - 64);
1094 mask = a[0] >> (64 - b);
1102 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1103 * use native multiplication on those, and then re-combine into the
1104 * resulting 128-bit value.
1106 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1113 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1115 uint64_t hi1, hi2, lo1, lo2;
1118 hi1 = factor1 >> 32;
1119 hi2 = factor2 >> 32;
1121 lo1 = factor1 & DT_MASK_LO;
1122 lo2 = factor2 & DT_MASK_LO;
1124 product[0] = lo1 * lo2;
1125 product[1] = hi1 * hi2;
1129 dtrace_shift_128(tmp, 32);
1130 dtrace_add_128(product, tmp, product);
1134 dtrace_shift_128(tmp, 32);
1135 dtrace_add_128(product, tmp, product);
1139 * This privilege check should be used by actions and subroutines to
1140 * verify that the user credentials of the process that enabled the
1141 * invoking ECB match the target credentials
1144 dtrace_priv_proc_common_user(dtrace_state_t *state)
1146 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1149 * We should always have a non-NULL state cred here, since if cred
1150 * is null (anonymous tracing), we fast-path bypass this routine.
1152 ASSERT(s_cr != NULL);
1154 if ((cr = CRED()) != NULL &&
1155 s_cr->cr_uid == cr->cr_uid &&
1156 s_cr->cr_uid == cr->cr_ruid &&
1157 s_cr->cr_uid == cr->cr_suid &&
1158 s_cr->cr_gid == cr->cr_gid &&
1159 s_cr->cr_gid == cr->cr_rgid &&
1160 s_cr->cr_gid == cr->cr_sgid)
1167 * This privilege check should be used by actions and subroutines to
1168 * verify that the zone of the process that enabled the invoking ECB
1169 * matches the target credentials
1172 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1175 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1178 * We should always have a non-NULL state cred here, since if cred
1179 * is null (anonymous tracing), we fast-path bypass this routine.
1181 ASSERT(s_cr != NULL);
1183 if ((cr = CRED()) != NULL &&
1184 s_cr->cr_zone == cr->cr_zone)
1194 * This privilege check should be used by actions and subroutines to
1195 * verify that the process has not setuid or changed credentials.
1198 dtrace_priv_proc_common_nocd(void)
1202 if ((proc = ttoproc(curthread)) != NULL &&
1203 !(proc->p_flag & SNOCD))
1210 dtrace_priv_proc_destructive(dtrace_state_t *state)
1212 int action = state->dts_cred.dcr_action;
1214 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1215 dtrace_priv_proc_common_zone(state) == 0)
1218 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1219 dtrace_priv_proc_common_user(state) == 0)
1222 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1223 dtrace_priv_proc_common_nocd() == 0)
1229 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1235 dtrace_priv_proc_control(dtrace_state_t *state)
1237 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1240 if (dtrace_priv_proc_common_zone(state) &&
1241 dtrace_priv_proc_common_user(state) &&
1242 dtrace_priv_proc_common_nocd())
1245 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1251 dtrace_priv_proc(dtrace_state_t *state)
1253 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1256 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1262 dtrace_priv_kernel(dtrace_state_t *state)
1264 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1267 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1273 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1275 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1278 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1284 * Note: not called from probe context. This function is called
1285 * asynchronously (and at a regular interval) from outside of probe context to
1286 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1287 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1290 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1292 dtrace_dynvar_t *dirty;
1293 dtrace_dstate_percpu_t *dcpu;
1296 for (i = 0; i < NCPU; i++) {
1297 dcpu = &dstate->dtds_percpu[i];
1299 ASSERT(dcpu->dtdsc_rinsing == NULL);
1302 * If the dirty list is NULL, there is no dirty work to do.
1304 if (dcpu->dtdsc_dirty == NULL)
1308 * If the clean list is non-NULL, then we're not going to do
1309 * any work for this CPU -- it means that there has not been
1310 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1311 * since the last time we cleaned house.
1313 if (dcpu->dtdsc_clean != NULL)
1319 * Atomically move the dirty list aside.
1322 dirty = dcpu->dtdsc_dirty;
1325 * Before we zap the dirty list, set the rinsing list.
1326 * (This allows for a potential assertion in
1327 * dtrace_dynvar(): if a free dynamic variable appears
1328 * on a hash chain, either the dirty list or the
1329 * rinsing list for some CPU must be non-NULL.)
1331 dcpu->dtdsc_rinsing = dirty;
1332 dtrace_membar_producer();
1333 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1334 dirty, NULL) != dirty);
1339 * We have no work to do; we can simply return.
1346 for (i = 0; i < NCPU; i++) {
1347 dcpu = &dstate->dtds_percpu[i];
1349 if (dcpu->dtdsc_rinsing == NULL)
1353 * We are now guaranteed that no hash chain contains a pointer
1354 * into this dirty list; we can make it clean.
1356 ASSERT(dcpu->dtdsc_clean == NULL);
1357 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1358 dcpu->dtdsc_rinsing = NULL;
1362 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1363 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1364 * This prevents a race whereby a CPU incorrectly decides that
1365 * the state should be something other than DTRACE_DSTATE_CLEAN
1366 * after dtrace_dynvar_clean() has completed.
1370 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1374 * Depending on the value of the op parameter, this function looks-up,
1375 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1376 * allocation is requested, this function will return a pointer to a
1377 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1378 * variable can be allocated. If NULL is returned, the appropriate counter
1379 * will be incremented.
1382 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1383 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1384 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1386 uint64_t hashval = DTRACE_DYNHASH_VALID;
1387 dtrace_dynhash_t *hash = dstate->dtds_hash;
1388 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1389 processorid_t me = curcpu, cpu = me;
1390 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1391 size_t bucket, ksize;
1392 size_t chunksize = dstate->dtds_chunksize;
1393 uintptr_t kdata, lock, nstate;
1399 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1400 * algorithm. For the by-value portions, we perform the algorithm in
1401 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1402 * bit, and seems to have only a minute effect on distribution. For
1403 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1404 * over each referenced byte. It's painful to do this, but it's much
1405 * better than pathological hash distribution. The efficacy of the
1406 * hashing algorithm (and a comparison with other algorithms) may be
1407 * found by running the ::dtrace_dynstat MDB dcmd.
1409 for (i = 0; i < nkeys; i++) {
1410 if (key[i].dttk_size == 0) {
1411 uint64_t val = key[i].dttk_value;
1413 hashval += (val >> 48) & 0xffff;
1414 hashval += (hashval << 10);
1415 hashval ^= (hashval >> 6);
1417 hashval += (val >> 32) & 0xffff;
1418 hashval += (hashval << 10);
1419 hashval ^= (hashval >> 6);
1421 hashval += (val >> 16) & 0xffff;
1422 hashval += (hashval << 10);
1423 hashval ^= (hashval >> 6);
1425 hashval += val & 0xffff;
1426 hashval += (hashval << 10);
1427 hashval ^= (hashval >> 6);
1430 * This is incredibly painful, but it beats the hell
1431 * out of the alternative.
1433 uint64_t j, size = key[i].dttk_size;
1434 uintptr_t base = (uintptr_t)key[i].dttk_value;
1436 if (!dtrace_canload(base, size, mstate, vstate))
1439 for (j = 0; j < size; j++) {
1440 hashval += dtrace_load8(base + j);
1441 hashval += (hashval << 10);
1442 hashval ^= (hashval >> 6);
1447 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1450 hashval += (hashval << 3);
1451 hashval ^= (hashval >> 11);
1452 hashval += (hashval << 15);
1455 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1456 * comes out to be one of our two sentinel hash values. If this
1457 * actually happens, we set the hashval to be a value known to be a
1458 * non-sentinel value.
1460 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1461 hashval = DTRACE_DYNHASH_VALID;
1464 * Yes, it's painful to do a divide here. If the cycle count becomes
1465 * important here, tricks can be pulled to reduce it. (However, it's
1466 * critical that hash collisions be kept to an absolute minimum;
1467 * they're much more painful than a divide.) It's better to have a
1468 * solution that generates few collisions and still keeps things
1469 * relatively simple.
1471 bucket = hashval % dstate->dtds_hashsize;
1473 if (op == DTRACE_DYNVAR_DEALLOC) {
1474 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1477 while ((lock = *lockp) & 1)
1480 if (dtrace_casptr((volatile void *)lockp,
1481 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1485 dtrace_membar_producer();
1490 lock = hash[bucket].dtdh_lock;
1492 dtrace_membar_consumer();
1494 start = hash[bucket].dtdh_chain;
1495 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1496 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1497 op != DTRACE_DYNVAR_DEALLOC));
1499 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1500 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1501 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1503 if (dvar->dtdv_hashval != hashval) {
1504 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1506 * We've reached the sink, and therefore the
1507 * end of the hash chain; we can kick out of
1508 * the loop knowing that we have seen a valid
1509 * snapshot of state.
1511 ASSERT(dvar->dtdv_next == NULL);
1512 ASSERT(dvar == &dtrace_dynhash_sink);
1516 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1518 * We've gone off the rails: somewhere along
1519 * the line, one of the members of this hash
1520 * chain was deleted. Note that we could also
1521 * detect this by simply letting this loop run
1522 * to completion, as we would eventually hit
1523 * the end of the dirty list. However, we
1524 * want to avoid running the length of the
1525 * dirty list unnecessarily (it might be quite
1526 * long), so we catch this as early as
1527 * possible by detecting the hash marker. In
1528 * this case, we simply set dvar to NULL and
1529 * break; the conditional after the loop will
1530 * send us back to top.
1539 if (dtuple->dtt_nkeys != nkeys)
1542 for (i = 0; i < nkeys; i++, dkey++) {
1543 if (dkey->dttk_size != key[i].dttk_size)
1544 goto next; /* size or type mismatch */
1546 if (dkey->dttk_size != 0) {
1548 (void *)(uintptr_t)key[i].dttk_value,
1549 (void *)(uintptr_t)dkey->dttk_value,
1553 if (dkey->dttk_value != key[i].dttk_value)
1558 if (op != DTRACE_DYNVAR_DEALLOC)
1561 ASSERT(dvar->dtdv_next == NULL ||
1562 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1565 ASSERT(hash[bucket].dtdh_chain != dvar);
1566 ASSERT(start != dvar);
1567 ASSERT(prev->dtdv_next == dvar);
1568 prev->dtdv_next = dvar->dtdv_next;
1570 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1571 start, dvar->dtdv_next) != start) {
1573 * We have failed to atomically swing the
1574 * hash table head pointer, presumably because
1575 * of a conflicting allocation on another CPU.
1576 * We need to reread the hash chain and try
1583 dtrace_membar_producer();
1586 * Now set the hash value to indicate that it's free.
1588 ASSERT(hash[bucket].dtdh_chain != dvar);
1589 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1591 dtrace_membar_producer();
1594 * Set the next pointer to point at the dirty list, and
1595 * atomically swing the dirty pointer to the newly freed dvar.
1598 next = dcpu->dtdsc_dirty;
1599 dvar->dtdv_next = next;
1600 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1603 * Finally, unlock this hash bucket.
1605 ASSERT(hash[bucket].dtdh_lock == lock);
1607 hash[bucket].dtdh_lock++;
1617 * If dvar is NULL, it is because we went off the rails:
1618 * one of the elements that we traversed in the hash chain
1619 * was deleted while we were traversing it. In this case,
1620 * we assert that we aren't doing a dealloc (deallocs lock
1621 * the hash bucket to prevent themselves from racing with
1622 * one another), and retry the hash chain traversal.
1624 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1628 if (op != DTRACE_DYNVAR_ALLOC) {
1630 * If we are not to allocate a new variable, we want to
1631 * return NULL now. Before we return, check that the value
1632 * of the lock word hasn't changed. If it has, we may have
1633 * seen an inconsistent snapshot.
1635 if (op == DTRACE_DYNVAR_NOALLOC) {
1636 if (hash[bucket].dtdh_lock != lock)
1639 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1640 ASSERT(hash[bucket].dtdh_lock == lock);
1642 hash[bucket].dtdh_lock++;
1649 * We need to allocate a new dynamic variable. The size we need is the
1650 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1651 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1652 * the size of any referred-to data (dsize). We then round the final
1653 * size up to the chunksize for allocation.
1655 for (ksize = 0, i = 0; i < nkeys; i++)
1656 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1659 * This should be pretty much impossible, but could happen if, say,
1660 * strange DIF specified the tuple. Ideally, this should be an
1661 * assertion and not an error condition -- but that requires that the
1662 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1663 * bullet-proof. (That is, it must not be able to be fooled by
1664 * malicious DIF.) Given the lack of backwards branches in DIF,
1665 * solving this would presumably not amount to solving the Halting
1666 * Problem -- but it still seems awfully hard.
1668 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1669 ksize + dsize > chunksize) {
1670 dcpu->dtdsc_drops++;
1674 nstate = DTRACE_DSTATE_EMPTY;
1678 free = dcpu->dtdsc_free;
1681 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1684 if (clean == NULL) {
1686 * We're out of dynamic variable space on
1687 * this CPU. Unless we have tried all CPUs,
1688 * we'll try to allocate from a different
1691 switch (dstate->dtds_state) {
1692 case DTRACE_DSTATE_CLEAN: {
1693 void *sp = &dstate->dtds_state;
1698 if (dcpu->dtdsc_dirty != NULL &&
1699 nstate == DTRACE_DSTATE_EMPTY)
1700 nstate = DTRACE_DSTATE_DIRTY;
1702 if (dcpu->dtdsc_rinsing != NULL)
1703 nstate = DTRACE_DSTATE_RINSING;
1705 dcpu = &dstate->dtds_percpu[cpu];
1710 (void) dtrace_cas32(sp,
1711 DTRACE_DSTATE_CLEAN, nstate);
1714 * To increment the correct bean
1715 * counter, take another lap.
1720 case DTRACE_DSTATE_DIRTY:
1721 dcpu->dtdsc_dirty_drops++;
1724 case DTRACE_DSTATE_RINSING:
1725 dcpu->dtdsc_rinsing_drops++;
1728 case DTRACE_DSTATE_EMPTY:
1729 dcpu->dtdsc_drops++;
1733 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1738 * The clean list appears to be non-empty. We want to
1739 * move the clean list to the free list; we start by
1740 * moving the clean pointer aside.
1742 if (dtrace_casptr(&dcpu->dtdsc_clean,
1743 clean, NULL) != clean) {
1745 * We are in one of two situations:
1747 * (a) The clean list was switched to the
1748 * free list by another CPU.
1750 * (b) The clean list was added to by the
1753 * In either of these situations, we can
1754 * just reattempt the free list allocation.
1759 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1762 * Now we'll move the clean list to the free list.
1763 * It's impossible for this to fail: the only way
1764 * the free list can be updated is through this
1765 * code path, and only one CPU can own the clean list.
1766 * Thus, it would only be possible for this to fail if
1767 * this code were racing with dtrace_dynvar_clean().
1768 * (That is, if dtrace_dynvar_clean() updated the clean
1769 * list, and we ended up racing to update the free
1770 * list.) This race is prevented by the dtrace_sync()
1771 * in dtrace_dynvar_clean() -- which flushes the
1772 * owners of the clean lists out before resetting
1775 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1776 ASSERT(rval == NULL);
1781 new_free = dvar->dtdv_next;
1782 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1785 * We have now allocated a new chunk. We copy the tuple keys into the
1786 * tuple array and copy any referenced key data into the data space
1787 * following the tuple array. As we do this, we relocate dttk_value
1788 * in the final tuple to point to the key data address in the chunk.
1790 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1791 dvar->dtdv_data = (void *)(kdata + ksize);
1792 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1794 for (i = 0; i < nkeys; i++) {
1795 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1796 size_t kesize = key[i].dttk_size;
1800 (const void *)(uintptr_t)key[i].dttk_value,
1801 (void *)kdata, kesize);
1802 dkey->dttk_value = kdata;
1803 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1805 dkey->dttk_value = key[i].dttk_value;
1808 dkey->dttk_size = kesize;
1811 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1812 dvar->dtdv_hashval = hashval;
1813 dvar->dtdv_next = start;
1815 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1819 * The cas has failed. Either another CPU is adding an element to
1820 * this hash chain, or another CPU is deleting an element from this
1821 * hash chain. The simplest way to deal with both of these cases
1822 * (though not necessarily the most efficient) is to free our
1823 * allocated block and tail-call ourselves. Note that the free is
1824 * to the dirty list and _not_ to the free list. This is to prevent
1825 * races with allocators, above.
1827 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1829 dtrace_membar_producer();
1832 free = dcpu->dtdsc_dirty;
1833 dvar->dtdv_next = free;
1834 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1836 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1841 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1843 if ((int64_t)nval < (int64_t)*oval)
1849 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1851 if ((int64_t)nval > (int64_t)*oval)
1856 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1858 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1859 int64_t val = (int64_t)nval;
1862 for (i = 0; i < zero; i++) {
1863 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1869 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1870 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1871 quanta[i - 1] += incr;
1876 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1884 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1886 uint64_t arg = *lquanta++;
1887 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1888 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1889 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1890 int32_t val = (int32_t)nval, level;
1893 ASSERT(levels != 0);
1897 * This is an underflow.
1903 level = (val - base) / step;
1905 if (level < levels) {
1906 lquanta[level + 1] += incr;
1911 * This is an overflow.
1913 lquanta[levels + 1] += incr;
1918 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
1926 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
1928 int64_t snval = (int64_t)nval;
1935 * What we want to say here is:
1937 * data[2] += nval * nval;
1939 * But given that nval is 64-bit, we could easily overflow, so
1940 * we do this as 128-bit arithmetic.
1945 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
1946 dtrace_add_128(data + 2, tmp, data + 2);
1951 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
1958 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
1964 * Aggregate given the tuple in the principal data buffer, and the aggregating
1965 * action denoted by the specified dtrace_aggregation_t. The aggregation
1966 * buffer is specified as the buf parameter. This routine does not return
1967 * failure; if there is no space in the aggregation buffer, the data will be
1968 * dropped, and a corresponding counter incremented.
1971 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
1972 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
1974 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
1975 uint32_t i, ndx, size, fsize;
1976 uint32_t align = sizeof (uint64_t) - 1;
1977 dtrace_aggbuffer_t *agb;
1978 dtrace_aggkey_t *key;
1979 uint32_t hashval = 0, limit, isstr;
1980 caddr_t tomax, data, kdata;
1981 dtrace_actkind_t action;
1982 dtrace_action_t *act;
1988 if (!agg->dtag_hasarg) {
1990 * Currently, only quantize() and lquantize() take additional
1991 * arguments, and they have the same semantics: an increment
1992 * value that defaults to 1 when not present. If additional
1993 * aggregating actions take arguments, the setting of the
1994 * default argument value will presumably have to become more
2000 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2001 size = rec->dtrd_offset - agg->dtag_base;
2002 fsize = size + rec->dtrd_size;
2004 ASSERT(dbuf->dtb_tomax != NULL);
2005 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2007 if ((tomax = buf->dtb_tomax) == NULL) {
2008 dtrace_buffer_drop(buf);
2013 * The metastructure is always at the bottom of the buffer.
2015 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2016 sizeof (dtrace_aggbuffer_t));
2018 if (buf->dtb_offset == 0) {
2020 * We just kludge up approximately 1/8th of the size to be
2021 * buckets. If this guess ends up being routinely
2022 * off-the-mark, we may need to dynamically readjust this
2023 * based on past performance.
2025 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2027 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2028 (uintptr_t)tomax || hashsize == 0) {
2030 * We've been given a ludicrously small buffer;
2031 * increment our drop count and leave.
2033 dtrace_buffer_drop(buf);
2038 * And now, a pathetic attempt to try to get a an odd (or
2039 * perchance, a prime) hash size for better hash distribution.
2041 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2042 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2044 agb->dtagb_hashsize = hashsize;
2045 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2046 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2047 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2049 for (i = 0; i < agb->dtagb_hashsize; i++)
2050 agb->dtagb_hash[i] = NULL;
2053 ASSERT(agg->dtag_first != NULL);
2054 ASSERT(agg->dtag_first->dta_intuple);
2057 * Calculate the hash value based on the key. Note that we _don't_
2058 * include the aggid in the hashing (but we will store it as part of
2059 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2060 * algorithm: a simple, quick algorithm that has no known funnels, and
2061 * gets good distribution in practice. The efficacy of the hashing
2062 * algorithm (and a comparison with other algorithms) may be found by
2063 * running the ::dtrace_aggstat MDB dcmd.
2065 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2066 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2067 limit = i + act->dta_rec.dtrd_size;
2068 ASSERT(limit <= size);
2069 isstr = DTRACEACT_ISSTRING(act);
2071 for (; i < limit; i++) {
2073 hashval += (hashval << 10);
2074 hashval ^= (hashval >> 6);
2076 if (isstr && data[i] == '\0')
2081 hashval += (hashval << 3);
2082 hashval ^= (hashval >> 11);
2083 hashval += (hashval << 15);
2086 * Yes, the divide here is expensive -- but it's generally the least
2087 * of the performance issues given the amount of data that we iterate
2088 * over to compute hash values, compare data, etc.
2090 ndx = hashval % agb->dtagb_hashsize;
2092 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2093 ASSERT((caddr_t)key >= tomax);
2094 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2096 if (hashval != key->dtak_hashval || key->dtak_size != size)
2099 kdata = key->dtak_data;
2100 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2102 for (act = agg->dtag_first; act->dta_intuple;
2103 act = act->dta_next) {
2104 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2105 limit = i + act->dta_rec.dtrd_size;
2106 ASSERT(limit <= size);
2107 isstr = DTRACEACT_ISSTRING(act);
2109 for (; i < limit; i++) {
2110 if (kdata[i] != data[i])
2113 if (isstr && data[i] == '\0')
2118 if (action != key->dtak_action) {
2120 * We are aggregating on the same value in the same
2121 * aggregation with two different aggregating actions.
2122 * (This should have been picked up in the compiler,
2123 * so we may be dealing with errant or devious DIF.)
2124 * This is an error condition; we indicate as much,
2127 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2132 * This is a hit: we need to apply the aggregator to
2133 * the value at this key.
2135 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2142 * We didn't find it. We need to allocate some zero-filled space,
2143 * link it into the hash table appropriately, and apply the aggregator
2144 * to the (zero-filled) value.
2146 offs = buf->dtb_offset;
2147 while (offs & (align - 1))
2148 offs += sizeof (uint32_t);
2151 * If we don't have enough room to both allocate a new key _and_
2152 * its associated data, increment the drop count and return.
2154 if ((uintptr_t)tomax + offs + fsize >
2155 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2156 dtrace_buffer_drop(buf);
2161 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2162 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2163 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2165 key->dtak_data = kdata = tomax + offs;
2166 buf->dtb_offset = offs + fsize;
2169 * Now copy the data across.
2171 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2173 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2177 * Because strings are not zeroed out by default, we need to iterate
2178 * looking for actions that store strings, and we need to explicitly
2179 * pad these strings out with zeroes.
2181 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2184 if (!DTRACEACT_ISSTRING(act))
2187 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2188 limit = i + act->dta_rec.dtrd_size;
2189 ASSERT(limit <= size);
2191 for (nul = 0; i < limit; i++) {
2197 if (data[i] != '\0')
2204 for (i = size; i < fsize; i++)
2207 key->dtak_hashval = hashval;
2208 key->dtak_size = size;
2209 key->dtak_action = action;
2210 key->dtak_next = agb->dtagb_hash[ndx];
2211 agb->dtagb_hash[ndx] = key;
2214 * Finally, apply the aggregator.
2216 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2217 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2221 * Given consumer state, this routine finds a speculation in the INACTIVE
2222 * state and transitions it into the ACTIVE state. If there is no speculation
2223 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2224 * incremented -- it is up to the caller to take appropriate action.
2227 dtrace_speculation(dtrace_state_t *state)
2230 dtrace_speculation_state_t current;
2231 uint32_t *stat = &state->dts_speculations_unavail, count;
2233 while (i < state->dts_nspeculations) {
2234 dtrace_speculation_t *spec = &state->dts_speculations[i];
2236 current = spec->dtsp_state;
2238 if (current != DTRACESPEC_INACTIVE) {
2239 if (current == DTRACESPEC_COMMITTINGMANY ||
2240 current == DTRACESPEC_COMMITTING ||
2241 current == DTRACESPEC_DISCARDING)
2242 stat = &state->dts_speculations_busy;
2247 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2248 current, DTRACESPEC_ACTIVE) == current)
2253 * We couldn't find a speculation. If we found as much as a single
2254 * busy speculation buffer, we'll attribute this failure as "busy"
2255 * instead of "unavail".
2259 } while (dtrace_cas32(stat, count, count + 1) != count);
2265 * This routine commits an active speculation. If the specified speculation
2266 * is not in a valid state to perform a commit(), this routine will silently do
2267 * nothing. The state of the specified speculation is transitioned according
2268 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2271 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2272 dtrace_specid_t which)
2274 dtrace_speculation_t *spec;
2275 dtrace_buffer_t *src, *dest;
2276 uintptr_t daddr, saddr, dlimit;
2277 dtrace_speculation_state_t current, new = 0;
2283 if (which > state->dts_nspeculations) {
2284 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2288 spec = &state->dts_speculations[which - 1];
2289 src = &spec->dtsp_buffer[cpu];
2290 dest = &state->dts_buffer[cpu];
2293 current = spec->dtsp_state;
2295 if (current == DTRACESPEC_COMMITTINGMANY)
2299 case DTRACESPEC_INACTIVE:
2300 case DTRACESPEC_DISCARDING:
2303 case DTRACESPEC_COMMITTING:
2305 * This is only possible if we are (a) commit()'ing
2306 * without having done a prior speculate() on this CPU
2307 * and (b) racing with another commit() on a different
2308 * CPU. There's nothing to do -- we just assert that
2311 ASSERT(src->dtb_offset == 0);
2314 case DTRACESPEC_ACTIVE:
2315 new = DTRACESPEC_COMMITTING;
2318 case DTRACESPEC_ACTIVEONE:
2320 * This speculation is active on one CPU. If our
2321 * buffer offset is non-zero, we know that the one CPU
2322 * must be us. Otherwise, we are committing on a
2323 * different CPU from the speculate(), and we must
2324 * rely on being asynchronously cleaned.
2326 if (src->dtb_offset != 0) {
2327 new = DTRACESPEC_COMMITTING;
2332 case DTRACESPEC_ACTIVEMANY:
2333 new = DTRACESPEC_COMMITTINGMANY;
2339 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2340 current, new) != current);
2343 * We have set the state to indicate that we are committing this
2344 * speculation. Now reserve the necessary space in the destination
2347 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2348 sizeof (uint64_t), state, NULL)) < 0) {
2349 dtrace_buffer_drop(dest);
2354 * We have the space; copy the buffer across. (Note that this is a
2355 * highly subobtimal bcopy(); in the unlikely event that this becomes
2356 * a serious performance issue, a high-performance DTrace-specific
2357 * bcopy() should obviously be invented.)
2359 daddr = (uintptr_t)dest->dtb_tomax + offs;
2360 dlimit = daddr + src->dtb_offset;
2361 saddr = (uintptr_t)src->dtb_tomax;
2364 * First, the aligned portion.
2366 while (dlimit - daddr >= sizeof (uint64_t)) {
2367 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2369 daddr += sizeof (uint64_t);
2370 saddr += sizeof (uint64_t);
2374 * Now any left-over bit...
2376 while (dlimit - daddr)
2377 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2380 * Finally, commit the reserved space in the destination buffer.
2382 dest->dtb_offset = offs + src->dtb_offset;
2386 * If we're lucky enough to be the only active CPU on this speculation
2387 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2389 if (current == DTRACESPEC_ACTIVE ||
2390 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2391 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2392 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2394 ASSERT(rval == DTRACESPEC_COMMITTING);
2397 src->dtb_offset = 0;
2398 src->dtb_xamot_drops += src->dtb_drops;
2403 * This routine discards an active speculation. If the specified speculation
2404 * is not in a valid state to perform a discard(), this routine will silently
2405 * do nothing. The state of the specified speculation is transitioned
2406 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2409 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2410 dtrace_specid_t which)
2412 dtrace_speculation_t *spec;
2413 dtrace_speculation_state_t current, new = 0;
2414 dtrace_buffer_t *buf;
2419 if (which > state->dts_nspeculations) {
2420 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2424 spec = &state->dts_speculations[which - 1];
2425 buf = &spec->dtsp_buffer[cpu];
2428 current = spec->dtsp_state;
2431 case DTRACESPEC_INACTIVE:
2432 case DTRACESPEC_COMMITTINGMANY:
2433 case DTRACESPEC_COMMITTING:
2434 case DTRACESPEC_DISCARDING:
2437 case DTRACESPEC_ACTIVE:
2438 case DTRACESPEC_ACTIVEMANY:
2439 new = DTRACESPEC_DISCARDING;
2442 case DTRACESPEC_ACTIVEONE:
2443 if (buf->dtb_offset != 0) {
2444 new = DTRACESPEC_INACTIVE;
2446 new = DTRACESPEC_DISCARDING;
2453 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2454 current, new) != current);
2456 buf->dtb_offset = 0;
2461 * Note: not called from probe context. This function is called
2462 * asynchronously from cross call context to clean any speculations that are
2463 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2464 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2468 dtrace_speculation_clean_here(dtrace_state_t *state)
2470 dtrace_icookie_t cookie;
2471 processorid_t cpu = curcpu;
2472 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2475 cookie = dtrace_interrupt_disable();
2477 if (dest->dtb_tomax == NULL) {
2478 dtrace_interrupt_enable(cookie);
2482 for (i = 0; i < state->dts_nspeculations; i++) {
2483 dtrace_speculation_t *spec = &state->dts_speculations[i];
2484 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2486 if (src->dtb_tomax == NULL)
2489 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2490 src->dtb_offset = 0;
2494 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2497 if (src->dtb_offset == 0)
2500 dtrace_speculation_commit(state, cpu, i + 1);
2503 dtrace_interrupt_enable(cookie);
2507 * Note: not called from probe context. This function is called
2508 * asynchronously (and at a regular interval) to clean any speculations that
2509 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2510 * is work to be done, it cross calls all CPUs to perform that work;
2511 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2512 * INACTIVE state until they have been cleaned by all CPUs.
2515 dtrace_speculation_clean(dtrace_state_t *state)
2520 for (i = 0; i < state->dts_nspeculations; i++) {
2521 dtrace_speculation_t *spec = &state->dts_speculations[i];
2523 ASSERT(!spec->dtsp_cleaning);
2525 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2526 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2530 spec->dtsp_cleaning = 1;
2536 dtrace_xcall(DTRACE_CPUALL,
2537 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2540 * We now know that all CPUs have committed or discarded their
2541 * speculation buffers, as appropriate. We can now set the state
2544 for (i = 0; i < state->dts_nspeculations; i++) {
2545 dtrace_speculation_t *spec = &state->dts_speculations[i];
2546 dtrace_speculation_state_t current, new;
2548 if (!spec->dtsp_cleaning)
2551 current = spec->dtsp_state;
2552 ASSERT(current == DTRACESPEC_DISCARDING ||
2553 current == DTRACESPEC_COMMITTINGMANY);
2555 new = DTRACESPEC_INACTIVE;
2557 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2558 ASSERT(rv == current);
2559 spec->dtsp_cleaning = 0;
2564 * Called as part of a speculate() to get the speculative buffer associated
2565 * with a given speculation. Returns NULL if the specified speculation is not
2566 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2567 * the active CPU is not the specified CPU -- the speculation will be
2568 * atomically transitioned into the ACTIVEMANY state.
2570 static dtrace_buffer_t *
2571 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2572 dtrace_specid_t which)
2574 dtrace_speculation_t *spec;
2575 dtrace_speculation_state_t current, new = 0;
2576 dtrace_buffer_t *buf;
2581 if (which > state->dts_nspeculations) {
2582 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2586 spec = &state->dts_speculations[which - 1];
2587 buf = &spec->dtsp_buffer[cpuid];
2590 current = spec->dtsp_state;
2593 case DTRACESPEC_INACTIVE:
2594 case DTRACESPEC_COMMITTINGMANY:
2595 case DTRACESPEC_DISCARDING:
2598 case DTRACESPEC_COMMITTING:
2599 ASSERT(buf->dtb_offset == 0);
2602 case DTRACESPEC_ACTIVEONE:
2604 * This speculation is currently active on one CPU.
2605 * Check the offset in the buffer; if it's non-zero,
2606 * that CPU must be us (and we leave the state alone).
2607 * If it's zero, assume that we're starting on a new
2608 * CPU -- and change the state to indicate that the
2609 * speculation is active on more than one CPU.
2611 if (buf->dtb_offset != 0)
2614 new = DTRACESPEC_ACTIVEMANY;
2617 case DTRACESPEC_ACTIVEMANY:
2620 case DTRACESPEC_ACTIVE:
2621 new = DTRACESPEC_ACTIVEONE;
2627 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2628 current, new) != current);
2630 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2635 * Return a string. In the event that the user lacks the privilege to access
2636 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2637 * don't fail access checking.
2639 * dtrace_dif_variable() uses this routine as a helper for various
2640 * builtin values such as 'execname' and 'probefunc.'
2643 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2644 dtrace_mstate_t *mstate)
2646 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2651 * The easy case: this probe is allowed to read all of memory, so
2652 * we can just return this as a vanilla pointer.
2654 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2658 * This is the tougher case: we copy the string in question from
2659 * kernel memory into scratch memory and return it that way: this
2660 * ensures that we won't trip up when access checking tests the
2661 * BYREF return value.
2663 strsz = dtrace_strlen((char *)addr, size) + 1;
2665 if (mstate->dtms_scratch_ptr + strsz >
2666 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2667 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2671 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2673 ret = mstate->dtms_scratch_ptr;
2674 mstate->dtms_scratch_ptr += strsz;
2679 * Return a string from a memoy address which is known to have one or
2680 * more concatenated, individually zero terminated, sub-strings.
2681 * In the event that the user lacks the privilege to access
2682 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2683 * don't fail access checking.
2685 * dtrace_dif_variable() uses this routine as a helper for various
2686 * builtin values such as 'execargs'.
2689 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2690 dtrace_mstate_t *mstate)
2696 if (mstate->dtms_scratch_ptr + strsz >
2697 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2698 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2702 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2705 /* Replace sub-string termination characters with a space. */
2706 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2711 ret = mstate->dtms_scratch_ptr;
2712 mstate->dtms_scratch_ptr += strsz;
2717 * This function implements the DIF emulator's variable lookups. The emulator
2718 * passes a reserved variable identifier and optional built-in array index.
2721 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2725 * If we're accessing one of the uncached arguments, we'll turn this
2726 * into a reference in the args array.
2728 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2729 ndx = v - DIF_VAR_ARG0;
2735 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2736 if (ndx >= sizeof (mstate->dtms_arg) /
2737 sizeof (mstate->dtms_arg[0])) {
2738 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2739 dtrace_provider_t *pv;
2742 pv = mstate->dtms_probe->dtpr_provider;
2743 if (pv->dtpv_pops.dtps_getargval != NULL)
2744 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2745 mstate->dtms_probe->dtpr_id,
2746 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2748 val = dtrace_getarg(ndx, aframes);
2751 * This is regrettably required to keep the compiler
2752 * from tail-optimizing the call to dtrace_getarg().
2753 * The condition always evaluates to true, but the
2754 * compiler has no way of figuring that out a priori.
2755 * (None of this would be necessary if the compiler
2756 * could be relied upon to _always_ tail-optimize
2757 * the call to dtrace_getarg() -- but it can't.)
2759 if (mstate->dtms_probe != NULL)
2765 return (mstate->dtms_arg[ndx]);
2768 case DIF_VAR_UREGS: {
2771 if (!dtrace_priv_proc(state))
2774 if ((lwp = curthread->t_lwp) == NULL) {
2775 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2776 cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2780 return (dtrace_getreg(lwp->lwp_regs, ndx));
2784 case DIF_VAR_UREGS: {
2785 struct trapframe *tframe;
2787 if (!dtrace_priv_proc(state))
2790 if ((tframe = curthread->td_frame) == NULL) {
2791 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2792 cpu_core[curcpu].cpuc_dtrace_illval = 0;
2796 return (dtrace_getreg(tframe, ndx));
2800 case DIF_VAR_CURTHREAD:
2801 if (!dtrace_priv_kernel(state))
2803 return ((uint64_t)(uintptr_t)curthread);
2805 case DIF_VAR_TIMESTAMP:
2806 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2807 mstate->dtms_timestamp = dtrace_gethrtime();
2808 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2810 return (mstate->dtms_timestamp);
2812 case DIF_VAR_VTIMESTAMP:
2813 ASSERT(dtrace_vtime_references != 0);
2814 return (curthread->t_dtrace_vtime);
2816 case DIF_VAR_WALLTIMESTAMP:
2817 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2818 mstate->dtms_walltimestamp = dtrace_gethrestime();
2819 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2821 return (mstate->dtms_walltimestamp);
2825 if (!dtrace_priv_kernel(state))
2827 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2828 mstate->dtms_ipl = dtrace_getipl();
2829 mstate->dtms_present |= DTRACE_MSTATE_IPL;
2831 return (mstate->dtms_ipl);
2835 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2836 return (mstate->dtms_epid);
2839 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2840 return (mstate->dtms_probe->dtpr_id);
2842 case DIF_VAR_STACKDEPTH:
2843 if (!dtrace_priv_kernel(state))
2845 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2846 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2848 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2849 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2851 return (mstate->dtms_stackdepth);
2853 case DIF_VAR_USTACKDEPTH:
2854 if (!dtrace_priv_proc(state))
2856 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2858 * See comment in DIF_VAR_PID.
2860 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2862 mstate->dtms_ustackdepth = 0;
2864 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2865 mstate->dtms_ustackdepth =
2866 dtrace_getustackdepth();
2867 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2869 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2871 return (mstate->dtms_ustackdepth);
2873 case DIF_VAR_CALLER:
2874 if (!dtrace_priv_kernel(state))
2876 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2877 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2879 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2881 * If this is an unanchored probe, we are
2882 * required to go through the slow path:
2883 * dtrace_caller() only guarantees correct
2884 * results for anchored probes.
2886 pc_t caller[2] = {0, 0};
2888 dtrace_getpcstack(caller, 2, aframes,
2889 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2890 mstate->dtms_caller = caller[1];
2891 } else if ((mstate->dtms_caller =
2892 dtrace_caller(aframes)) == -1) {
2894 * We have failed to do this the quick way;
2895 * we must resort to the slower approach of
2896 * calling dtrace_getpcstack().
2900 dtrace_getpcstack(&caller, 1, aframes, NULL);
2901 mstate->dtms_caller = caller;
2904 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
2906 return (mstate->dtms_caller);
2908 case DIF_VAR_UCALLER:
2909 if (!dtrace_priv_proc(state))
2912 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
2916 * dtrace_getupcstack() fills in the first uint64_t
2917 * with the current PID. The second uint64_t will
2918 * be the program counter at user-level. The third
2919 * uint64_t will contain the caller, which is what
2923 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2924 dtrace_getupcstack(ustack, 3);
2925 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2926 mstate->dtms_ucaller = ustack[2];
2927 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
2930 return (mstate->dtms_ucaller);
2932 case DIF_VAR_PROBEPROV:
2933 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2934 return (dtrace_dif_varstr(
2935 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
2938 case DIF_VAR_PROBEMOD:
2939 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2940 return (dtrace_dif_varstr(
2941 (uintptr_t)mstate->dtms_probe->dtpr_mod,
2944 case DIF_VAR_PROBEFUNC:
2945 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2946 return (dtrace_dif_varstr(
2947 (uintptr_t)mstate->dtms_probe->dtpr_func,
2950 case DIF_VAR_PROBENAME:
2951 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2952 return (dtrace_dif_varstr(
2953 (uintptr_t)mstate->dtms_probe->dtpr_name,
2957 if (!dtrace_priv_proc(state))
2962 * Note that we are assuming that an unanchored probe is
2963 * always due to a high-level interrupt. (And we're assuming
2964 * that there is only a single high level interrupt.)
2966 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2967 return (pid0.pid_id);
2970 * It is always safe to dereference one's own t_procp pointer:
2971 * it always points to a valid, allocated proc structure.
2972 * Further, it is always safe to dereference the p_pidp member
2973 * of one's own proc structure. (These are truisms becuase
2974 * threads and processes don't clean up their own state --
2975 * they leave that task to whomever reaps them.)
2977 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
2979 return ((uint64_t)curproc->p_pid);
2983 if (!dtrace_priv_proc(state))
2988 * See comment in DIF_VAR_PID.
2990 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2991 return (pid0.pid_id);
2994 * It is always safe to dereference one's own t_procp pointer:
2995 * it always points to a valid, allocated proc structure.
2996 * (This is true because threads don't clean up their own
2997 * state -- they leave that task to whomever reaps them.)
2999 return ((uint64_t)curthread->t_procp->p_ppid);
3001 return ((uint64_t)curproc->p_pptr->p_pid);
3007 * See comment in DIF_VAR_PID.
3009 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3013 return ((uint64_t)curthread->t_tid);
3015 case DIF_VAR_EXECARGS: {
3016 struct pargs *p_args = curthread->td_proc->p_args;
3021 return (dtrace_dif_varstrz(
3022 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3025 case DIF_VAR_EXECNAME:
3027 if (!dtrace_priv_proc(state))
3031 * See comment in DIF_VAR_PID.
3033 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3034 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3037 * It is always safe to dereference one's own t_procp pointer:
3038 * it always points to a valid, allocated proc structure.
3039 * (This is true because threads don't clean up their own
3040 * state -- they leave that task to whomever reaps them.)
3042 return (dtrace_dif_varstr(
3043 (uintptr_t)curthread->t_procp->p_user.u_comm,
3046 return (dtrace_dif_varstr(
3047 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3050 case DIF_VAR_ZONENAME:
3052 if (!dtrace_priv_proc(state))
3056 * See comment in DIF_VAR_PID.
3058 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3059 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3062 * It is always safe to dereference one's own t_procp pointer:
3063 * it always points to a valid, allocated proc structure.
3064 * (This is true because threads don't clean up their own
3065 * state -- they leave that task to whomever reaps them.)
3067 return (dtrace_dif_varstr(
3068 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3075 if (!dtrace_priv_proc(state))
3080 * See comment in DIF_VAR_PID.
3082 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3083 return ((uint64_t)p0.p_cred->cr_uid);
3087 * It is always safe to dereference one's own t_procp pointer:
3088 * it always points to a valid, allocated proc structure.
3089 * (This is true because threads don't clean up their own
3090 * state -- they leave that task to whomever reaps them.)
3092 * Additionally, it is safe to dereference one's own process
3093 * credential, since this is never NULL after process birth.
3095 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3098 if (!dtrace_priv_proc(state))
3103 * See comment in DIF_VAR_PID.
3105 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3106 return ((uint64_t)p0.p_cred->cr_gid);
3110 * It is always safe to dereference one's own t_procp pointer:
3111 * it always points to a valid, allocated proc structure.
3112 * (This is true because threads don't clean up their own
3113 * state -- they leave that task to whomever reaps them.)
3115 * Additionally, it is safe to dereference one's own process
3116 * credential, since this is never NULL after process birth.
3118 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3120 case DIF_VAR_ERRNO: {
3123 if (!dtrace_priv_proc(state))
3127 * See comment in DIF_VAR_PID.
3129 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3133 * It is always safe to dereference one's own t_lwp pointer in
3134 * the event that this pointer is non-NULL. (This is true
3135 * because threads and lwps don't clean up their own state --
3136 * they leave that task to whomever reaps them.)
3138 if ((lwp = curthread->t_lwp) == NULL)
3141 return ((uint64_t)lwp->lwp_errno);
3143 return (curthread->td_errno);
3152 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3158 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3159 * Notice that we don't bother validating the proper number of arguments or
3160 * their types in the tuple stack. This isn't needed because all argument
3161 * interpretation is safe because of our load safety -- the worst that can
3162 * happen is that a bogus program can obtain bogus results.
3165 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3166 dtrace_key_t *tupregs, int nargs,
3167 dtrace_mstate_t *mstate, dtrace_state_t *state)
3169 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3170 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3171 dtrace_vstate_t *vstate = &state->dts_vstate;
3184 struct thread *lowner;
3186 struct lock_object *li;
3193 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3197 case DIF_SUBR_MUTEX_OWNED:
3198 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3204 m.mx = dtrace_load64(tupregs[0].dttk_value);
3205 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3206 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3208 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3211 case DIF_SUBR_MUTEX_OWNER:
3212 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3218 m.mx = dtrace_load64(tupregs[0].dttk_value);
3219 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3220 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3221 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3226 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3227 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3233 m.mx = dtrace_load64(tupregs[0].dttk_value);
3234 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3237 case DIF_SUBR_MUTEX_TYPE_SPIN:
3238 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3244 m.mx = dtrace_load64(tupregs[0].dttk_value);
3245 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3248 case DIF_SUBR_RW_READ_HELD: {
3251 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3257 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3258 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3262 case DIF_SUBR_RW_WRITE_HELD:
3263 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3269 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3270 regs[rd] = _RW_WRITE_HELD(&r.ri);
3273 case DIF_SUBR_RW_ISWRITER:
3274 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3280 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3281 regs[rd] = _RW_ISWRITER(&r.ri);
3285 case DIF_SUBR_MUTEX_OWNED:
3286 if (!dtrace_canload(tupregs[0].dttk_value,
3287 sizeof (struct lock_object), mstate, vstate)) {
3291 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3292 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3295 case DIF_SUBR_MUTEX_OWNER:
3296 if (!dtrace_canload(tupregs[0].dttk_value,
3297 sizeof (struct lock_object), mstate, vstate)) {
3301 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3302 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3303 regs[rd] = (uintptr_t)lowner;
3306 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3307 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3312 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3313 /* XXX - should be only LC_SLEEPABLE? */
3314 regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3315 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3318 case DIF_SUBR_MUTEX_TYPE_SPIN:
3319 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3324 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3325 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3328 case DIF_SUBR_RW_READ_HELD:
3329 case DIF_SUBR_SX_SHARED_HELD:
3330 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3335 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3336 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3340 case DIF_SUBR_RW_WRITE_HELD:
3341 case DIF_SUBR_SX_EXCLUSIVE_HELD:
3342 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3347 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3348 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3349 regs[rd] = (lowner == curthread);
3352 case DIF_SUBR_RW_ISWRITER:
3353 case DIF_SUBR_SX_ISEXCLUSIVE:
3354 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3359 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3360 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3363 #endif /* ! defined(sun) */
3365 case DIF_SUBR_BCOPY: {
3367 * We need to be sure that the destination is in the scratch
3368 * region -- no other region is allowed.
3370 uintptr_t src = tupregs[0].dttk_value;
3371 uintptr_t dest = tupregs[1].dttk_value;
3372 size_t size = tupregs[2].dttk_value;
3374 if (!dtrace_inscratch(dest, size, mstate)) {
3375 *flags |= CPU_DTRACE_BADADDR;
3380 if (!dtrace_canload(src, size, mstate, vstate)) {
3385 dtrace_bcopy((void *)src, (void *)dest, size);
3389 case DIF_SUBR_ALLOCA:
3390 case DIF_SUBR_COPYIN: {
3391 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3393 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3394 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3397 * This action doesn't require any credential checks since
3398 * probes will not activate in user contexts to which the
3399 * enabling user does not have permissions.
3403 * Rounding up the user allocation size could have overflowed
3404 * a large, bogus allocation (like -1ULL) to 0.
3406 if (scratch_size < size ||
3407 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3408 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3413 if (subr == DIF_SUBR_COPYIN) {
3414 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3415 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3416 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3419 mstate->dtms_scratch_ptr += scratch_size;
3424 case DIF_SUBR_COPYINTO: {
3425 uint64_t size = tupregs[1].dttk_value;
3426 uintptr_t dest = tupregs[2].dttk_value;
3429 * This action doesn't require any credential checks since
3430 * probes will not activate in user contexts to which the
3431 * enabling user does not have permissions.
3433 if (!dtrace_inscratch(dest, size, mstate)) {
3434 *flags |= CPU_DTRACE_BADADDR;
3439 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3440 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3441 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3445 case DIF_SUBR_COPYINSTR: {
3446 uintptr_t dest = mstate->dtms_scratch_ptr;
3447 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3449 if (nargs > 1 && tupregs[1].dttk_value < size)
3450 size = tupregs[1].dttk_value + 1;
3453 * This action doesn't require any credential checks since
3454 * probes will not activate in user contexts to which the
3455 * enabling user does not have permissions.
3457 if (!DTRACE_INSCRATCH(mstate, size)) {
3458 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3463 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3464 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3465 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3467 ((char *)dest)[size - 1] = '\0';
3468 mstate->dtms_scratch_ptr += size;
3474 case DIF_SUBR_MSGSIZE:
3475 case DIF_SUBR_MSGDSIZE: {
3476 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3477 uintptr_t wptr, rptr;
3481 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3483 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3489 wptr = dtrace_loadptr(baddr +
3490 offsetof(mblk_t, b_wptr));
3492 rptr = dtrace_loadptr(baddr +
3493 offsetof(mblk_t, b_rptr));
3496 *flags |= CPU_DTRACE_BADADDR;
3497 *illval = tupregs[0].dttk_value;
3501 daddr = dtrace_loadptr(baddr +
3502 offsetof(mblk_t, b_datap));
3504 baddr = dtrace_loadptr(baddr +
3505 offsetof(mblk_t, b_cont));
3508 * We want to prevent against denial-of-service here,
3509 * so we're only going to search the list for
3510 * dtrace_msgdsize_max mblks.
3512 if (cont++ > dtrace_msgdsize_max) {
3513 *flags |= CPU_DTRACE_ILLOP;
3517 if (subr == DIF_SUBR_MSGDSIZE) {
3518 if (dtrace_load8(daddr +
3519 offsetof(dblk_t, db_type)) != M_DATA)
3523 count += wptr - rptr;
3526 if (!(*flags & CPU_DTRACE_FAULT))
3533 case DIF_SUBR_PROGENYOF: {
3534 pid_t pid = tupregs[0].dttk_value;
3538 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3540 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3542 if (p->p_pidp->pid_id == pid) {
3544 if (p->p_pid == pid) {
3551 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3557 case DIF_SUBR_SPECULATION:
3558 regs[rd] = dtrace_speculation(state);
3561 case DIF_SUBR_COPYOUT: {
3562 uintptr_t kaddr = tupregs[0].dttk_value;
3563 uintptr_t uaddr = tupregs[1].dttk_value;
3564 uint64_t size = tupregs[2].dttk_value;
3566 if (!dtrace_destructive_disallow &&
3567 dtrace_priv_proc_control(state) &&
3568 !dtrace_istoxic(kaddr, size)) {
3569 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3570 dtrace_copyout(kaddr, uaddr, size, flags);
3571 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3576 case DIF_SUBR_COPYOUTSTR: {
3577 uintptr_t kaddr = tupregs[0].dttk_value;
3578 uintptr_t uaddr = tupregs[1].dttk_value;
3579 uint64_t size = tupregs[2].dttk_value;
3581 if (!dtrace_destructive_disallow &&
3582 dtrace_priv_proc_control(state) &&
3583 !dtrace_istoxic(kaddr, size)) {
3584 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3585 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3586 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3591 case DIF_SUBR_STRLEN: {
3593 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3594 sz = dtrace_strlen((char *)addr,
3595 state->dts_options[DTRACEOPT_STRSIZE]);
3597 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3607 case DIF_SUBR_STRCHR:
3608 case DIF_SUBR_STRRCHR: {
3610 * We're going to iterate over the string looking for the
3611 * specified character. We will iterate until we have reached
3612 * the string length or we have found the character. If this
3613 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3614 * of the specified character instead of the first.
3616 uintptr_t saddr = tupregs[0].dttk_value;
3617 uintptr_t addr = tupregs[0].dttk_value;
3618 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3619 char c, target = (char)tupregs[1].dttk_value;
3621 for (regs[rd] = 0; addr < limit; addr++) {
3622 if ((c = dtrace_load8(addr)) == target) {
3625 if (subr == DIF_SUBR_STRCHR)
3633 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3641 case DIF_SUBR_STRSTR:
3642 case DIF_SUBR_INDEX:
3643 case DIF_SUBR_RINDEX: {
3645 * We're going to iterate over the string looking for the
3646 * specified string. We will iterate until we have reached
3647 * the string length or we have found the string. (Yes, this
3648 * is done in the most naive way possible -- but considering
3649 * that the string we're searching for is likely to be
3650 * relatively short, the complexity of Rabin-Karp or similar
3651 * hardly seems merited.)
3653 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3654 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3655 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3656 size_t len = dtrace_strlen(addr, size);
3657 size_t sublen = dtrace_strlen(substr, size);
3658 char *limit = addr + len, *orig = addr;
3659 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3662 regs[rd] = notfound;
3664 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3669 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3676 * strstr() and index()/rindex() have similar semantics if
3677 * both strings are the empty string: strstr() returns a
3678 * pointer to the (empty) string, and index() and rindex()
3679 * both return index 0 (regardless of any position argument).
3681 if (sublen == 0 && len == 0) {
3682 if (subr == DIF_SUBR_STRSTR)
3683 regs[rd] = (uintptr_t)addr;
3689 if (subr != DIF_SUBR_STRSTR) {
3690 if (subr == DIF_SUBR_RINDEX) {
3697 * Both index() and rindex() take an optional position
3698 * argument that denotes the starting position.
3701 int64_t pos = (int64_t)tupregs[2].dttk_value;
3704 * If the position argument to index() is
3705 * negative, Perl implicitly clamps it at
3706 * zero. This semantic is a little surprising
3707 * given the special meaning of negative
3708 * positions to similar Perl functions like
3709 * substr(), but it appears to reflect a
3710 * notion that index() can start from a
3711 * negative index and increment its way up to
3712 * the string. Given this notion, Perl's
3713 * rindex() is at least self-consistent in
3714 * that it implicitly clamps positions greater
3715 * than the string length to be the string
3716 * length. Where Perl completely loses
3717 * coherence, however, is when the specified
3718 * substring is the empty string (""). In
3719 * this case, even if the position is
3720 * negative, rindex() returns 0 -- and even if
3721 * the position is greater than the length,
3722 * index() returns the string length. These
3723 * semantics violate the notion that index()
3724 * should never return a value less than the
3725 * specified position and that rindex() should
3726 * never return a value greater than the
3727 * specified position. (One assumes that
3728 * these semantics are artifacts of Perl's
3729 * implementation and not the results of
3730 * deliberate design -- it beggars belief that
3731 * even Larry Wall could desire such oddness.)
3732 * While in the abstract one would wish for
3733 * consistent position semantics across
3734 * substr(), index() and rindex() -- or at the
3735 * very least self-consistent position
3736 * semantics for index() and rindex() -- we
3737 * instead opt to keep with the extant Perl
3738 * semantics, in all their broken glory. (Do
3739 * we have more desire to maintain Perl's
3740 * semantics than Perl does? Probably.)
3742 if (subr == DIF_SUBR_RINDEX) {
3766 for (regs[rd] = notfound; addr != limit; addr += inc) {
3767 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3768 if (subr != DIF_SUBR_STRSTR) {
3770 * As D index() and rindex() are
3771 * modeled on Perl (and not on awk),
3772 * we return a zero-based (and not a
3773 * one-based) index. (For you Perl
3774 * weenies: no, we're not going to add
3775 * $[ -- and shouldn't you be at a con
3778 regs[rd] = (uintptr_t)(addr - orig);
3782 ASSERT(subr == DIF_SUBR_STRSTR);
3783 regs[rd] = (uintptr_t)addr;
3791 case DIF_SUBR_STRTOK: {
3792 uintptr_t addr = tupregs[0].dttk_value;
3793 uintptr_t tokaddr = tupregs[1].dttk_value;
3794 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3795 uintptr_t limit, toklimit = tokaddr + size;
3796 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
3797 char *dest = (char *)mstate->dtms_scratch_ptr;
3801 * Check both the token buffer and (later) the input buffer,
3802 * since both could be non-scratch addresses.
3804 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3809 if (!DTRACE_INSCRATCH(mstate, size)) {
3810 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3817 * If the address specified is NULL, we use our saved
3818 * strtok pointer from the mstate. Note that this
3819 * means that the saved strtok pointer is _only_
3820 * valid within multiple enablings of the same probe --
3821 * it behaves like an implicit clause-local variable.
3823 addr = mstate->dtms_strtok;
3826 * If the user-specified address is non-NULL we must
3827 * access check it. This is the only time we have
3828 * a chance to do so, since this address may reside
3829 * in the string table of this clause-- future calls
3830 * (when we fetch addr from mstate->dtms_strtok)
3831 * would fail this access check.
3833 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3840 * First, zero the token map, and then process the token
3841 * string -- setting a bit in the map for every character
3842 * found in the token string.
3844 for (i = 0; i < sizeof (tokmap); i++)
3847 for (; tokaddr < toklimit; tokaddr++) {
3848 if ((c = dtrace_load8(tokaddr)) == '\0')
3851 ASSERT((c >> 3) < sizeof (tokmap));
3852 tokmap[c >> 3] |= (1 << (c & 0x7));
3855 for (limit = addr + size; addr < limit; addr++) {
3857 * We're looking for a character that is _not_ contained
3858 * in the token string.
3860 if ((c = dtrace_load8(addr)) == '\0')
3863 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3869 * We reached the end of the string without finding
3870 * any character that was not in the token string.
3871 * We return NULL in this case, and we set the saved
3872 * address to NULL as well.
3875 mstate->dtms_strtok = 0;
3880 * From here on, we're copying into the destination string.
3882 for (i = 0; addr < limit && i < size - 1; addr++) {
3883 if ((c = dtrace_load8(addr)) == '\0')
3886 if (tokmap[c >> 3] & (1 << (c & 0x7)))
3895 regs[rd] = (uintptr_t)dest;
3896 mstate->dtms_scratch_ptr += size;
3897 mstate->dtms_strtok = addr;
3901 case DIF_SUBR_SUBSTR: {
3902 uintptr_t s = tupregs[0].dttk_value;
3903 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3904 char *d = (char *)mstate->dtms_scratch_ptr;
3905 int64_t index = (int64_t)tupregs[1].dttk_value;
3906 int64_t remaining = (int64_t)tupregs[2].dttk_value;
3907 size_t len = dtrace_strlen((char *)s, size);
3910 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3915 if (!DTRACE_INSCRATCH(mstate, size)) {
3916 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3922 remaining = (int64_t)size;
3927 if (index < 0 && index + remaining > 0) {
3933 if (index >= len || index < 0) {
3935 } else if (remaining < 0) {
3936 remaining += len - index;
3937 } else if (index + remaining > size) {
3938 remaining = size - index;
3941 for (i = 0; i < remaining; i++) {
3942 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
3948 mstate->dtms_scratch_ptr += size;
3949 regs[rd] = (uintptr_t)d;
3954 case DIF_SUBR_GETMAJOR:
3956 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
3958 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
3962 case DIF_SUBR_GETMINOR:
3964 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
3966 regs[rd] = tupregs[0].dttk_value & MAXMIN;
3970 case DIF_SUBR_DDI_PATHNAME: {
3972 * This one is a galactic mess. We are going to roughly
3973 * emulate ddi_pathname(), but it's made more complicated
3974 * by the fact that we (a) want to include the minor name and
3975 * (b) must proceed iteratively instead of recursively.
3977 uintptr_t dest = mstate->dtms_scratch_ptr;
3978 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3979 char *start = (char *)dest, *end = start + size - 1;
3980 uintptr_t daddr = tupregs[0].dttk_value;
3981 int64_t minor = (int64_t)tupregs[1].dttk_value;
3983 int i, len, depth = 0;
3986 * Due to all the pointer jumping we do and context we must
3987 * rely upon, we just mandate that the user must have kernel
3988 * read privileges to use this routine.
3990 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
3991 *flags |= CPU_DTRACE_KPRIV;
3996 if (!DTRACE_INSCRATCH(mstate, size)) {
3997 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4005 * We want to have a name for the minor. In order to do this,
4006 * we need to walk the minor list from the devinfo. We want
4007 * to be sure that we don't infinitely walk a circular list,
4008 * so we check for circularity by sending a scout pointer
4009 * ahead two elements for every element that we iterate over;
4010 * if the list is circular, these will ultimately point to the
4011 * same element. You may recognize this little trick as the
4012 * answer to a stupid interview question -- one that always
4013 * seems to be asked by those who had to have it laboriously
4014 * explained to them, and who can't even concisely describe
4015 * the conditions under which one would be forced to resort to
4016 * this technique. Needless to say, those conditions are
4017 * found here -- and probably only here. Is this the only use
4018 * of this infamous trick in shipping, production code? If it
4019 * isn't, it probably should be...
4022 uintptr_t maddr = dtrace_loadptr(daddr +
4023 offsetof(struct dev_info, devi_minor));
4025 uintptr_t next = offsetof(struct ddi_minor_data, next);
4026 uintptr_t name = offsetof(struct ddi_minor_data,
4027 d_minor) + offsetof(struct ddi_minor, name);
4028 uintptr_t dev = offsetof(struct ddi_minor_data,
4029 d_minor) + offsetof(struct ddi_minor, dev);
4033 scout = dtrace_loadptr(maddr + next);
4035 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4038 m = dtrace_load64(maddr + dev) & MAXMIN64;
4040 m = dtrace_load32(maddr + dev) & MAXMIN;
4043 maddr = dtrace_loadptr(maddr + next);
4048 scout = dtrace_loadptr(scout + next);
4053 scout = dtrace_loadptr(scout + next);
4058 if (scout == maddr) {
4059 *flags |= CPU_DTRACE_ILLOP;
4067 * We have the minor data. Now we need to
4068 * copy the minor's name into the end of the
4071 s = (char *)dtrace_loadptr(maddr + name);
4072 len = dtrace_strlen(s, size);
4074 if (*flags & CPU_DTRACE_FAULT)
4078 if ((end -= (len + 1)) < start)
4084 for (i = 1; i <= len; i++)
4085 end[i] = dtrace_load8((uintptr_t)s++);
4090 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4091 ddi_node_state_t devi_state;
4093 devi_state = dtrace_load32(daddr +
4094 offsetof(struct dev_info, devi_node_state));
4096 if (*flags & CPU_DTRACE_FAULT)
4099 if (devi_state >= DS_INITIALIZED) {
4100 s = (char *)dtrace_loadptr(daddr +
4101 offsetof(struct dev_info, devi_addr));
4102 len = dtrace_strlen(s, size);
4104 if (*flags & CPU_DTRACE_FAULT)
4108 if ((end -= (len + 1)) < start)
4114 for (i = 1; i <= len; i++)
4115 end[i] = dtrace_load8((uintptr_t)s++);
4119 * Now for the node name...
4121 s = (char *)dtrace_loadptr(daddr +
4122 offsetof(struct dev_info, devi_node_name));
4124 daddr = dtrace_loadptr(daddr +
4125 offsetof(struct dev_info, devi_parent));
4128 * If our parent is NULL (that is, if we're the root
4129 * node), we're going to use the special path
4135 len = dtrace_strlen(s, size);
4136 if (*flags & CPU_DTRACE_FAULT)
4139 if ((end -= (len + 1)) < start)
4142 for (i = 1; i <= len; i++)
4143 end[i] = dtrace_load8((uintptr_t)s++);
4146 if (depth++ > dtrace_devdepth_max) {
4147 *flags |= CPU_DTRACE_ILLOP;
4153 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4156 regs[rd] = (uintptr_t)end;
4157 mstate->dtms_scratch_ptr += size;
4164 case DIF_SUBR_STRJOIN: {
4165 char *d = (char *)mstate->dtms_scratch_ptr;
4166 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4167 uintptr_t s1 = tupregs[0].dttk_value;
4168 uintptr_t s2 = tupregs[1].dttk_value;
4171 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4172 !dtrace_strcanload(s2, size, mstate, vstate)) {
4177 if (!DTRACE_INSCRATCH(mstate, size)) {
4178 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4185 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4190 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4198 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4203 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4208 mstate->dtms_scratch_ptr += i;
4209 regs[rd] = (uintptr_t)d;
4215 case DIF_SUBR_LLTOSTR: {
4216 int64_t i = (int64_t)tupregs[0].dttk_value;
4217 int64_t val = i < 0 ? i * -1 : i;
4218 uint64_t size = 22; /* enough room for 2^64 in decimal */
4219 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4221 if (!DTRACE_INSCRATCH(mstate, size)) {
4222 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4227 for (*end-- = '\0'; val; val /= 10)
4228 *end-- = '0' + (val % 10);
4236 regs[rd] = (uintptr_t)end + 1;
4237 mstate->dtms_scratch_ptr += size;
4241 case DIF_SUBR_HTONS:
4242 case DIF_SUBR_NTOHS:
4243 #if BYTE_ORDER == BIG_ENDIAN
4244 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4246 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4251 case DIF_SUBR_HTONL:
4252 case DIF_SUBR_NTOHL:
4253 #if BYTE_ORDER == BIG_ENDIAN
4254 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4256 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4261 case DIF_SUBR_HTONLL:
4262 case DIF_SUBR_NTOHLL:
4263 #if BYTE_ORDER == BIG_ENDIAN
4264 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4266 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4271 case DIF_SUBR_DIRNAME:
4272 case DIF_SUBR_BASENAME: {
4273 char *dest = (char *)mstate->dtms_scratch_ptr;
4274 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4275 uintptr_t src = tupregs[0].dttk_value;
4276 int i, j, len = dtrace_strlen((char *)src, size);
4277 int lastbase = -1, firstbase = -1, lastdir = -1;
4280 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4285 if (!DTRACE_INSCRATCH(mstate, size)) {
4286 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4292 * The basename and dirname for a zero-length string is
4297 src = (uintptr_t)".";
4301 * Start from the back of the string, moving back toward the
4302 * front until we see a character that isn't a slash. That
4303 * character is the last character in the basename.
4305 for (i = len - 1; i >= 0; i--) {
4306 if (dtrace_load8(src + i) != '/')
4314 * Starting from the last character in the basename, move
4315 * towards the front until we find a slash. The character
4316 * that we processed immediately before that is the first
4317 * character in the basename.
4319 for (; i >= 0; i--) {
4320 if (dtrace_load8(src + i) == '/')
4328 * Now keep going until we find a non-slash character. That
4329 * character is the last character in the dirname.
4331 for (; i >= 0; i--) {
4332 if (dtrace_load8(src + i) != '/')
4339 ASSERT(!(lastbase == -1 && firstbase != -1));
4340 ASSERT(!(firstbase == -1 && lastdir != -1));
4342 if (lastbase == -1) {
4344 * We didn't find a non-slash character. We know that
4345 * the length is non-zero, so the whole string must be
4346 * slashes. In either the dirname or the basename
4347 * case, we return '/'.
4349 ASSERT(firstbase == -1);
4350 firstbase = lastbase = lastdir = 0;
4353 if (firstbase == -1) {
4355 * The entire string consists only of a basename
4356 * component. If we're looking for dirname, we need
4357 * to change our string to be just "."; if we're
4358 * looking for a basename, we'll just set the first
4359 * character of the basename to be 0.
4361 if (subr == DIF_SUBR_DIRNAME) {
4362 ASSERT(lastdir == -1);
4363 src = (uintptr_t)".";
4370 if (subr == DIF_SUBR_DIRNAME) {
4371 if (lastdir == -1) {
4373 * We know that we have a slash in the name --
4374 * or lastdir would be set to 0, above. And
4375 * because lastdir is -1, we know that this
4376 * slash must be the first character. (That
4377 * is, the full string must be of the form
4378 * "/basename".) In this case, the last
4379 * character of the directory name is 0.
4387 ASSERT(subr == DIF_SUBR_BASENAME);
4388 ASSERT(firstbase != -1 && lastbase != -1);
4393 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4394 dest[j] = dtrace_load8(src + i);
4397 regs[rd] = (uintptr_t)dest;
4398 mstate->dtms_scratch_ptr += size;
4402 case DIF_SUBR_CLEANPATH: {
4403 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4404 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4405 uintptr_t src = tupregs[0].dttk_value;
4408 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4413 if (!DTRACE_INSCRATCH(mstate, size)) {
4414 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4420 * Move forward, loading each character.
4423 c = dtrace_load8(src + i++);
4425 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4433 c = dtrace_load8(src + i++);
4437 * We have two slashes -- we can just advance
4438 * to the next character.
4445 * This is not "." and it's not ".." -- we can
4446 * just store the "/" and this character and
4454 c = dtrace_load8(src + i++);
4458 * This is a "/./" component. We're not going
4459 * to store anything in the destination buffer;
4460 * we're just going to go to the next component.
4467 * This is not ".." -- we can just store the
4468 * "/." and this character and continue
4477 c = dtrace_load8(src + i++);
4479 if (c != '/' && c != '\0') {
4481 * This is not ".." -- it's "..[mumble]".
4482 * We'll store the "/.." and this character
4483 * and continue processing.
4493 * This is "/../" or "/..\0". We need to back up
4494 * our destination pointer until we find a "/".
4497 while (j != 0 && dest[--j] != '/')
4502 } while (c != '\0');
4505 regs[rd] = (uintptr_t)dest;
4506 mstate->dtms_scratch_ptr += size;
4510 case DIF_SUBR_INET_NTOA:
4511 case DIF_SUBR_INET_NTOA6:
4512 case DIF_SUBR_INET_NTOP: {
4517 if (subr == DIF_SUBR_INET_NTOP) {
4518 af = (int)tupregs[0].dttk_value;
4521 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4525 if (af == AF_INET) {
4530 * Safely load the IPv4 address.
4532 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4535 * Check an IPv4 string will fit in scratch.
4537 size = INET_ADDRSTRLEN;
4538 if (!DTRACE_INSCRATCH(mstate, size)) {
4539 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4543 base = (char *)mstate->dtms_scratch_ptr;
4544 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4547 * Stringify as a dotted decimal quad.
4550 ptr8 = (uint8_t *)&ip4;
4551 for (i = 3; i >= 0; i--) {
4557 for (; val; val /= 10) {
4558 *end-- = '0' + (val % 10);
4565 ASSERT(end + 1 >= base);
4567 } else if (af == AF_INET6) {
4568 struct in6_addr ip6;
4569 int firstzero, tryzero, numzero, v6end;
4571 const char digits[] = "0123456789abcdef";
4574 * Stringify using RFC 1884 convention 2 - 16 bit
4575 * hexadecimal values with a zero-run compression.
4576 * Lower case hexadecimal digits are used.
4577 * eg, fe80::214:4fff:fe0b:76c8.
4578 * The IPv4 embedded form is returned for inet_ntop,
4579 * just the IPv4 string is returned for inet_ntoa6.
4583 * Safely load the IPv6 address.
4586 (void *)(uintptr_t)tupregs[argi].dttk_value,
4587 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4590 * Check an IPv6 string will fit in scratch.
4592 size = INET6_ADDRSTRLEN;
4593 if (!DTRACE_INSCRATCH(mstate, size)) {
4594 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4598 base = (char *)mstate->dtms_scratch_ptr;
4599 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4603 * Find the longest run of 16 bit zero values
4604 * for the single allowed zero compression - "::".
4609 for (i = 0; i < sizeof (struct in6_addr); i++) {
4611 if (ip6._S6_un._S6_u8[i] == 0 &&
4613 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4615 tryzero == -1 && i % 2 == 0) {
4620 if (tryzero != -1 &&
4622 (ip6._S6_un._S6_u8[i] != 0 ||
4624 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4626 i == sizeof (struct in6_addr) - 1)) {
4628 if (i - tryzero <= numzero) {
4633 firstzero = tryzero;
4634 numzero = i - i % 2 - tryzero;
4638 if (ip6._S6_un._S6_u8[i] == 0 &&
4640 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4642 i == sizeof (struct in6_addr) - 1)
4646 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4649 * Check for an IPv4 embedded address.
4651 v6end = sizeof (struct in6_addr) - 2;
4652 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4653 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4654 for (i = sizeof (struct in6_addr) - 1;
4655 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4656 ASSERT(end >= base);
4659 val = ip6._S6_un._S6_u8[i];
4661 val = ip6.__u6_addr.__u6_addr8[i];
4667 for (; val; val /= 10) {
4668 *end-- = '0' + val % 10;
4672 if (i > DTRACE_V4MAPPED_OFFSET)
4676 if (subr == DIF_SUBR_INET_NTOA6)
4680 * Set v6end to skip the IPv4 address that
4681 * we have already stringified.
4687 * Build the IPv6 string by working through the
4688 * address in reverse.
4690 for (i = v6end; i >= 0; i -= 2) {
4691 ASSERT(end >= base);
4693 if (i == firstzero + numzero - 2) {
4700 if (i < 14 && i != firstzero - 2)
4704 val = (ip6._S6_un._S6_u8[i] << 8) +
4705 ip6._S6_un._S6_u8[i + 1];
4707 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4708 ip6.__u6_addr.__u6_addr8[i + 1];
4714 for (; val; val /= 16) {
4715 *end-- = digits[val % 16];
4719 ASSERT(end + 1 >= base);
4723 * The user didn't use AH_INET or AH_INET6.
4725 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4730 inetout: regs[rd] = (uintptr_t)end + 1;
4731 mstate->dtms_scratch_ptr += size;
4735 case DIF_SUBR_MEMREF: {
4736 uintptr_t size = 2 * sizeof(uintptr_t);
4737 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4738 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4740 /* address and length */
4741 memref[0] = tupregs[0].dttk_value;
4742 memref[1] = tupregs[1].dttk_value;
4744 regs[rd] = (uintptr_t) memref;
4745 mstate->dtms_scratch_ptr += scratch_size;
4749 case DIF_SUBR_TYPEREF: {
4750 uintptr_t size = 4 * sizeof(uintptr_t);
4751 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4752 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4754 /* address, num_elements, type_str, type_len */
4755 typeref[0] = tupregs[0].dttk_value;
4756 typeref[1] = tupregs[1].dttk_value;
4757 typeref[2] = tupregs[2].dttk_value;
4758 typeref[3] = tupregs[3].dttk_value;
4760 regs[rd] = (uintptr_t) typeref;
4761 mstate->dtms_scratch_ptr += scratch_size;
4768 * Emulate the execution of DTrace IR instructions specified by the given
4769 * DIF object. This function is deliberately void of assertions as all of
4770 * the necessary checks are handled by a call to dtrace_difo_validate().
4773 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4774 dtrace_vstate_t *vstate, dtrace_state_t *state)
4776 const dif_instr_t *text = difo->dtdo_buf;
4777 const uint_t textlen = difo->dtdo_len;
4778 const char *strtab = difo->dtdo_strtab;
4779 const uint64_t *inttab = difo->dtdo_inttab;
4782 dtrace_statvar_t *svar;
4783 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4785 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4786 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4788 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4789 uint64_t regs[DIF_DIR_NREGS];
4792 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4794 uint_t pc = 0, id, opc = 0;
4800 * We stash the current DIF object into the machine state: we need it
4801 * for subsequent access checking.
4803 mstate->dtms_difo = difo;
4805 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
4807 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4811 r1 = DIF_INSTR_R1(instr);
4812 r2 = DIF_INSTR_R2(instr);
4813 rd = DIF_INSTR_RD(instr);
4815 switch (DIF_INSTR_OP(instr)) {
4817 regs[rd] = regs[r1] | regs[r2];
4820 regs[rd] = regs[r1] ^ regs[r2];
4823 regs[rd] = regs[r1] & regs[r2];
4826 regs[rd] = regs[r1] << regs[r2];
4829 regs[rd] = regs[r1] >> regs[r2];
4832 regs[rd] = regs[r1] - regs[r2];
4835 regs[rd] = regs[r1] + regs[r2];
4838 regs[rd] = regs[r1] * regs[r2];
4841 if (regs[r2] == 0) {
4843 *flags |= CPU_DTRACE_DIVZERO;
4845 regs[rd] = (int64_t)regs[r1] /
4851 if (regs[r2] == 0) {
4853 *flags |= CPU_DTRACE_DIVZERO;
4855 regs[rd] = regs[r1] / regs[r2];
4860 if (regs[r2] == 0) {
4862 *flags |= CPU_DTRACE_DIVZERO;
4864 regs[rd] = (int64_t)regs[r1] %
4870 if (regs[r2] == 0) {
4872 *flags |= CPU_DTRACE_DIVZERO;
4874 regs[rd] = regs[r1] % regs[r2];
4879 regs[rd] = ~regs[r1];
4882 regs[rd] = regs[r1];
4885 cc_r = regs[r1] - regs[r2];
4889 cc_c = regs[r1] < regs[r2];
4892 cc_n = cc_v = cc_c = 0;
4893 cc_z = regs[r1] == 0;
4896 pc = DIF_INSTR_LABEL(instr);
4900 pc = DIF_INSTR_LABEL(instr);
4904 pc = DIF_INSTR_LABEL(instr);
4907 if ((cc_z | (cc_n ^ cc_v)) == 0)
4908 pc = DIF_INSTR_LABEL(instr);
4911 if ((cc_c | cc_z) == 0)
4912 pc = DIF_INSTR_LABEL(instr);
4915 if ((cc_n ^ cc_v) == 0)
4916 pc = DIF_INSTR_LABEL(instr);
4920 pc = DIF_INSTR_LABEL(instr);
4924 pc = DIF_INSTR_LABEL(instr);
4928 pc = DIF_INSTR_LABEL(instr);
4931 if (cc_z | (cc_n ^ cc_v))
4932 pc = DIF_INSTR_LABEL(instr);
4936 pc = DIF_INSTR_LABEL(instr);
4939 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4940 *flags |= CPU_DTRACE_KPRIV;
4946 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
4949 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4950 *flags |= CPU_DTRACE_KPRIV;
4956 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
4959 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4960 *flags |= CPU_DTRACE_KPRIV;
4966 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
4969 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4970 *flags |= CPU_DTRACE_KPRIV;
4976 regs[rd] = dtrace_load8(regs[r1]);
4979 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4980 *flags |= CPU_DTRACE_KPRIV;
4986 regs[rd] = dtrace_load16(regs[r1]);
4989 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4990 *flags |= CPU_DTRACE_KPRIV;
4996 regs[rd] = dtrace_load32(regs[r1]);
4999 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5000 *flags |= CPU_DTRACE_KPRIV;
5006 regs[rd] = dtrace_load64(regs[r1]);
5010 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5013 regs[rd] = (int16_t)
5014 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5017 regs[rd] = (int32_t)
5018 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5022 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5026 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5030 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5034 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5043 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5046 regs[rd] = (uint64_t)(uintptr_t)
5047 (strtab + DIF_INSTR_STRING(instr));
5050 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5051 uintptr_t s1 = regs[r1];
5052 uintptr_t s2 = regs[r2];
5055 !dtrace_strcanload(s1, sz, mstate, vstate))
5058 !dtrace_strcanload(s2, sz, mstate, vstate))
5061 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5069 regs[rd] = dtrace_dif_variable(mstate, state,
5073 id = DIF_INSTR_VAR(instr);
5075 if (id >= DIF_VAR_OTHER_UBASE) {
5078 id -= DIF_VAR_OTHER_UBASE;
5079 svar = vstate->dtvs_globals[id];
5080 ASSERT(svar != NULL);
5081 v = &svar->dtsv_var;
5083 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5084 regs[rd] = svar->dtsv_data;
5088 a = (uintptr_t)svar->dtsv_data;
5090 if (*(uint8_t *)a == UINT8_MAX) {
5092 * If the 0th byte is set to UINT8_MAX
5093 * then this is to be treated as a
5094 * reference to a NULL variable.
5098 regs[rd] = a + sizeof (uint64_t);
5104 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5108 id = DIF_INSTR_VAR(instr);
5110 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5111 id -= DIF_VAR_OTHER_UBASE;
5113 svar = vstate->dtvs_globals[id];
5114 ASSERT(svar != NULL);
5115 v = &svar->dtsv_var;
5117 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5118 uintptr_t a = (uintptr_t)svar->dtsv_data;
5121 ASSERT(svar->dtsv_size != 0);
5123 if (regs[rd] == 0) {
5124 *(uint8_t *)a = UINT8_MAX;
5128 a += sizeof (uint64_t);
5130 if (!dtrace_vcanload(
5131 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5135 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5136 (void *)a, &v->dtdv_type);
5140 svar->dtsv_data = regs[rd];
5145 * There are no DTrace built-in thread-local arrays at
5146 * present. This opcode is saved for future work.
5148 *flags |= CPU_DTRACE_ILLOP;
5153 id = DIF_INSTR_VAR(instr);
5155 if (id < DIF_VAR_OTHER_UBASE) {
5157 * For now, this has no meaning.
5163 id -= DIF_VAR_OTHER_UBASE;
5165 ASSERT(id < vstate->dtvs_nlocals);
5166 ASSERT(vstate->dtvs_locals != NULL);
5168 svar = vstate->dtvs_locals[id];
5169 ASSERT(svar != NULL);
5170 v = &svar->dtsv_var;
5172 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5173 uintptr_t a = (uintptr_t)svar->dtsv_data;
5174 size_t sz = v->dtdv_type.dtdt_size;
5176 sz += sizeof (uint64_t);
5177 ASSERT(svar->dtsv_size == NCPU * sz);
5180 if (*(uint8_t *)a == UINT8_MAX) {
5182 * If the 0th byte is set to UINT8_MAX
5183 * then this is to be treated as a
5184 * reference to a NULL variable.
5188 regs[rd] = a + sizeof (uint64_t);
5194 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5195 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5196 regs[rd] = tmp[curcpu];
5200 id = DIF_INSTR_VAR(instr);
5202 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5203 id -= DIF_VAR_OTHER_UBASE;
5204 ASSERT(id < vstate->dtvs_nlocals);
5206 ASSERT(vstate->dtvs_locals != NULL);
5207 svar = vstate->dtvs_locals[id];
5208 ASSERT(svar != NULL);
5209 v = &svar->dtsv_var;
5211 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5212 uintptr_t a = (uintptr_t)svar->dtsv_data;
5213 size_t sz = v->dtdv_type.dtdt_size;
5215 sz += sizeof (uint64_t);
5216 ASSERT(svar->dtsv_size == NCPU * sz);
5219 if (regs[rd] == 0) {
5220 *(uint8_t *)a = UINT8_MAX;
5224 a += sizeof (uint64_t);
5227 if (!dtrace_vcanload(
5228 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5232 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5233 (void *)a, &v->dtdv_type);
5237 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5238 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5239 tmp[curcpu] = regs[rd];
5243 dtrace_dynvar_t *dvar;
5246 id = DIF_INSTR_VAR(instr);
5247 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5248 id -= DIF_VAR_OTHER_UBASE;
5249 v = &vstate->dtvs_tlocals[id];
5251 key = &tupregs[DIF_DTR_NREGS];
5252 key[0].dttk_value = (uint64_t)id;
5253 key[0].dttk_size = 0;
5254 DTRACE_TLS_THRKEY(key[1].dttk_value);
5255 key[1].dttk_size = 0;
5257 dvar = dtrace_dynvar(dstate, 2, key,
5258 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5266 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5267 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5269 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5276 dtrace_dynvar_t *dvar;
5279 id = DIF_INSTR_VAR(instr);
5280 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5281 id -= DIF_VAR_OTHER_UBASE;
5283 key = &tupregs[DIF_DTR_NREGS];
5284 key[0].dttk_value = (uint64_t)id;
5285 key[0].dttk_size = 0;
5286 DTRACE_TLS_THRKEY(key[1].dttk_value);
5287 key[1].dttk_size = 0;
5288 v = &vstate->dtvs_tlocals[id];
5290 dvar = dtrace_dynvar(dstate, 2, key,
5291 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5292 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5293 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5294 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5297 * Given that we're storing to thread-local data,
5298 * we need to flush our predicate cache.
5300 curthread->t_predcache = 0;
5305 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5306 if (!dtrace_vcanload(
5307 (void *)(uintptr_t)regs[rd],
5308 &v->dtdv_type, mstate, vstate))
5311 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5312 dvar->dtdv_data, &v->dtdv_type);
5314 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5321 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5325 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5326 regs, tupregs, ttop, mstate, state);
5330 if (ttop == DIF_DTR_NREGS) {
5331 *flags |= CPU_DTRACE_TUPOFLOW;
5335 if (r1 == DIF_TYPE_STRING) {
5337 * If this is a string type and the size is 0,
5338 * we'll use the system-wide default string
5339 * size. Note that we are _not_ looking at
5340 * the value of the DTRACEOPT_STRSIZE option;
5341 * had this been set, we would expect to have
5342 * a non-zero size value in the "pushtr".
5344 tupregs[ttop].dttk_size =
5345 dtrace_strlen((char *)(uintptr_t)regs[rd],
5346 regs[r2] ? regs[r2] :
5347 dtrace_strsize_default) + 1;
5349 tupregs[ttop].dttk_size = regs[r2];
5352 tupregs[ttop++].dttk_value = regs[rd];
5356 if (ttop == DIF_DTR_NREGS) {
5357 *flags |= CPU_DTRACE_TUPOFLOW;
5361 tupregs[ttop].dttk_value = regs[rd];
5362 tupregs[ttop++].dttk_size = 0;
5370 case DIF_OP_FLUSHTS:
5375 case DIF_OP_LDTAA: {
5376 dtrace_dynvar_t *dvar;
5377 dtrace_key_t *key = tupregs;
5378 uint_t nkeys = ttop;
5380 id = DIF_INSTR_VAR(instr);
5381 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5382 id -= DIF_VAR_OTHER_UBASE;
5384 key[nkeys].dttk_value = (uint64_t)id;
5385 key[nkeys++].dttk_size = 0;
5387 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5388 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5389 key[nkeys++].dttk_size = 0;
5390 v = &vstate->dtvs_tlocals[id];
5392 v = &vstate->dtvs_globals[id]->dtsv_var;
5395 dvar = dtrace_dynvar(dstate, nkeys, key,
5396 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5397 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5398 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5405 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5406 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5408 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5415 case DIF_OP_STTAA: {
5416 dtrace_dynvar_t *dvar;
5417 dtrace_key_t *key = tupregs;
5418 uint_t nkeys = ttop;
5420 id = DIF_INSTR_VAR(instr);
5421 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5422 id -= DIF_VAR_OTHER_UBASE;
5424 key[nkeys].dttk_value = (uint64_t)id;
5425 key[nkeys++].dttk_size = 0;
5427 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5428 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5429 key[nkeys++].dttk_size = 0;
5430 v = &vstate->dtvs_tlocals[id];
5432 v = &vstate->dtvs_globals[id]->dtsv_var;
5435 dvar = dtrace_dynvar(dstate, nkeys, key,
5436 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5437 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5438 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5439 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5444 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5445 if (!dtrace_vcanload(
5446 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5450 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5451 dvar->dtdv_data, &v->dtdv_type);
5453 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5459 case DIF_OP_ALLOCS: {
5460 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5461 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5464 * Rounding up the user allocation size could have
5465 * overflowed large, bogus allocations (like -1ULL) to
5468 if (size < regs[r1] ||
5469 !DTRACE_INSCRATCH(mstate, size)) {
5470 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5475 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5476 mstate->dtms_scratch_ptr += size;
5482 if (!dtrace_canstore(regs[rd], regs[r2],
5484 *flags |= CPU_DTRACE_BADADDR;
5489 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5492 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5493 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5497 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5498 *flags |= CPU_DTRACE_BADADDR;
5502 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5506 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5507 *flags |= CPU_DTRACE_BADADDR;
5512 *flags |= CPU_DTRACE_BADALIGN;
5516 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5520 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5521 *flags |= CPU_DTRACE_BADADDR;
5526 *flags |= CPU_DTRACE_BADALIGN;
5530 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5534 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5535 *flags |= CPU_DTRACE_BADADDR;
5540 *flags |= CPU_DTRACE_BADALIGN;
5544 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5549 if (!(*flags & CPU_DTRACE_FAULT))
5552 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5553 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5559 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5561 dtrace_probe_t *probe = ecb->dte_probe;
5562 dtrace_provider_t *prov = probe->dtpr_provider;
5563 char c[DTRACE_FULLNAMELEN + 80], *str;
5564 char *msg = "dtrace: breakpoint action at probe ";
5565 char *ecbmsg = " (ecb ";
5566 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5567 uintptr_t val = (uintptr_t)ecb;
5568 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5570 if (dtrace_destructive_disallow)
5574 * It's impossible to be taking action on the NULL probe.
5576 ASSERT(probe != NULL);
5579 * This is a poor man's (destitute man's?) sprintf(): we want to
5580 * print the provider name, module name, function name and name of
5581 * the probe, along with the hex address of the ECB with the breakpoint
5582 * action -- all of which we must place in the character buffer by
5585 while (*msg != '\0')
5588 for (str = prov->dtpv_name; *str != '\0'; str++)
5592 for (str = probe->dtpr_mod; *str != '\0'; str++)
5596 for (str = probe->dtpr_func; *str != '\0'; str++)
5600 for (str = probe->dtpr_name; *str != '\0'; str++)
5603 while (*ecbmsg != '\0')
5606 while (shift >= 0) {
5607 mask = (uintptr_t)0xf << shift;
5609 if (val >= ((uintptr_t)1 << shift))
5610 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5620 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5625 dtrace_action_panic(dtrace_ecb_t *ecb)
5627 dtrace_probe_t *probe = ecb->dte_probe;
5630 * It's impossible to be taking action on the NULL probe.
5632 ASSERT(probe != NULL);
5634 if (dtrace_destructive_disallow)
5637 if (dtrace_panicked != NULL)
5640 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5644 * We won the right to panic. (We want to be sure that only one
5645 * thread calls panic() from dtrace_probe(), and that panic() is
5646 * called exactly once.)
5648 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5649 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5650 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5654 dtrace_action_raise(uint64_t sig)
5656 if (dtrace_destructive_disallow)
5660 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5666 * raise() has a queue depth of 1 -- we ignore all subsequent
5667 * invocations of the raise() action.
5669 if (curthread->t_dtrace_sig == 0)
5670 curthread->t_dtrace_sig = (uint8_t)sig;
5672 curthread->t_sig_check = 1;
5675 struct proc *p = curproc;
5683 dtrace_action_stop(void)
5685 if (dtrace_destructive_disallow)
5689 if (!curthread->t_dtrace_stop) {
5690 curthread->t_dtrace_stop = 1;
5691 curthread->t_sig_check = 1;
5695 struct proc *p = curproc;
5697 psignal(p, SIGSTOP);
5703 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5706 volatile uint16_t *flags;
5710 cpu_t *cpu = &solaris_cpu[curcpu];
5713 if (dtrace_destructive_disallow)
5716 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5718 now = dtrace_gethrtime();
5720 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5722 * We need to advance the mark to the current time.
5724 cpu->cpu_dtrace_chillmark = now;
5725 cpu->cpu_dtrace_chilled = 0;
5729 * Now check to see if the requested chill time would take us over
5730 * the maximum amount of time allowed in the chill interval. (Or
5731 * worse, if the calculation itself induces overflow.)
5733 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5734 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5735 *flags |= CPU_DTRACE_ILLOP;
5739 while (dtrace_gethrtime() - now < val)
5743 * Normally, we assure that the value of the variable "timestamp" does
5744 * not change within an ECB. The presence of chill() represents an
5745 * exception to this rule, however.
5747 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5748 cpu->cpu_dtrace_chilled += val;
5752 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5753 uint64_t *buf, uint64_t arg)
5755 int nframes = DTRACE_USTACK_NFRAMES(arg);
5756 int strsize = DTRACE_USTACK_STRSIZE(arg);
5757 uint64_t *pcs = &buf[1], *fps;
5758 char *str = (char *)&pcs[nframes];
5759 int size, offs = 0, i, j;
5760 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5761 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5765 * Should be taking a faster path if string space has not been
5768 ASSERT(strsize != 0);
5771 * We will first allocate some temporary space for the frame pointers.
5773 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5774 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5775 (nframes * sizeof (uint64_t));
5777 if (!DTRACE_INSCRATCH(mstate, size)) {
5779 * Not enough room for our frame pointers -- need to indicate
5780 * that we ran out of scratch space.
5782 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5786 mstate->dtms_scratch_ptr += size;
5787 saved = mstate->dtms_scratch_ptr;
5790 * Now get a stack with both program counters and frame pointers.
5792 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5793 dtrace_getufpstack(buf, fps, nframes + 1);
5794 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5797 * If that faulted, we're cooked.
5799 if (*flags & CPU_DTRACE_FAULT)
5803 * Now we want to walk up the stack, calling the USTACK helper. For
5804 * each iteration, we restore the scratch pointer.
5806 for (i = 0; i < nframes; i++) {
5807 mstate->dtms_scratch_ptr = saved;
5809 if (offs >= strsize)
5812 sym = (char *)(uintptr_t)dtrace_helper(
5813 DTRACE_HELPER_ACTION_USTACK,
5814 mstate, state, pcs[i], fps[i]);
5817 * If we faulted while running the helper, we're going to
5818 * clear the fault and null out the corresponding string.
5820 if (*flags & CPU_DTRACE_FAULT) {
5821 *flags &= ~CPU_DTRACE_FAULT;
5831 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5834 * Now copy in the string that the helper returned to us.
5836 for (j = 0; offs + j < strsize; j++) {
5837 if ((str[offs + j] = sym[j]) == '\0')
5841 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5846 if (offs >= strsize) {
5848 * If we didn't have room for all of the strings, we don't
5849 * abort processing -- this needn't be a fatal error -- but we
5850 * still want to increment a counter (dts_stkstroverflows) to
5851 * allow this condition to be warned about. (If this is from
5852 * a jstack() action, it is easily tuned via jstackstrsize.)
5854 dtrace_error(&state->dts_stkstroverflows);
5857 while (offs < strsize)
5861 mstate->dtms_scratch_ptr = old;
5865 * If you're looking for the epicenter of DTrace, you just found it. This
5866 * is the function called by the provider to fire a probe -- from which all
5867 * subsequent probe-context DTrace activity emanates.
5870 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5871 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5873 processorid_t cpuid;
5874 dtrace_icookie_t cookie;
5875 dtrace_probe_t *probe;
5876 dtrace_mstate_t mstate;
5878 dtrace_action_t *act;
5882 volatile uint16_t *flags;
5885 if (panicstr != NULL)
5890 * Kick out immediately if this CPU is still being born (in which case
5891 * curthread will be set to -1) or the current thread can't allow
5892 * probes in its current context.
5894 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
5898 cookie = dtrace_interrupt_disable();
5899 probe = dtrace_probes[id - 1];
5901 onintr = CPU_ON_INTR(CPU);
5903 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
5904 probe->dtpr_predcache == curthread->t_predcache) {
5906 * We have hit in the predicate cache; we know that
5907 * this predicate would evaluate to be false.
5909 dtrace_interrupt_enable(cookie);
5914 if (panic_quiesce) {
5916 if (panicstr != NULL) {
5919 * We don't trace anything if we're panicking.
5921 dtrace_interrupt_enable(cookie);
5925 now = dtrace_gethrtime();
5926 vtime = dtrace_vtime_references != 0;
5928 if (vtime && curthread->t_dtrace_start)
5929 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
5931 mstate.dtms_difo = NULL;
5932 mstate.dtms_probe = probe;
5933 mstate.dtms_strtok = 0;
5934 mstate.dtms_arg[0] = arg0;
5935 mstate.dtms_arg[1] = arg1;
5936 mstate.dtms_arg[2] = arg2;
5937 mstate.dtms_arg[3] = arg3;
5938 mstate.dtms_arg[4] = arg4;
5940 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
5942 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
5943 dtrace_predicate_t *pred = ecb->dte_predicate;
5944 dtrace_state_t *state = ecb->dte_state;
5945 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
5946 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
5947 dtrace_vstate_t *vstate = &state->dts_vstate;
5948 dtrace_provider_t *prov = probe->dtpr_provider;
5953 * A little subtlety with the following (seemingly innocuous)
5954 * declaration of the automatic 'val': by looking at the
5955 * code, you might think that it could be declared in the
5956 * action processing loop, below. (That is, it's only used in
5957 * the action processing loop.) However, it must be declared
5958 * out of that scope because in the case of DIF expression
5959 * arguments to aggregating actions, one iteration of the
5960 * action loop will use the last iteration's value.
5964 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
5965 *flags &= ~CPU_DTRACE_ERROR;
5967 if (prov == dtrace_provider) {
5969 * If dtrace itself is the provider of this probe,
5970 * we're only going to continue processing the ECB if
5971 * arg0 (the dtrace_state_t) is equal to the ECB's
5972 * creating state. (This prevents disjoint consumers
5973 * from seeing one another's metaprobes.)
5975 if (arg0 != (uint64_t)(uintptr_t)state)
5979 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
5981 * We're not currently active. If our provider isn't
5982 * the dtrace pseudo provider, we're not interested.
5984 if (prov != dtrace_provider)
5988 * Now we must further check if we are in the BEGIN
5989 * probe. If we are, we will only continue processing
5990 * if we're still in WARMUP -- if one BEGIN enabling
5991 * has invoked the exit() action, we don't want to
5992 * evaluate subsequent BEGIN enablings.
5994 if (probe->dtpr_id == dtrace_probeid_begin &&
5995 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
5996 ASSERT(state->dts_activity ==
5997 DTRACE_ACTIVITY_DRAINING);
6002 if (ecb->dte_cond) {
6004 * If the dte_cond bits indicate that this
6005 * consumer is only allowed to see user-mode firings
6006 * of this probe, call the provider's dtps_usermode()
6007 * entry point to check that the probe was fired
6008 * while in a user context. Skip this ECB if that's
6011 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6012 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6013 probe->dtpr_id, probe->dtpr_arg) == 0)
6018 * This is more subtle than it looks. We have to be
6019 * absolutely certain that CRED() isn't going to
6020 * change out from under us so it's only legit to
6021 * examine that structure if we're in constrained
6022 * situations. Currently, the only times we'll this
6023 * check is if a non-super-user has enabled the
6024 * profile or syscall providers -- providers that
6025 * allow visibility of all processes. For the
6026 * profile case, the check above will ensure that
6027 * we're examining a user context.
6029 if (ecb->dte_cond & DTRACE_COND_OWNER) {
6032 ecb->dte_state->dts_cred.dcr_cred;
6035 ASSERT(s_cr != NULL);
6037 if ((cr = CRED()) == NULL ||
6038 s_cr->cr_uid != cr->cr_uid ||
6039 s_cr->cr_uid != cr->cr_ruid ||
6040 s_cr->cr_uid != cr->cr_suid ||
6041 s_cr->cr_gid != cr->cr_gid ||
6042 s_cr->cr_gid != cr->cr_rgid ||
6043 s_cr->cr_gid != cr->cr_sgid ||
6044 (proc = ttoproc(curthread)) == NULL ||
6045 (proc->p_flag & SNOCD))
6049 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6052 ecb->dte_state->dts_cred.dcr_cred;
6054 ASSERT(s_cr != NULL);
6056 if ((cr = CRED()) == NULL ||
6057 s_cr->cr_zone->zone_id !=
6058 cr->cr_zone->zone_id)
6064 if (now - state->dts_alive > dtrace_deadman_timeout) {
6066 * We seem to be dead. Unless we (a) have kernel
6067 * destructive permissions (b) have expicitly enabled
6068 * destructive actions and (c) destructive actions have
6069 * not been disabled, we're going to transition into
6070 * the KILLED state, from which no further processing
6071 * on this state will be performed.
6073 if (!dtrace_priv_kernel_destructive(state) ||
6074 !state->dts_cred.dcr_destructive ||
6075 dtrace_destructive_disallow) {
6076 void *activity = &state->dts_activity;
6077 dtrace_activity_t current;
6080 current = state->dts_activity;
6081 } while (dtrace_cas32(activity, current,
6082 DTRACE_ACTIVITY_KILLED) != current);
6088 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6089 ecb->dte_alignment, state, &mstate)) < 0)
6092 tomax = buf->dtb_tomax;
6093 ASSERT(tomax != NULL);
6095 if (ecb->dte_size != 0)
6096 DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
6098 mstate.dtms_epid = ecb->dte_epid;
6099 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6101 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6102 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6104 mstate.dtms_access = 0;
6107 dtrace_difo_t *dp = pred->dtp_difo;
6110 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6112 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6113 dtrace_cacheid_t cid = probe->dtpr_predcache;
6115 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6117 * Update the predicate cache...
6119 ASSERT(cid == pred->dtp_cacheid);
6120 curthread->t_predcache = cid;
6127 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6128 act != NULL; act = act->dta_next) {
6131 dtrace_recdesc_t *rec = &act->dta_rec;
6133 size = rec->dtrd_size;
6134 valoffs = offs + rec->dtrd_offset;
6136 if (DTRACEACT_ISAGG(act->dta_kind)) {
6138 dtrace_aggregation_t *agg;
6140 agg = (dtrace_aggregation_t *)act;
6142 if ((dp = act->dta_difo) != NULL)
6143 v = dtrace_dif_emulate(dp,
6144 &mstate, vstate, state);
6146 if (*flags & CPU_DTRACE_ERROR)
6150 * Note that we always pass the expression
6151 * value from the previous iteration of the
6152 * action loop. This value will only be used
6153 * if there is an expression argument to the
6154 * aggregating action, denoted by the
6155 * dtag_hasarg field.
6157 dtrace_aggregate(agg, buf,
6158 offs, aggbuf, v, val);
6162 switch (act->dta_kind) {
6163 case DTRACEACT_STOP:
6164 if (dtrace_priv_proc_destructive(state))
6165 dtrace_action_stop();
6168 case DTRACEACT_BREAKPOINT:
6169 if (dtrace_priv_kernel_destructive(state))
6170 dtrace_action_breakpoint(ecb);
6173 case DTRACEACT_PANIC:
6174 if (dtrace_priv_kernel_destructive(state))
6175 dtrace_action_panic(ecb);
6178 case DTRACEACT_STACK:
6179 if (!dtrace_priv_kernel(state))
6182 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6183 size / sizeof (pc_t), probe->dtpr_aframes,
6184 DTRACE_ANCHORED(probe) ? NULL :
6188 case DTRACEACT_JSTACK:
6189 case DTRACEACT_USTACK:
6190 if (!dtrace_priv_proc(state))
6194 * See comment in DIF_VAR_PID.
6196 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6198 int depth = DTRACE_USTACK_NFRAMES(
6201 dtrace_bzero((void *)(tomax + valoffs),
6202 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6203 + depth * sizeof (uint64_t));
6208 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6209 curproc->p_dtrace_helpers != NULL) {
6211 * This is the slow path -- we have
6212 * allocated string space, and we're
6213 * getting the stack of a process that
6214 * has helpers. Call into a separate
6215 * routine to perform this processing.
6217 dtrace_action_ustack(&mstate, state,
6218 (uint64_t *)(tomax + valoffs),
6223 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6224 dtrace_getupcstack((uint64_t *)
6226 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6227 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6237 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6239 if (*flags & CPU_DTRACE_ERROR)
6242 switch (act->dta_kind) {
6243 case DTRACEACT_SPECULATE:
6244 ASSERT(buf == &state->dts_buffer[cpuid]);
6245 buf = dtrace_speculation_buffer(state,
6249 *flags |= CPU_DTRACE_DROP;
6253 offs = dtrace_buffer_reserve(buf,
6254 ecb->dte_needed, ecb->dte_alignment,
6258 *flags |= CPU_DTRACE_DROP;
6262 tomax = buf->dtb_tomax;
6263 ASSERT(tomax != NULL);
6265 if (ecb->dte_size != 0)
6266 DTRACE_STORE(uint32_t, tomax, offs,
6270 case DTRACEACT_PRINTM: {
6271 /* The DIF returns a 'memref'. */
6272 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6274 /* Get the size from the memref. */
6278 * Check if the size exceeds the allocated
6281 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6283 *flags |= CPU_DTRACE_DROP;
6287 /* Store the size in the buffer first. */
6288 DTRACE_STORE(uintptr_t, tomax,
6292 * Offset the buffer address to the start
6295 valoffs += sizeof(uintptr_t);
6298 * Reset to the memory address rather than
6299 * the memref array, then let the BYREF
6300 * code below do the work to store the
6301 * memory data in the buffer.
6307 case DTRACEACT_PRINTT: {
6308 /* The DIF returns a 'typeref'. */
6309 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6314 * Get the type string length and round it
6315 * up so that the data that follows is
6316 * aligned for easy access.
6318 size_t typs = strlen((char *) typeref[2]) + 1;
6319 typs = roundup(typs, sizeof(uintptr_t));
6322 *Get the size from the typeref using the
6323 * number of elements and the type size.
6325 size = typeref[1] * typeref[3];
6328 * Check if the size exceeds the allocated
6331 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6333 *flags |= CPU_DTRACE_DROP;
6337 /* Store the size in the buffer first. */
6338 DTRACE_STORE(uintptr_t, tomax,
6340 valoffs += sizeof(uintptr_t);
6342 /* Store the type size in the buffer. */
6343 DTRACE_STORE(uintptr_t, tomax,
6344 valoffs, typeref[3]);
6345 valoffs += sizeof(uintptr_t);
6349 for (s = 0; s < typs; s++) {
6351 c = dtrace_load8(val++);
6353 DTRACE_STORE(uint8_t, tomax,
6358 * Reset to the memory address rather than
6359 * the typeref array, then let the BYREF
6360 * code below do the work to store the
6361 * memory data in the buffer.
6367 case DTRACEACT_CHILL:
6368 if (dtrace_priv_kernel_destructive(state))
6369 dtrace_action_chill(&mstate, val);
6372 case DTRACEACT_RAISE:
6373 if (dtrace_priv_proc_destructive(state))
6374 dtrace_action_raise(val);
6377 case DTRACEACT_COMMIT:
6381 * We need to commit our buffer state.
6384 buf->dtb_offset = offs + ecb->dte_size;
6385 buf = &state->dts_buffer[cpuid];
6386 dtrace_speculation_commit(state, cpuid, val);
6390 case DTRACEACT_DISCARD:
6391 dtrace_speculation_discard(state, cpuid, val);
6394 case DTRACEACT_DIFEXPR:
6395 case DTRACEACT_LIBACT:
6396 case DTRACEACT_PRINTF:
6397 case DTRACEACT_PRINTA:
6398 case DTRACEACT_SYSTEM:
6399 case DTRACEACT_FREOPEN:
6404 if (!dtrace_priv_kernel(state))
6408 case DTRACEACT_USYM:
6409 case DTRACEACT_UMOD:
6410 case DTRACEACT_UADDR: {
6412 struct pid *pid = curthread->t_procp->p_pidp;
6415 if (!dtrace_priv_proc(state))
6418 DTRACE_STORE(uint64_t, tomax,
6420 valoffs, (uint64_t)pid->pid_id);
6422 valoffs, (uint64_t) curproc->p_pid);
6424 DTRACE_STORE(uint64_t, tomax,
6425 valoffs + sizeof (uint64_t), val);
6430 case DTRACEACT_EXIT: {
6432 * For the exit action, we are going to attempt
6433 * to atomically set our activity to be
6434 * draining. If this fails (either because
6435 * another CPU has beat us to the exit action,
6436 * or because our current activity is something
6437 * other than ACTIVE or WARMUP), we will
6438 * continue. This assures that the exit action
6439 * can be successfully recorded at most once
6440 * when we're in the ACTIVE state. If we're
6441 * encountering the exit() action while in
6442 * COOLDOWN, however, we want to honor the new
6443 * status code. (We know that we're the only
6444 * thread in COOLDOWN, so there is no race.)
6446 void *activity = &state->dts_activity;
6447 dtrace_activity_t current = state->dts_activity;
6449 if (current == DTRACE_ACTIVITY_COOLDOWN)
6452 if (current != DTRACE_ACTIVITY_WARMUP)
6453 current = DTRACE_ACTIVITY_ACTIVE;
6455 if (dtrace_cas32(activity, current,
6456 DTRACE_ACTIVITY_DRAINING) != current) {
6457 *flags |= CPU_DTRACE_DROP;
6468 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6469 uintptr_t end = valoffs + size;
6471 if (!dtrace_vcanload((void *)(uintptr_t)val,
6472 &dp->dtdo_rtype, &mstate, vstate))
6476 * If this is a string, we're going to only
6477 * load until we find the zero byte -- after
6478 * which we'll store zero bytes.
6480 if (dp->dtdo_rtype.dtdt_kind ==
6483 int intuple = act->dta_intuple;
6486 for (s = 0; s < size; s++) {
6488 c = dtrace_load8(val++);
6490 DTRACE_STORE(uint8_t, tomax,
6493 if (c == '\0' && intuple)
6500 while (valoffs < end) {
6501 DTRACE_STORE(uint8_t, tomax, valoffs++,
6502 dtrace_load8(val++));
6512 case sizeof (uint8_t):
6513 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6515 case sizeof (uint16_t):
6516 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6518 case sizeof (uint32_t):
6519 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6521 case sizeof (uint64_t):
6522 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6526 * Any other size should have been returned by
6527 * reference, not by value.
6534 if (*flags & CPU_DTRACE_DROP)
6537 if (*flags & CPU_DTRACE_FAULT) {
6539 dtrace_action_t *err;
6543 if (probe->dtpr_id == dtrace_probeid_error) {
6545 * There's nothing we can do -- we had an
6546 * error on the error probe. We bump an
6547 * error counter to at least indicate that
6548 * this condition happened.
6550 dtrace_error(&state->dts_dblerrors);
6556 * Before recursing on dtrace_probe(), we
6557 * need to explicitly clear out our start
6558 * time to prevent it from being accumulated
6559 * into t_dtrace_vtime.
6561 curthread->t_dtrace_start = 0;
6565 * Iterate over the actions to figure out which action
6566 * we were processing when we experienced the error.
6567 * Note that act points _past_ the faulting action; if
6568 * act is ecb->dte_action, the fault was in the
6569 * predicate, if it's ecb->dte_action->dta_next it's
6570 * in action #1, and so on.
6572 for (err = ecb->dte_action, ndx = 0;
6573 err != act; err = err->dta_next, ndx++)
6576 dtrace_probe_error(state, ecb->dte_epid, ndx,
6577 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6578 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6579 cpu_core[cpuid].cpuc_dtrace_illval);
6585 buf->dtb_offset = offs + ecb->dte_size;
6589 curthread->t_dtrace_start = dtrace_gethrtime();
6591 dtrace_interrupt_enable(cookie);
6595 * DTrace Probe Hashing Functions
6597 * The functions in this section (and indeed, the functions in remaining
6598 * sections) are not _called_ from probe context. (Any exceptions to this are
6599 * marked with a "Note:".) Rather, they are called from elsewhere in the
6600 * DTrace framework to look-up probes in, add probes to and remove probes from
6601 * the DTrace probe hashes. (Each probe is hashed by each element of the
6602 * probe tuple -- allowing for fast lookups, regardless of what was
6606 dtrace_hash_str(const char *p)
6612 hval = (hval << 4) + *p++;
6613 if ((g = (hval & 0xf0000000)) != 0)
6620 static dtrace_hash_t *
6621 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6623 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6625 hash->dth_stroffs = stroffs;
6626 hash->dth_nextoffs = nextoffs;
6627 hash->dth_prevoffs = prevoffs;
6630 hash->dth_mask = hash->dth_size - 1;
6632 hash->dth_tab = kmem_zalloc(hash->dth_size *
6633 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6639 dtrace_hash_destroy(dtrace_hash_t *hash)
6644 for (i = 0; i < hash->dth_size; i++)
6645 ASSERT(hash->dth_tab[i] == NULL);
6648 kmem_free(hash->dth_tab,
6649 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6650 kmem_free(hash, sizeof (dtrace_hash_t));
6654 dtrace_hash_resize(dtrace_hash_t *hash)
6656 int size = hash->dth_size, i, ndx;
6657 int new_size = hash->dth_size << 1;
6658 int new_mask = new_size - 1;
6659 dtrace_hashbucket_t **new_tab, *bucket, *next;
6661 ASSERT((new_size & new_mask) == 0);
6663 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6665 for (i = 0; i < size; i++) {
6666 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6667 dtrace_probe_t *probe = bucket->dthb_chain;
6669 ASSERT(probe != NULL);
6670 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6672 next = bucket->dthb_next;
6673 bucket->dthb_next = new_tab[ndx];
6674 new_tab[ndx] = bucket;
6678 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6679 hash->dth_tab = new_tab;
6680 hash->dth_size = new_size;
6681 hash->dth_mask = new_mask;
6685 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6687 int hashval = DTRACE_HASHSTR(hash, new);
6688 int ndx = hashval & hash->dth_mask;
6689 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6690 dtrace_probe_t **nextp, **prevp;
6692 for (; bucket != NULL; bucket = bucket->dthb_next) {
6693 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6697 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6698 dtrace_hash_resize(hash);
6699 dtrace_hash_add(hash, new);
6703 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6704 bucket->dthb_next = hash->dth_tab[ndx];
6705 hash->dth_tab[ndx] = bucket;
6706 hash->dth_nbuckets++;
6709 nextp = DTRACE_HASHNEXT(hash, new);
6710 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6711 *nextp = bucket->dthb_chain;
6713 if (bucket->dthb_chain != NULL) {
6714 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6715 ASSERT(*prevp == NULL);
6719 bucket->dthb_chain = new;
6723 static dtrace_probe_t *
6724 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6726 int hashval = DTRACE_HASHSTR(hash, template);
6727 int ndx = hashval & hash->dth_mask;
6728 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6730 for (; bucket != NULL; bucket = bucket->dthb_next) {
6731 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6732 return (bucket->dthb_chain);
6739 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6741 int hashval = DTRACE_HASHSTR(hash, template);
6742 int ndx = hashval & hash->dth_mask;
6743 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6745 for (; bucket != NULL; bucket = bucket->dthb_next) {
6746 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6747 return (bucket->dthb_len);
6754 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6756 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6757 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6759 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6760 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6763 * Find the bucket that we're removing this probe from.
6765 for (; bucket != NULL; bucket = bucket->dthb_next) {
6766 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6770 ASSERT(bucket != NULL);
6772 if (*prevp == NULL) {
6773 if (*nextp == NULL) {
6775 * The removed probe was the only probe on this
6776 * bucket; we need to remove the bucket.
6778 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6780 ASSERT(bucket->dthb_chain == probe);
6784 hash->dth_tab[ndx] = bucket->dthb_next;
6786 while (b->dthb_next != bucket)
6788 b->dthb_next = bucket->dthb_next;
6791 ASSERT(hash->dth_nbuckets > 0);
6792 hash->dth_nbuckets--;
6793 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6797 bucket->dthb_chain = *nextp;
6799 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6803 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6807 * DTrace Utility Functions
6809 * These are random utility functions that are _not_ called from probe context.
6812 dtrace_badattr(const dtrace_attribute_t *a)
6814 return (a->dtat_name > DTRACE_STABILITY_MAX ||
6815 a->dtat_data > DTRACE_STABILITY_MAX ||
6816 a->dtat_class > DTRACE_CLASS_MAX);
6820 * Return a duplicate copy of a string. If the specified string is NULL,
6821 * this function returns a zero-length string.
6824 dtrace_strdup(const char *str)
6826 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6829 (void) strcpy(new, str);
6834 #define DTRACE_ISALPHA(c) \
6835 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6838 dtrace_badname(const char *s)
6842 if (s == NULL || (c = *s++) == '\0')
6845 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6848 while ((c = *s++) != '\0') {
6849 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6850 c != '-' && c != '_' && c != '.' && c != '`')
6858 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6863 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6865 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6867 priv = DTRACE_PRIV_ALL;
6869 *uidp = crgetuid(cr);
6870 *zoneidp = crgetzoneid(cr);
6873 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6874 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6875 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6876 priv |= DTRACE_PRIV_USER;
6877 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6878 priv |= DTRACE_PRIV_PROC;
6879 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6880 priv |= DTRACE_PRIV_OWNER;
6881 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6882 priv |= DTRACE_PRIV_ZONEOWNER;
6885 priv = DTRACE_PRIV_ALL;
6891 #ifdef DTRACE_ERRDEBUG
6893 dtrace_errdebug(const char *str)
6895 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
6898 mutex_enter(&dtrace_errlock);
6899 dtrace_errlast = str;
6900 dtrace_errthread = curthread;
6902 while (occupied++ < DTRACE_ERRHASHSZ) {
6903 if (dtrace_errhash[hval].dter_msg == str) {
6904 dtrace_errhash[hval].dter_count++;
6908 if (dtrace_errhash[hval].dter_msg != NULL) {
6909 hval = (hval + 1) % DTRACE_ERRHASHSZ;
6913 dtrace_errhash[hval].dter_msg = str;
6914 dtrace_errhash[hval].dter_count = 1;
6918 panic("dtrace: undersized error hash");
6920 mutex_exit(&dtrace_errlock);
6925 * DTrace Matching Functions
6927 * These functions are used to match groups of probes, given some elements of
6928 * a probe tuple, or some globbed expressions for elements of a probe tuple.
6931 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
6934 if (priv != DTRACE_PRIV_ALL) {
6935 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
6936 uint32_t match = priv & ppriv;
6939 * No PRIV_DTRACE_* privileges...
6941 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
6942 DTRACE_PRIV_KERNEL)) == 0)
6946 * No matching bits, but there were bits to match...
6948 if (match == 0 && ppriv != 0)
6952 * Need to have permissions to the process, but don't...
6954 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
6955 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
6960 * Need to be in the same zone unless we possess the
6961 * privilege to examine all zones.
6963 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
6964 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
6973 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
6974 * consists of input pattern strings and an ops-vector to evaluate them.
6975 * This function returns >0 for match, 0 for no match, and <0 for error.
6978 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
6979 uint32_t priv, uid_t uid, zoneid_t zoneid)
6981 dtrace_provider_t *pvp = prp->dtpr_provider;
6984 if (pvp->dtpv_defunct)
6987 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
6990 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
6993 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
6996 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
6999 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7006 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7007 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7008 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7009 * In addition, all of the recursion cases except for '*' matching have been
7010 * unwound. For '*', we still implement recursive evaluation, but a depth
7011 * counter is maintained and matching is aborted if we recurse too deep.
7012 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7015 dtrace_match_glob(const char *s, const char *p, int depth)
7021 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7025 s = ""; /* treat NULL as empty string */
7034 if ((c = *p++) == '\0')
7035 return (s1 == '\0');
7039 int ok = 0, notflag = 0;
7050 if ((c = *p++) == '\0')
7054 if (c == '-' && lc != '\0' && *p != ']') {
7055 if ((c = *p++) == '\0')
7057 if (c == '\\' && (c = *p++) == '\0')
7061 if (s1 < lc || s1 > c)
7065 } else if (lc <= s1 && s1 <= c)
7068 } else if (c == '\\' && (c = *p++) == '\0')
7071 lc = c; /* save left-hand 'c' for next iteration */
7081 if ((c = *p++) == '\0')
7093 if ((c = *p++) == '\0')
7109 p++; /* consecutive *'s are identical to a single one */
7114 for (s = olds; *s != '\0'; s++) {
7115 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7125 dtrace_match_string(const char *s, const char *p, int depth)
7127 return (s != NULL && strcmp(s, p) == 0);
7132 dtrace_match_nul(const char *s, const char *p, int depth)
7134 return (1); /* always match the empty pattern */
7139 dtrace_match_nonzero(const char *s, const char *p, int depth)
7141 return (s != NULL && s[0] != '\0');
7145 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7146 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7148 dtrace_probe_t template, *probe;
7149 dtrace_hash_t *hash = NULL;
7150 int len, best = INT_MAX, nmatched = 0;
7153 ASSERT(MUTEX_HELD(&dtrace_lock));
7156 * If the probe ID is specified in the key, just lookup by ID and
7157 * invoke the match callback once if a matching probe is found.
7159 if (pkp->dtpk_id != DTRACE_IDNONE) {
7160 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7161 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7162 (void) (*matched)(probe, arg);
7168 template.dtpr_mod = (char *)pkp->dtpk_mod;
7169 template.dtpr_func = (char *)pkp->dtpk_func;
7170 template.dtpr_name = (char *)pkp->dtpk_name;
7173 * We want to find the most distinct of the module name, function
7174 * name, and name. So for each one that is not a glob pattern or
7175 * empty string, we perform a lookup in the corresponding hash and
7176 * use the hash table with the fewest collisions to do our search.
7178 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7179 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7181 hash = dtrace_bymod;
7184 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7185 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7187 hash = dtrace_byfunc;
7190 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7191 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7193 hash = dtrace_byname;
7197 * If we did not select a hash table, iterate over every probe and
7198 * invoke our callback for each one that matches our input probe key.
7201 for (i = 0; i < dtrace_nprobes; i++) {
7202 if ((probe = dtrace_probes[i]) == NULL ||
7203 dtrace_match_probe(probe, pkp, priv, uid,
7209 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7217 * If we selected a hash table, iterate over each probe of the same key
7218 * name and invoke the callback for every probe that matches the other
7219 * attributes of our input probe key.
7221 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7222 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7224 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7229 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7237 * Return the function pointer dtrace_probecmp() should use to compare the
7238 * specified pattern with a string. For NULL or empty patterns, we select
7239 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7240 * For non-empty non-glob strings, we use dtrace_match_string().
7242 static dtrace_probekey_f *
7243 dtrace_probekey_func(const char *p)
7247 if (p == NULL || *p == '\0')
7248 return (&dtrace_match_nul);
7250 while ((c = *p++) != '\0') {
7251 if (c == '[' || c == '?' || c == '*' || c == '\\')
7252 return (&dtrace_match_glob);
7255 return (&dtrace_match_string);
7259 * Build a probe comparison key for use with dtrace_match_probe() from the
7260 * given probe description. By convention, a null key only matches anchored
7261 * probes: if each field is the empty string, reset dtpk_fmatch to
7262 * dtrace_match_nonzero().
7265 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7267 pkp->dtpk_prov = pdp->dtpd_provider;
7268 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7270 pkp->dtpk_mod = pdp->dtpd_mod;
7271 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7273 pkp->dtpk_func = pdp->dtpd_func;
7274 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7276 pkp->dtpk_name = pdp->dtpd_name;
7277 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7279 pkp->dtpk_id = pdp->dtpd_id;
7281 if (pkp->dtpk_id == DTRACE_IDNONE &&
7282 pkp->dtpk_pmatch == &dtrace_match_nul &&
7283 pkp->dtpk_mmatch == &dtrace_match_nul &&
7284 pkp->dtpk_fmatch == &dtrace_match_nul &&
7285 pkp->dtpk_nmatch == &dtrace_match_nul)
7286 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7290 * DTrace Provider-to-Framework API Functions
7292 * These functions implement much of the Provider-to-Framework API, as
7293 * described in <sys/dtrace.h>. The parts of the API not in this section are
7294 * the functions in the API for probe management (found below), and
7295 * dtrace_probe() itself (found above).
7299 * Register the calling provider with the DTrace framework. This should
7300 * generally be called by DTrace providers in their attach(9E) entry point.
7303 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7304 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7306 dtrace_provider_t *provider;
7308 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7309 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7310 "arguments", name ? name : "<NULL>");
7314 if (name[0] == '\0' || dtrace_badname(name)) {
7315 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7316 "provider name", name);
7320 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7321 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7322 pops->dtps_destroy == NULL ||
7323 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7324 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7325 "provider ops", name);
7329 if (dtrace_badattr(&pap->dtpa_provider) ||
7330 dtrace_badattr(&pap->dtpa_mod) ||
7331 dtrace_badattr(&pap->dtpa_func) ||
7332 dtrace_badattr(&pap->dtpa_name) ||
7333 dtrace_badattr(&pap->dtpa_args)) {
7334 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7335 "provider attributes", name);
7339 if (priv & ~DTRACE_PRIV_ALL) {
7340 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7341 "privilege attributes", name);
7345 if ((priv & DTRACE_PRIV_KERNEL) &&
7346 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7347 pops->dtps_usermode == NULL) {
7348 cmn_err(CE_WARN, "failed to register provider '%s': need "
7349 "dtps_usermode() op for given privilege attributes", name);
7353 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7354 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7355 (void) strcpy(provider->dtpv_name, name);
7357 provider->dtpv_attr = *pap;
7358 provider->dtpv_priv.dtpp_flags = priv;
7360 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7361 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7363 provider->dtpv_pops = *pops;
7365 if (pops->dtps_provide == NULL) {
7366 ASSERT(pops->dtps_provide_module != NULL);
7367 provider->dtpv_pops.dtps_provide =
7368 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7371 if (pops->dtps_provide_module == NULL) {
7372 ASSERT(pops->dtps_provide != NULL);
7373 provider->dtpv_pops.dtps_provide_module =
7374 (void (*)(void *, modctl_t *))dtrace_nullop;
7377 if (pops->dtps_suspend == NULL) {
7378 ASSERT(pops->dtps_resume == NULL);
7379 provider->dtpv_pops.dtps_suspend =
7380 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7381 provider->dtpv_pops.dtps_resume =
7382 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7385 provider->dtpv_arg = arg;
7386 *idp = (dtrace_provider_id_t)provider;
7388 if (pops == &dtrace_provider_ops) {
7389 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7390 ASSERT(MUTEX_HELD(&dtrace_lock));
7391 ASSERT(dtrace_anon.dta_enabling == NULL);
7394 * We make sure that the DTrace provider is at the head of
7395 * the provider chain.
7397 provider->dtpv_next = dtrace_provider;
7398 dtrace_provider = provider;
7402 mutex_enter(&dtrace_provider_lock);
7403 mutex_enter(&dtrace_lock);
7406 * If there is at least one provider registered, we'll add this
7407 * provider after the first provider.
7409 if (dtrace_provider != NULL) {
7410 provider->dtpv_next = dtrace_provider->dtpv_next;
7411 dtrace_provider->dtpv_next = provider;
7413 dtrace_provider = provider;
7416 if (dtrace_retained != NULL) {
7417 dtrace_enabling_provide(provider);
7420 * Now we need to call dtrace_enabling_matchall() -- which
7421 * will acquire cpu_lock and dtrace_lock. We therefore need
7422 * to drop all of our locks before calling into it...
7424 mutex_exit(&dtrace_lock);
7425 mutex_exit(&dtrace_provider_lock);
7426 dtrace_enabling_matchall();
7431 mutex_exit(&dtrace_lock);
7432 mutex_exit(&dtrace_provider_lock);
7438 * Unregister the specified provider from the DTrace framework. This should
7439 * generally be called by DTrace providers in their detach(9E) entry point.
7442 dtrace_unregister(dtrace_provider_id_t id)
7444 dtrace_provider_t *old = (dtrace_provider_t *)id;
7445 dtrace_provider_t *prev = NULL;
7447 dtrace_probe_t *probe, *first = NULL;
7449 if (old->dtpv_pops.dtps_enable ==
7450 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7452 * If DTrace itself is the provider, we're called with locks
7455 ASSERT(old == dtrace_provider);
7457 ASSERT(dtrace_devi != NULL);
7459 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7460 ASSERT(MUTEX_HELD(&dtrace_lock));
7463 if (dtrace_provider->dtpv_next != NULL) {
7465 * There's another provider here; return failure.
7470 mutex_enter(&dtrace_provider_lock);
7471 mutex_enter(&mod_lock);
7472 mutex_enter(&dtrace_lock);
7476 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7477 * probes, we refuse to let providers slither away, unless this
7478 * provider has already been explicitly invalidated.
7480 if (!old->dtpv_defunct &&
7481 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7482 dtrace_anon.dta_state->dts_necbs > 0))) {
7484 mutex_exit(&dtrace_lock);
7485 mutex_exit(&mod_lock);
7486 mutex_exit(&dtrace_provider_lock);
7492 * Attempt to destroy the probes associated with this provider.
7494 for (i = 0; i < dtrace_nprobes; i++) {
7495 if ((probe = dtrace_probes[i]) == NULL)
7498 if (probe->dtpr_provider != old)
7501 if (probe->dtpr_ecb == NULL)
7505 * We have at least one ECB; we can't remove this provider.
7508 mutex_exit(&dtrace_lock);
7509 mutex_exit(&mod_lock);
7510 mutex_exit(&dtrace_provider_lock);
7516 * All of the probes for this provider are disabled; we can safely
7517 * remove all of them from their hash chains and from the probe array.
7519 for (i = 0; i < dtrace_nprobes; i++) {
7520 if ((probe = dtrace_probes[i]) == NULL)
7523 if (probe->dtpr_provider != old)
7526 dtrace_probes[i] = NULL;
7528 dtrace_hash_remove(dtrace_bymod, probe);
7529 dtrace_hash_remove(dtrace_byfunc, probe);
7530 dtrace_hash_remove(dtrace_byname, probe);
7532 if (first == NULL) {
7534 probe->dtpr_nextmod = NULL;
7536 probe->dtpr_nextmod = first;
7542 * The provider's probes have been removed from the hash chains and
7543 * from the probe array. Now issue a dtrace_sync() to be sure that
7544 * everyone has cleared out from any probe array processing.
7548 for (probe = first; probe != NULL; probe = first) {
7549 first = probe->dtpr_nextmod;
7551 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7553 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7554 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7555 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7557 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7559 free_unr(dtrace_arena, probe->dtpr_id);
7561 kmem_free(probe, sizeof (dtrace_probe_t));
7564 if ((prev = dtrace_provider) == old) {
7566 ASSERT(self || dtrace_devi == NULL);
7567 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7569 dtrace_provider = old->dtpv_next;
7571 while (prev != NULL && prev->dtpv_next != old)
7572 prev = prev->dtpv_next;
7575 panic("attempt to unregister non-existent "
7576 "dtrace provider %p\n", (void *)id);
7579 prev->dtpv_next = old->dtpv_next;
7583 mutex_exit(&dtrace_lock);
7584 mutex_exit(&mod_lock);
7585 mutex_exit(&dtrace_provider_lock);
7588 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7589 kmem_free(old, sizeof (dtrace_provider_t));
7595 * Invalidate the specified provider. All subsequent probe lookups for the
7596 * specified provider will fail, but its probes will not be removed.
7599 dtrace_invalidate(dtrace_provider_id_t id)
7601 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7603 ASSERT(pvp->dtpv_pops.dtps_enable !=
7604 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7606 mutex_enter(&dtrace_provider_lock);
7607 mutex_enter(&dtrace_lock);
7609 pvp->dtpv_defunct = 1;
7611 mutex_exit(&dtrace_lock);
7612 mutex_exit(&dtrace_provider_lock);
7616 * Indicate whether or not DTrace has attached.
7619 dtrace_attached(void)
7622 * dtrace_provider will be non-NULL iff the DTrace driver has
7623 * attached. (It's non-NULL because DTrace is always itself a
7626 return (dtrace_provider != NULL);
7630 * Remove all the unenabled probes for the given provider. This function is
7631 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7632 * -- just as many of its associated probes as it can.
7635 dtrace_condense(dtrace_provider_id_t id)
7637 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7639 dtrace_probe_t *probe;
7642 * Make sure this isn't the dtrace provider itself.
7644 ASSERT(prov->dtpv_pops.dtps_enable !=
7645 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7647 mutex_enter(&dtrace_provider_lock);
7648 mutex_enter(&dtrace_lock);
7651 * Attempt to destroy the probes associated with this provider.
7653 for (i = 0; i < dtrace_nprobes; i++) {
7654 if ((probe = dtrace_probes[i]) == NULL)
7657 if (probe->dtpr_provider != prov)
7660 if (probe->dtpr_ecb != NULL)
7663 dtrace_probes[i] = NULL;
7665 dtrace_hash_remove(dtrace_bymod, probe);
7666 dtrace_hash_remove(dtrace_byfunc, probe);
7667 dtrace_hash_remove(dtrace_byname, probe);
7669 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7671 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7672 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7673 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7674 kmem_free(probe, sizeof (dtrace_probe_t));
7676 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7678 free_unr(dtrace_arena, i + 1);
7682 mutex_exit(&dtrace_lock);
7683 mutex_exit(&dtrace_provider_lock);
7689 * DTrace Probe Management Functions
7691 * The functions in this section perform the DTrace probe management,
7692 * including functions to create probes, look-up probes, and call into the
7693 * providers to request that probes be provided. Some of these functions are
7694 * in the Provider-to-Framework API; these functions can be identified by the
7695 * fact that they are not declared "static".
7699 * Create a probe with the specified module name, function name, and name.
7702 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7703 const char *func, const char *name, int aframes, void *arg)
7705 dtrace_probe_t *probe, **probes;
7706 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7709 if (provider == dtrace_provider) {
7710 ASSERT(MUTEX_HELD(&dtrace_lock));
7712 mutex_enter(&dtrace_lock);
7716 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7717 VM_BESTFIT | VM_SLEEP);
7719 id = alloc_unr(dtrace_arena);
7721 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7723 probe->dtpr_id = id;
7724 probe->dtpr_gen = dtrace_probegen++;
7725 probe->dtpr_mod = dtrace_strdup(mod);
7726 probe->dtpr_func = dtrace_strdup(func);
7727 probe->dtpr_name = dtrace_strdup(name);
7728 probe->dtpr_arg = arg;
7729 probe->dtpr_aframes = aframes;
7730 probe->dtpr_provider = provider;
7732 dtrace_hash_add(dtrace_bymod, probe);
7733 dtrace_hash_add(dtrace_byfunc, probe);
7734 dtrace_hash_add(dtrace_byname, probe);
7736 if (id - 1 >= dtrace_nprobes) {
7737 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7738 size_t nsize = osize << 1;
7742 ASSERT(dtrace_probes == NULL);
7743 nsize = sizeof (dtrace_probe_t *);
7746 probes = kmem_zalloc(nsize, KM_SLEEP);
7748 if (dtrace_probes == NULL) {
7750 dtrace_probes = probes;
7753 dtrace_probe_t **oprobes = dtrace_probes;
7755 bcopy(oprobes, probes, osize);
7756 dtrace_membar_producer();
7757 dtrace_probes = probes;
7762 * All CPUs are now seeing the new probes array; we can
7763 * safely free the old array.
7765 kmem_free(oprobes, osize);
7766 dtrace_nprobes <<= 1;
7769 ASSERT(id - 1 < dtrace_nprobes);
7772 ASSERT(dtrace_probes[id - 1] == NULL);
7773 dtrace_probes[id - 1] = probe;
7775 if (provider != dtrace_provider)
7776 mutex_exit(&dtrace_lock);
7781 static dtrace_probe_t *
7782 dtrace_probe_lookup_id(dtrace_id_t id)
7784 ASSERT(MUTEX_HELD(&dtrace_lock));
7786 if (id == 0 || id > dtrace_nprobes)
7789 return (dtrace_probes[id - 1]);
7793 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7795 *((dtrace_id_t *)arg) = probe->dtpr_id;
7797 return (DTRACE_MATCH_DONE);
7801 * Look up a probe based on provider and one or more of module name, function
7802 * name and probe name.
7805 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
7806 char *func, char *name)
7808 dtrace_probekey_t pkey;
7812 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7813 pkey.dtpk_pmatch = &dtrace_match_string;
7814 pkey.dtpk_mod = mod;
7815 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7816 pkey.dtpk_func = func;
7817 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7818 pkey.dtpk_name = name;
7819 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7820 pkey.dtpk_id = DTRACE_IDNONE;
7822 mutex_enter(&dtrace_lock);
7823 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7824 dtrace_probe_lookup_match, &id);
7825 mutex_exit(&dtrace_lock);
7827 ASSERT(match == 1 || match == 0);
7828 return (match ? id : 0);
7832 * Returns the probe argument associated with the specified probe.
7835 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7837 dtrace_probe_t *probe;
7840 mutex_enter(&dtrace_lock);
7842 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7843 probe->dtpr_provider == (dtrace_provider_t *)id)
7844 rval = probe->dtpr_arg;
7846 mutex_exit(&dtrace_lock);
7852 * Copy a probe into a probe description.
7855 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7857 bzero(pdp, sizeof (dtrace_probedesc_t));
7858 pdp->dtpd_id = prp->dtpr_id;
7860 (void) strncpy(pdp->dtpd_provider,
7861 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7863 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7864 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7865 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7870 dtrace_probe_provide_cb(linker_file_t lf, void *arg)
7872 dtrace_provider_t *prv = (dtrace_provider_t *) arg;
7874 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, lf);
7882 * Called to indicate that a probe -- or probes -- should be provided by a
7883 * specfied provider. If the specified description is NULL, the provider will
7884 * be told to provide all of its probes. (This is done whenever a new
7885 * consumer comes along, or whenever a retained enabling is to be matched.) If
7886 * the specified description is non-NULL, the provider is given the
7887 * opportunity to dynamically provide the specified probe, allowing providers
7888 * to support the creation of probes on-the-fly. (So-called _autocreated_
7889 * probes.) If the provider is NULL, the operations will be applied to all
7890 * providers; if the provider is non-NULL the operations will only be applied
7891 * to the specified provider. The dtrace_provider_lock must be held, and the
7892 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7893 * will need to grab the dtrace_lock when it reenters the framework through
7894 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7897 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7904 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7908 prv = dtrace_provider;
7913 * First, call the blanket provide operation.
7915 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7918 * Now call the per-module provide operation. We will grab
7919 * mod_lock to prevent the list from being modified. Note
7920 * that this also prevents the mod_busy bits from changing.
7921 * (mod_busy can only be changed with mod_lock held.)
7923 mutex_enter(&mod_lock);
7928 if (ctl->mod_busy || ctl->mod_mp == NULL)
7931 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
7933 } while ((ctl = ctl->mod_next) != &modules);
7935 (void) linker_file_foreach(dtrace_probe_provide_cb, prv);
7938 mutex_exit(&mod_lock);
7939 } while (all && (prv = prv->dtpv_next) != NULL);
7944 * Iterate over each probe, and call the Framework-to-Provider API function
7948 dtrace_probe_foreach(uintptr_t offs)
7950 dtrace_provider_t *prov;
7951 void (*func)(void *, dtrace_id_t, void *);
7952 dtrace_probe_t *probe;
7953 dtrace_icookie_t cookie;
7957 * We disable interrupts to walk through the probe array. This is
7958 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
7959 * won't see stale data.
7961 cookie = dtrace_interrupt_disable();
7963 for (i = 0; i < dtrace_nprobes; i++) {
7964 if ((probe = dtrace_probes[i]) == NULL)
7967 if (probe->dtpr_ecb == NULL) {
7969 * This probe isn't enabled -- don't call the function.
7974 prov = probe->dtpr_provider;
7975 func = *((void(**)(void *, dtrace_id_t, void *))
7976 ((uintptr_t)&prov->dtpv_pops + offs));
7978 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
7981 dtrace_interrupt_enable(cookie);
7986 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
7988 dtrace_probekey_t pkey;
7993 ASSERT(MUTEX_HELD(&dtrace_lock));
7994 dtrace_ecb_create_cache = NULL;
7998 * If we're passed a NULL description, we're being asked to
7999 * create an ECB with a NULL probe.
8001 (void) dtrace_ecb_create_enable(NULL, enab);
8005 dtrace_probekey(desc, &pkey);
8006 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8007 &priv, &uid, &zoneid);
8009 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8014 * DTrace Helper Provider Functions
8017 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8019 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8020 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8021 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8025 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8026 const dof_provider_t *dofprov, char *strtab)
8028 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8029 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8030 dofprov->dofpv_provattr);
8031 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8032 dofprov->dofpv_modattr);
8033 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8034 dofprov->dofpv_funcattr);
8035 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8036 dofprov->dofpv_nameattr);
8037 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8038 dofprov->dofpv_argsattr);
8042 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8044 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8045 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8046 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8047 dof_provider_t *provider;
8049 uint32_t *off, *enoff;
8053 dtrace_helper_provdesc_t dhpv;
8054 dtrace_helper_probedesc_t dhpb;
8055 dtrace_meta_t *meta = dtrace_meta_pid;
8056 dtrace_mops_t *mops = &meta->dtm_mops;
8059 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8060 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8061 provider->dofpv_strtab * dof->dofh_secsize);
8062 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8063 provider->dofpv_probes * dof->dofh_secsize);
8064 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8065 provider->dofpv_prargs * dof->dofh_secsize);
8066 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8067 provider->dofpv_proffs * dof->dofh_secsize);
8069 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8070 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8071 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8075 * See dtrace_helper_provider_validate().
8077 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8078 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8079 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8080 provider->dofpv_prenoffs * dof->dofh_secsize);
8081 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8084 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8087 * Create the provider.
8089 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8091 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8097 * Create the probes.
8099 for (i = 0; i < nprobes; i++) {
8100 probe = (dof_probe_t *)(uintptr_t)(daddr +
8101 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8103 dhpb.dthpb_mod = dhp->dofhp_mod;
8104 dhpb.dthpb_func = strtab + probe->dofpr_func;
8105 dhpb.dthpb_name = strtab + probe->dofpr_name;
8106 dhpb.dthpb_base = probe->dofpr_addr;
8107 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8108 dhpb.dthpb_noffs = probe->dofpr_noffs;
8109 if (enoff != NULL) {
8110 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8111 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8113 dhpb.dthpb_enoffs = NULL;
8114 dhpb.dthpb_nenoffs = 0;
8116 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8117 dhpb.dthpb_nargc = probe->dofpr_nargc;
8118 dhpb.dthpb_xargc = probe->dofpr_xargc;
8119 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8120 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8122 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8127 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8129 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8130 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8133 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8135 for (i = 0; i < dof->dofh_secnum; i++) {
8136 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8137 dof->dofh_secoff + i * dof->dofh_secsize);
8139 if (sec->dofs_type != DOF_SECT_PROVIDER)
8142 dtrace_helper_provide_one(dhp, sec, pid);
8146 * We may have just created probes, so we must now rematch against
8147 * any retained enablings. Note that this call will acquire both
8148 * cpu_lock and dtrace_lock; the fact that we are holding
8149 * dtrace_meta_lock now is what defines the ordering with respect to
8150 * these three locks.
8152 dtrace_enabling_matchall();
8156 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8158 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8159 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8161 dof_provider_t *provider;
8163 dtrace_helper_provdesc_t dhpv;
8164 dtrace_meta_t *meta = dtrace_meta_pid;
8165 dtrace_mops_t *mops = &meta->dtm_mops;
8167 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8168 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8169 provider->dofpv_strtab * dof->dofh_secsize);
8171 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8174 * Create the provider.
8176 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8178 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8184 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8186 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8187 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8190 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8192 for (i = 0; i < dof->dofh_secnum; i++) {
8193 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8194 dof->dofh_secoff + i * dof->dofh_secsize);
8196 if (sec->dofs_type != DOF_SECT_PROVIDER)
8199 dtrace_helper_provider_remove_one(dhp, sec, pid);
8204 * DTrace Meta Provider-to-Framework API Functions
8206 * These functions implement the Meta Provider-to-Framework API, as described
8207 * in <sys/dtrace.h>.
8210 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8211 dtrace_meta_provider_id_t *idp)
8213 dtrace_meta_t *meta;
8214 dtrace_helpers_t *help, *next;
8217 *idp = DTRACE_METAPROVNONE;
8220 * We strictly don't need the name, but we hold onto it for
8221 * debuggability. All hail error queues!
8224 cmn_err(CE_WARN, "failed to register meta-provider: "
8230 mops->dtms_create_probe == NULL ||
8231 mops->dtms_provide_pid == NULL ||
8232 mops->dtms_remove_pid == NULL) {
8233 cmn_err(CE_WARN, "failed to register meta-register %s: "
8234 "invalid ops", name);
8238 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8239 meta->dtm_mops = *mops;
8240 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8241 (void) strcpy(meta->dtm_name, name);
8242 meta->dtm_arg = arg;
8244 mutex_enter(&dtrace_meta_lock);
8245 mutex_enter(&dtrace_lock);
8247 if (dtrace_meta_pid != NULL) {
8248 mutex_exit(&dtrace_lock);
8249 mutex_exit(&dtrace_meta_lock);
8250 cmn_err(CE_WARN, "failed to register meta-register %s: "
8251 "user-land meta-provider exists", name);
8252 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8253 kmem_free(meta, sizeof (dtrace_meta_t));
8257 dtrace_meta_pid = meta;
8258 *idp = (dtrace_meta_provider_id_t)meta;
8261 * If there are providers and probes ready to go, pass them
8262 * off to the new meta provider now.
8265 help = dtrace_deferred_pid;
8266 dtrace_deferred_pid = NULL;
8268 mutex_exit(&dtrace_lock);
8270 while (help != NULL) {
8271 for (i = 0; i < help->dthps_nprovs; i++) {
8272 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8276 next = help->dthps_next;
8277 help->dthps_next = NULL;
8278 help->dthps_prev = NULL;
8279 help->dthps_deferred = 0;
8283 mutex_exit(&dtrace_meta_lock);
8289 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8291 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8293 mutex_enter(&dtrace_meta_lock);
8294 mutex_enter(&dtrace_lock);
8296 if (old == dtrace_meta_pid) {
8297 pp = &dtrace_meta_pid;
8299 panic("attempt to unregister non-existent "
8300 "dtrace meta-provider %p\n", (void *)old);
8303 if (old->dtm_count != 0) {
8304 mutex_exit(&dtrace_lock);
8305 mutex_exit(&dtrace_meta_lock);
8311 mutex_exit(&dtrace_lock);
8312 mutex_exit(&dtrace_meta_lock);
8314 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8315 kmem_free(old, sizeof (dtrace_meta_t));
8322 * DTrace DIF Object Functions
8325 dtrace_difo_err(uint_t pc, const char *format, ...)
8327 if (dtrace_err_verbose) {
8330 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8331 va_start(alist, format);
8332 (void) vuprintf(format, alist);
8336 #ifdef DTRACE_ERRDEBUG
8337 dtrace_errdebug(format);
8343 * Validate a DTrace DIF object by checking the IR instructions. The following
8344 * rules are currently enforced by dtrace_difo_validate():
8346 * 1. Each instruction must have a valid opcode
8347 * 2. Each register, string, variable, or subroutine reference must be valid
8348 * 3. No instruction can modify register %r0 (must be zero)
8349 * 4. All instruction reserved bits must be set to zero
8350 * 5. The last instruction must be a "ret" instruction
8351 * 6. All branch targets must reference a valid instruction _after_ the branch
8354 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8358 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8362 kcheckload = cr == NULL ||
8363 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8365 dp->dtdo_destructive = 0;
8367 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8368 dif_instr_t instr = dp->dtdo_buf[pc];
8370 uint_t r1 = DIF_INSTR_R1(instr);
8371 uint_t r2 = DIF_INSTR_R2(instr);
8372 uint_t rd = DIF_INSTR_RD(instr);
8373 uint_t rs = DIF_INSTR_RS(instr);
8374 uint_t label = DIF_INSTR_LABEL(instr);
8375 uint_t v = DIF_INSTR_VAR(instr);
8376 uint_t subr = DIF_INSTR_SUBR(instr);
8377 uint_t type = DIF_INSTR_TYPE(instr);
8378 uint_t op = DIF_INSTR_OP(instr);
8396 err += efunc(pc, "invalid register %u\n", r1);
8398 err += efunc(pc, "invalid register %u\n", r2);
8400 err += efunc(pc, "invalid register %u\n", rd);
8402 err += efunc(pc, "cannot write to %r0\n");
8408 err += efunc(pc, "invalid register %u\n", r1);
8410 err += efunc(pc, "non-zero reserved bits\n");
8412 err += efunc(pc, "invalid register %u\n", rd);
8414 err += efunc(pc, "cannot write to %r0\n");
8424 err += efunc(pc, "invalid register %u\n", r1);
8426 err += efunc(pc, "non-zero reserved bits\n");
8428 err += efunc(pc, "invalid register %u\n", rd);
8430 err += efunc(pc, "cannot write to %r0\n");
8432 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8433 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8443 err += efunc(pc, "invalid register %u\n", r1);
8445 err += efunc(pc, "non-zero reserved bits\n");
8447 err += efunc(pc, "invalid register %u\n", rd);
8449 err += efunc(pc, "cannot write to %r0\n");
8459 err += efunc(pc, "invalid register %u\n", r1);
8461 err += efunc(pc, "non-zero reserved bits\n");
8463 err += efunc(pc, "invalid register %u\n", rd);
8465 err += efunc(pc, "cannot write to %r0\n");
8472 err += efunc(pc, "invalid register %u\n", r1);
8474 err += efunc(pc, "non-zero reserved bits\n");
8476 err += efunc(pc, "invalid register %u\n", rd);
8478 err += efunc(pc, "cannot write to 0 address\n");
8483 err += efunc(pc, "invalid register %u\n", r1);
8485 err += efunc(pc, "invalid register %u\n", r2);
8487 err += efunc(pc, "non-zero reserved bits\n");
8491 err += efunc(pc, "invalid register %u\n", r1);
8492 if (r2 != 0 || rd != 0)
8493 err += efunc(pc, "non-zero reserved bits\n");
8506 if (label >= dp->dtdo_len) {
8507 err += efunc(pc, "invalid branch target %u\n",
8511 err += efunc(pc, "backward branch to %u\n",
8516 if (r1 != 0 || r2 != 0)
8517 err += efunc(pc, "non-zero reserved bits\n");
8519 err += efunc(pc, "invalid register %u\n", rd);
8523 case DIF_OP_FLUSHTS:
8524 if (r1 != 0 || r2 != 0 || rd != 0)
8525 err += efunc(pc, "non-zero reserved bits\n");
8528 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8529 err += efunc(pc, "invalid integer ref %u\n",
8530 DIF_INSTR_INTEGER(instr));
8533 err += efunc(pc, "invalid register %u\n", rd);
8535 err += efunc(pc, "cannot write to %r0\n");
8538 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8539 err += efunc(pc, "invalid string ref %u\n",
8540 DIF_INSTR_STRING(instr));
8543 err += efunc(pc, "invalid register %u\n", rd);
8545 err += efunc(pc, "cannot write to %r0\n");
8549 if (r1 > DIF_VAR_ARRAY_MAX)
8550 err += efunc(pc, "invalid array %u\n", r1);
8552 err += efunc(pc, "invalid register %u\n", r2);
8554 err += efunc(pc, "invalid register %u\n", rd);
8556 err += efunc(pc, "cannot write to %r0\n");
8563 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8564 err += efunc(pc, "invalid variable %u\n", v);
8566 err += efunc(pc, "invalid register %u\n", rd);
8568 err += efunc(pc, "cannot write to %r0\n");
8575 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8576 err += efunc(pc, "invalid variable %u\n", v);
8578 err += efunc(pc, "invalid register %u\n", rd);
8581 if (subr > DIF_SUBR_MAX)
8582 err += efunc(pc, "invalid subr %u\n", subr);
8584 err += efunc(pc, "invalid register %u\n", rd);
8586 err += efunc(pc, "cannot write to %r0\n");
8588 if (subr == DIF_SUBR_COPYOUT ||
8589 subr == DIF_SUBR_COPYOUTSTR) {
8590 dp->dtdo_destructive = 1;
8594 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8595 err += efunc(pc, "invalid ref type %u\n", type);
8597 err += efunc(pc, "invalid register %u\n", r2);
8599 err += efunc(pc, "invalid register %u\n", rs);
8602 if (type != DIF_TYPE_CTF)
8603 err += efunc(pc, "invalid val type %u\n", type);
8605 err += efunc(pc, "invalid register %u\n", r2);
8607 err += efunc(pc, "invalid register %u\n", rs);
8610 err += efunc(pc, "invalid opcode %u\n",
8611 DIF_INSTR_OP(instr));
8615 if (dp->dtdo_len != 0 &&
8616 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8617 err += efunc(dp->dtdo_len - 1,
8618 "expected 'ret' as last DIF instruction\n");
8621 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8623 * If we're not returning by reference, the size must be either
8624 * 0 or the size of one of the base types.
8626 switch (dp->dtdo_rtype.dtdt_size) {
8628 case sizeof (uint8_t):
8629 case sizeof (uint16_t):
8630 case sizeof (uint32_t):
8631 case sizeof (uint64_t):
8635 err += efunc(dp->dtdo_len - 1, "bad return size");
8639 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8640 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8641 dtrace_diftype_t *vt, *et;
8644 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8645 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8646 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8647 err += efunc(i, "unrecognized variable scope %d\n",
8652 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8653 v->dtdv_kind != DIFV_KIND_SCALAR) {
8654 err += efunc(i, "unrecognized variable type %d\n",
8659 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8660 err += efunc(i, "%d exceeds variable id limit\n", id);
8664 if (id < DIF_VAR_OTHER_UBASE)
8668 * For user-defined variables, we need to check that this
8669 * definition is identical to any previous definition that we
8672 ndx = id - DIF_VAR_OTHER_UBASE;
8674 switch (v->dtdv_scope) {
8675 case DIFV_SCOPE_GLOBAL:
8676 if (ndx < vstate->dtvs_nglobals) {
8677 dtrace_statvar_t *svar;
8679 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8680 existing = &svar->dtsv_var;
8685 case DIFV_SCOPE_THREAD:
8686 if (ndx < vstate->dtvs_ntlocals)
8687 existing = &vstate->dtvs_tlocals[ndx];
8690 case DIFV_SCOPE_LOCAL:
8691 if (ndx < vstate->dtvs_nlocals) {
8692 dtrace_statvar_t *svar;
8694 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8695 existing = &svar->dtsv_var;
8703 if (vt->dtdt_flags & DIF_TF_BYREF) {
8704 if (vt->dtdt_size == 0) {
8705 err += efunc(i, "zero-sized variable\n");
8709 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8710 vt->dtdt_size > dtrace_global_maxsize) {
8711 err += efunc(i, "oversized by-ref global\n");
8716 if (existing == NULL || existing->dtdv_id == 0)
8719 ASSERT(existing->dtdv_id == v->dtdv_id);
8720 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8722 if (existing->dtdv_kind != v->dtdv_kind)
8723 err += efunc(i, "%d changed variable kind\n", id);
8725 et = &existing->dtdv_type;
8727 if (vt->dtdt_flags != et->dtdt_flags) {
8728 err += efunc(i, "%d changed variable type flags\n", id);
8732 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8733 err += efunc(i, "%d changed variable type size\n", id);
8742 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
8743 * are much more constrained than normal DIFOs. Specifically, they may
8746 * 1. Make calls to subroutines other than copyin(), copyinstr() or
8747 * miscellaneous string routines
8748 * 2. Access DTrace variables other than the args[] array, and the
8749 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8750 * 3. Have thread-local variables.
8751 * 4. Have dynamic variables.
8754 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8756 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8760 for (pc = 0; pc < dp->dtdo_len; pc++) {
8761 dif_instr_t instr = dp->dtdo_buf[pc];
8763 uint_t v = DIF_INSTR_VAR(instr);
8764 uint_t subr = DIF_INSTR_SUBR(instr);
8765 uint_t op = DIF_INSTR_OP(instr);
8820 case DIF_OP_FLUSHTS:
8832 if (v >= DIF_VAR_OTHER_UBASE)
8835 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8838 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8839 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8840 v == DIF_VAR_EXECARGS ||
8841 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8842 v == DIF_VAR_UID || v == DIF_VAR_GID)
8845 err += efunc(pc, "illegal variable %u\n", v);
8852 err += efunc(pc, "illegal dynamic variable load\n");
8858 err += efunc(pc, "illegal dynamic variable store\n");
8862 if (subr == DIF_SUBR_ALLOCA ||
8863 subr == DIF_SUBR_BCOPY ||
8864 subr == DIF_SUBR_COPYIN ||
8865 subr == DIF_SUBR_COPYINTO ||
8866 subr == DIF_SUBR_COPYINSTR ||
8867 subr == DIF_SUBR_INDEX ||
8868 subr == DIF_SUBR_INET_NTOA ||
8869 subr == DIF_SUBR_INET_NTOA6 ||
8870 subr == DIF_SUBR_INET_NTOP ||
8871 subr == DIF_SUBR_LLTOSTR ||
8872 subr == DIF_SUBR_RINDEX ||
8873 subr == DIF_SUBR_STRCHR ||
8874 subr == DIF_SUBR_STRJOIN ||
8875 subr == DIF_SUBR_STRRCHR ||
8876 subr == DIF_SUBR_STRSTR ||
8877 subr == DIF_SUBR_HTONS ||
8878 subr == DIF_SUBR_HTONL ||
8879 subr == DIF_SUBR_HTONLL ||
8880 subr == DIF_SUBR_NTOHS ||
8881 subr == DIF_SUBR_NTOHL ||
8882 subr == DIF_SUBR_NTOHLL ||
8883 subr == DIF_SUBR_MEMREF ||
8884 subr == DIF_SUBR_TYPEREF)
8887 err += efunc(pc, "invalid subr %u\n", subr);
8891 err += efunc(pc, "invalid opcode %u\n",
8892 DIF_INSTR_OP(instr));
8900 * Returns 1 if the expression in the DIF object can be cached on a per-thread
8904 dtrace_difo_cacheable(dtrace_difo_t *dp)
8911 for (i = 0; i < dp->dtdo_varlen; i++) {
8912 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8914 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8917 switch (v->dtdv_id) {
8918 case DIF_VAR_CURTHREAD:
8921 case DIF_VAR_EXECARGS:
8922 case DIF_VAR_EXECNAME:
8923 case DIF_VAR_ZONENAME:
8932 * This DIF object may be cacheable. Now we need to look for any
8933 * array loading instructions, any memory loading instructions, or
8934 * any stores to thread-local variables.
8936 for (i = 0; i < dp->dtdo_len; i++) {
8937 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
8939 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
8940 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
8941 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
8942 op == DIF_OP_LDGA || op == DIF_OP_STTS)
8950 dtrace_difo_hold(dtrace_difo_t *dp)
8954 ASSERT(MUTEX_HELD(&dtrace_lock));
8957 ASSERT(dp->dtdo_refcnt != 0);
8960 * We need to check this DIF object for references to the variable
8961 * DIF_VAR_VTIMESTAMP.
8963 for (i = 0; i < dp->dtdo_varlen; i++) {
8964 dtrace_difv_t *v = &dp->dtdo_vartab[i];
8966 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8969 if (dtrace_vtime_references++ == 0)
8970 dtrace_vtime_enable();
8975 * This routine calculates the dynamic variable chunksize for a given DIF
8976 * object. The calculation is not fool-proof, and can probably be tricked by
8977 * malicious DIF -- but it works for all compiler-generated DIF. Because this
8978 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
8979 * if a dynamic variable size exceeds the chunksize.
8982 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8985 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
8986 const dif_instr_t *text = dp->dtdo_buf;
8992 for (pc = 0; pc < dp->dtdo_len; pc++) {
8993 dif_instr_t instr = text[pc];
8994 uint_t op = DIF_INSTR_OP(instr);
8995 uint_t rd = DIF_INSTR_RD(instr);
8996 uint_t r1 = DIF_INSTR_R1(instr);
9000 dtrace_key_t *key = tupregs;
9004 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9009 key = &tupregs[DIF_DTR_NREGS];
9010 key[0].dttk_size = 0;
9011 key[1].dttk_size = 0;
9013 scope = DIFV_SCOPE_THREAD;
9020 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9021 key[nkeys++].dttk_size = 0;
9023 key[nkeys++].dttk_size = 0;
9025 if (op == DIF_OP_STTAA) {
9026 scope = DIFV_SCOPE_THREAD;
9028 scope = DIFV_SCOPE_GLOBAL;
9034 if (ttop == DIF_DTR_NREGS)
9037 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9039 * If the register for the size of the "pushtr"
9040 * is %r0 (or the value is 0) and the type is
9041 * a string, we'll use the system-wide default
9044 tupregs[ttop++].dttk_size =
9045 dtrace_strsize_default;
9050 tupregs[ttop++].dttk_size = sval;
9056 if (ttop == DIF_DTR_NREGS)
9059 tupregs[ttop++].dttk_size = 0;
9062 case DIF_OP_FLUSHTS:
9079 * We have a dynamic variable allocation; calculate its size.
9081 for (ksize = 0, i = 0; i < nkeys; i++)
9082 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9084 size = sizeof (dtrace_dynvar_t);
9085 size += sizeof (dtrace_key_t) * (nkeys - 1);
9089 * Now we need to determine the size of the stored data.
9091 id = DIF_INSTR_VAR(instr);
9093 for (i = 0; i < dp->dtdo_varlen; i++) {
9094 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9096 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9097 size += v->dtdv_type.dtdt_size;
9102 if (i == dp->dtdo_varlen)
9106 * We have the size. If this is larger than the chunk size
9107 * for our dynamic variable state, reset the chunk size.
9109 size = P2ROUNDUP(size, sizeof (uint64_t));
9111 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9112 vstate->dtvs_dynvars.dtds_chunksize = size;
9117 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9119 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9122 ASSERT(MUTEX_HELD(&dtrace_lock));
9123 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9125 for (i = 0; i < dp->dtdo_varlen; i++) {
9126 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9127 dtrace_statvar_t *svar, ***svarp = NULL;
9129 uint8_t scope = v->dtdv_scope;
9132 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9135 id -= DIF_VAR_OTHER_UBASE;
9138 case DIFV_SCOPE_THREAD:
9139 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9140 dtrace_difv_t *tlocals;
9142 if ((ntlocals = (otlocals << 1)) == 0)
9145 osz = otlocals * sizeof (dtrace_difv_t);
9146 nsz = ntlocals * sizeof (dtrace_difv_t);
9148 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9151 bcopy(vstate->dtvs_tlocals,
9153 kmem_free(vstate->dtvs_tlocals, osz);
9156 vstate->dtvs_tlocals = tlocals;
9157 vstate->dtvs_ntlocals = ntlocals;
9160 vstate->dtvs_tlocals[id] = *v;
9163 case DIFV_SCOPE_LOCAL:
9164 np = &vstate->dtvs_nlocals;
9165 svarp = &vstate->dtvs_locals;
9167 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9168 dsize = NCPU * (v->dtdv_type.dtdt_size +
9171 dsize = NCPU * sizeof (uint64_t);
9175 case DIFV_SCOPE_GLOBAL:
9176 np = &vstate->dtvs_nglobals;
9177 svarp = &vstate->dtvs_globals;
9179 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9180 dsize = v->dtdv_type.dtdt_size +
9189 while (id >= (oldsvars = *np)) {
9190 dtrace_statvar_t **statics;
9191 int newsvars, oldsize, newsize;
9193 if ((newsvars = (oldsvars << 1)) == 0)
9196 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9197 newsize = newsvars * sizeof (dtrace_statvar_t *);
9199 statics = kmem_zalloc(newsize, KM_SLEEP);
9202 bcopy(*svarp, statics, oldsize);
9203 kmem_free(*svarp, oldsize);
9210 if ((svar = (*svarp)[id]) == NULL) {
9211 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9212 svar->dtsv_var = *v;
9214 if ((svar->dtsv_size = dsize) != 0) {
9215 svar->dtsv_data = (uint64_t)(uintptr_t)
9216 kmem_zalloc(dsize, KM_SLEEP);
9219 (*svarp)[id] = svar;
9222 svar->dtsv_refcnt++;
9225 dtrace_difo_chunksize(dp, vstate);
9226 dtrace_difo_hold(dp);
9229 static dtrace_difo_t *
9230 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9235 ASSERT(dp->dtdo_buf != NULL);
9236 ASSERT(dp->dtdo_refcnt != 0);
9238 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9240 ASSERT(dp->dtdo_buf != NULL);
9241 sz = dp->dtdo_len * sizeof (dif_instr_t);
9242 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9243 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9244 new->dtdo_len = dp->dtdo_len;
9246 if (dp->dtdo_strtab != NULL) {
9247 ASSERT(dp->dtdo_strlen != 0);
9248 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9249 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9250 new->dtdo_strlen = dp->dtdo_strlen;
9253 if (dp->dtdo_inttab != NULL) {
9254 ASSERT(dp->dtdo_intlen != 0);
9255 sz = dp->dtdo_intlen * sizeof (uint64_t);
9256 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9257 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9258 new->dtdo_intlen = dp->dtdo_intlen;
9261 if (dp->dtdo_vartab != NULL) {
9262 ASSERT(dp->dtdo_varlen != 0);
9263 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9264 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9265 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9266 new->dtdo_varlen = dp->dtdo_varlen;
9269 dtrace_difo_init(new, vstate);
9274 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9278 ASSERT(dp->dtdo_refcnt == 0);
9280 for (i = 0; i < dp->dtdo_varlen; i++) {
9281 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9282 dtrace_statvar_t *svar, **svarp = NULL;
9284 uint8_t scope = v->dtdv_scope;
9288 case DIFV_SCOPE_THREAD:
9291 case DIFV_SCOPE_LOCAL:
9292 np = &vstate->dtvs_nlocals;
9293 svarp = vstate->dtvs_locals;
9296 case DIFV_SCOPE_GLOBAL:
9297 np = &vstate->dtvs_nglobals;
9298 svarp = vstate->dtvs_globals;
9305 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9308 id -= DIF_VAR_OTHER_UBASE;
9312 ASSERT(svar != NULL);
9313 ASSERT(svar->dtsv_refcnt > 0);
9315 if (--svar->dtsv_refcnt > 0)
9318 if (svar->dtsv_size != 0) {
9319 ASSERT(svar->dtsv_data != 0);
9320 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9324 kmem_free(svar, sizeof (dtrace_statvar_t));
9328 if (dp->dtdo_buf != NULL)
9329 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9330 if (dp->dtdo_inttab != NULL)
9331 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9332 if (dp->dtdo_strtab != NULL)
9333 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9334 if (dp->dtdo_vartab != NULL)
9335 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9337 kmem_free(dp, sizeof (dtrace_difo_t));
9341 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9345 ASSERT(MUTEX_HELD(&dtrace_lock));
9346 ASSERT(dp->dtdo_refcnt != 0);
9348 for (i = 0; i < dp->dtdo_varlen; i++) {
9349 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9351 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9354 ASSERT(dtrace_vtime_references > 0);
9355 if (--dtrace_vtime_references == 0)
9356 dtrace_vtime_disable();
9359 if (--dp->dtdo_refcnt == 0)
9360 dtrace_difo_destroy(dp, vstate);
9364 * DTrace Format Functions
9367 dtrace_format_add(dtrace_state_t *state, char *str)
9370 uint16_t ndx, len = strlen(str) + 1;
9372 fmt = kmem_zalloc(len, KM_SLEEP);
9373 bcopy(str, fmt, len);
9375 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9376 if (state->dts_formats[ndx] == NULL) {
9377 state->dts_formats[ndx] = fmt;
9382 if (state->dts_nformats == USHRT_MAX) {
9384 * This is only likely if a denial-of-service attack is being
9385 * attempted. As such, it's okay to fail silently here.
9387 kmem_free(fmt, len);
9392 * For simplicity, we always resize the formats array to be exactly the
9393 * number of formats.
9395 ndx = state->dts_nformats++;
9396 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9398 if (state->dts_formats != NULL) {
9400 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9401 kmem_free(state->dts_formats, ndx * sizeof (char *));
9404 state->dts_formats = new;
9405 state->dts_formats[ndx] = fmt;
9411 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9415 ASSERT(state->dts_formats != NULL);
9416 ASSERT(format <= state->dts_nformats);
9417 ASSERT(state->dts_formats[format - 1] != NULL);
9419 fmt = state->dts_formats[format - 1];
9420 kmem_free(fmt, strlen(fmt) + 1);
9421 state->dts_formats[format - 1] = NULL;
9425 dtrace_format_destroy(dtrace_state_t *state)
9429 if (state->dts_nformats == 0) {
9430 ASSERT(state->dts_formats == NULL);
9434 ASSERT(state->dts_formats != NULL);
9436 for (i = 0; i < state->dts_nformats; i++) {
9437 char *fmt = state->dts_formats[i];
9442 kmem_free(fmt, strlen(fmt) + 1);
9445 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9446 state->dts_nformats = 0;
9447 state->dts_formats = NULL;
9451 * DTrace Predicate Functions
9453 static dtrace_predicate_t *
9454 dtrace_predicate_create(dtrace_difo_t *dp)
9456 dtrace_predicate_t *pred;
9458 ASSERT(MUTEX_HELD(&dtrace_lock));
9459 ASSERT(dp->dtdo_refcnt != 0);
9461 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9462 pred->dtp_difo = dp;
9463 pred->dtp_refcnt = 1;
9465 if (!dtrace_difo_cacheable(dp))
9468 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9470 * This is only theoretically possible -- we have had 2^32
9471 * cacheable predicates on this machine. We cannot allow any
9472 * more predicates to become cacheable: as unlikely as it is,
9473 * there may be a thread caching a (now stale) predicate cache
9474 * ID. (N.B.: the temptation is being successfully resisted to
9475 * have this cmn_err() "Holy shit -- we executed this code!")
9480 pred->dtp_cacheid = dtrace_predcache_id++;
9486 dtrace_predicate_hold(dtrace_predicate_t *pred)
9488 ASSERT(MUTEX_HELD(&dtrace_lock));
9489 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9490 ASSERT(pred->dtp_refcnt > 0);
9496 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9498 dtrace_difo_t *dp = pred->dtp_difo;
9500 ASSERT(MUTEX_HELD(&dtrace_lock));
9501 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9502 ASSERT(pred->dtp_refcnt > 0);
9504 if (--pred->dtp_refcnt == 0) {
9505 dtrace_difo_release(pred->dtp_difo, vstate);
9506 kmem_free(pred, sizeof (dtrace_predicate_t));
9511 * DTrace Action Description Functions
9513 static dtrace_actdesc_t *
9514 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9515 uint64_t uarg, uint64_t arg)
9517 dtrace_actdesc_t *act;
9520 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9521 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9524 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9525 act->dtad_kind = kind;
9526 act->dtad_ntuple = ntuple;
9527 act->dtad_uarg = uarg;
9528 act->dtad_arg = arg;
9529 act->dtad_refcnt = 1;
9535 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9537 ASSERT(act->dtad_refcnt >= 1);
9542 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9544 dtrace_actkind_t kind = act->dtad_kind;
9547 ASSERT(act->dtad_refcnt >= 1);
9549 if (--act->dtad_refcnt != 0)
9552 if ((dp = act->dtad_difo) != NULL)
9553 dtrace_difo_release(dp, vstate);
9555 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9556 char *str = (char *)(uintptr_t)act->dtad_arg;
9559 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9560 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9564 kmem_free(str, strlen(str) + 1);
9567 kmem_free(act, sizeof (dtrace_actdesc_t));
9571 * DTrace ECB Functions
9573 static dtrace_ecb_t *
9574 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9579 ASSERT(MUTEX_HELD(&dtrace_lock));
9581 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9582 ecb->dte_predicate = NULL;
9583 ecb->dte_probe = probe;
9586 * The default size is the size of the default action: recording
9589 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9590 ecb->dte_alignment = sizeof (dtrace_epid_t);
9592 epid = state->dts_epid++;
9594 if (epid - 1 >= state->dts_necbs) {
9595 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9596 int necbs = state->dts_necbs << 1;
9598 ASSERT(epid == state->dts_necbs + 1);
9601 ASSERT(oecbs == NULL);
9605 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9608 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9610 dtrace_membar_producer();
9611 state->dts_ecbs = ecbs;
9613 if (oecbs != NULL) {
9615 * If this state is active, we must dtrace_sync()
9616 * before we can free the old dts_ecbs array: we're
9617 * coming in hot, and there may be active ring
9618 * buffer processing (which indexes into the dts_ecbs
9619 * array) on another CPU.
9621 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9624 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9627 dtrace_membar_producer();
9628 state->dts_necbs = necbs;
9631 ecb->dte_state = state;
9633 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9634 dtrace_membar_producer();
9635 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9641 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9643 dtrace_probe_t *probe = ecb->dte_probe;
9645 ASSERT(MUTEX_HELD(&cpu_lock));
9646 ASSERT(MUTEX_HELD(&dtrace_lock));
9647 ASSERT(ecb->dte_next == NULL);
9649 if (probe == NULL) {
9651 * This is the NULL probe -- there's nothing to do.
9656 if (probe->dtpr_ecb == NULL) {
9657 dtrace_provider_t *prov = probe->dtpr_provider;
9660 * We're the first ECB on this probe.
9662 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9664 if (ecb->dte_predicate != NULL)
9665 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9667 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9668 probe->dtpr_id, probe->dtpr_arg);
9671 * This probe is already active. Swing the last pointer to
9672 * point to the new ECB, and issue a dtrace_sync() to assure
9673 * that all CPUs have seen the change.
9675 ASSERT(probe->dtpr_ecb_last != NULL);
9676 probe->dtpr_ecb_last->dte_next = ecb;
9677 probe->dtpr_ecb_last = ecb;
9678 probe->dtpr_predcache = 0;
9685 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9687 uint32_t maxalign = sizeof (dtrace_epid_t);
9688 uint32_t align = sizeof (uint8_t), offs, diff;
9689 dtrace_action_t *act;
9691 uint32_t aggbase = UINT32_MAX;
9692 dtrace_state_t *state = ecb->dte_state;
9695 * If we record anything, we always record the epid. (And we always
9698 offs = sizeof (dtrace_epid_t);
9699 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9701 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9702 dtrace_recdesc_t *rec = &act->dta_rec;
9704 if ((align = rec->dtrd_alignment) > maxalign)
9707 if (!wastuple && act->dta_intuple) {
9709 * This is the first record in a tuple. Align the
9710 * offset to be at offset 4 in an 8-byte aligned
9713 diff = offs + sizeof (dtrace_aggid_t);
9715 if ((diff = (diff & (sizeof (uint64_t) - 1))))
9716 offs += sizeof (uint64_t) - diff;
9718 aggbase = offs - sizeof (dtrace_aggid_t);
9719 ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9723 if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9725 * The current offset is not properly aligned; align it.
9727 offs += align - diff;
9730 rec->dtrd_offset = offs;
9732 if (offs + rec->dtrd_size > ecb->dte_needed) {
9733 ecb->dte_needed = offs + rec->dtrd_size;
9735 if (ecb->dte_needed > state->dts_needed)
9736 state->dts_needed = ecb->dte_needed;
9739 if (DTRACEACT_ISAGG(act->dta_kind)) {
9740 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9741 dtrace_action_t *first = agg->dtag_first, *prev;
9743 ASSERT(rec->dtrd_size != 0 && first != NULL);
9745 ASSERT(aggbase != UINT32_MAX);
9747 agg->dtag_base = aggbase;
9749 while ((prev = first->dta_prev) != NULL &&
9750 DTRACEACT_ISAGG(prev->dta_kind)) {
9751 agg = (dtrace_aggregation_t *)prev;
9752 first = agg->dtag_first;
9756 offs = prev->dta_rec.dtrd_offset +
9757 prev->dta_rec.dtrd_size;
9759 offs = sizeof (dtrace_epid_t);
9763 if (!act->dta_intuple)
9764 ecb->dte_size = offs + rec->dtrd_size;
9766 offs += rec->dtrd_size;
9769 wastuple = act->dta_intuple;
9772 if ((act = ecb->dte_action) != NULL &&
9773 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9774 ecb->dte_size == sizeof (dtrace_epid_t)) {
9776 * If the size is still sizeof (dtrace_epid_t), then all
9777 * actions store no data; set the size to 0.
9779 ecb->dte_alignment = maxalign;
9783 * If the needed space is still sizeof (dtrace_epid_t), then
9784 * all actions need no additional space; set the needed
9787 if (ecb->dte_needed == sizeof (dtrace_epid_t))
9788 ecb->dte_needed = 0;
9794 * Set our alignment, and make sure that the dte_size and dte_needed
9795 * are aligned to the size of an EPID.
9797 ecb->dte_alignment = maxalign;
9798 ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9799 ~(sizeof (dtrace_epid_t) - 1);
9800 ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9801 ~(sizeof (dtrace_epid_t) - 1);
9802 ASSERT(ecb->dte_size <= ecb->dte_needed);
9805 static dtrace_action_t *
9806 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9808 dtrace_aggregation_t *agg;
9809 size_t size = sizeof (uint64_t);
9810 int ntuple = desc->dtad_ntuple;
9811 dtrace_action_t *act;
9812 dtrace_recdesc_t *frec;
9813 dtrace_aggid_t aggid;
9814 dtrace_state_t *state = ecb->dte_state;
9816 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9817 agg->dtag_ecb = ecb;
9819 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9821 switch (desc->dtad_kind) {
9823 agg->dtag_initial = INT64_MAX;
9824 agg->dtag_aggregate = dtrace_aggregate_min;
9828 agg->dtag_initial = INT64_MIN;
9829 agg->dtag_aggregate = dtrace_aggregate_max;
9832 case DTRACEAGG_COUNT:
9833 agg->dtag_aggregate = dtrace_aggregate_count;
9836 case DTRACEAGG_QUANTIZE:
9837 agg->dtag_aggregate = dtrace_aggregate_quantize;
9838 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9842 case DTRACEAGG_LQUANTIZE: {
9843 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9844 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9846 agg->dtag_initial = desc->dtad_arg;
9847 agg->dtag_aggregate = dtrace_aggregate_lquantize;
9849 if (step == 0 || levels == 0)
9852 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9857 agg->dtag_aggregate = dtrace_aggregate_avg;
9858 size = sizeof (uint64_t) * 2;
9861 case DTRACEAGG_STDDEV:
9862 agg->dtag_aggregate = dtrace_aggregate_stddev;
9863 size = sizeof (uint64_t) * 4;
9867 agg->dtag_aggregate = dtrace_aggregate_sum;
9874 agg->dtag_action.dta_rec.dtrd_size = size;
9880 * We must make sure that we have enough actions for the n-tuple.
9882 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9883 if (DTRACEACT_ISAGG(act->dta_kind))
9886 if (--ntuple == 0) {
9888 * This is the action with which our n-tuple begins.
9890 agg->dtag_first = act;
9896 * This n-tuple is short by ntuple elements. Return failure.
9898 ASSERT(ntuple != 0);
9900 kmem_free(agg, sizeof (dtrace_aggregation_t));
9905 * If the last action in the tuple has a size of zero, it's actually
9906 * an expression argument for the aggregating action.
9908 ASSERT(ecb->dte_action_last != NULL);
9909 act = ecb->dte_action_last;
9911 if (act->dta_kind == DTRACEACT_DIFEXPR) {
9912 ASSERT(act->dta_difo != NULL);
9914 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
9915 agg->dtag_hasarg = 1;
9919 * We need to allocate an id for this aggregation.
9922 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
9923 VM_BESTFIT | VM_SLEEP);
9925 aggid = alloc_unr(state->dts_aggid_arena);
9928 if (aggid - 1 >= state->dts_naggregations) {
9929 dtrace_aggregation_t **oaggs = state->dts_aggregations;
9930 dtrace_aggregation_t **aggs;
9931 int naggs = state->dts_naggregations << 1;
9932 int onaggs = state->dts_naggregations;
9934 ASSERT(aggid == state->dts_naggregations + 1);
9937 ASSERT(oaggs == NULL);
9941 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
9943 if (oaggs != NULL) {
9944 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
9945 kmem_free(oaggs, onaggs * sizeof (*aggs));
9948 state->dts_aggregations = aggs;
9949 state->dts_naggregations = naggs;
9952 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
9953 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
9955 frec = &agg->dtag_first->dta_rec;
9956 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
9957 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
9959 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
9960 ASSERT(!act->dta_intuple);
9961 act->dta_intuple = 1;
9964 return (&agg->dtag_action);
9968 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
9970 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9971 dtrace_state_t *state = ecb->dte_state;
9972 dtrace_aggid_t aggid = agg->dtag_id;
9974 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
9976 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
9978 free_unr(state->dts_aggid_arena, aggid);
9981 ASSERT(state->dts_aggregations[aggid - 1] == agg);
9982 state->dts_aggregations[aggid - 1] = NULL;
9984 kmem_free(agg, sizeof (dtrace_aggregation_t));
9988 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9990 dtrace_action_t *action, *last;
9991 dtrace_difo_t *dp = desc->dtad_difo;
9992 uint32_t size = 0, align = sizeof (uint8_t), mask;
9993 uint16_t format = 0;
9994 dtrace_recdesc_t *rec;
9995 dtrace_state_t *state = ecb->dte_state;
9996 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
9997 uint64_t arg = desc->dtad_arg;
9999 ASSERT(MUTEX_HELD(&dtrace_lock));
10000 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10002 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10004 * If this is an aggregating action, there must be neither
10005 * a speculate nor a commit on the action chain.
10007 dtrace_action_t *act;
10009 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10010 if (act->dta_kind == DTRACEACT_COMMIT)
10013 if (act->dta_kind == DTRACEACT_SPECULATE)
10017 action = dtrace_ecb_aggregation_create(ecb, desc);
10019 if (action == NULL)
10022 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10023 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10024 dp != NULL && dp->dtdo_destructive)) {
10025 state->dts_destructive = 1;
10028 switch (desc->dtad_kind) {
10029 case DTRACEACT_PRINTF:
10030 case DTRACEACT_PRINTA:
10031 case DTRACEACT_SYSTEM:
10032 case DTRACEACT_FREOPEN:
10034 * We know that our arg is a string -- turn it into a
10038 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA);
10043 ASSERT(arg > KERNELBASE);
10045 format = dtrace_format_add(state,
10046 (char *)(uintptr_t)arg);
10050 case DTRACEACT_LIBACT:
10051 case DTRACEACT_DIFEXPR:
10055 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10058 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10059 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10062 size = opt[DTRACEOPT_STRSIZE];
10067 case DTRACEACT_STACK:
10068 if ((nframes = arg) == 0) {
10069 nframes = opt[DTRACEOPT_STACKFRAMES];
10070 ASSERT(nframes > 0);
10074 size = nframes * sizeof (pc_t);
10077 case DTRACEACT_JSTACK:
10078 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10079 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10081 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10082 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10084 arg = DTRACE_USTACK_ARG(nframes, strsize);
10087 case DTRACEACT_USTACK:
10088 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10089 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10090 strsize = DTRACE_USTACK_STRSIZE(arg);
10091 nframes = opt[DTRACEOPT_USTACKFRAMES];
10092 ASSERT(nframes > 0);
10093 arg = DTRACE_USTACK_ARG(nframes, strsize);
10097 * Save a slot for the pid.
10099 size = (nframes + 1) * sizeof (uint64_t);
10100 size += DTRACE_USTACK_STRSIZE(arg);
10101 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10105 case DTRACEACT_SYM:
10106 case DTRACEACT_MOD:
10107 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10108 sizeof (uint64_t)) ||
10109 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10113 case DTRACEACT_USYM:
10114 case DTRACEACT_UMOD:
10115 case DTRACEACT_UADDR:
10117 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10118 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10122 * We have a slot for the pid, plus a slot for the
10123 * argument. To keep things simple (aligned with
10124 * bitness-neutral sizing), we store each as a 64-bit
10127 size = 2 * sizeof (uint64_t);
10130 case DTRACEACT_STOP:
10131 case DTRACEACT_BREAKPOINT:
10132 case DTRACEACT_PANIC:
10135 case DTRACEACT_CHILL:
10136 case DTRACEACT_DISCARD:
10137 case DTRACEACT_RAISE:
10142 case DTRACEACT_EXIT:
10144 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10145 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10149 case DTRACEACT_SPECULATE:
10150 if (ecb->dte_size > sizeof (dtrace_epid_t))
10156 state->dts_speculates = 1;
10159 case DTRACEACT_PRINTM:
10160 size = dp->dtdo_rtype.dtdt_size;
10163 case DTRACEACT_PRINTT:
10164 size = dp->dtdo_rtype.dtdt_size;
10167 case DTRACEACT_COMMIT: {
10168 dtrace_action_t *act = ecb->dte_action;
10170 for (; act != NULL; act = act->dta_next) {
10171 if (act->dta_kind == DTRACEACT_COMMIT)
10184 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10186 * If this is a data-storing action or a speculate,
10187 * we must be sure that there isn't a commit on the
10190 dtrace_action_t *act = ecb->dte_action;
10192 for (; act != NULL; act = act->dta_next) {
10193 if (act->dta_kind == DTRACEACT_COMMIT)
10198 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10199 action->dta_rec.dtrd_size = size;
10202 action->dta_refcnt = 1;
10203 rec = &action->dta_rec;
10204 size = rec->dtrd_size;
10206 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10207 if (!(size & mask)) {
10213 action->dta_kind = desc->dtad_kind;
10215 if ((action->dta_difo = dp) != NULL)
10216 dtrace_difo_hold(dp);
10218 rec->dtrd_action = action->dta_kind;
10219 rec->dtrd_arg = arg;
10220 rec->dtrd_uarg = desc->dtad_uarg;
10221 rec->dtrd_alignment = (uint16_t)align;
10222 rec->dtrd_format = format;
10224 if ((last = ecb->dte_action_last) != NULL) {
10225 ASSERT(ecb->dte_action != NULL);
10226 action->dta_prev = last;
10227 last->dta_next = action;
10229 ASSERT(ecb->dte_action == NULL);
10230 ecb->dte_action = action;
10233 ecb->dte_action_last = action;
10239 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10241 dtrace_action_t *act = ecb->dte_action, *next;
10242 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10246 if (act != NULL && act->dta_refcnt > 1) {
10247 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10250 for (; act != NULL; act = next) {
10251 next = act->dta_next;
10252 ASSERT(next != NULL || act == ecb->dte_action_last);
10253 ASSERT(act->dta_refcnt == 1);
10255 if ((format = act->dta_rec.dtrd_format) != 0)
10256 dtrace_format_remove(ecb->dte_state, format);
10258 if ((dp = act->dta_difo) != NULL)
10259 dtrace_difo_release(dp, vstate);
10261 if (DTRACEACT_ISAGG(act->dta_kind)) {
10262 dtrace_ecb_aggregation_destroy(ecb, act);
10264 kmem_free(act, sizeof (dtrace_action_t));
10269 ecb->dte_action = NULL;
10270 ecb->dte_action_last = NULL;
10271 ecb->dte_size = sizeof (dtrace_epid_t);
10275 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10278 * We disable the ECB by removing it from its probe.
10280 dtrace_ecb_t *pecb, *prev = NULL;
10281 dtrace_probe_t *probe = ecb->dte_probe;
10283 ASSERT(MUTEX_HELD(&dtrace_lock));
10285 if (probe == NULL) {
10287 * This is the NULL probe; there is nothing to disable.
10292 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10298 ASSERT(pecb != NULL);
10300 if (prev == NULL) {
10301 probe->dtpr_ecb = ecb->dte_next;
10303 prev->dte_next = ecb->dte_next;
10306 if (ecb == probe->dtpr_ecb_last) {
10307 ASSERT(ecb->dte_next == NULL);
10308 probe->dtpr_ecb_last = prev;
10312 * The ECB has been disconnected from the probe; now sync to assure
10313 * that all CPUs have seen the change before returning.
10317 if (probe->dtpr_ecb == NULL) {
10319 * That was the last ECB on the probe; clear the predicate
10320 * cache ID for the probe, disable it and sync one more time
10321 * to assure that we'll never hit it again.
10323 dtrace_provider_t *prov = probe->dtpr_provider;
10325 ASSERT(ecb->dte_next == NULL);
10326 ASSERT(probe->dtpr_ecb_last == NULL);
10327 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10328 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10329 probe->dtpr_id, probe->dtpr_arg);
10333 * There is at least one ECB remaining on the probe. If there
10334 * is _exactly_ one, set the probe's predicate cache ID to be
10335 * the predicate cache ID of the remaining ECB.
10337 ASSERT(probe->dtpr_ecb_last != NULL);
10338 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10340 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10341 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10343 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10346 probe->dtpr_predcache = p->dtp_cacheid;
10349 ecb->dte_next = NULL;
10354 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10356 dtrace_state_t *state = ecb->dte_state;
10357 dtrace_vstate_t *vstate = &state->dts_vstate;
10358 dtrace_predicate_t *pred;
10359 dtrace_epid_t epid = ecb->dte_epid;
10361 ASSERT(MUTEX_HELD(&dtrace_lock));
10362 ASSERT(ecb->dte_next == NULL);
10363 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10365 if ((pred = ecb->dte_predicate) != NULL)
10366 dtrace_predicate_release(pred, vstate);
10368 dtrace_ecb_action_remove(ecb);
10370 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10371 state->dts_ecbs[epid - 1] = NULL;
10373 kmem_free(ecb, sizeof (dtrace_ecb_t));
10376 static dtrace_ecb_t *
10377 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10378 dtrace_enabling_t *enab)
10381 dtrace_predicate_t *pred;
10382 dtrace_actdesc_t *act;
10383 dtrace_provider_t *prov;
10384 dtrace_ecbdesc_t *desc = enab->dten_current;
10386 ASSERT(MUTEX_HELD(&dtrace_lock));
10387 ASSERT(state != NULL);
10389 ecb = dtrace_ecb_add(state, probe);
10390 ecb->dte_uarg = desc->dted_uarg;
10392 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10393 dtrace_predicate_hold(pred);
10394 ecb->dte_predicate = pred;
10397 if (probe != NULL) {
10399 * If the provider shows more leg than the consumer is old
10400 * enough to see, we need to enable the appropriate implicit
10401 * predicate bits to prevent the ecb from activating at
10404 * Providers specifying DTRACE_PRIV_USER at register time
10405 * are stating that they need the /proc-style privilege
10406 * model to be enforced, and this is what DTRACE_COND_OWNER
10407 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10409 prov = probe->dtpr_provider;
10410 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10411 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10412 ecb->dte_cond |= DTRACE_COND_OWNER;
10414 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10415 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10416 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10419 * If the provider shows us kernel innards and the user
10420 * is lacking sufficient privilege, enable the
10421 * DTRACE_COND_USERMODE implicit predicate.
10423 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10424 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10425 ecb->dte_cond |= DTRACE_COND_USERMODE;
10428 if (dtrace_ecb_create_cache != NULL) {
10430 * If we have a cached ecb, we'll use its action list instead
10431 * of creating our own (saving both time and space).
10433 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10434 dtrace_action_t *act = cached->dte_action;
10437 ASSERT(act->dta_refcnt > 0);
10439 ecb->dte_action = act;
10440 ecb->dte_action_last = cached->dte_action_last;
10441 ecb->dte_needed = cached->dte_needed;
10442 ecb->dte_size = cached->dte_size;
10443 ecb->dte_alignment = cached->dte_alignment;
10449 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10450 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10451 dtrace_ecb_destroy(ecb);
10456 dtrace_ecb_resize(ecb);
10458 return (dtrace_ecb_create_cache = ecb);
10462 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10465 dtrace_enabling_t *enab = arg;
10466 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10468 ASSERT(state != NULL);
10470 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10472 * This probe was created in a generation for which this
10473 * enabling has previously created ECBs; we don't want to
10474 * enable it again, so just kick out.
10476 return (DTRACE_MATCH_NEXT);
10479 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10480 return (DTRACE_MATCH_DONE);
10482 dtrace_ecb_enable(ecb);
10483 return (DTRACE_MATCH_NEXT);
10486 static dtrace_ecb_t *
10487 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10491 ASSERT(MUTEX_HELD(&dtrace_lock));
10493 if (id == 0 || id > state->dts_necbs)
10496 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10497 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10499 return (state->dts_ecbs[id - 1]);
10502 static dtrace_aggregation_t *
10503 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10505 dtrace_aggregation_t *agg;
10507 ASSERT(MUTEX_HELD(&dtrace_lock));
10509 if (id == 0 || id > state->dts_naggregations)
10512 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10513 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10514 agg->dtag_id == id);
10516 return (state->dts_aggregations[id - 1]);
10520 * DTrace Buffer Functions
10522 * The following functions manipulate DTrace buffers. Most of these functions
10523 * are called in the context of establishing or processing consumer state;
10524 * exceptions are explicitly noted.
10528 * Note: called from cross call context. This function switches the two
10529 * buffers on a given CPU. The atomicity of this operation is assured by
10530 * disabling interrupts while the actual switch takes place; the disabling of
10531 * interrupts serializes the execution with any execution of dtrace_probe() on
10535 dtrace_buffer_switch(dtrace_buffer_t *buf)
10537 caddr_t tomax = buf->dtb_tomax;
10538 caddr_t xamot = buf->dtb_xamot;
10539 dtrace_icookie_t cookie;
10541 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10542 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10544 cookie = dtrace_interrupt_disable();
10545 buf->dtb_tomax = xamot;
10546 buf->dtb_xamot = tomax;
10547 buf->dtb_xamot_drops = buf->dtb_drops;
10548 buf->dtb_xamot_offset = buf->dtb_offset;
10549 buf->dtb_xamot_errors = buf->dtb_errors;
10550 buf->dtb_xamot_flags = buf->dtb_flags;
10551 buf->dtb_offset = 0;
10552 buf->dtb_drops = 0;
10553 buf->dtb_errors = 0;
10554 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10555 dtrace_interrupt_enable(cookie);
10559 * Note: called from cross call context. This function activates a buffer
10560 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10561 * is guaranteed by the disabling of interrupts.
10564 dtrace_buffer_activate(dtrace_state_t *state)
10566 dtrace_buffer_t *buf;
10567 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10569 buf = &state->dts_buffer[curcpu];
10571 if (buf->dtb_tomax != NULL) {
10573 * We might like to assert that the buffer is marked inactive,
10574 * but this isn't necessarily true: the buffer for the CPU
10575 * that processes the BEGIN probe has its buffer activated
10576 * manually. In this case, we take the (harmless) action
10577 * re-clearing the bit INACTIVE bit.
10579 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10582 dtrace_interrupt_enable(cookie);
10586 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10592 dtrace_buffer_t *buf;
10595 ASSERT(MUTEX_HELD(&cpu_lock));
10596 ASSERT(MUTEX_HELD(&dtrace_lock));
10598 if (size > dtrace_nonroot_maxsize &&
10599 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10605 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10608 buf = &bufs[cp->cpu_id];
10611 * If there is already a buffer allocated for this CPU, it
10612 * is only possible that this is a DR event. In this case,
10614 if (buf->dtb_tomax != NULL) {
10615 ASSERT(buf->dtb_size == size);
10619 ASSERT(buf->dtb_xamot == NULL);
10621 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10624 buf->dtb_size = size;
10625 buf->dtb_flags = flags;
10626 buf->dtb_offset = 0;
10627 buf->dtb_drops = 0;
10629 if (flags & DTRACEBUF_NOSWITCH)
10632 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10634 } while ((cp = cp->cpu_next) != cpu_list);
10642 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10645 buf = &bufs[cp->cpu_id];
10647 if (buf->dtb_xamot != NULL) {
10648 ASSERT(buf->dtb_tomax != NULL);
10649 ASSERT(buf->dtb_size == size);
10650 kmem_free(buf->dtb_xamot, size);
10653 if (buf->dtb_tomax != NULL) {
10654 ASSERT(buf->dtb_size == size);
10655 kmem_free(buf->dtb_tomax, size);
10658 buf->dtb_tomax = NULL;
10659 buf->dtb_xamot = NULL;
10661 } while ((cp = cp->cpu_next) != cpu_list);
10667 #if defined(__amd64__)
10669 * FreeBSD isn't good at limiting the amount of memory we
10670 * ask to malloc, so let's place a limit here before trying
10671 * to do something that might well end in tears at bedtime.
10673 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10677 ASSERT(MUTEX_HELD(&dtrace_lock));
10679 if (cpu != DTRACE_CPUALL && cpu != i)
10685 * If there is already a buffer allocated for this CPU, it
10686 * is only possible that this is a DR event. In this case,
10687 * the buffer size must match our specified size.
10689 if (buf->dtb_tomax != NULL) {
10690 ASSERT(buf->dtb_size == size);
10694 ASSERT(buf->dtb_xamot == NULL);
10696 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10699 buf->dtb_size = size;
10700 buf->dtb_flags = flags;
10701 buf->dtb_offset = 0;
10702 buf->dtb_drops = 0;
10704 if (flags & DTRACEBUF_NOSWITCH)
10707 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10715 * Error allocating memory, so free the buffers that were
10716 * allocated before the failed allocation.
10719 if (cpu != DTRACE_CPUALL && cpu != i)
10724 if (buf->dtb_xamot != NULL) {
10725 ASSERT(buf->dtb_tomax != NULL);
10726 ASSERT(buf->dtb_size == size);
10727 kmem_free(buf->dtb_xamot, size);
10730 if (buf->dtb_tomax != NULL) {
10731 ASSERT(buf->dtb_size == size);
10732 kmem_free(buf->dtb_tomax, size);
10735 buf->dtb_tomax = NULL;
10736 buf->dtb_xamot = NULL;
10746 * Note: called from probe context. This function just increments the drop
10747 * count on a buffer. It has been made a function to allow for the
10748 * possibility of understanding the source of mysterious drop counts. (A
10749 * problem for which one may be particularly disappointed that DTrace cannot
10750 * be used to understand DTrace.)
10753 dtrace_buffer_drop(dtrace_buffer_t *buf)
10759 * Note: called from probe context. This function is called to reserve space
10760 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
10761 * mstate. Returns the new offset in the buffer, or a negative value if an
10762 * error has occurred.
10765 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10766 dtrace_state_t *state, dtrace_mstate_t *mstate)
10768 intptr_t offs = buf->dtb_offset, soffs;
10773 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10776 if ((tomax = buf->dtb_tomax) == NULL) {
10777 dtrace_buffer_drop(buf);
10781 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10782 while (offs & (align - 1)) {
10784 * Assert that our alignment is off by a number which
10785 * is itself sizeof (uint32_t) aligned.
10787 ASSERT(!((align - (offs & (align - 1))) &
10788 (sizeof (uint32_t) - 1)));
10789 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10790 offs += sizeof (uint32_t);
10793 if ((soffs = offs + needed) > buf->dtb_size) {
10794 dtrace_buffer_drop(buf);
10798 if (mstate == NULL)
10801 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10802 mstate->dtms_scratch_size = buf->dtb_size - soffs;
10803 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10808 if (buf->dtb_flags & DTRACEBUF_FILL) {
10809 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10810 (buf->dtb_flags & DTRACEBUF_FULL))
10815 total = needed + (offs & (align - 1));
10818 * For a ring buffer, life is quite a bit more complicated. Before
10819 * we can store any padding, we need to adjust our wrapping offset.
10820 * (If we've never before wrapped or we're not about to, no adjustment
10823 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10824 offs + total > buf->dtb_size) {
10825 woffs = buf->dtb_xamot_offset;
10827 if (offs + total > buf->dtb_size) {
10829 * We can't fit in the end of the buffer. First, a
10830 * sanity check that we can fit in the buffer at all.
10832 if (total > buf->dtb_size) {
10833 dtrace_buffer_drop(buf);
10838 * We're going to be storing at the top of the buffer,
10839 * so now we need to deal with the wrapped offset. We
10840 * only reset our wrapped offset to 0 if it is
10841 * currently greater than the current offset. If it
10842 * is less than the current offset, it is because a
10843 * previous allocation induced a wrap -- but the
10844 * allocation didn't subsequently take the space due
10845 * to an error or false predicate evaluation. In this
10846 * case, we'll just leave the wrapped offset alone: if
10847 * the wrapped offset hasn't been advanced far enough
10848 * for this allocation, it will be adjusted in the
10851 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10859 * Now we know that we're going to be storing to the
10860 * top of the buffer and that there is room for us
10861 * there. We need to clear the buffer from the current
10862 * offset to the end (there may be old gunk there).
10864 while (offs < buf->dtb_size)
10868 * We need to set our offset to zero. And because we
10869 * are wrapping, we need to set the bit indicating as
10870 * much. We can also adjust our needed space back
10871 * down to the space required by the ECB -- we know
10872 * that the top of the buffer is aligned.
10876 buf->dtb_flags |= DTRACEBUF_WRAPPED;
10879 * There is room for us in the buffer, so we simply
10880 * need to check the wrapped offset.
10882 if (woffs < offs) {
10884 * The wrapped offset is less than the offset.
10885 * This can happen if we allocated buffer space
10886 * that induced a wrap, but then we didn't
10887 * subsequently take the space due to an error
10888 * or false predicate evaluation. This is
10889 * okay; we know that _this_ allocation isn't
10890 * going to induce a wrap. We still can't
10891 * reset the wrapped offset to be zero,
10892 * however: the space may have been trashed in
10893 * the previous failed probe attempt. But at
10894 * least the wrapped offset doesn't need to
10895 * be adjusted at all...
10901 while (offs + total > woffs) {
10902 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
10905 if (epid == DTRACE_EPIDNONE) {
10906 size = sizeof (uint32_t);
10908 ASSERT(epid <= state->dts_necbs);
10909 ASSERT(state->dts_ecbs[epid - 1] != NULL);
10911 size = state->dts_ecbs[epid - 1]->dte_size;
10914 ASSERT(woffs + size <= buf->dtb_size);
10917 if (woffs + size == buf->dtb_size) {
10919 * We've reached the end of the buffer; we want
10920 * to set the wrapped offset to 0 and break
10921 * out. However, if the offs is 0, then we're
10922 * in a strange edge-condition: the amount of
10923 * space that we want to reserve plus the size
10924 * of the record that we're overwriting is
10925 * greater than the size of the buffer. This
10926 * is problematic because if we reserve the
10927 * space but subsequently don't consume it (due
10928 * to a failed predicate or error) the wrapped
10929 * offset will be 0 -- yet the EPID at offset 0
10930 * will not be committed. This situation is
10931 * relatively easy to deal with: if we're in
10932 * this case, the buffer is indistinguishable
10933 * from one that hasn't wrapped; we need only
10934 * finish the job by clearing the wrapped bit,
10935 * explicitly setting the offset to be 0, and
10936 * zero'ing out the old data in the buffer.
10939 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
10940 buf->dtb_offset = 0;
10943 while (woffs < buf->dtb_size)
10944 tomax[woffs++] = 0;
10955 * We have a wrapped offset. It may be that the wrapped offset
10956 * has become zero -- that's okay.
10958 buf->dtb_xamot_offset = woffs;
10963 * Now we can plow the buffer with any necessary padding.
10965 while (offs & (align - 1)) {
10967 * Assert that our alignment is off by a number which
10968 * is itself sizeof (uint32_t) aligned.
10970 ASSERT(!((align - (offs & (align - 1))) &
10971 (sizeof (uint32_t) - 1)));
10972 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10973 offs += sizeof (uint32_t);
10976 if (buf->dtb_flags & DTRACEBUF_FILL) {
10977 if (offs + needed > buf->dtb_size - state->dts_reserve) {
10978 buf->dtb_flags |= DTRACEBUF_FULL;
10983 if (mstate == NULL)
10987 * For ring buffers and fill buffers, the scratch space is always
10988 * the inactive buffer.
10990 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
10991 mstate->dtms_scratch_size = buf->dtb_size;
10992 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10998 dtrace_buffer_polish(dtrace_buffer_t *buf)
11000 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11001 ASSERT(MUTEX_HELD(&dtrace_lock));
11003 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11007 * We need to polish the ring buffer. There are three cases:
11009 * - The first (and presumably most common) is that there is no gap
11010 * between the buffer offset and the wrapped offset. In this case,
11011 * there is nothing in the buffer that isn't valid data; we can
11012 * mark the buffer as polished and return.
11014 * - The second (less common than the first but still more common
11015 * than the third) is that there is a gap between the buffer offset
11016 * and the wrapped offset, and the wrapped offset is larger than the
11017 * buffer offset. This can happen because of an alignment issue, or
11018 * can happen because of a call to dtrace_buffer_reserve() that
11019 * didn't subsequently consume the buffer space. In this case,
11020 * we need to zero the data from the buffer offset to the wrapped
11023 * - The third (and least common) is that there is a gap between the
11024 * buffer offset and the wrapped offset, but the wrapped offset is
11025 * _less_ than the buffer offset. This can only happen because a
11026 * call to dtrace_buffer_reserve() induced a wrap, but the space
11027 * was not subsequently consumed. In this case, we need to zero the
11028 * space from the offset to the end of the buffer _and_ from the
11029 * top of the buffer to the wrapped offset.
11031 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11032 bzero(buf->dtb_tomax + buf->dtb_offset,
11033 buf->dtb_xamot_offset - buf->dtb_offset);
11036 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11037 bzero(buf->dtb_tomax + buf->dtb_offset,
11038 buf->dtb_size - buf->dtb_offset);
11039 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11044 dtrace_buffer_free(dtrace_buffer_t *bufs)
11048 for (i = 0; i < NCPU; i++) {
11049 dtrace_buffer_t *buf = &bufs[i];
11051 if (buf->dtb_tomax == NULL) {
11052 ASSERT(buf->dtb_xamot == NULL);
11053 ASSERT(buf->dtb_size == 0);
11057 if (buf->dtb_xamot != NULL) {
11058 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11059 kmem_free(buf->dtb_xamot, buf->dtb_size);
11062 kmem_free(buf->dtb_tomax, buf->dtb_size);
11064 buf->dtb_tomax = NULL;
11065 buf->dtb_xamot = NULL;
11070 * DTrace Enabling Functions
11072 static dtrace_enabling_t *
11073 dtrace_enabling_create(dtrace_vstate_t *vstate)
11075 dtrace_enabling_t *enab;
11077 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11078 enab->dten_vstate = vstate;
11084 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11086 dtrace_ecbdesc_t **ndesc;
11087 size_t osize, nsize;
11090 * We can't add to enablings after we've enabled them, or after we've
11093 ASSERT(enab->dten_probegen == 0);
11094 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11096 if (enab->dten_ndesc < enab->dten_maxdesc) {
11097 enab->dten_desc[enab->dten_ndesc++] = ecb;
11101 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11103 if (enab->dten_maxdesc == 0) {
11104 enab->dten_maxdesc = 1;
11106 enab->dten_maxdesc <<= 1;
11109 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11111 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11112 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11113 bcopy(enab->dten_desc, ndesc, osize);
11114 if (enab->dten_desc != NULL)
11115 kmem_free(enab->dten_desc, osize);
11117 enab->dten_desc = ndesc;
11118 enab->dten_desc[enab->dten_ndesc++] = ecb;
11122 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11123 dtrace_probedesc_t *pd)
11125 dtrace_ecbdesc_t *new;
11126 dtrace_predicate_t *pred;
11127 dtrace_actdesc_t *act;
11130 * We're going to create a new ECB description that matches the
11131 * specified ECB in every way, but has the specified probe description.
11133 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11135 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11136 dtrace_predicate_hold(pred);
11138 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11139 dtrace_actdesc_hold(act);
11141 new->dted_action = ecb->dted_action;
11142 new->dted_pred = ecb->dted_pred;
11143 new->dted_probe = *pd;
11144 new->dted_uarg = ecb->dted_uarg;
11146 dtrace_enabling_add(enab, new);
11150 dtrace_enabling_dump(dtrace_enabling_t *enab)
11154 for (i = 0; i < enab->dten_ndesc; i++) {
11155 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11157 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11158 desc->dtpd_provider, desc->dtpd_mod,
11159 desc->dtpd_func, desc->dtpd_name);
11164 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11167 dtrace_ecbdesc_t *ep;
11168 dtrace_vstate_t *vstate = enab->dten_vstate;
11170 ASSERT(MUTEX_HELD(&dtrace_lock));
11172 for (i = 0; i < enab->dten_ndesc; i++) {
11173 dtrace_actdesc_t *act, *next;
11174 dtrace_predicate_t *pred;
11176 ep = enab->dten_desc[i];
11178 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11179 dtrace_predicate_release(pred, vstate);
11181 for (act = ep->dted_action; act != NULL; act = next) {
11182 next = act->dtad_next;
11183 dtrace_actdesc_release(act, vstate);
11186 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11189 if (enab->dten_desc != NULL)
11190 kmem_free(enab->dten_desc,
11191 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11194 * If this was a retained enabling, decrement the dts_nretained count
11195 * and take it off of the dtrace_retained list.
11197 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11198 dtrace_retained == enab) {
11199 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11200 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11201 enab->dten_vstate->dtvs_state->dts_nretained--;
11204 if (enab->dten_prev == NULL) {
11205 if (dtrace_retained == enab) {
11206 dtrace_retained = enab->dten_next;
11208 if (dtrace_retained != NULL)
11209 dtrace_retained->dten_prev = NULL;
11212 ASSERT(enab != dtrace_retained);
11213 ASSERT(dtrace_retained != NULL);
11214 enab->dten_prev->dten_next = enab->dten_next;
11217 if (enab->dten_next != NULL) {
11218 ASSERT(dtrace_retained != NULL);
11219 enab->dten_next->dten_prev = enab->dten_prev;
11222 kmem_free(enab, sizeof (dtrace_enabling_t));
11226 dtrace_enabling_retain(dtrace_enabling_t *enab)
11228 dtrace_state_t *state;
11230 ASSERT(MUTEX_HELD(&dtrace_lock));
11231 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11232 ASSERT(enab->dten_vstate != NULL);
11234 state = enab->dten_vstate->dtvs_state;
11235 ASSERT(state != NULL);
11238 * We only allow each state to retain dtrace_retain_max enablings.
11240 if (state->dts_nretained >= dtrace_retain_max)
11243 state->dts_nretained++;
11245 if (dtrace_retained == NULL) {
11246 dtrace_retained = enab;
11250 enab->dten_next = dtrace_retained;
11251 dtrace_retained->dten_prev = enab;
11252 dtrace_retained = enab;
11258 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11259 dtrace_probedesc_t *create)
11261 dtrace_enabling_t *new, *enab;
11262 int found = 0, err = ENOENT;
11264 ASSERT(MUTEX_HELD(&dtrace_lock));
11265 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11266 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11267 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11268 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11270 new = dtrace_enabling_create(&state->dts_vstate);
11273 * Iterate over all retained enablings, looking for enablings that
11274 * match the specified state.
11276 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11280 * dtvs_state can only be NULL for helper enablings -- and
11281 * helper enablings can't be retained.
11283 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11285 if (enab->dten_vstate->dtvs_state != state)
11289 * Now iterate over each probe description; we're looking for
11290 * an exact match to the specified probe description.
11292 for (i = 0; i < enab->dten_ndesc; i++) {
11293 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11294 dtrace_probedesc_t *pd = &ep->dted_probe;
11296 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11299 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11302 if (strcmp(pd->dtpd_func, match->dtpd_func))
11305 if (strcmp(pd->dtpd_name, match->dtpd_name))
11309 * We have a winning probe! Add it to our growing
11313 dtrace_enabling_addlike(new, ep, create);
11317 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11318 dtrace_enabling_destroy(new);
11326 dtrace_enabling_retract(dtrace_state_t *state)
11328 dtrace_enabling_t *enab, *next;
11330 ASSERT(MUTEX_HELD(&dtrace_lock));
11333 * Iterate over all retained enablings, destroy the enablings retained
11334 * for the specified state.
11336 for (enab = dtrace_retained; enab != NULL; enab = next) {
11337 next = enab->dten_next;
11340 * dtvs_state can only be NULL for helper enablings -- and
11341 * helper enablings can't be retained.
11343 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11345 if (enab->dten_vstate->dtvs_state == state) {
11346 ASSERT(state->dts_nretained > 0);
11347 dtrace_enabling_destroy(enab);
11351 ASSERT(state->dts_nretained == 0);
11355 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11360 ASSERT(MUTEX_HELD(&cpu_lock));
11361 ASSERT(MUTEX_HELD(&dtrace_lock));
11363 for (i = 0; i < enab->dten_ndesc; i++) {
11364 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11366 enab->dten_current = ep;
11367 enab->dten_error = 0;
11369 matched += dtrace_probe_enable(&ep->dted_probe, enab);
11371 if (enab->dten_error != 0) {
11373 * If we get an error half-way through enabling the
11374 * probes, we kick out -- perhaps with some number of
11375 * them enabled. Leaving enabled probes enabled may
11376 * be slightly confusing for user-level, but we expect
11377 * that no one will attempt to actually drive on in
11378 * the face of such errors. If this is an anonymous
11379 * enabling (indicated with a NULL nmatched pointer),
11380 * we cmn_err() a message. We aren't expecting to
11381 * get such an error -- such as it can exist at all,
11382 * it would be a result of corrupted DOF in the driver
11385 if (nmatched == NULL) {
11386 cmn_err(CE_WARN, "dtrace_enabling_match() "
11387 "error on %p: %d", (void *)ep,
11391 return (enab->dten_error);
11395 enab->dten_probegen = dtrace_probegen;
11396 if (nmatched != NULL)
11397 *nmatched = matched;
11403 dtrace_enabling_matchall(void)
11405 dtrace_enabling_t *enab;
11407 mutex_enter(&cpu_lock);
11408 mutex_enter(&dtrace_lock);
11411 * Iterate over all retained enablings to see if any probes match
11412 * against them. We only perform this operation on enablings for which
11413 * we have sufficient permissions by virtue of being in the global zone
11414 * or in the same zone as the DTrace client. Because we can be called
11415 * after dtrace_detach() has been called, we cannot assert that there
11416 * are retained enablings. We can safely load from dtrace_retained,
11417 * however: the taskq_destroy() at the end of dtrace_detach() will
11418 * block pending our completion.
11420 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11422 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11424 if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11426 (void) dtrace_enabling_match(enab, NULL);
11429 mutex_exit(&dtrace_lock);
11430 mutex_exit(&cpu_lock);
11434 * If an enabling is to be enabled without having matched probes (that is, if
11435 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11436 * enabling must be _primed_ by creating an ECB for every ECB description.
11437 * This must be done to assure that we know the number of speculations, the
11438 * number of aggregations, the minimum buffer size needed, etc. before we
11439 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11440 * enabling any probes, we create ECBs for every ECB decription, but with a
11441 * NULL probe -- which is exactly what this function does.
11444 dtrace_enabling_prime(dtrace_state_t *state)
11446 dtrace_enabling_t *enab;
11449 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11450 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11452 if (enab->dten_vstate->dtvs_state != state)
11456 * We don't want to prime an enabling more than once, lest
11457 * we allow a malicious user to induce resource exhaustion.
11458 * (The ECBs that result from priming an enabling aren't
11459 * leaked -- but they also aren't deallocated until the
11460 * consumer state is destroyed.)
11462 if (enab->dten_primed)
11465 for (i = 0; i < enab->dten_ndesc; i++) {
11466 enab->dten_current = enab->dten_desc[i];
11467 (void) dtrace_probe_enable(NULL, enab);
11470 enab->dten_primed = 1;
11475 * Called to indicate that probes should be provided due to retained
11476 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11477 * must take an initial lap through the enabling calling the dtps_provide()
11478 * entry point explicitly to allow for autocreated probes.
11481 dtrace_enabling_provide(dtrace_provider_t *prv)
11484 dtrace_probedesc_t desc;
11486 ASSERT(MUTEX_HELD(&dtrace_lock));
11487 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11491 prv = dtrace_provider;
11495 dtrace_enabling_t *enab = dtrace_retained;
11496 void *parg = prv->dtpv_arg;
11498 for (; enab != NULL; enab = enab->dten_next) {
11499 for (i = 0; i < enab->dten_ndesc; i++) {
11500 desc = enab->dten_desc[i]->dted_probe;
11501 mutex_exit(&dtrace_lock);
11502 prv->dtpv_pops.dtps_provide(parg, &desc);
11503 mutex_enter(&dtrace_lock);
11506 } while (all && (prv = prv->dtpv_next) != NULL);
11508 mutex_exit(&dtrace_lock);
11509 dtrace_probe_provide(NULL, all ? NULL : prv);
11510 mutex_enter(&dtrace_lock);
11514 * DTrace DOF Functions
11518 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11520 if (dtrace_err_verbose)
11521 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11523 #ifdef DTRACE_ERRDEBUG
11524 dtrace_errdebug(str);
11529 * Create DOF out of a currently enabled state. Right now, we only create
11530 * DOF containing the run-time options -- but this could be expanded to create
11531 * complete DOF representing the enabled state.
11534 dtrace_dof_create(dtrace_state_t *state)
11538 dof_optdesc_t *opt;
11539 int i, len = sizeof (dof_hdr_t) +
11540 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11541 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11543 ASSERT(MUTEX_HELD(&dtrace_lock));
11545 dof = kmem_zalloc(len, KM_SLEEP);
11546 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11547 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11548 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11549 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11551 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11552 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11553 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11554 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11555 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11556 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11558 dof->dofh_flags = 0;
11559 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11560 dof->dofh_secsize = sizeof (dof_sec_t);
11561 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11562 dof->dofh_secoff = sizeof (dof_hdr_t);
11563 dof->dofh_loadsz = len;
11564 dof->dofh_filesz = len;
11568 * Fill in the option section header...
11570 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11571 sec->dofs_type = DOF_SECT_OPTDESC;
11572 sec->dofs_align = sizeof (uint64_t);
11573 sec->dofs_flags = DOF_SECF_LOAD;
11574 sec->dofs_entsize = sizeof (dof_optdesc_t);
11576 opt = (dof_optdesc_t *)((uintptr_t)sec +
11577 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11579 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11580 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11582 for (i = 0; i < DTRACEOPT_MAX; i++) {
11583 opt[i].dofo_option = i;
11584 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11585 opt[i].dofo_value = state->dts_options[i];
11592 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11594 dof_hdr_t hdr, *dof;
11596 ASSERT(!MUTEX_HELD(&dtrace_lock));
11599 * First, we're going to copyin() the sizeof (dof_hdr_t).
11601 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11602 dtrace_dof_error(NULL, "failed to copyin DOF header");
11608 * Now we'll allocate the entire DOF and copy it in -- provided
11609 * that the length isn't outrageous.
11611 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11612 dtrace_dof_error(&hdr, "load size exceeds maximum");
11617 if (hdr.dofh_loadsz < sizeof (hdr)) {
11618 dtrace_dof_error(&hdr, "invalid load size");
11623 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11625 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
11626 kmem_free(dof, hdr.dofh_loadsz);
11635 static __inline uchar_t
11636 dtrace_dof_char(char c) {
11655 return (c - 'A' + 10);
11662 return (c - 'a' + 10);
11664 /* Should not reach here. */
11670 dtrace_dof_property(const char *name)
11674 unsigned int len, i;
11679 * Unfortunately, array of values in .conf files are always (and
11680 * only) interpreted to be integer arrays. We must read our DOF
11681 * as an integer array, and then squeeze it into a byte array.
11683 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11684 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11687 for (i = 0; i < len; i++)
11688 buf[i] = (uchar_t)(((int *)buf)[i]);
11690 if (len < sizeof (dof_hdr_t)) {
11691 ddi_prop_free(buf);
11692 dtrace_dof_error(NULL, "truncated header");
11696 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11697 ddi_prop_free(buf);
11698 dtrace_dof_error(NULL, "truncated DOF");
11702 if (loadsz >= dtrace_dof_maxsize) {
11703 ddi_prop_free(buf);
11704 dtrace_dof_error(NULL, "oversized DOF");
11708 dof = kmem_alloc(loadsz, KM_SLEEP);
11709 bcopy(buf, dof, loadsz);
11710 ddi_prop_free(buf);
11715 if ((p_env = getenv(name)) == NULL)
11718 len = strlen(p_env) / 2;
11720 buf = kmem_alloc(len, KM_SLEEP);
11722 dof = (dof_hdr_t *) buf;
11726 for (i = 0; i < len; i++) {
11727 buf[i] = (dtrace_dof_char(p[0]) << 4) |
11728 dtrace_dof_char(p[1]);
11734 if (len < sizeof (dof_hdr_t)) {
11736 dtrace_dof_error(NULL, "truncated header");
11740 if (len < (loadsz = dof->dofh_loadsz)) {
11742 dtrace_dof_error(NULL, "truncated DOF");
11746 if (loadsz >= dtrace_dof_maxsize) {
11748 dtrace_dof_error(NULL, "oversized DOF");
11757 dtrace_dof_destroy(dof_hdr_t *dof)
11759 kmem_free(dof, dof->dofh_loadsz);
11763 * Return the dof_sec_t pointer corresponding to a given section index. If the
11764 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
11765 * a type other than DOF_SECT_NONE is specified, the header is checked against
11766 * this type and NULL is returned if the types do not match.
11769 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11771 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11772 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11774 if (i >= dof->dofh_secnum) {
11775 dtrace_dof_error(dof, "referenced section index is invalid");
11779 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11780 dtrace_dof_error(dof, "referenced section is not loadable");
11784 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11785 dtrace_dof_error(dof, "referenced section is the wrong type");
11792 static dtrace_probedesc_t *
11793 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11795 dof_probedesc_t *probe;
11797 uintptr_t daddr = (uintptr_t)dof;
11801 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11802 dtrace_dof_error(dof, "invalid probe section");
11806 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11807 dtrace_dof_error(dof, "bad alignment in probe description");
11811 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11812 dtrace_dof_error(dof, "truncated probe description");
11816 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11817 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11819 if (strtab == NULL)
11822 str = daddr + strtab->dofs_offset;
11823 size = strtab->dofs_size;
11825 if (probe->dofp_provider >= strtab->dofs_size) {
11826 dtrace_dof_error(dof, "corrupt probe provider");
11830 (void) strncpy(desc->dtpd_provider,
11831 (char *)(str + probe->dofp_provider),
11832 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11834 if (probe->dofp_mod >= strtab->dofs_size) {
11835 dtrace_dof_error(dof, "corrupt probe module");
11839 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11840 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11842 if (probe->dofp_func >= strtab->dofs_size) {
11843 dtrace_dof_error(dof, "corrupt probe function");
11847 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11848 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11850 if (probe->dofp_name >= strtab->dofs_size) {
11851 dtrace_dof_error(dof, "corrupt probe name");
11855 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11856 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11861 static dtrace_difo_t *
11862 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11867 dof_difohdr_t *dofd;
11868 uintptr_t daddr = (uintptr_t)dof;
11869 size_t max = dtrace_difo_maxsize;
11872 static const struct {
11880 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11881 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11882 sizeof (dif_instr_t), "multiple DIF sections" },
11884 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11885 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11886 sizeof (uint64_t), "multiple integer tables" },
11888 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11889 offsetof(dtrace_difo_t, dtdo_strlen), 0,
11890 sizeof (char), "multiple string tables" },
11892 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11893 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11894 sizeof (uint_t), "multiple variable tables" },
11896 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
11899 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11900 dtrace_dof_error(dof, "invalid DIFO header section");
11904 if (sec->dofs_align != sizeof (dof_secidx_t)) {
11905 dtrace_dof_error(dof, "bad alignment in DIFO header");
11909 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
11910 sec->dofs_size % sizeof (dof_secidx_t)) {
11911 dtrace_dof_error(dof, "bad size in DIFO header");
11915 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11916 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
11918 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
11919 dp->dtdo_rtype = dofd->dofd_rtype;
11921 for (l = 0; l < n; l++) {
11926 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
11927 dofd->dofd_links[l])) == NULL)
11928 goto err; /* invalid section link */
11930 if (ttl + subsec->dofs_size > max) {
11931 dtrace_dof_error(dof, "exceeds maximum size");
11935 ttl += subsec->dofs_size;
11937 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
11938 if (subsec->dofs_type != difo[i].section)
11941 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
11942 dtrace_dof_error(dof, "section not loaded");
11946 if (subsec->dofs_align != difo[i].align) {
11947 dtrace_dof_error(dof, "bad alignment");
11951 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
11952 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
11954 if (*bufp != NULL) {
11955 dtrace_dof_error(dof, difo[i].msg);
11959 if (difo[i].entsize != subsec->dofs_entsize) {
11960 dtrace_dof_error(dof, "entry size mismatch");
11964 if (subsec->dofs_entsize != 0 &&
11965 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
11966 dtrace_dof_error(dof, "corrupt entry size");
11970 *lenp = subsec->dofs_size;
11971 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
11972 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
11973 *bufp, subsec->dofs_size);
11975 if (subsec->dofs_entsize != 0)
11976 *lenp /= subsec->dofs_entsize;
11982 * If we encounter a loadable DIFO sub-section that is not
11983 * known to us, assume this is a broken program and fail.
11985 if (difo[i].section == DOF_SECT_NONE &&
11986 (subsec->dofs_flags & DOF_SECF_LOAD)) {
11987 dtrace_dof_error(dof, "unrecognized DIFO subsection");
11992 if (dp->dtdo_buf == NULL) {
11994 * We can't have a DIF object without DIF text.
11996 dtrace_dof_error(dof, "missing DIF text");
12001 * Before we validate the DIF object, run through the variable table
12002 * looking for the strings -- if any of their size are under, we'll set
12003 * their size to be the system-wide default string size. Note that
12004 * this should _not_ happen if the "strsize" option has been set --
12005 * in this case, the compiler should have set the size to reflect the
12006 * setting of the option.
12008 for (i = 0; i < dp->dtdo_varlen; i++) {
12009 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12010 dtrace_diftype_t *t = &v->dtdv_type;
12012 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12015 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12016 t->dtdt_size = dtrace_strsize_default;
12019 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12022 dtrace_difo_init(dp, vstate);
12026 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12027 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12028 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12029 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12031 kmem_free(dp, sizeof (dtrace_difo_t));
12035 static dtrace_predicate_t *
12036 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12041 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12044 return (dtrace_predicate_create(dp));
12047 static dtrace_actdesc_t *
12048 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12051 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12052 dof_actdesc_t *desc;
12053 dof_sec_t *difosec;
12055 uintptr_t daddr = (uintptr_t)dof;
12057 dtrace_actkind_t kind;
12059 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12060 dtrace_dof_error(dof, "invalid action section");
12064 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12065 dtrace_dof_error(dof, "truncated action description");
12069 if (sec->dofs_align != sizeof (uint64_t)) {
12070 dtrace_dof_error(dof, "bad alignment in action description");
12074 if (sec->dofs_size < sec->dofs_entsize) {
12075 dtrace_dof_error(dof, "section entry size exceeds total size");
12079 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12080 dtrace_dof_error(dof, "bad entry size in action description");
12084 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12085 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12089 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12090 desc = (dof_actdesc_t *)(daddr +
12091 (uintptr_t)sec->dofs_offset + offs);
12092 kind = (dtrace_actkind_t)desc->dofa_kind;
12094 if (DTRACEACT_ISPRINTFLIKE(kind) &&
12095 (kind != DTRACEACT_PRINTA ||
12096 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12102 * printf()-like actions must have a format string.
12104 if ((strtab = dtrace_dof_sect(dof,
12105 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12108 str = (char *)((uintptr_t)dof +
12109 (uintptr_t)strtab->dofs_offset);
12111 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12112 if (str[i] == '\0')
12116 if (i >= strtab->dofs_size) {
12117 dtrace_dof_error(dof, "bogus format string");
12121 if (i == desc->dofa_arg) {
12122 dtrace_dof_error(dof, "empty format string");
12126 i -= desc->dofa_arg;
12127 fmt = kmem_alloc(i + 1, KM_SLEEP);
12128 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12129 arg = (uint64_t)(uintptr_t)fmt;
12131 if (kind == DTRACEACT_PRINTA) {
12132 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12135 arg = desc->dofa_arg;
12139 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12140 desc->dofa_uarg, arg);
12142 if (last != NULL) {
12143 last->dtad_next = act;
12150 if (desc->dofa_difo == DOF_SECIDX_NONE)
12153 if ((difosec = dtrace_dof_sect(dof,
12154 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12157 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12159 if (act->dtad_difo == NULL)
12163 ASSERT(first != NULL);
12167 for (act = first; act != NULL; act = next) {
12168 next = act->dtad_next;
12169 dtrace_actdesc_release(act, vstate);
12175 static dtrace_ecbdesc_t *
12176 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12179 dtrace_ecbdesc_t *ep;
12180 dof_ecbdesc_t *ecb;
12181 dtrace_probedesc_t *desc;
12182 dtrace_predicate_t *pred = NULL;
12184 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12185 dtrace_dof_error(dof, "truncated ECB description");
12189 if (sec->dofs_align != sizeof (uint64_t)) {
12190 dtrace_dof_error(dof, "bad alignment in ECB description");
12194 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12195 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12200 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12201 ep->dted_uarg = ecb->dofe_uarg;
12202 desc = &ep->dted_probe;
12204 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12207 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12208 if ((sec = dtrace_dof_sect(dof,
12209 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12212 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12215 ep->dted_pred.dtpdd_predicate = pred;
12218 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12219 if ((sec = dtrace_dof_sect(dof,
12220 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12223 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12225 if (ep->dted_action == NULL)
12233 dtrace_predicate_release(pred, vstate);
12234 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12239 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12240 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12241 * site of any user SETX relocations to account for load object base address.
12242 * In the future, if we need other relocations, this function can be extended.
12245 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12247 uintptr_t daddr = (uintptr_t)dof;
12248 dof_relohdr_t *dofr =
12249 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12250 dof_sec_t *ss, *rs, *ts;
12254 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12255 sec->dofs_align != sizeof (dof_secidx_t)) {
12256 dtrace_dof_error(dof, "invalid relocation header");
12260 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12261 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12262 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12264 if (ss == NULL || rs == NULL || ts == NULL)
12265 return (-1); /* dtrace_dof_error() has been called already */
12267 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12268 rs->dofs_align != sizeof (uint64_t)) {
12269 dtrace_dof_error(dof, "invalid relocation section");
12273 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12274 n = rs->dofs_size / rs->dofs_entsize;
12276 for (i = 0; i < n; i++) {
12277 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12279 switch (r->dofr_type) {
12280 case DOF_RELO_NONE:
12282 case DOF_RELO_SETX:
12283 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12284 sizeof (uint64_t) > ts->dofs_size) {
12285 dtrace_dof_error(dof, "bad relocation offset");
12289 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12290 dtrace_dof_error(dof, "misaligned setx relo");
12294 *(uint64_t *)taddr += ubase;
12297 dtrace_dof_error(dof, "invalid relocation type");
12301 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12308 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12309 * header: it should be at the front of a memory region that is at least
12310 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12311 * size. It need not be validated in any other way.
12314 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12315 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12317 uint64_t len = dof->dofh_loadsz, seclen;
12318 uintptr_t daddr = (uintptr_t)dof;
12319 dtrace_ecbdesc_t *ep;
12320 dtrace_enabling_t *enab;
12323 ASSERT(MUTEX_HELD(&dtrace_lock));
12324 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12327 * Check the DOF header identification bytes. In addition to checking
12328 * valid settings, we also verify that unused bits/bytes are zeroed so
12329 * we can use them later without fear of regressing existing binaries.
12331 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12332 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12333 dtrace_dof_error(dof, "DOF magic string mismatch");
12337 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12338 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12339 dtrace_dof_error(dof, "DOF has invalid data model");
12343 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12344 dtrace_dof_error(dof, "DOF encoding mismatch");
12348 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12349 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12350 dtrace_dof_error(dof, "DOF version mismatch");
12354 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12355 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12359 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12360 dtrace_dof_error(dof, "DOF uses too many integer registers");
12364 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12365 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12369 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12370 if (dof->dofh_ident[i] != 0) {
12371 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12376 if (dof->dofh_flags & ~DOF_FL_VALID) {
12377 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12381 if (dof->dofh_secsize == 0) {
12382 dtrace_dof_error(dof, "zero section header size");
12387 * Check that the section headers don't exceed the amount of DOF
12388 * data. Note that we cast the section size and number of sections
12389 * to uint64_t's to prevent possible overflow in the multiplication.
12391 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12393 if (dof->dofh_secoff > len || seclen > len ||
12394 dof->dofh_secoff + seclen > len) {
12395 dtrace_dof_error(dof, "truncated section headers");
12399 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12400 dtrace_dof_error(dof, "misaligned section headers");
12404 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12405 dtrace_dof_error(dof, "misaligned section size");
12410 * Take an initial pass through the section headers to be sure that
12411 * the headers don't have stray offsets. If the 'noprobes' flag is
12412 * set, do not permit sections relating to providers, probes, or args.
12414 for (i = 0; i < dof->dofh_secnum; i++) {
12415 dof_sec_t *sec = (dof_sec_t *)(daddr +
12416 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12419 switch (sec->dofs_type) {
12420 case DOF_SECT_PROVIDER:
12421 case DOF_SECT_PROBES:
12422 case DOF_SECT_PRARGS:
12423 case DOF_SECT_PROFFS:
12424 dtrace_dof_error(dof, "illegal sections "
12430 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12431 continue; /* just ignore non-loadable sections */
12433 if (sec->dofs_align & (sec->dofs_align - 1)) {
12434 dtrace_dof_error(dof, "bad section alignment");
12438 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12439 dtrace_dof_error(dof, "misaligned section");
12443 if (sec->dofs_offset > len || sec->dofs_size > len ||
12444 sec->dofs_offset + sec->dofs_size > len) {
12445 dtrace_dof_error(dof, "corrupt section header");
12449 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12450 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12451 dtrace_dof_error(dof, "non-terminating string table");
12457 * Take a second pass through the sections and locate and perform any
12458 * relocations that are present. We do this after the first pass to
12459 * be sure that all sections have had their headers validated.
12461 for (i = 0; i < dof->dofh_secnum; i++) {
12462 dof_sec_t *sec = (dof_sec_t *)(daddr +
12463 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12465 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12466 continue; /* skip sections that are not loadable */
12468 switch (sec->dofs_type) {
12469 case DOF_SECT_URELHDR:
12470 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12476 if ((enab = *enabp) == NULL)
12477 enab = *enabp = dtrace_enabling_create(vstate);
12479 for (i = 0; i < dof->dofh_secnum; i++) {
12480 dof_sec_t *sec = (dof_sec_t *)(daddr +
12481 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12483 if (sec->dofs_type != DOF_SECT_ECBDESC)
12486 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12487 dtrace_enabling_destroy(enab);
12492 dtrace_enabling_add(enab, ep);
12499 * Process DOF for any options. This routine assumes that the DOF has been
12500 * at least processed by dtrace_dof_slurp().
12503 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12508 dof_optdesc_t *desc;
12510 for (i = 0; i < dof->dofh_secnum; i++) {
12511 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12512 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12514 if (sec->dofs_type != DOF_SECT_OPTDESC)
12517 if (sec->dofs_align != sizeof (uint64_t)) {
12518 dtrace_dof_error(dof, "bad alignment in "
12519 "option description");
12523 if ((entsize = sec->dofs_entsize) == 0) {
12524 dtrace_dof_error(dof, "zeroed option entry size");
12528 if (entsize < sizeof (dof_optdesc_t)) {
12529 dtrace_dof_error(dof, "bad option entry size");
12533 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12534 desc = (dof_optdesc_t *)((uintptr_t)dof +
12535 (uintptr_t)sec->dofs_offset + offs);
12537 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12538 dtrace_dof_error(dof, "non-zero option string");
12542 if (desc->dofo_value == DTRACEOPT_UNSET) {
12543 dtrace_dof_error(dof, "unset option");
12547 if ((rval = dtrace_state_option(state,
12548 desc->dofo_option, desc->dofo_value)) != 0) {
12549 dtrace_dof_error(dof, "rejected option");
12559 * DTrace Consumer State Functions
12562 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12564 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12567 dtrace_dynvar_t *dvar, *next, *start;
12570 ASSERT(MUTEX_HELD(&dtrace_lock));
12571 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12573 bzero(dstate, sizeof (dtrace_dstate_t));
12575 if ((dstate->dtds_chunksize = chunksize) == 0)
12576 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12578 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12581 if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
12584 dstate->dtds_size = size;
12585 dstate->dtds_base = base;
12586 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12587 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12589 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12591 if (hashsize != 1 && (hashsize & 1))
12594 dstate->dtds_hashsize = hashsize;
12595 dstate->dtds_hash = dstate->dtds_base;
12598 * Set all of our hash buckets to point to the single sink, and (if
12599 * it hasn't already been set), set the sink's hash value to be the
12600 * sink sentinel value. The sink is needed for dynamic variable
12601 * lookups to know that they have iterated over an entire, valid hash
12604 for (i = 0; i < hashsize; i++)
12605 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12607 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12608 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12611 * Determine number of active CPUs. Divide free list evenly among
12614 start = (dtrace_dynvar_t *)
12615 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12616 limit = (uintptr_t)base + size;
12618 maxper = (limit - (uintptr_t)start) / NCPU;
12619 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12624 for (i = 0; i < NCPU; i++) {
12626 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12629 * If we don't even have enough chunks to make it once through
12630 * NCPUs, we're just going to allocate everything to the first
12631 * CPU. And if we're on the last CPU, we're going to allocate
12632 * whatever is left over. In either case, we set the limit to
12633 * be the limit of the dynamic variable space.
12635 if (maxper == 0 || i == NCPU - 1) {
12636 limit = (uintptr_t)base + size;
12639 limit = (uintptr_t)start + maxper;
12640 start = (dtrace_dynvar_t *)limit;
12643 ASSERT(limit <= (uintptr_t)base + size);
12646 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12647 dstate->dtds_chunksize);
12649 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12652 dvar->dtdv_next = next;
12664 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12666 ASSERT(MUTEX_HELD(&cpu_lock));
12668 if (dstate->dtds_base == NULL)
12671 kmem_free(dstate->dtds_base, dstate->dtds_size);
12672 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12676 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12679 * Logical XOR, where are you?
12681 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12683 if (vstate->dtvs_nglobals > 0) {
12684 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12685 sizeof (dtrace_statvar_t *));
12688 if (vstate->dtvs_ntlocals > 0) {
12689 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12690 sizeof (dtrace_difv_t));
12693 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
12695 if (vstate->dtvs_nlocals > 0) {
12696 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
12697 sizeof (dtrace_statvar_t *));
12703 dtrace_state_clean(dtrace_state_t *state)
12705 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12708 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12709 dtrace_speculation_clean(state);
12713 dtrace_state_deadman(dtrace_state_t *state)
12719 now = dtrace_gethrtime();
12721 if (state != dtrace_anon.dta_state &&
12722 now - state->dts_laststatus >= dtrace_deadman_user)
12726 * We must be sure that dts_alive never appears to be less than the
12727 * value upon entry to dtrace_state_deadman(), and because we lack a
12728 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12729 * store INT64_MAX to it, followed by a memory barrier, followed by
12730 * the new value. This assures that dts_alive never appears to be
12731 * less than its true value, regardless of the order in which the
12732 * stores to the underlying storage are issued.
12734 state->dts_alive = INT64_MAX;
12735 dtrace_membar_producer();
12736 state->dts_alive = now;
12740 dtrace_state_clean(void *arg)
12742 dtrace_state_t *state = arg;
12743 dtrace_optval_t *opt = state->dts_options;
12745 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12748 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12749 dtrace_speculation_clean(state);
12751 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
12752 dtrace_state_clean, state);
12756 dtrace_state_deadman(void *arg)
12758 dtrace_state_t *state = arg;
12763 dtrace_debug_output();
12765 now = dtrace_gethrtime();
12767 if (state != dtrace_anon.dta_state &&
12768 now - state->dts_laststatus >= dtrace_deadman_user)
12772 * We must be sure that dts_alive never appears to be less than the
12773 * value upon entry to dtrace_state_deadman(), and because we lack a
12774 * dtrace_cas64(), we cannot store to it atomically. We thus instead
12775 * store INT64_MAX to it, followed by a memory barrier, followed by
12776 * the new value. This assures that dts_alive never appears to be
12777 * less than its true value, regardless of the order in which the
12778 * stores to the underlying storage are issued.
12780 state->dts_alive = INT64_MAX;
12781 dtrace_membar_producer();
12782 state->dts_alive = now;
12784 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
12785 dtrace_state_deadman, state);
12789 static dtrace_state_t *
12791 dtrace_state_create(dev_t *devp, cred_t *cr)
12793 dtrace_state_create(struct cdev *dev)
12804 dtrace_state_t *state;
12805 dtrace_optval_t *opt;
12806 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12808 ASSERT(MUTEX_HELD(&dtrace_lock));
12809 ASSERT(MUTEX_HELD(&cpu_lock));
12812 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12813 VM_BESTFIT | VM_SLEEP);
12815 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12816 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12820 state = ddi_get_soft_state(dtrace_softstate, minor);
12827 /* Allocate memory for the state. */
12828 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
12831 state->dts_epid = DTRACE_EPIDNONE + 1;
12833 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
12835 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12836 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12838 if (devp != NULL) {
12839 major = getemajor(*devp);
12841 major = ddi_driver_major(dtrace_devi);
12844 state->dts_dev = makedevice(major, minor);
12847 *devp = state->dts_dev;
12849 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
12850 state->dts_dev = dev;
12854 * We allocate NCPU buffers. On the one hand, this can be quite
12855 * a bit of memory per instance (nearly 36K on a Starcat). On the
12856 * other hand, it saves an additional memory reference in the probe
12859 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12860 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12863 state->dts_cleaner = CYCLIC_NONE;
12864 state->dts_deadman = CYCLIC_NONE;
12866 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
12867 callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
12869 state->dts_vstate.dtvs_state = state;
12871 for (i = 0; i < DTRACEOPT_MAX; i++)
12872 state->dts_options[i] = DTRACEOPT_UNSET;
12875 * Set the default options.
12877 opt = state->dts_options;
12878 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12879 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12880 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12881 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12882 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12883 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12884 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12885 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12886 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12887 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12888 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12889 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12890 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12891 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12893 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12896 * Depending on the user credentials, we set flag bits which alter probe
12897 * visibility or the amount of destructiveness allowed. In the case of
12898 * actual anonymous tracing, or the possession of all privileges, all of
12899 * the normal checks are bypassed.
12901 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
12902 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
12903 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
12906 * Set up the credentials for this instantiation. We take a
12907 * hold on the credential to prevent it from disappearing on
12908 * us; this in turn prevents the zone_t referenced by this
12909 * credential from disappearing. This means that we can
12910 * examine the credential and the zone from probe context.
12913 state->dts_cred.dcr_cred = cr;
12916 * CRA_PROC means "we have *some* privilege for dtrace" and
12917 * unlocks the use of variables like pid, zonename, etc.
12919 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
12920 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12921 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
12925 * dtrace_user allows use of syscall and profile providers.
12926 * If the user also has proc_owner and/or proc_zone, we
12927 * extend the scope to include additional visibility and
12928 * destructive power.
12930 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
12931 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
12932 state->dts_cred.dcr_visible |=
12933 DTRACE_CRV_ALLPROC;
12935 state->dts_cred.dcr_action |=
12936 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12939 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
12940 state->dts_cred.dcr_visible |=
12941 DTRACE_CRV_ALLZONE;
12943 state->dts_cred.dcr_action |=
12944 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12948 * If we have all privs in whatever zone this is,
12949 * we can do destructive things to processes which
12950 * have altered credentials.
12953 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12954 cr->cr_zone->zone_privset)) {
12955 state->dts_cred.dcr_action |=
12956 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12962 * Holding the dtrace_kernel privilege also implies that
12963 * the user has the dtrace_user privilege from a visibility
12964 * perspective. But without further privileges, some
12965 * destructive actions are not available.
12967 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
12969 * Make all probes in all zones visible. However,
12970 * this doesn't mean that all actions become available
12973 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
12974 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
12976 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
12979 * Holding proc_owner means that destructive actions
12980 * for *this* zone are allowed.
12982 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12983 state->dts_cred.dcr_action |=
12984 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12987 * Holding proc_zone means that destructive actions
12988 * for this user/group ID in all zones is allowed.
12990 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12991 state->dts_cred.dcr_action |=
12992 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12996 * If we have all privs in whatever zone this is,
12997 * we can do destructive things to processes which
12998 * have altered credentials.
13000 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13001 cr->cr_zone->zone_privset)) {
13002 state->dts_cred.dcr_action |=
13003 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13009 * Holding the dtrace_proc privilege gives control over fasttrap
13010 * and pid providers. We need to grant wider destructive
13011 * privileges in the event that the user has proc_owner and/or
13014 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13015 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13016 state->dts_cred.dcr_action |=
13017 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13019 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13020 state->dts_cred.dcr_action |=
13021 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13029 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13031 dtrace_optval_t *opt = state->dts_options, size;
13032 processorid_t cpu = 0;;
13033 int flags = 0, rval;
13035 ASSERT(MUTEX_HELD(&dtrace_lock));
13036 ASSERT(MUTEX_HELD(&cpu_lock));
13037 ASSERT(which < DTRACEOPT_MAX);
13038 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13039 (state == dtrace_anon.dta_state &&
13040 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13042 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13045 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13046 cpu = opt[DTRACEOPT_CPU];
13048 if (which == DTRACEOPT_SPECSIZE)
13049 flags |= DTRACEBUF_NOSWITCH;
13051 if (which == DTRACEOPT_BUFSIZE) {
13052 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13053 flags |= DTRACEBUF_RING;
13055 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13056 flags |= DTRACEBUF_FILL;
13058 if (state != dtrace_anon.dta_state ||
13059 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13060 flags |= DTRACEBUF_INACTIVE;
13063 for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
13065 * The size must be 8-byte aligned. If the size is not 8-byte
13066 * aligned, drop it down by the difference.
13068 if (size & (sizeof (uint64_t) - 1))
13069 size -= size & (sizeof (uint64_t) - 1);
13071 if (size < state->dts_reserve) {
13073 * Buffers always must be large enough to accommodate
13074 * their prereserved space. We return E2BIG instead
13075 * of ENOMEM in this case to allow for user-level
13076 * software to differentiate the cases.
13081 rval = dtrace_buffer_alloc(buf, size, flags, cpu);
13083 if (rval != ENOMEM) {
13088 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13096 dtrace_state_buffers(dtrace_state_t *state)
13098 dtrace_speculation_t *spec = state->dts_speculations;
13101 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13102 DTRACEOPT_BUFSIZE)) != 0)
13105 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13106 DTRACEOPT_AGGSIZE)) != 0)
13109 for (i = 0; i < state->dts_nspeculations; i++) {
13110 if ((rval = dtrace_state_buffer(state,
13111 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13119 dtrace_state_prereserve(dtrace_state_t *state)
13122 dtrace_probe_t *probe;
13124 state->dts_reserve = 0;
13126 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13130 * If our buffer policy is a "fill" buffer policy, we need to set the
13131 * prereserved space to be the space required by the END probes.
13133 probe = dtrace_probes[dtrace_probeid_end - 1];
13134 ASSERT(probe != NULL);
13136 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13137 if (ecb->dte_state != state)
13140 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13145 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13147 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13148 dtrace_speculation_t *spec;
13149 dtrace_buffer_t *buf;
13151 cyc_handler_t hdlr;
13154 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13155 dtrace_icookie_t cookie;
13157 mutex_enter(&cpu_lock);
13158 mutex_enter(&dtrace_lock);
13160 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13166 * Before we can perform any checks, we must prime all of the
13167 * retained enablings that correspond to this state.
13169 dtrace_enabling_prime(state);
13171 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13176 dtrace_state_prereserve(state);
13179 * Now we want to do is try to allocate our speculations.
13180 * We do not automatically resize the number of speculations; if
13181 * this fails, we will fail the operation.
13183 nspec = opt[DTRACEOPT_NSPEC];
13184 ASSERT(nspec != DTRACEOPT_UNSET);
13186 if (nspec > INT_MAX) {
13191 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
13193 if (spec == NULL) {
13198 state->dts_speculations = spec;
13199 state->dts_nspeculations = (int)nspec;
13201 for (i = 0; i < nspec; i++) {
13202 if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
13207 spec[i].dtsp_buffer = buf;
13210 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13211 if (dtrace_anon.dta_state == NULL) {
13216 if (state->dts_necbs != 0) {
13221 state->dts_anon = dtrace_anon_grab();
13222 ASSERT(state->dts_anon != NULL);
13223 state = state->dts_anon;
13226 * We want "grabanon" to be set in the grabbed state, so we'll
13227 * copy that option value from the grabbing state into the
13230 state->dts_options[DTRACEOPT_GRABANON] =
13231 opt[DTRACEOPT_GRABANON];
13233 *cpu = dtrace_anon.dta_beganon;
13236 * If the anonymous state is active (as it almost certainly
13237 * is if the anonymous enabling ultimately matched anything),
13238 * we don't allow any further option processing -- but we
13239 * don't return failure.
13241 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13245 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13246 opt[DTRACEOPT_AGGSIZE] != 0) {
13247 if (state->dts_aggregations == NULL) {
13249 * We're not going to create an aggregation buffer
13250 * because we don't have any ECBs that contain
13251 * aggregations -- set this option to 0.
13253 opt[DTRACEOPT_AGGSIZE] = 0;
13256 * If we have an aggregation buffer, we must also have
13257 * a buffer to use as scratch.
13259 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13260 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13261 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13266 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13267 opt[DTRACEOPT_SPECSIZE] != 0) {
13268 if (!state->dts_speculates) {
13270 * We're not going to create speculation buffers
13271 * because we don't have any ECBs that actually
13272 * speculate -- set the speculation size to 0.
13274 opt[DTRACEOPT_SPECSIZE] = 0;
13279 * The bare minimum size for any buffer that we're actually going to
13280 * do anything to is sizeof (uint64_t).
13282 sz = sizeof (uint64_t);
13284 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13285 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13286 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13288 * A buffer size has been explicitly set to 0 (or to a size
13289 * that will be adjusted to 0) and we need the space -- we
13290 * need to return failure. We return ENOSPC to differentiate
13291 * it from failing to allocate a buffer due to failure to meet
13292 * the reserve (for which we return E2BIG).
13298 if ((rval = dtrace_state_buffers(state)) != 0)
13301 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13302 sz = dtrace_dstate_defsize;
13305 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13310 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13312 } while (sz >>= 1);
13314 opt[DTRACEOPT_DYNVARSIZE] = sz;
13319 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13320 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13322 if (opt[DTRACEOPT_CLEANRATE] == 0)
13323 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13325 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13326 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13328 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13329 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13331 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13333 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13334 hdlr.cyh_arg = state;
13335 hdlr.cyh_level = CY_LOW_LEVEL;
13338 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13340 state->dts_cleaner = cyclic_add(&hdlr, &when);
13342 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13343 hdlr.cyh_arg = state;
13344 hdlr.cyh_level = CY_LOW_LEVEL;
13347 when.cyt_interval = dtrace_deadman_interval;
13349 state->dts_deadman = cyclic_add(&hdlr, &when);
13351 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13352 dtrace_state_clean, state);
13353 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13354 dtrace_state_deadman, state);
13357 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13360 * Now it's time to actually fire the BEGIN probe. We need to disable
13361 * interrupts here both to record the CPU on which we fired the BEGIN
13362 * probe (the data from this CPU will be processed first at user
13363 * level) and to manually activate the buffer for this CPU.
13365 cookie = dtrace_interrupt_disable();
13367 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13368 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13370 dtrace_probe(dtrace_probeid_begin,
13371 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13372 dtrace_interrupt_enable(cookie);
13374 * We may have had an exit action from a BEGIN probe; only change our
13375 * state to ACTIVE if we're still in WARMUP.
13377 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13378 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13380 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13381 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13384 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13385 * want each CPU to transition its principal buffer out of the
13386 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13387 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13388 * atomically transition from processing none of a state's ECBs to
13389 * processing all of them.
13391 dtrace_xcall(DTRACE_CPUALL,
13392 (dtrace_xcall_t)dtrace_buffer_activate, state);
13396 dtrace_buffer_free(state->dts_buffer);
13397 dtrace_buffer_free(state->dts_aggbuffer);
13399 if ((nspec = state->dts_nspeculations) == 0) {
13400 ASSERT(state->dts_speculations == NULL);
13404 spec = state->dts_speculations;
13405 ASSERT(spec != NULL);
13407 for (i = 0; i < state->dts_nspeculations; i++) {
13408 if ((buf = spec[i].dtsp_buffer) == NULL)
13411 dtrace_buffer_free(buf);
13412 kmem_free(buf, bufsize);
13415 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13416 state->dts_nspeculations = 0;
13417 state->dts_speculations = NULL;
13420 mutex_exit(&dtrace_lock);
13421 mutex_exit(&cpu_lock);
13427 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13429 dtrace_icookie_t cookie;
13431 ASSERT(MUTEX_HELD(&dtrace_lock));
13433 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13434 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13438 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13439 * to be sure that every CPU has seen it. See below for the details
13440 * on why this is done.
13442 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13446 * By this point, it is impossible for any CPU to be still processing
13447 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13448 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13449 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13450 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13451 * iff we're in the END probe.
13453 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13455 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13458 * Finally, we can release the reserve and call the END probe. We
13459 * disable interrupts across calling the END probe to allow us to
13460 * return the CPU on which we actually called the END probe. This
13461 * allows user-land to be sure that this CPU's principal buffer is
13464 state->dts_reserve = 0;
13466 cookie = dtrace_interrupt_disable();
13468 dtrace_probe(dtrace_probeid_end,
13469 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13470 dtrace_interrupt_enable(cookie);
13472 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13479 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13480 dtrace_optval_t val)
13482 ASSERT(MUTEX_HELD(&dtrace_lock));
13484 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13487 if (option >= DTRACEOPT_MAX)
13490 if (option != DTRACEOPT_CPU && val < 0)
13494 case DTRACEOPT_DESTRUCTIVE:
13495 if (dtrace_destructive_disallow)
13498 state->dts_cred.dcr_destructive = 1;
13501 case DTRACEOPT_BUFSIZE:
13502 case DTRACEOPT_DYNVARSIZE:
13503 case DTRACEOPT_AGGSIZE:
13504 case DTRACEOPT_SPECSIZE:
13505 case DTRACEOPT_STRSIZE:
13509 if (val >= LONG_MAX) {
13511 * If this is an otherwise negative value, set it to
13512 * the highest multiple of 128m less than LONG_MAX.
13513 * Technically, we're adjusting the size without
13514 * regard to the buffer resizing policy, but in fact,
13515 * this has no effect -- if we set the buffer size to
13516 * ~LONG_MAX and the buffer policy is ultimately set to
13517 * be "manual", the buffer allocation is guaranteed to
13518 * fail, if only because the allocation requires two
13519 * buffers. (We set the the size to the highest
13520 * multiple of 128m because it ensures that the size
13521 * will remain a multiple of a megabyte when
13522 * repeatedly halved -- all the way down to 15m.)
13524 val = LONG_MAX - (1 << 27) + 1;
13528 state->dts_options[option] = val;
13534 dtrace_state_destroy(dtrace_state_t *state)
13537 dtrace_vstate_t *vstate = &state->dts_vstate;
13539 minor_t minor = getminor(state->dts_dev);
13541 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13542 dtrace_speculation_t *spec = state->dts_speculations;
13543 int nspec = state->dts_nspeculations;
13546 ASSERT(MUTEX_HELD(&dtrace_lock));
13547 ASSERT(MUTEX_HELD(&cpu_lock));
13550 * First, retract any retained enablings for this state.
13552 dtrace_enabling_retract(state);
13553 ASSERT(state->dts_nretained == 0);
13555 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13556 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13558 * We have managed to come into dtrace_state_destroy() on a
13559 * hot enabling -- almost certainly because of a disorderly
13560 * shutdown of a consumer. (That is, a consumer that is
13561 * exiting without having called dtrace_stop().) In this case,
13562 * we're going to set our activity to be KILLED, and then
13563 * issue a sync to be sure that everyone is out of probe
13564 * context before we start blowing away ECBs.
13566 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13571 * Release the credential hold we took in dtrace_state_create().
13573 if (state->dts_cred.dcr_cred != NULL)
13574 crfree(state->dts_cred.dcr_cred);
13577 * Now we can safely disable and destroy any enabled probes. Because
13578 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13579 * (especially if they're all enabled), we take two passes through the
13580 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13581 * in the second we disable whatever is left over.
13583 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13584 for (i = 0; i < state->dts_necbs; i++) {
13585 if ((ecb = state->dts_ecbs[i]) == NULL)
13588 if (match && ecb->dte_probe != NULL) {
13589 dtrace_probe_t *probe = ecb->dte_probe;
13590 dtrace_provider_t *prov = probe->dtpr_provider;
13592 if (!(prov->dtpv_priv.dtpp_flags & match))
13596 dtrace_ecb_disable(ecb);
13597 dtrace_ecb_destroy(ecb);
13605 * Before we free the buffers, perform one more sync to assure that
13606 * every CPU is out of probe context.
13610 dtrace_buffer_free(state->dts_buffer);
13611 dtrace_buffer_free(state->dts_aggbuffer);
13613 for (i = 0; i < nspec; i++)
13614 dtrace_buffer_free(spec[i].dtsp_buffer);
13617 if (state->dts_cleaner != CYCLIC_NONE)
13618 cyclic_remove(state->dts_cleaner);
13620 if (state->dts_deadman != CYCLIC_NONE)
13621 cyclic_remove(state->dts_deadman);
13623 callout_stop(&state->dts_cleaner);
13624 callout_drain(&state->dts_cleaner);
13625 callout_stop(&state->dts_deadman);
13626 callout_drain(&state->dts_deadman);
13629 dtrace_dstate_fini(&vstate->dtvs_dynvars);
13630 dtrace_vstate_fini(vstate);
13631 if (state->dts_ecbs != NULL)
13632 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13634 if (state->dts_aggregations != NULL) {
13636 for (i = 0; i < state->dts_naggregations; i++)
13637 ASSERT(state->dts_aggregations[i] == NULL);
13639 ASSERT(state->dts_naggregations > 0);
13640 kmem_free(state->dts_aggregations,
13641 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13644 kmem_free(state->dts_buffer, bufsize);
13645 kmem_free(state->dts_aggbuffer, bufsize);
13647 for (i = 0; i < nspec; i++)
13648 kmem_free(spec[i].dtsp_buffer, bufsize);
13651 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13653 dtrace_format_destroy(state);
13655 if (state->dts_aggid_arena != NULL) {
13657 vmem_destroy(state->dts_aggid_arena);
13659 delete_unrhdr(state->dts_aggid_arena);
13661 state->dts_aggid_arena = NULL;
13664 ddi_soft_state_free(dtrace_softstate, minor);
13665 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13670 * DTrace Anonymous Enabling Functions
13672 static dtrace_state_t *
13673 dtrace_anon_grab(void)
13675 dtrace_state_t *state;
13677 ASSERT(MUTEX_HELD(&dtrace_lock));
13679 if ((state = dtrace_anon.dta_state) == NULL) {
13680 ASSERT(dtrace_anon.dta_enabling == NULL);
13684 ASSERT(dtrace_anon.dta_enabling != NULL);
13685 ASSERT(dtrace_retained != NULL);
13687 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13688 dtrace_anon.dta_enabling = NULL;
13689 dtrace_anon.dta_state = NULL;
13695 dtrace_anon_property(void)
13698 dtrace_state_t *state;
13700 char c[32]; /* enough for "dof-data-" + digits */
13702 ASSERT(MUTEX_HELD(&dtrace_lock));
13703 ASSERT(MUTEX_HELD(&cpu_lock));
13705 for (i = 0; ; i++) {
13706 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
13708 dtrace_err_verbose = 1;
13710 if ((dof = dtrace_dof_property(c)) == NULL) {
13711 dtrace_err_verbose = 0;
13717 * We want to create anonymous state, so we need to transition
13718 * the kernel debugger to indicate that DTrace is active. If
13719 * this fails (e.g. because the debugger has modified text in
13720 * some way), we won't continue with the processing.
13722 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13723 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
13724 "enabling ignored.");
13725 dtrace_dof_destroy(dof);
13731 * If we haven't allocated an anonymous state, we'll do so now.
13733 if ((state = dtrace_anon.dta_state) == NULL) {
13735 state = dtrace_state_create(NULL, NULL);
13737 state = dtrace_state_create(NULL);
13739 dtrace_anon.dta_state = state;
13741 if (state == NULL) {
13743 * This basically shouldn't happen: the only
13744 * failure mode from dtrace_state_create() is a
13745 * failure of ddi_soft_state_zalloc() that
13746 * itself should never happen. Still, the
13747 * interface allows for a failure mode, and
13748 * we want to fail as gracefully as possible:
13749 * we'll emit an error message and cease
13750 * processing anonymous state in this case.
13752 cmn_err(CE_WARN, "failed to create "
13753 "anonymous state");
13754 dtrace_dof_destroy(dof);
13759 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
13760 &dtrace_anon.dta_enabling, 0, B_TRUE);
13763 rv = dtrace_dof_options(dof, state);
13765 dtrace_err_verbose = 0;
13766 dtrace_dof_destroy(dof);
13770 * This is malformed DOF; chuck any anonymous state
13773 ASSERT(dtrace_anon.dta_enabling == NULL);
13774 dtrace_state_destroy(state);
13775 dtrace_anon.dta_state = NULL;
13779 ASSERT(dtrace_anon.dta_enabling != NULL);
13782 if (dtrace_anon.dta_enabling != NULL) {
13786 * dtrace_enabling_retain() can only fail because we are
13787 * trying to retain more enablings than are allowed -- but
13788 * we only have one anonymous enabling, and we are guaranteed
13789 * to be allowed at least one retained enabling; we assert
13790 * that dtrace_enabling_retain() returns success.
13792 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
13795 dtrace_enabling_dump(dtrace_anon.dta_enabling);
13800 * DTrace Helper Functions
13803 dtrace_helper_trace(dtrace_helper_action_t *helper,
13804 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
13806 uint32_t size, next, nnext, i;
13807 dtrace_helptrace_t *ent;
13808 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
13810 if (!dtrace_helptrace_enabled)
13813 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
13816 * What would a tracing framework be without its own tracing
13817 * framework? (Well, a hell of a lot simpler, for starters...)
13819 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
13820 sizeof (uint64_t) - sizeof (uint64_t);
13823 * Iterate until we can allocate a slot in the trace buffer.
13826 next = dtrace_helptrace_next;
13828 if (next + size < dtrace_helptrace_bufsize) {
13829 nnext = next + size;
13833 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
13836 * We have our slot; fill it in.
13841 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
13842 ent->dtht_helper = helper;
13843 ent->dtht_where = where;
13844 ent->dtht_nlocals = vstate->dtvs_nlocals;
13846 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13847 mstate->dtms_fltoffs : -1;
13848 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13849 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
13851 for (i = 0; i < vstate->dtvs_nlocals; i++) {
13852 dtrace_statvar_t *svar;
13854 if ((svar = vstate->dtvs_locals[i]) == NULL)
13857 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13858 ent->dtht_locals[i] =
13859 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
13864 dtrace_helper(int which, dtrace_mstate_t *mstate,
13865 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13867 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
13868 uint64_t sarg0 = mstate->dtms_arg[0];
13869 uint64_t sarg1 = mstate->dtms_arg[1];
13871 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13872 dtrace_helper_action_t *helper;
13873 dtrace_vstate_t *vstate;
13874 dtrace_difo_t *pred;
13875 int i, trace = dtrace_helptrace_enabled;
13877 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13879 if (helpers == NULL)
13882 if ((helper = helpers->dthps_actions[which]) == NULL)
13885 vstate = &helpers->dthps_vstate;
13886 mstate->dtms_arg[0] = arg0;
13887 mstate->dtms_arg[1] = arg1;
13890 * Now iterate over each helper. If its predicate evaluates to 'true',
13891 * we'll call the corresponding actions. Note that the below calls
13892 * to dtrace_dif_emulate() may set faults in machine state. This is
13893 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
13894 * the stored DIF offset with its own (which is the desired behavior).
13895 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13896 * from machine state; this is okay, too.
13898 for (; helper != NULL; helper = helper->dtha_next) {
13899 if ((pred = helper->dtha_predicate) != NULL) {
13901 dtrace_helper_trace(helper, mstate, vstate, 0);
13903 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
13906 if (*flags & CPU_DTRACE_FAULT)
13910 for (i = 0; i < helper->dtha_nactions; i++) {
13912 dtrace_helper_trace(helper,
13913 mstate, vstate, i + 1);
13915 rval = dtrace_dif_emulate(helper->dtha_actions[i],
13916 mstate, vstate, state);
13918 if (*flags & CPU_DTRACE_FAULT)
13924 dtrace_helper_trace(helper, mstate, vstate,
13925 DTRACE_HELPTRACE_NEXT);
13929 dtrace_helper_trace(helper, mstate, vstate,
13930 DTRACE_HELPTRACE_DONE);
13933 * Restore the arg0 that we saved upon entry.
13935 mstate->dtms_arg[0] = sarg0;
13936 mstate->dtms_arg[1] = sarg1;
13942 dtrace_helper_trace(helper, mstate, vstate,
13943 DTRACE_HELPTRACE_ERR);
13946 * Restore the arg0 that we saved upon entry.
13948 mstate->dtms_arg[0] = sarg0;
13949 mstate->dtms_arg[1] = sarg1;
13955 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
13956 dtrace_vstate_t *vstate)
13960 if (helper->dtha_predicate != NULL)
13961 dtrace_difo_release(helper->dtha_predicate, vstate);
13963 for (i = 0; i < helper->dtha_nactions; i++) {
13964 ASSERT(helper->dtha_actions[i] != NULL);
13965 dtrace_difo_release(helper->dtha_actions[i], vstate);
13968 kmem_free(helper->dtha_actions,
13969 helper->dtha_nactions * sizeof (dtrace_difo_t *));
13970 kmem_free(helper, sizeof (dtrace_helper_action_t));
13974 dtrace_helper_destroygen(int gen)
13976 proc_t *p = curproc;
13977 dtrace_helpers_t *help = p->p_dtrace_helpers;
13978 dtrace_vstate_t *vstate;
13981 ASSERT(MUTEX_HELD(&dtrace_lock));
13983 if (help == NULL || gen > help->dthps_generation)
13986 vstate = &help->dthps_vstate;
13988 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13989 dtrace_helper_action_t *last = NULL, *h, *next;
13991 for (h = help->dthps_actions[i]; h != NULL; h = next) {
13992 next = h->dtha_next;
13994 if (h->dtha_generation == gen) {
13995 if (last != NULL) {
13996 last->dtha_next = next;
13998 help->dthps_actions[i] = next;
14001 dtrace_helper_action_destroy(h, vstate);
14009 * Interate until we've cleared out all helper providers with the
14010 * given generation number.
14013 dtrace_helper_provider_t *prov;
14016 * Look for a helper provider with the right generation. We
14017 * have to start back at the beginning of the list each time
14018 * because we drop dtrace_lock. It's unlikely that we'll make
14019 * more than two passes.
14021 for (i = 0; i < help->dthps_nprovs; i++) {
14022 prov = help->dthps_provs[i];
14024 if (prov->dthp_generation == gen)
14029 * If there were no matches, we're done.
14031 if (i == help->dthps_nprovs)
14035 * Move the last helper provider into this slot.
14037 help->dthps_nprovs--;
14038 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14039 help->dthps_provs[help->dthps_nprovs] = NULL;
14041 mutex_exit(&dtrace_lock);
14044 * If we have a meta provider, remove this helper provider.
14046 mutex_enter(&dtrace_meta_lock);
14047 if (dtrace_meta_pid != NULL) {
14048 ASSERT(dtrace_deferred_pid == NULL);
14049 dtrace_helper_provider_remove(&prov->dthp_prov,
14052 mutex_exit(&dtrace_meta_lock);
14054 dtrace_helper_provider_destroy(prov);
14056 mutex_enter(&dtrace_lock);
14063 dtrace_helper_validate(dtrace_helper_action_t *helper)
14068 if ((dp = helper->dtha_predicate) != NULL)
14069 err += dtrace_difo_validate_helper(dp);
14071 for (i = 0; i < helper->dtha_nactions; i++)
14072 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14078 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14080 dtrace_helpers_t *help;
14081 dtrace_helper_action_t *helper, *last;
14082 dtrace_actdesc_t *act;
14083 dtrace_vstate_t *vstate;
14084 dtrace_predicate_t *pred;
14085 int count = 0, nactions = 0, i;
14087 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14090 help = curproc->p_dtrace_helpers;
14091 last = help->dthps_actions[which];
14092 vstate = &help->dthps_vstate;
14094 for (count = 0; last != NULL; last = last->dtha_next) {
14096 if (last->dtha_next == NULL)
14101 * If we already have dtrace_helper_actions_max helper actions for this
14102 * helper action type, we'll refuse to add a new one.
14104 if (count >= dtrace_helper_actions_max)
14107 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14108 helper->dtha_generation = help->dthps_generation;
14110 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14111 ASSERT(pred->dtp_difo != NULL);
14112 dtrace_difo_hold(pred->dtp_difo);
14113 helper->dtha_predicate = pred->dtp_difo;
14116 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14117 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14120 if (act->dtad_difo == NULL)
14126 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14127 (helper->dtha_nactions = nactions), KM_SLEEP);
14129 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14130 dtrace_difo_hold(act->dtad_difo);
14131 helper->dtha_actions[i++] = act->dtad_difo;
14134 if (!dtrace_helper_validate(helper))
14137 if (last == NULL) {
14138 help->dthps_actions[which] = helper;
14140 last->dtha_next = helper;
14143 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14144 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14145 dtrace_helptrace_next = 0;
14150 dtrace_helper_action_destroy(helper, vstate);
14155 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14156 dof_helper_t *dofhp)
14158 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14160 mutex_enter(&dtrace_meta_lock);
14161 mutex_enter(&dtrace_lock);
14163 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14165 * If the dtrace module is loaded but not attached, or if
14166 * there aren't isn't a meta provider registered to deal with
14167 * these provider descriptions, we need to postpone creating
14168 * the actual providers until later.
14171 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14172 dtrace_deferred_pid != help) {
14173 help->dthps_deferred = 1;
14174 help->dthps_pid = p->p_pid;
14175 help->dthps_next = dtrace_deferred_pid;
14176 help->dthps_prev = NULL;
14177 if (dtrace_deferred_pid != NULL)
14178 dtrace_deferred_pid->dthps_prev = help;
14179 dtrace_deferred_pid = help;
14182 mutex_exit(&dtrace_lock);
14184 } else if (dofhp != NULL) {
14186 * If the dtrace module is loaded and we have a particular
14187 * helper provider description, pass that off to the
14191 mutex_exit(&dtrace_lock);
14193 dtrace_helper_provide(dofhp, p->p_pid);
14197 * Otherwise, just pass all the helper provider descriptions
14198 * off to the meta provider.
14202 mutex_exit(&dtrace_lock);
14204 for (i = 0; i < help->dthps_nprovs; i++) {
14205 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14210 mutex_exit(&dtrace_meta_lock);
14214 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14216 dtrace_helpers_t *help;
14217 dtrace_helper_provider_t *hprov, **tmp_provs;
14218 uint_t tmp_maxprovs, i;
14220 ASSERT(MUTEX_HELD(&dtrace_lock));
14222 help = curproc->p_dtrace_helpers;
14223 ASSERT(help != NULL);
14226 * If we already have dtrace_helper_providers_max helper providers,
14227 * we're refuse to add a new one.
14229 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14233 * Check to make sure this isn't a duplicate.
14235 for (i = 0; i < help->dthps_nprovs; i++) {
14236 if (dofhp->dofhp_addr ==
14237 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14241 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14242 hprov->dthp_prov = *dofhp;
14243 hprov->dthp_ref = 1;
14244 hprov->dthp_generation = gen;
14247 * Allocate a bigger table for helper providers if it's already full.
14249 if (help->dthps_maxprovs == help->dthps_nprovs) {
14250 tmp_maxprovs = help->dthps_maxprovs;
14251 tmp_provs = help->dthps_provs;
14253 if (help->dthps_maxprovs == 0)
14254 help->dthps_maxprovs = 2;
14256 help->dthps_maxprovs *= 2;
14257 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14258 help->dthps_maxprovs = dtrace_helper_providers_max;
14260 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14262 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14263 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14265 if (tmp_provs != NULL) {
14266 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14267 sizeof (dtrace_helper_provider_t *));
14268 kmem_free(tmp_provs, tmp_maxprovs *
14269 sizeof (dtrace_helper_provider_t *));
14273 help->dthps_provs[help->dthps_nprovs] = hprov;
14274 help->dthps_nprovs++;
14280 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14282 mutex_enter(&dtrace_lock);
14284 if (--hprov->dthp_ref == 0) {
14286 mutex_exit(&dtrace_lock);
14287 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14288 dtrace_dof_destroy(dof);
14289 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14291 mutex_exit(&dtrace_lock);
14296 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14298 uintptr_t daddr = (uintptr_t)dof;
14299 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14300 dof_provider_t *provider;
14301 dof_probe_t *probe;
14303 char *strtab, *typestr;
14304 dof_stridx_t typeidx;
14306 uint_t nprobes, j, k;
14308 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14310 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14311 dtrace_dof_error(dof, "misaligned section offset");
14316 * The section needs to be large enough to contain the DOF provider
14317 * structure appropriate for the given version.
14319 if (sec->dofs_size <
14320 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14321 offsetof(dof_provider_t, dofpv_prenoffs) :
14322 sizeof (dof_provider_t))) {
14323 dtrace_dof_error(dof, "provider section too small");
14327 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14328 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14329 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14330 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14331 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14333 if (str_sec == NULL || prb_sec == NULL ||
14334 arg_sec == NULL || off_sec == NULL)
14339 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14340 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14341 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14342 provider->dofpv_prenoffs)) == NULL)
14345 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14347 if (provider->dofpv_name >= str_sec->dofs_size ||
14348 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14349 dtrace_dof_error(dof, "invalid provider name");
14353 if (prb_sec->dofs_entsize == 0 ||
14354 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14355 dtrace_dof_error(dof, "invalid entry size");
14359 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14360 dtrace_dof_error(dof, "misaligned entry size");
14364 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14365 dtrace_dof_error(dof, "invalid entry size");
14369 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14370 dtrace_dof_error(dof, "misaligned section offset");
14374 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14375 dtrace_dof_error(dof, "invalid entry size");
14379 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14381 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14384 * Take a pass through the probes to check for errors.
14386 for (j = 0; j < nprobes; j++) {
14387 probe = (dof_probe_t *)(uintptr_t)(daddr +
14388 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14390 if (probe->dofpr_func >= str_sec->dofs_size) {
14391 dtrace_dof_error(dof, "invalid function name");
14395 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14396 dtrace_dof_error(dof, "function name too long");
14400 if (probe->dofpr_name >= str_sec->dofs_size ||
14401 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14402 dtrace_dof_error(dof, "invalid probe name");
14407 * The offset count must not wrap the index, and the offsets
14408 * must also not overflow the section's data.
14410 if (probe->dofpr_offidx + probe->dofpr_noffs <
14411 probe->dofpr_offidx ||
14412 (probe->dofpr_offidx + probe->dofpr_noffs) *
14413 off_sec->dofs_entsize > off_sec->dofs_size) {
14414 dtrace_dof_error(dof, "invalid probe offset");
14418 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14420 * If there's no is-enabled offset section, make sure
14421 * there aren't any is-enabled offsets. Otherwise
14422 * perform the same checks as for probe offsets
14423 * (immediately above).
14425 if (enoff_sec == NULL) {
14426 if (probe->dofpr_enoffidx != 0 ||
14427 probe->dofpr_nenoffs != 0) {
14428 dtrace_dof_error(dof, "is-enabled "
14429 "offsets with null section");
14432 } else if (probe->dofpr_enoffidx +
14433 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14434 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14435 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14436 dtrace_dof_error(dof, "invalid is-enabled "
14441 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14442 dtrace_dof_error(dof, "zero probe and "
14443 "is-enabled offsets");
14446 } else if (probe->dofpr_noffs == 0) {
14447 dtrace_dof_error(dof, "zero probe offsets");
14451 if (probe->dofpr_argidx + probe->dofpr_xargc <
14452 probe->dofpr_argidx ||
14453 (probe->dofpr_argidx + probe->dofpr_xargc) *
14454 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14455 dtrace_dof_error(dof, "invalid args");
14459 typeidx = probe->dofpr_nargv;
14460 typestr = strtab + probe->dofpr_nargv;
14461 for (k = 0; k < probe->dofpr_nargc; k++) {
14462 if (typeidx >= str_sec->dofs_size) {
14463 dtrace_dof_error(dof, "bad "
14464 "native argument type");
14468 typesz = strlen(typestr) + 1;
14469 if (typesz > DTRACE_ARGTYPELEN) {
14470 dtrace_dof_error(dof, "native "
14471 "argument type too long");
14478 typeidx = probe->dofpr_xargv;
14479 typestr = strtab + probe->dofpr_xargv;
14480 for (k = 0; k < probe->dofpr_xargc; k++) {
14481 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14482 dtrace_dof_error(dof, "bad "
14483 "native argument index");
14487 if (typeidx >= str_sec->dofs_size) {
14488 dtrace_dof_error(dof, "bad "
14489 "translated argument type");
14493 typesz = strlen(typestr) + 1;
14494 if (typesz > DTRACE_ARGTYPELEN) {
14495 dtrace_dof_error(dof, "translated argument "
14509 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14511 dtrace_helpers_t *help;
14512 dtrace_vstate_t *vstate;
14513 dtrace_enabling_t *enab = NULL;
14514 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14515 uintptr_t daddr = (uintptr_t)dof;
14517 ASSERT(MUTEX_HELD(&dtrace_lock));
14519 if ((help = curproc->p_dtrace_helpers) == NULL)
14520 help = dtrace_helpers_create(curproc);
14522 vstate = &help->dthps_vstate;
14524 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14525 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14526 dtrace_dof_destroy(dof);
14531 * Look for helper providers and validate their descriptions.
14534 for (i = 0; i < dof->dofh_secnum; i++) {
14535 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14536 dof->dofh_secoff + i * dof->dofh_secsize);
14538 if (sec->dofs_type != DOF_SECT_PROVIDER)
14541 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14542 dtrace_enabling_destroy(enab);
14543 dtrace_dof_destroy(dof);
14552 * Now we need to walk through the ECB descriptions in the enabling.
14554 for (i = 0; i < enab->dten_ndesc; i++) {
14555 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14556 dtrace_probedesc_t *desc = &ep->dted_probe;
14558 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14561 if (strcmp(desc->dtpd_mod, "helper") != 0)
14564 if (strcmp(desc->dtpd_func, "ustack") != 0)
14567 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14570 * Adding this helper action failed -- we are now going
14571 * to rip out the entire generation and return failure.
14573 (void) dtrace_helper_destroygen(help->dthps_generation);
14574 dtrace_enabling_destroy(enab);
14575 dtrace_dof_destroy(dof);
14582 if (nhelpers < enab->dten_ndesc)
14583 dtrace_dof_error(dof, "unmatched helpers");
14585 gen = help->dthps_generation++;
14586 dtrace_enabling_destroy(enab);
14588 if (dhp != NULL && nprovs > 0) {
14589 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14590 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14591 mutex_exit(&dtrace_lock);
14592 dtrace_helper_provider_register(curproc, help, dhp);
14593 mutex_enter(&dtrace_lock);
14600 dtrace_dof_destroy(dof);
14605 static dtrace_helpers_t *
14606 dtrace_helpers_create(proc_t *p)
14608 dtrace_helpers_t *help;
14610 ASSERT(MUTEX_HELD(&dtrace_lock));
14611 ASSERT(p->p_dtrace_helpers == NULL);
14613 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14614 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14615 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14617 p->p_dtrace_helpers = help;
14627 dtrace_helpers_destroy(proc_t *p)
14629 dtrace_helpers_t *help;
14630 dtrace_vstate_t *vstate;
14632 proc_t *p = curproc;
14636 mutex_enter(&dtrace_lock);
14638 ASSERT(p->p_dtrace_helpers != NULL);
14639 ASSERT(dtrace_helpers > 0);
14641 help = p->p_dtrace_helpers;
14642 vstate = &help->dthps_vstate;
14645 * We're now going to lose the help from this process.
14647 p->p_dtrace_helpers = NULL;
14651 * Destory the helper actions.
14653 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14654 dtrace_helper_action_t *h, *next;
14656 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14657 next = h->dtha_next;
14658 dtrace_helper_action_destroy(h, vstate);
14663 mutex_exit(&dtrace_lock);
14666 * Destroy the helper providers.
14668 if (help->dthps_maxprovs > 0) {
14669 mutex_enter(&dtrace_meta_lock);
14670 if (dtrace_meta_pid != NULL) {
14671 ASSERT(dtrace_deferred_pid == NULL);
14673 for (i = 0; i < help->dthps_nprovs; i++) {
14674 dtrace_helper_provider_remove(
14675 &help->dthps_provs[i]->dthp_prov, p->p_pid);
14678 mutex_enter(&dtrace_lock);
14679 ASSERT(help->dthps_deferred == 0 ||
14680 help->dthps_next != NULL ||
14681 help->dthps_prev != NULL ||
14682 help == dtrace_deferred_pid);
14685 * Remove the helper from the deferred list.
14687 if (help->dthps_next != NULL)
14688 help->dthps_next->dthps_prev = help->dthps_prev;
14689 if (help->dthps_prev != NULL)
14690 help->dthps_prev->dthps_next = help->dthps_next;
14691 if (dtrace_deferred_pid == help) {
14692 dtrace_deferred_pid = help->dthps_next;
14693 ASSERT(help->dthps_prev == NULL);
14696 mutex_exit(&dtrace_lock);
14699 mutex_exit(&dtrace_meta_lock);
14701 for (i = 0; i < help->dthps_nprovs; i++) {
14702 dtrace_helper_provider_destroy(help->dthps_provs[i]);
14705 kmem_free(help->dthps_provs, help->dthps_maxprovs *
14706 sizeof (dtrace_helper_provider_t *));
14709 mutex_enter(&dtrace_lock);
14711 dtrace_vstate_fini(&help->dthps_vstate);
14712 kmem_free(help->dthps_actions,
14713 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
14714 kmem_free(help, sizeof (dtrace_helpers_t));
14717 mutex_exit(&dtrace_lock);
14724 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
14726 dtrace_helpers_t *help, *newhelp;
14727 dtrace_helper_action_t *helper, *new, *last;
14729 dtrace_vstate_t *vstate;
14730 int i, j, sz, hasprovs = 0;
14732 mutex_enter(&dtrace_lock);
14733 ASSERT(from->p_dtrace_helpers != NULL);
14734 ASSERT(dtrace_helpers > 0);
14736 help = from->p_dtrace_helpers;
14737 newhelp = dtrace_helpers_create(to);
14738 ASSERT(to->p_dtrace_helpers != NULL);
14740 newhelp->dthps_generation = help->dthps_generation;
14741 vstate = &newhelp->dthps_vstate;
14744 * Duplicate the helper actions.
14746 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14747 if ((helper = help->dthps_actions[i]) == NULL)
14750 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
14751 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
14753 new->dtha_generation = helper->dtha_generation;
14755 if ((dp = helper->dtha_predicate) != NULL) {
14756 dp = dtrace_difo_duplicate(dp, vstate);
14757 new->dtha_predicate = dp;
14760 new->dtha_nactions = helper->dtha_nactions;
14761 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
14762 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
14764 for (j = 0; j < new->dtha_nactions; j++) {
14765 dtrace_difo_t *dp = helper->dtha_actions[j];
14767 ASSERT(dp != NULL);
14768 dp = dtrace_difo_duplicate(dp, vstate);
14769 new->dtha_actions[j] = dp;
14772 if (last != NULL) {
14773 last->dtha_next = new;
14775 newhelp->dthps_actions[i] = new;
14783 * Duplicate the helper providers and register them with the
14784 * DTrace framework.
14786 if (help->dthps_nprovs > 0) {
14787 newhelp->dthps_nprovs = help->dthps_nprovs;
14788 newhelp->dthps_maxprovs = help->dthps_nprovs;
14789 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
14790 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14791 for (i = 0; i < newhelp->dthps_nprovs; i++) {
14792 newhelp->dthps_provs[i] = help->dthps_provs[i];
14793 newhelp->dthps_provs[i]->dthp_ref++;
14799 mutex_exit(&dtrace_lock);
14802 dtrace_helper_provider_register(to, newhelp, NULL);
14807 * DTrace Hook Functions
14810 dtrace_module_loaded(modctl_t *ctl)
14812 dtrace_provider_t *prv;
14814 mutex_enter(&dtrace_provider_lock);
14815 mutex_enter(&mod_lock);
14817 ASSERT(ctl->mod_busy);
14820 * We're going to call each providers per-module provide operation
14821 * specifying only this module.
14823 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
14824 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
14826 mutex_exit(&mod_lock);
14827 mutex_exit(&dtrace_provider_lock);
14830 * If we have any retained enablings, we need to match against them.
14831 * Enabling probes requires that cpu_lock be held, and we cannot hold
14832 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
14833 * module. (In particular, this happens when loading scheduling
14834 * classes.) So if we have any retained enablings, we need to dispatch
14835 * our task queue to do the match for us.
14837 mutex_enter(&dtrace_lock);
14839 if (dtrace_retained == NULL) {
14840 mutex_exit(&dtrace_lock);
14844 (void) taskq_dispatch(dtrace_taskq,
14845 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
14847 mutex_exit(&dtrace_lock);
14850 * And now, for a little heuristic sleaze: in general, we want to
14851 * match modules as soon as they load. However, we cannot guarantee
14852 * this, because it would lead us to the lock ordering violation
14853 * outlined above. The common case, of course, is that cpu_lock is
14854 * _not_ held -- so we delay here for a clock tick, hoping that that's
14855 * long enough for the task queue to do its work. If it's not, it's
14856 * not a serious problem -- it just means that the module that we
14857 * just loaded may not be immediately instrumentable.
14863 dtrace_module_unloaded(modctl_t *ctl)
14865 dtrace_probe_t template, *probe, *first, *next;
14866 dtrace_provider_t *prov;
14868 template.dtpr_mod = ctl->mod_modname;
14870 mutex_enter(&dtrace_provider_lock);
14871 mutex_enter(&mod_lock);
14872 mutex_enter(&dtrace_lock);
14874 if (dtrace_bymod == NULL) {
14876 * The DTrace module is loaded (obviously) but not attached;
14877 * we don't have any work to do.
14879 mutex_exit(&dtrace_provider_lock);
14880 mutex_exit(&mod_lock);
14881 mutex_exit(&dtrace_lock);
14885 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14886 probe != NULL; probe = probe->dtpr_nextmod) {
14887 if (probe->dtpr_ecb != NULL) {
14888 mutex_exit(&dtrace_provider_lock);
14889 mutex_exit(&mod_lock);
14890 mutex_exit(&dtrace_lock);
14893 * This shouldn't _actually_ be possible -- we're
14894 * unloading a module that has an enabled probe in it.
14895 * (It's normally up to the provider to make sure that
14896 * this can't happen.) However, because dtps_enable()
14897 * doesn't have a failure mode, there can be an
14898 * enable/unload race. Upshot: we don't want to
14899 * assert, but we're not going to disable the
14902 if (dtrace_err_verbose) {
14903 cmn_err(CE_WARN, "unloaded module '%s' had "
14904 "enabled probes", ctl->mod_modname);
14913 for (first = NULL; probe != NULL; probe = next) {
14914 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
14916 dtrace_probes[probe->dtpr_id - 1] = NULL;
14918 next = probe->dtpr_nextmod;
14919 dtrace_hash_remove(dtrace_bymod, probe);
14920 dtrace_hash_remove(dtrace_byfunc, probe);
14921 dtrace_hash_remove(dtrace_byname, probe);
14923 if (first == NULL) {
14925 probe->dtpr_nextmod = NULL;
14927 probe->dtpr_nextmod = first;
14933 * We've removed all of the module's probes from the hash chains and
14934 * from the probe array. Now issue a dtrace_sync() to be sure that
14935 * everyone has cleared out from any probe array processing.
14939 for (probe = first; probe != NULL; probe = first) {
14940 first = probe->dtpr_nextmod;
14941 prov = probe->dtpr_provider;
14942 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
14944 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
14945 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
14946 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
14947 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
14948 kmem_free(probe, sizeof (dtrace_probe_t));
14951 mutex_exit(&dtrace_lock);
14952 mutex_exit(&mod_lock);
14953 mutex_exit(&dtrace_provider_lock);
14957 dtrace_suspend(void)
14959 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
14963 dtrace_resume(void)
14965 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
14970 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
14972 ASSERT(MUTEX_HELD(&cpu_lock));
14973 mutex_enter(&dtrace_lock);
14977 dtrace_state_t *state;
14978 dtrace_optval_t *opt, rs, c;
14981 * For now, we only allocate a new buffer for anonymous state.
14983 if ((state = dtrace_anon.dta_state) == NULL)
14986 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14989 opt = state->dts_options;
14990 c = opt[DTRACEOPT_CPU];
14992 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
14996 * Regardless of what the actual policy is, we're going to
14997 * temporarily set our resize policy to be manual. We're
14998 * also going to temporarily set our CPU option to denote
14999 * the newly configured CPU.
15001 rs = opt[DTRACEOPT_BUFRESIZE];
15002 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15003 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15005 (void) dtrace_state_buffers(state);
15007 opt[DTRACEOPT_BUFRESIZE] = rs;
15008 opt[DTRACEOPT_CPU] = c;
15015 * We don't free the buffer in the CPU_UNCONFIG case. (The
15016 * buffer will be freed when the consumer exits.)
15024 mutex_exit(&dtrace_lock);
15030 dtrace_cpu_setup_initial(processorid_t cpu)
15032 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15037 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15039 if (dtrace_toxranges >= dtrace_toxranges_max) {
15041 dtrace_toxrange_t *range;
15043 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15046 ASSERT(dtrace_toxrange == NULL);
15047 ASSERT(dtrace_toxranges_max == 0);
15048 dtrace_toxranges_max = 1;
15050 dtrace_toxranges_max <<= 1;
15053 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15054 range = kmem_zalloc(nsize, KM_SLEEP);
15056 if (dtrace_toxrange != NULL) {
15057 ASSERT(osize != 0);
15058 bcopy(dtrace_toxrange, range, osize);
15059 kmem_free(dtrace_toxrange, osize);
15062 dtrace_toxrange = range;
15065 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15066 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15068 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15069 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15070 dtrace_toxranges++;
15074 * DTrace Driver Cookbook Functions
15079 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15081 dtrace_provider_id_t id;
15082 dtrace_state_t *state = NULL;
15083 dtrace_enabling_t *enab;
15085 mutex_enter(&cpu_lock);
15086 mutex_enter(&dtrace_provider_lock);
15087 mutex_enter(&dtrace_lock);
15089 if (ddi_soft_state_init(&dtrace_softstate,
15090 sizeof (dtrace_state_t), 0) != 0) {
15091 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15092 mutex_exit(&cpu_lock);
15093 mutex_exit(&dtrace_provider_lock);
15094 mutex_exit(&dtrace_lock);
15095 return (DDI_FAILURE);
15098 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15099 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15100 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15101 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15102 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15103 ddi_remove_minor_node(devi, NULL);
15104 ddi_soft_state_fini(&dtrace_softstate);
15105 mutex_exit(&cpu_lock);
15106 mutex_exit(&dtrace_provider_lock);
15107 mutex_exit(&dtrace_lock);
15108 return (DDI_FAILURE);
15111 ddi_report_dev(devi);
15112 dtrace_devi = devi;
15114 dtrace_modload = dtrace_module_loaded;
15115 dtrace_modunload = dtrace_module_unloaded;
15116 dtrace_cpu_init = dtrace_cpu_setup_initial;
15117 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15118 dtrace_helpers_fork = dtrace_helpers_duplicate;
15119 dtrace_cpustart_init = dtrace_suspend;
15120 dtrace_cpustart_fini = dtrace_resume;
15121 dtrace_debugger_init = dtrace_suspend;
15122 dtrace_debugger_fini = dtrace_resume;
15124 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15126 ASSERT(MUTEX_HELD(&cpu_lock));
15128 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15129 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15130 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15131 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15132 VM_SLEEP | VMC_IDENTIFIER);
15133 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15136 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15137 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15138 NULL, NULL, NULL, NULL, NULL, 0);
15140 ASSERT(MUTEX_HELD(&cpu_lock));
15141 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15142 offsetof(dtrace_probe_t, dtpr_nextmod),
15143 offsetof(dtrace_probe_t, dtpr_prevmod));
15145 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15146 offsetof(dtrace_probe_t, dtpr_nextfunc),
15147 offsetof(dtrace_probe_t, dtpr_prevfunc));
15149 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15150 offsetof(dtrace_probe_t, dtpr_nextname),
15151 offsetof(dtrace_probe_t, dtpr_prevname));
15153 if (dtrace_retain_max < 1) {
15154 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15155 "setting to 1", dtrace_retain_max);
15156 dtrace_retain_max = 1;
15160 * Now discover our toxic ranges.
15162 dtrace_toxic_ranges(dtrace_toxrange_add);
15165 * Before we register ourselves as a provider to our own framework,
15166 * we would like to assert that dtrace_provider is NULL -- but that's
15167 * not true if we were loaded as a dependency of a DTrace provider.
15168 * Once we've registered, we can assert that dtrace_provider is our
15171 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15172 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15174 ASSERT(dtrace_provider != NULL);
15175 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15177 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15178 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15179 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15180 dtrace_provider, NULL, NULL, "END", 0, NULL);
15181 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15182 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15184 dtrace_anon_property();
15185 mutex_exit(&cpu_lock);
15188 * If DTrace helper tracing is enabled, we need to allocate the
15189 * trace buffer and initialize the values.
15191 if (dtrace_helptrace_enabled) {
15192 ASSERT(dtrace_helptrace_buffer == NULL);
15193 dtrace_helptrace_buffer =
15194 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15195 dtrace_helptrace_next = 0;
15199 * If there are already providers, we must ask them to provide their
15200 * probes, and then match any anonymous enabling against them. Note
15201 * that there should be no other retained enablings at this time:
15202 * the only retained enablings at this time should be the anonymous
15205 if (dtrace_anon.dta_enabling != NULL) {
15206 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15208 dtrace_enabling_provide(NULL);
15209 state = dtrace_anon.dta_state;
15212 * We couldn't hold cpu_lock across the above call to
15213 * dtrace_enabling_provide(), but we must hold it to actually
15214 * enable the probes. We have to drop all of our locks, pick
15215 * up cpu_lock, and regain our locks before matching the
15216 * retained anonymous enabling.
15218 mutex_exit(&dtrace_lock);
15219 mutex_exit(&dtrace_provider_lock);
15221 mutex_enter(&cpu_lock);
15222 mutex_enter(&dtrace_provider_lock);
15223 mutex_enter(&dtrace_lock);
15225 if ((enab = dtrace_anon.dta_enabling) != NULL)
15226 (void) dtrace_enabling_match(enab, NULL);
15228 mutex_exit(&cpu_lock);
15231 mutex_exit(&dtrace_lock);
15232 mutex_exit(&dtrace_provider_lock);
15234 if (state != NULL) {
15236 * If we created any anonymous state, set it going now.
15238 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15241 return (DDI_SUCCESS);
15246 #if __FreeBSD_version >= 800039
15248 dtrace_dtr(void *data __unused)
15257 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15259 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15262 dtrace_state_t *state;
15268 if (getminor(*devp) == DTRACEMNRN_HELPER)
15272 * If this wasn't an open with the "helper" minor, then it must be
15273 * the "dtrace" minor.
15275 ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15277 cred_t *cred_p = NULL;
15279 #if __FreeBSD_version < 800039
15281 * The first minor device is the one that is cloned so there is
15282 * nothing more to do here.
15284 if (dev2unit(dev) == 0)
15288 * Devices are cloned, so if the DTrace state has already
15289 * been allocated, that means this device belongs to a
15290 * different client. Each client should open '/dev/dtrace'
15291 * to get a cloned device.
15293 if (dev->si_drv1 != NULL)
15297 cred_p = dev->si_cred;
15301 * If no DTRACE_PRIV_* bits are set in the credential, then the
15302 * caller lacks sufficient permission to do anything with DTrace.
15304 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15305 if (priv == DTRACE_PRIV_NONE) {
15307 #if __FreeBSD_version < 800039
15308 /* Destroy the cloned device. */
15317 * Ask all providers to provide all their probes.
15319 mutex_enter(&dtrace_provider_lock);
15320 dtrace_probe_provide(NULL, NULL);
15321 mutex_exit(&dtrace_provider_lock);
15323 mutex_enter(&cpu_lock);
15324 mutex_enter(&dtrace_lock);
15326 dtrace_membar_producer();
15330 * If the kernel debugger is active (that is, if the kernel debugger
15331 * modified text in some way), we won't allow the open.
15333 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15335 mutex_exit(&cpu_lock);
15336 mutex_exit(&dtrace_lock);
15340 state = dtrace_state_create(devp, cred_p);
15342 state = dtrace_state_create(dev);
15343 #if __FreeBSD_version < 800039
15344 dev->si_drv1 = state;
15346 devfs_set_cdevpriv(state, dtrace_dtr);
15350 mutex_exit(&cpu_lock);
15352 if (state == NULL) {
15354 if (--dtrace_opens == 0)
15355 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15359 mutex_exit(&dtrace_lock);
15361 #if __FreeBSD_version < 800039
15362 /* Destroy the cloned device. */
15369 mutex_exit(&dtrace_lock);
15377 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15379 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15383 minor_t minor = getminor(dev);
15384 dtrace_state_t *state;
15386 if (minor == DTRACEMNRN_HELPER)
15389 state = ddi_get_soft_state(dtrace_softstate, minor);
15391 #if __FreeBSD_version < 800039
15392 dtrace_state_t *state = dev->si_drv1;
15394 /* Check if this is not a cloned device. */
15395 if (dev2unit(dev) == 0)
15398 dtrace_state_t *state;
15399 devfs_get_cdevpriv((void **) &state);
15404 mutex_enter(&cpu_lock);
15405 mutex_enter(&dtrace_lock);
15407 if (state != NULL) {
15408 if (state->dts_anon) {
15410 * There is anonymous state. Destroy that first.
15412 ASSERT(dtrace_anon.dta_state == NULL);
15413 dtrace_state_destroy(state->dts_anon);
15416 dtrace_state_destroy(state);
15419 kmem_free(state, 0);
15420 #if __FreeBSD_version < 800039
15421 dev->si_drv1 = NULL;
15423 devfs_clear_cdevpriv();
15428 ASSERT(dtrace_opens > 0);
15430 if (--dtrace_opens == 0)
15431 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15436 mutex_exit(&dtrace_lock);
15437 mutex_exit(&cpu_lock);
15439 #if __FreeBSD_version < 800039
15440 /* Schedule this cloned device to be destroyed. */
15441 destroy_dev_sched(dev);
15450 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15453 dof_helper_t help, *dhp = NULL;
15456 case DTRACEHIOC_ADDDOF:
15457 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15458 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15463 arg = (intptr_t)help.dofhp_dof;
15466 case DTRACEHIOC_ADD: {
15467 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15472 mutex_enter(&dtrace_lock);
15475 * dtrace_helper_slurp() takes responsibility for the dof --
15476 * it may free it now or it may save it and free it later.
15478 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15485 mutex_exit(&dtrace_lock);
15489 case DTRACEHIOC_REMOVE: {
15490 mutex_enter(&dtrace_lock);
15491 rval = dtrace_helper_destroygen(arg);
15492 mutex_exit(&dtrace_lock);
15506 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15508 minor_t minor = getminor(dev);
15509 dtrace_state_t *state;
15512 if (minor == DTRACEMNRN_HELPER)
15513 return (dtrace_ioctl_helper(cmd, arg, rv));
15515 state = ddi_get_soft_state(dtrace_softstate, minor);
15517 if (state->dts_anon) {
15518 ASSERT(dtrace_anon.dta_state == NULL);
15519 state = state->dts_anon;
15523 case DTRACEIOC_PROVIDER: {
15524 dtrace_providerdesc_t pvd;
15525 dtrace_provider_t *pvp;
15527 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15530 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15531 mutex_enter(&dtrace_provider_lock);
15533 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15534 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15538 mutex_exit(&dtrace_provider_lock);
15543 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15544 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15546 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15552 case DTRACEIOC_EPROBE: {
15553 dtrace_eprobedesc_t epdesc;
15555 dtrace_action_t *act;
15561 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15564 mutex_enter(&dtrace_lock);
15566 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15567 mutex_exit(&dtrace_lock);
15571 if (ecb->dte_probe == NULL) {
15572 mutex_exit(&dtrace_lock);
15576 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15577 epdesc.dtepd_uarg = ecb->dte_uarg;
15578 epdesc.dtepd_size = ecb->dte_size;
15580 nrecs = epdesc.dtepd_nrecs;
15581 epdesc.dtepd_nrecs = 0;
15582 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15583 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15586 epdesc.dtepd_nrecs++;
15590 * Now that we have the size, we need to allocate a temporary
15591 * buffer in which to store the complete description. We need
15592 * the temporary buffer to be able to drop dtrace_lock()
15593 * across the copyout(), below.
15595 size = sizeof (dtrace_eprobedesc_t) +
15596 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15598 buf = kmem_alloc(size, KM_SLEEP);
15599 dest = (uintptr_t)buf;
15601 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15602 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15604 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15605 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15611 bcopy(&act->dta_rec, (void *)dest,
15612 sizeof (dtrace_recdesc_t));
15613 dest += sizeof (dtrace_recdesc_t);
15616 mutex_exit(&dtrace_lock);
15618 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15619 kmem_free(buf, size);
15623 kmem_free(buf, size);
15627 case DTRACEIOC_AGGDESC: {
15628 dtrace_aggdesc_t aggdesc;
15629 dtrace_action_t *act;
15630 dtrace_aggregation_t *agg;
15633 dtrace_recdesc_t *lrec;
15638 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15641 mutex_enter(&dtrace_lock);
15643 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15644 mutex_exit(&dtrace_lock);
15648 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15650 nrecs = aggdesc.dtagd_nrecs;
15651 aggdesc.dtagd_nrecs = 0;
15653 offs = agg->dtag_base;
15654 lrec = &agg->dtag_action.dta_rec;
15655 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15657 for (act = agg->dtag_first; ; act = act->dta_next) {
15658 ASSERT(act->dta_intuple ||
15659 DTRACEACT_ISAGG(act->dta_kind));
15662 * If this action has a record size of zero, it
15663 * denotes an argument to the aggregating action.
15664 * Because the presence of this record doesn't (or
15665 * shouldn't) affect the way the data is interpreted,
15666 * we don't copy it out to save user-level the
15667 * confusion of dealing with a zero-length record.
15669 if (act->dta_rec.dtrd_size == 0) {
15670 ASSERT(agg->dtag_hasarg);
15674 aggdesc.dtagd_nrecs++;
15676 if (act == &agg->dtag_action)
15681 * Now that we have the size, we need to allocate a temporary
15682 * buffer in which to store the complete description. We need
15683 * the temporary buffer to be able to drop dtrace_lock()
15684 * across the copyout(), below.
15686 size = sizeof (dtrace_aggdesc_t) +
15687 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15689 buf = kmem_alloc(size, KM_SLEEP);
15690 dest = (uintptr_t)buf;
15692 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15693 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15695 for (act = agg->dtag_first; ; act = act->dta_next) {
15696 dtrace_recdesc_t rec = act->dta_rec;
15699 * See the comment in the above loop for why we pass
15700 * over zero-length records.
15702 if (rec.dtrd_size == 0) {
15703 ASSERT(agg->dtag_hasarg);
15710 rec.dtrd_offset -= offs;
15711 bcopy(&rec, (void *)dest, sizeof (rec));
15712 dest += sizeof (dtrace_recdesc_t);
15714 if (act == &agg->dtag_action)
15718 mutex_exit(&dtrace_lock);
15720 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15721 kmem_free(buf, size);
15725 kmem_free(buf, size);
15729 case DTRACEIOC_ENABLE: {
15731 dtrace_enabling_t *enab = NULL;
15732 dtrace_vstate_t *vstate;
15738 * If a NULL argument has been passed, we take this as our
15739 * cue to reevaluate our enablings.
15742 dtrace_enabling_matchall();
15747 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
15750 mutex_enter(&cpu_lock);
15751 mutex_enter(&dtrace_lock);
15752 vstate = &state->dts_vstate;
15754 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
15755 mutex_exit(&dtrace_lock);
15756 mutex_exit(&cpu_lock);
15757 dtrace_dof_destroy(dof);
15761 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
15762 mutex_exit(&dtrace_lock);
15763 mutex_exit(&cpu_lock);
15764 dtrace_dof_destroy(dof);
15768 if ((rval = dtrace_dof_options(dof, state)) != 0) {
15769 dtrace_enabling_destroy(enab);
15770 mutex_exit(&dtrace_lock);
15771 mutex_exit(&cpu_lock);
15772 dtrace_dof_destroy(dof);
15776 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
15777 err = dtrace_enabling_retain(enab);
15779 dtrace_enabling_destroy(enab);
15782 mutex_exit(&cpu_lock);
15783 mutex_exit(&dtrace_lock);
15784 dtrace_dof_destroy(dof);
15789 case DTRACEIOC_REPLICATE: {
15790 dtrace_repldesc_t desc;
15791 dtrace_probedesc_t *match = &desc.dtrpd_match;
15792 dtrace_probedesc_t *create = &desc.dtrpd_create;
15795 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15798 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15799 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15800 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15801 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15803 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15804 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15805 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15806 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15808 mutex_enter(&dtrace_lock);
15809 err = dtrace_enabling_replicate(state, match, create);
15810 mutex_exit(&dtrace_lock);
15815 case DTRACEIOC_PROBEMATCH:
15816 case DTRACEIOC_PROBES: {
15817 dtrace_probe_t *probe = NULL;
15818 dtrace_probedesc_t desc;
15819 dtrace_probekey_t pkey;
15826 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15829 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15830 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15831 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15832 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15835 * Before we attempt to match this probe, we want to give
15836 * all providers the opportunity to provide it.
15838 if (desc.dtpd_id == DTRACE_IDNONE) {
15839 mutex_enter(&dtrace_provider_lock);
15840 dtrace_probe_provide(&desc, NULL);
15841 mutex_exit(&dtrace_provider_lock);
15845 if (cmd == DTRACEIOC_PROBEMATCH) {
15846 dtrace_probekey(&desc, &pkey);
15847 pkey.dtpk_id = DTRACE_IDNONE;
15850 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
15852 mutex_enter(&dtrace_lock);
15854 if (cmd == DTRACEIOC_PROBEMATCH) {
15855 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15856 if ((probe = dtrace_probes[i - 1]) != NULL &&
15857 (m = dtrace_match_probe(probe, &pkey,
15858 priv, uid, zoneid)) != 0)
15863 mutex_exit(&dtrace_lock);
15868 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15869 if ((probe = dtrace_probes[i - 1]) != NULL &&
15870 dtrace_match_priv(probe, priv, uid, zoneid))
15875 if (probe == NULL) {
15876 mutex_exit(&dtrace_lock);
15880 dtrace_probe_description(probe, &desc);
15881 mutex_exit(&dtrace_lock);
15883 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15889 case DTRACEIOC_PROBEARG: {
15890 dtrace_argdesc_t desc;
15891 dtrace_probe_t *probe;
15892 dtrace_provider_t *prov;
15894 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15897 if (desc.dtargd_id == DTRACE_IDNONE)
15900 if (desc.dtargd_ndx == DTRACE_ARGNONE)
15903 mutex_enter(&dtrace_provider_lock);
15904 mutex_enter(&mod_lock);
15905 mutex_enter(&dtrace_lock);
15907 if (desc.dtargd_id > dtrace_nprobes) {
15908 mutex_exit(&dtrace_lock);
15909 mutex_exit(&mod_lock);
15910 mutex_exit(&dtrace_provider_lock);
15914 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
15915 mutex_exit(&dtrace_lock);
15916 mutex_exit(&mod_lock);
15917 mutex_exit(&dtrace_provider_lock);
15921 mutex_exit(&dtrace_lock);
15923 prov = probe->dtpr_provider;
15925 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
15927 * There isn't any typed information for this probe.
15928 * Set the argument number to DTRACE_ARGNONE.
15930 desc.dtargd_ndx = DTRACE_ARGNONE;
15932 desc.dtargd_native[0] = '\0';
15933 desc.dtargd_xlate[0] = '\0';
15934 desc.dtargd_mapping = desc.dtargd_ndx;
15936 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
15937 probe->dtpr_id, probe->dtpr_arg, &desc);
15940 mutex_exit(&mod_lock);
15941 mutex_exit(&dtrace_provider_lock);
15943 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15949 case DTRACEIOC_GO: {
15950 processorid_t cpuid;
15951 rval = dtrace_state_go(state, &cpuid);
15956 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15962 case DTRACEIOC_STOP: {
15963 processorid_t cpuid;
15965 mutex_enter(&dtrace_lock);
15966 rval = dtrace_state_stop(state, &cpuid);
15967 mutex_exit(&dtrace_lock);
15972 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15978 case DTRACEIOC_DOFGET: {
15979 dof_hdr_t hdr, *dof;
15982 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
15985 mutex_enter(&dtrace_lock);
15986 dof = dtrace_dof_create(state);
15987 mutex_exit(&dtrace_lock);
15989 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
15990 rval = copyout(dof, (void *)arg, len);
15991 dtrace_dof_destroy(dof);
15993 return (rval == 0 ? 0 : EFAULT);
15996 case DTRACEIOC_AGGSNAP:
15997 case DTRACEIOC_BUFSNAP: {
15998 dtrace_bufdesc_t desc;
16000 dtrace_buffer_t *buf;
16002 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16005 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16008 mutex_enter(&dtrace_lock);
16010 if (cmd == DTRACEIOC_BUFSNAP) {
16011 buf = &state->dts_buffer[desc.dtbd_cpu];
16013 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16016 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16017 size_t sz = buf->dtb_offset;
16019 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16020 mutex_exit(&dtrace_lock);
16025 * If this buffer has already been consumed, we're
16026 * going to indicate that there's nothing left here
16029 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16030 mutex_exit(&dtrace_lock);
16032 desc.dtbd_size = 0;
16033 desc.dtbd_drops = 0;
16034 desc.dtbd_errors = 0;
16035 desc.dtbd_oldest = 0;
16036 sz = sizeof (desc);
16038 if (copyout(&desc, (void *)arg, sz) != 0)
16045 * If this is a ring buffer that has wrapped, we want
16046 * to copy the whole thing out.
16048 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16049 dtrace_buffer_polish(buf);
16050 sz = buf->dtb_size;
16053 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16054 mutex_exit(&dtrace_lock);
16058 desc.dtbd_size = sz;
16059 desc.dtbd_drops = buf->dtb_drops;
16060 desc.dtbd_errors = buf->dtb_errors;
16061 desc.dtbd_oldest = buf->dtb_xamot_offset;
16063 mutex_exit(&dtrace_lock);
16065 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16068 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16073 if (buf->dtb_tomax == NULL) {
16074 ASSERT(buf->dtb_xamot == NULL);
16075 mutex_exit(&dtrace_lock);
16079 cached = buf->dtb_tomax;
16080 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16082 dtrace_xcall(desc.dtbd_cpu,
16083 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16085 state->dts_errors += buf->dtb_xamot_errors;
16088 * If the buffers did not actually switch, then the cross call
16089 * did not take place -- presumably because the given CPU is
16090 * not in the ready set. If this is the case, we'll return
16093 if (buf->dtb_tomax == cached) {
16094 ASSERT(buf->dtb_xamot != cached);
16095 mutex_exit(&dtrace_lock);
16099 ASSERT(cached == buf->dtb_xamot);
16102 * We have our snapshot; now copy it out.
16104 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16105 buf->dtb_xamot_offset) != 0) {
16106 mutex_exit(&dtrace_lock);
16110 desc.dtbd_size = buf->dtb_xamot_offset;
16111 desc.dtbd_drops = buf->dtb_xamot_drops;
16112 desc.dtbd_errors = buf->dtb_xamot_errors;
16113 desc.dtbd_oldest = 0;
16115 mutex_exit(&dtrace_lock);
16118 * Finally, copy out the buffer description.
16120 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16126 case DTRACEIOC_CONF: {
16127 dtrace_conf_t conf;
16129 bzero(&conf, sizeof (conf));
16130 conf.dtc_difversion = DIF_VERSION;
16131 conf.dtc_difintregs = DIF_DIR_NREGS;
16132 conf.dtc_diftupregs = DIF_DTR_NREGS;
16133 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16135 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16141 case DTRACEIOC_STATUS: {
16142 dtrace_status_t stat;
16143 dtrace_dstate_t *dstate;
16148 * See the comment in dtrace_state_deadman() for the reason
16149 * for setting dts_laststatus to INT64_MAX before setting
16150 * it to the correct value.
16152 state->dts_laststatus = INT64_MAX;
16153 dtrace_membar_producer();
16154 state->dts_laststatus = dtrace_gethrtime();
16156 bzero(&stat, sizeof (stat));
16158 mutex_enter(&dtrace_lock);
16160 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16161 mutex_exit(&dtrace_lock);
16165 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16166 stat.dtst_exiting = 1;
16168 nerrs = state->dts_errors;
16169 dstate = &state->dts_vstate.dtvs_dynvars;
16171 for (i = 0; i < NCPU; i++) {
16172 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16174 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16175 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16176 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16178 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16179 stat.dtst_filled++;
16181 nerrs += state->dts_buffer[i].dtb_errors;
16183 for (j = 0; j < state->dts_nspeculations; j++) {
16184 dtrace_speculation_t *spec;
16185 dtrace_buffer_t *buf;
16187 spec = &state->dts_speculations[j];
16188 buf = &spec->dtsp_buffer[i];
16189 stat.dtst_specdrops += buf->dtb_xamot_drops;
16193 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16194 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16195 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16196 stat.dtst_dblerrors = state->dts_dblerrors;
16198 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16199 stat.dtst_errors = nerrs;
16201 mutex_exit(&dtrace_lock);
16203 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16209 case DTRACEIOC_FORMAT: {
16210 dtrace_fmtdesc_t fmt;
16214 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16217 mutex_enter(&dtrace_lock);
16219 if (fmt.dtfd_format == 0 ||
16220 fmt.dtfd_format > state->dts_nformats) {
16221 mutex_exit(&dtrace_lock);
16226 * Format strings are allocated contiguously and they are
16227 * never freed; if a format index is less than the number
16228 * of formats, we can assert that the format map is non-NULL
16229 * and that the format for the specified index is non-NULL.
16231 ASSERT(state->dts_formats != NULL);
16232 str = state->dts_formats[fmt.dtfd_format - 1];
16233 ASSERT(str != NULL);
16235 len = strlen(str) + 1;
16237 if (len > fmt.dtfd_length) {
16238 fmt.dtfd_length = len;
16240 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16241 mutex_exit(&dtrace_lock);
16245 if (copyout(str, fmt.dtfd_string, len) != 0) {
16246 mutex_exit(&dtrace_lock);
16251 mutex_exit(&dtrace_lock);
16264 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16266 dtrace_state_t *state;
16273 return (DDI_SUCCESS);
16276 return (DDI_FAILURE);
16279 mutex_enter(&cpu_lock);
16280 mutex_enter(&dtrace_provider_lock);
16281 mutex_enter(&dtrace_lock);
16283 ASSERT(dtrace_opens == 0);
16285 if (dtrace_helpers > 0) {
16286 mutex_exit(&dtrace_provider_lock);
16287 mutex_exit(&dtrace_lock);
16288 mutex_exit(&cpu_lock);
16289 return (DDI_FAILURE);
16292 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16293 mutex_exit(&dtrace_provider_lock);
16294 mutex_exit(&dtrace_lock);
16295 mutex_exit(&cpu_lock);
16296 return (DDI_FAILURE);
16299 dtrace_provider = NULL;
16301 if ((state = dtrace_anon_grab()) != NULL) {
16303 * If there were ECBs on this state, the provider should
16304 * have not been allowed to detach; assert that there is
16307 ASSERT(state->dts_necbs == 0);
16308 dtrace_state_destroy(state);
16311 * If we're being detached with anonymous state, we need to
16312 * indicate to the kernel debugger that DTrace is now inactive.
16314 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16317 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16318 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16319 dtrace_cpu_init = NULL;
16320 dtrace_helpers_cleanup = NULL;
16321 dtrace_helpers_fork = NULL;
16322 dtrace_cpustart_init = NULL;
16323 dtrace_cpustart_fini = NULL;
16324 dtrace_debugger_init = NULL;
16325 dtrace_debugger_fini = NULL;
16326 dtrace_modload = NULL;
16327 dtrace_modunload = NULL;
16329 mutex_exit(&cpu_lock);
16331 if (dtrace_helptrace_enabled) {
16332 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16333 dtrace_helptrace_buffer = NULL;
16336 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16337 dtrace_probes = NULL;
16338 dtrace_nprobes = 0;
16340 dtrace_hash_destroy(dtrace_bymod);
16341 dtrace_hash_destroy(dtrace_byfunc);
16342 dtrace_hash_destroy(dtrace_byname);
16343 dtrace_bymod = NULL;
16344 dtrace_byfunc = NULL;
16345 dtrace_byname = NULL;
16347 kmem_cache_destroy(dtrace_state_cache);
16348 vmem_destroy(dtrace_minor);
16349 vmem_destroy(dtrace_arena);
16351 if (dtrace_toxrange != NULL) {
16352 kmem_free(dtrace_toxrange,
16353 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16354 dtrace_toxrange = NULL;
16355 dtrace_toxranges = 0;
16356 dtrace_toxranges_max = 0;
16359 ddi_remove_minor_node(dtrace_devi, NULL);
16360 dtrace_devi = NULL;
16362 ddi_soft_state_fini(&dtrace_softstate);
16364 ASSERT(dtrace_vtime_references == 0);
16365 ASSERT(dtrace_opens == 0);
16366 ASSERT(dtrace_retained == NULL);
16368 mutex_exit(&dtrace_lock);
16369 mutex_exit(&dtrace_provider_lock);
16372 * We don't destroy the task queue until after we have dropped our
16373 * locks (taskq_destroy() may block on running tasks). To prevent
16374 * attempting to do work after we have effectively detached but before
16375 * the task queue has been destroyed, all tasks dispatched via the
16376 * task queue must check that DTrace is still attached before
16377 * performing any operation.
16379 taskq_destroy(dtrace_taskq);
16380 dtrace_taskq = NULL;
16382 return (DDI_SUCCESS);
16389 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16394 case DDI_INFO_DEVT2DEVINFO:
16395 *result = (void *)dtrace_devi;
16396 error = DDI_SUCCESS;
16398 case DDI_INFO_DEVT2INSTANCE:
16399 *result = (void *)0;
16400 error = DDI_SUCCESS;
16403 error = DDI_FAILURE;
16410 static struct cb_ops dtrace_cb_ops = {
16411 dtrace_open, /* open */
16412 dtrace_close, /* close */
16413 nulldev, /* strategy */
16414 nulldev, /* print */
16418 dtrace_ioctl, /* ioctl */
16419 nodev, /* devmap */
16421 nodev, /* segmap */
16422 nochpoll, /* poll */
16423 ddi_prop_op, /* cb_prop_op */
16425 D_NEW | D_MP /* Driver compatibility flag */
16428 static struct dev_ops dtrace_ops = {
16429 DEVO_REV, /* devo_rev */
16431 dtrace_info, /* get_dev_info */
16432 nulldev, /* identify */
16433 nulldev, /* probe */
16434 dtrace_attach, /* attach */
16435 dtrace_detach, /* detach */
16437 &dtrace_cb_ops, /* driver operations */
16438 NULL, /* bus operations */
16439 nodev /* dev power */
16442 static struct modldrv modldrv = {
16443 &mod_driverops, /* module type (this is a pseudo driver) */
16444 "Dynamic Tracing", /* name of module */
16445 &dtrace_ops, /* driver ops */
16448 static struct modlinkage modlinkage = {
16457 return (mod_install(&modlinkage));
16461 _info(struct modinfo *modinfop)
16463 return (mod_info(&modlinkage, modinfop));
16469 return (mod_remove(&modlinkage));
16473 static d_ioctl_t dtrace_ioctl;
16474 static d_ioctl_t dtrace_ioctl_helper;
16475 static void dtrace_load(void *);
16476 static int dtrace_unload(void);
16477 #if __FreeBSD_version < 800039
16478 static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16479 static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */
16480 static eventhandler_tag eh_tag; /* Event handler tag. */
16482 static struct cdev *dtrace_dev;
16483 static struct cdev *helper_dev;
16486 void dtrace_invop_init(void);
16487 void dtrace_invop_uninit(void);
16489 static struct cdevsw dtrace_cdevsw = {
16490 .d_version = D_VERSION,
16491 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16492 .d_close = dtrace_close,
16493 .d_ioctl = dtrace_ioctl,
16494 .d_open = dtrace_open,
16495 .d_name = "dtrace",
16498 static struct cdevsw helper_cdevsw = {
16499 .d_version = D_VERSION,
16500 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16501 .d_ioctl = dtrace_ioctl_helper,
16502 .d_name = "helper",
16505 #include <dtrace_anon.c>
16506 #if __FreeBSD_version < 800039
16507 #include <dtrace_clone.c>
16509 #include <dtrace_ioctl.c>
16510 #include <dtrace_load.c>
16511 #include <dtrace_modevent.c>
16512 #include <dtrace_sysctl.c>
16513 #include <dtrace_unload.c>
16514 #include <dtrace_vtime.c>
16515 #include <dtrace_hacks.c>
16516 #include <dtrace_isa.c>
16518 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16519 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16520 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16522 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16523 MODULE_VERSION(dtrace, 1);
16524 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
16525 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);